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 2007-2008 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     DWORD numPreloadStreams = This->stateBlock->streamIsUP ? 0 : vertexDeclaration->num_streams;
 160     DWORD *streams = vertexDeclaration->streams;
 161
 162     /* Check for transformed vertices, disable vertex shader if present */
 163     strided->u.s.position_transformed = vertexDeclaration->position_transformed;
 164     if(vertexDeclaration->position_transformed) {
 165         useVertexShaderFunction = FALSE;
 166     }
 167
 168     /* Translate the declaration into strided data */
 169     for (i = 0 ; i < vertexDeclaration->declarationWNumElements - 1; ++i) {
 170         GLint streamVBO = 0;
 171         BOOL stride_used;
 172         unsigned int idx;
 173
 174         element = vertexDeclaration->pDeclarationWine + i;
 175         TRACE("%p Element %p (%d of %d)\n", vertexDeclaration->pDeclarationWine,
 176             element,  i + 1, vertexDeclaration->declarationWNumElements - 1);
 177
 178         if (This->stateBlock->streamSource[element->Stream] == NULL)
 179             continue;
 180
 181         stride  = This->stateBlock->streamStride[element->Stream];
 182         if (This->stateBlock->streamIsUP) {
 183             TRACE("Stream is up %d, %p\n", element->Stream, This->stateBlock->streamSource[element->Stream]);
 184             streamVBO = 0;
 185             data    = (BYTE *)This->stateBlock->streamSource[element->Stream];
 186         } else {
 187             TRACE("Stream isn't up %d, %p\n", element->Stream, This->stateBlock->streamSource[element->Stream]);
 188             data    = IWineD3DVertexBufferImpl_GetMemory(This->stateBlock->streamSource[element->Stream], 0, &streamVBO);
 189
 190             /* Can't use vbo's if the base vertex index is negative. OpenGL doesn't accept negative offsets
 191              * (or rather offsets bigger than the vbo, because the pointer is unsigned), so use system memory
 192              * sources. In most sane cases the pointer - offset will still be > 0, otherwise it will wrap
 193              * around to some big value. Hope that with the indices, the driver wraps it back internally. If
 194              * not, drawStridedSlow is needed, including a vertex buffer path.
 195              */
 196             if(This->stateBlock->loadBaseVertexIndex < 0) {
 197                 WARN("loadBaseVertexIndex is < 0 (%d), not using vbos\n", This->stateBlock->loadBaseVertexIndex);
 198                 streamVBO = 0;
 199                 data = ((IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[element->Stream])->resource.allocatedMemory;
 200                 if((UINT_PTR)data < -This->stateBlock->loadBaseVertexIndex * stride) {
 201                     FIXME("System memory vertex data load offset is negative!\n");
 202                 }
 203             }
 204
 205             if(fixup) {
 206                 if( streamVBO != 0) *fixup = TRUE;
 207                 else if(*fixup && !useVertexShaderFunction &&
 208                        (element->Usage == WINED3DDECLUSAGE_COLOR ||
 209                         element->Usage == WINED3DDECLUSAGE_POSITIONT)) {
 210                     /* This may be bad with the fixed function pipeline */
 211                     FIXME("Missing vbo streams with unfixed colors or transformed position, expect problems\n");
 212                 }
 213             }
 214         }
 215         data += element->Offset;
 216
 217         TRACE("Offset %d Stream %d UsageIndex %d\n", element->Offset, element->Stream, element->UsageIndex);
 218
 219         if (useVertexShaderFunction)
 220             stride_used = vshader_get_input(This->stateBlock->vertexShader,
 221                 element->Usage, element->UsageIndex, &idx);
 222         else
 223             stride_used = fixed_get_input(element->Usage, element->UsageIndex, &idx);
 224
 225         if (stride_used) {
 226             TRACE("Loaded %s array %u [usage=%s, usage_idx=%u, "
 227                     "stream=%u, offset=%u, stride=%u, type=%s, VBO=%u]\n",
 228                     useVertexShaderFunction? "shader": "fixed function", idx,
 229                     debug_d3ddeclusage(element->Usage), element->UsageIndex,
 230                     element->Stream, element->Offset, stride, debug_d3ddecltype(element->Type), streamVBO);
 231
 232             strided->u.input[idx].lpData = data;
 233             strided->u.input[idx].dwType = element->Type;
 234             strided->u.input[idx].dwStride = stride;
 235             strided->u.input[idx].VBO = streamVBO;
 236             strided->u.input[idx].streamNo = element->Stream;
 237         }
 238     }
 239     /* Now call PreLoad on all the vertex buffers. In the very rare case
 240      * that the buffers stopps converting PreLoad will dirtify the VDECL again.
 241      * The vertex buffer can now use the strided structure in the device instead of finding its
 242      * own again.
 243      *
 244      * NULL streams won't be recorded in the array, UP streams won't be either. A stream is only
 245      * once in there.
 246      */
 247     for(i=0; i < numPreloadStreams; i++) {
 248         IWineD3DVertexBuffer *vb = This->stateBlock->streamSource[streams[i]];
 249         if(vb) {
 250             IWineD3DVertexBuffer_PreLoad(vb);
 251         }
 252     }
 253 }
 254
 255 static void drawStridedFast(IWineD3DDevice *iface,UINT numberOfVertices, GLenum glPrimitiveType,
 256                      const void *idxData, short idxSize, ULONG minIndex, ULONG startIdx, ULONG startVertex) {
 257     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
 258
 259     if (idxSize != 0 /* This crashes sometimes!*/) {
 260         TRACE("(%p) : glElements(%x, %d, %d, ...)\n", This, glPrimitiveType, numberOfVertices, minIndex);
 261         idxData = idxData == (void *)-1 ? NULL : idxData;
 262 #if 1
 263         glDrawElements(glPrimitiveType, numberOfVertices, idxSize == 2 ? GL_UNSIGNED_SHORT : GL_UNSIGNED_INT,
 264                      (const char *)idxData+(idxSize * startIdx));
 265         checkGLcall("glDrawElements");
 266 #else /* using drawRangeElements may be faster */
 267
 268         glDrawRangeElements(glPrimitiveType, minIndex, minIndex + numberOfVertices - 1, numberOfVertices,
 269                       idxSize == 2 ? GL_UNSIGNED_SHORT : GL_UNSIGNED_INT,
 270                       (const char *)idxData+(idxSize * startIdx));
 271         checkGLcall("glDrawRangeElements");
 272 #endif
 273
 274     } else {
 275
 276         /* Note first is now zero as we shuffled along earlier */
 277         TRACE("(%p) : glDrawArrays(%x, 0, %d)\n", This, glPrimitiveType, numberOfVertices);
 278         glDrawArrays(glPrimitiveType, startVertex, numberOfVertices);
 279         checkGLcall("glDrawArrays");
 280
 281     }
 282
 283     return;
 284 }
 285
 286 /*
 287  * Actually draw using the supplied information.
 288  * Slower GL version which extracts info about each vertex in turn
 289  */
 290
 291 static void drawStridedSlow(IWineD3DDevice *iface, WineDirect3DVertexStridedData *sd,
 292                      UINT NumVertexes, GLenum glPrimType,
 293                      const void *idxData, short idxSize, ULONG minIndex, ULONG startIdx, ULONG startVertex) {
 294
 295     unsigned int               textureNo    = 0;
 296     const WORD                *pIdxBufS     = NULL;
 297     const DWORD               *pIdxBufL     = NULL;
 298     LONG                       vx_index;
 299     DWORD diffuseColor = 0xFFFFFFFF;       /* Diffuse Color              */
 300     DWORD specularColor = 0;               /* Specular Color             */
 301     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
 302     UINT *streamOffset = This->stateBlock->streamOffset;
 303     long                      SkipnStrides = startVertex + This->stateBlock->loadBaseVertexIndex;
 304     BOOL                      pixelShader = use_ps(This);
 305
 306     BYTE *texCoords[WINED3DDP_MAXTEXCOORD];
 307     BYTE *diffuse = NULL, *specular = NULL, *normal = NULL, *position = NULL;
 308
 309     TRACE("Using slow vertex array code\n");
 310
 311     /* Variable Initialization */
 312     if (idxSize != 0) {
 313         /* Immediate mode drawing can't make use of indices in a vbo - get the data from the index buffer.
 314          * If the index buffer has no vbo(not supported or other reason), or with user pointer drawing
 315          * idxData will be != NULL
 316          */
 317         if(idxData == NULL) {
 318             idxData = ((IWineD3DIndexBufferImpl *) This->stateBlock->pIndexData)->resource.allocatedMemory;
 319         }
 320
 321         if (idxSize == 2) pIdxBufS = (const WORD *) idxData;
 322         else pIdxBufL = (const DWORD *) idxData;
 323     }
 324
 325     /* Adding the stream offset once is cheaper than doing it every iteration. Do not modify the strided data, it is a pointer
 326      * to the strided Data in the device and might be needed intact on the next draw
 327      */
 328     for (textureNo = 0; textureNo < GL_LIMITS(texture_stages); ++textureNo) {
 329         if(sd->u.s.texCoords[textureNo].lpData) {
 330             texCoords[textureNo] = sd->u.s.texCoords[textureNo].lpData + streamOffset[sd->u.s.texCoords[textureNo].streamNo];
 331         } else {
 332             texCoords[textureNo] = NULL;
 333         }
 334     }
 335     if(sd->u.s.diffuse.lpData) {
 336         diffuse = sd->u.s.diffuse.lpData + streamOffset[sd->u.s.diffuse.streamNo];
 337     }
 338     if(sd->u.s.specular.lpData) {
 339         specular = sd->u.s.specular.lpData + streamOffset[sd->u.s.specular.streamNo];
 340     }
 341     if(sd->u.s.normal.lpData) {
 342         normal = sd->u.s.normal.lpData + streamOffset[sd->u.s.normal.streamNo];
 343     }
 344     if(sd->u.s.position.lpData) {
 345         position = sd->u.s.position.lpData + streamOffset[sd->u.s.position.streamNo];
 346     }
 347
 348     /* The texture coordinate types are not so easy to map into a common function signature - we're
 349      * not using the vector functions here
 350      */
 351     if(FIXME_ON(d3d_draw)) {
 352         for (textureNo = 0; textureNo < GL_LIMITS(textures); ++textureNo) {
 353             DWORD type = sd->u.s.texCoords[textureNo].dwType;
 354             if (sd->u.s.texCoords[textureNo].lpData &&
 355                 type != WINED3DDECLTYPE_FLOAT1 &&
 356                 type != WINED3DDECLTYPE_FLOAT2 &&
 357                 type != WINED3DDECLTYPE_FLOAT3 &&
 358                 type != WINED3DDECLTYPE_FLOAT4) {
 359                 FIXME("Implement fixed function texture coordinates from %s\n", debug_d3ddecltype(type));
 360             }
 361         }
 362         if(specular && This->stateBlock->renderState[WINED3DRS_FOGENABLE] &&
 363            (This->stateBlock->renderState[WINED3DRS_FOGVERTEXMODE] == WINED3DFOG_NONE || sd->u.s.position_transformed )&&
 364            This->stateBlock->renderState[WINED3DRS_FOGTABLEMODE] == WINED3DFOG_NONE) {
 365             if(GL_SUPPORT(EXT_FOG_COORD) && sd->u.s.specular.dwType != WINED3DDECLTYPE_D3DCOLOR) {
 366                 FIXME("Implement fog coordinates from %s\n", debug_d3ddecltype(sd->u.s.specular.dwType));
 367             }
 368         }
 369         if(This->activeContext->num_untracked_materials && sd->u.s.diffuse.dwType != WINED3DDECLTYPE_D3DCOLOR) {
 370             FIXME("Implement diffuse color tracking from %s\n", debug_d3ddecltype(sd->u.s.diffuse.dwType));
 371         }
 372     }
 373
 374     /* Start drawing in GL */
 375     VTRACE(("glBegin(%x)\n", glPrimType));
 376     glBegin(glPrimType);
 377
 378     /* Default settings for data that is not passed */
 379     if (sd->u.s.normal.lpData == NULL) {
 380         glNormal3f(0, 0, 0);
 381     }
 382     if(sd->u.s.diffuse.lpData == NULL) {
 383         glColor4f(1.0f, 1.0f, 1.0f, 1.0f);
 384     }
 385     if(sd->u.s.specular.lpData == NULL) {
 386         if (GL_SUPPORT(EXT_SECONDARY_COLOR)) {
 387             GL_EXTCALL(glSecondaryColor3fEXT)(0, 0, 0);
 388         }
 389     }
 390
 391     /* We shouldn't start this function if any VBO is involved. Should I put a safety check here?
 392      * Guess it's not necessary(we crash then anyway) and would only eat CPU time
 393      */
 394
 395     /* For each primitive */
 396     for (vx_index = 0; vx_index < NumVertexes; ++vx_index) {
 397
 398         /* Initialize diffuse color */
 399         diffuseColor = 0xFFFFFFFF;
 400
 401         /* Blending data and Point sizes are not supported by this function. They are not supported by the fixed
 402          * function pipeline at all. A Fixme for them is printed after decoding the vertex declaration
 403          */
 404
 405         /* For indexed data, we need to go a few more strides in */
 406         if (idxData != NULL) {
 407
 408             /* Indexed so work out the number of strides to skip */
 409             if (idxSize == 2) {
 410                 VTRACE(("Idx for vertex %d = %d\n", vx_index, pIdxBufS[startIdx+vx_index]));
 411                 SkipnStrides = pIdxBufS[startIdx + vx_index] + This->stateBlock->loadBaseVertexIndex;
 412             } else {
 413                 VTRACE(("Idx for vertex %d = %d\n", vx_index, pIdxBufL[startIdx+vx_index]));
 414                 SkipnStrides = pIdxBufL[startIdx + vx_index] + This->stateBlock->loadBaseVertexIndex;
 415             }
 416         }
 417
 418         /* Texture coords --------------------------- */
 419         for (textureNo = 0; textureNo < GL_LIMITS(texture_stages); ++textureNo) {
 420
 421             if (!GL_SUPPORT(ARB_MULTITEXTURE) && textureNo > 0) {
 422                 FIXME("Program using multiple concurrent textures which this opengl implementation doesn't support\n");
 423                 continue ;
 424             }
 425
 426             /* Query tex coords */
 427             if (This->stateBlock->textures[textureNo] != NULL || pixelShader) {
 428
 429                 int    coordIdx = This->stateBlock->textureState[textureNo][WINED3DTSS_TEXCOORDINDEX];
 430                 int texture_idx = This->texUnitMap[textureNo];
 431                 float *ptrToCoords = NULL;
 432                 float  s = 0.0, t = 0.0, r = 0.0, q = 0.0;
 433
 434                 if (coordIdx > 7) {
 435                     VTRACE(("tex: %d - Skip tex coords, as being system generated\n", textureNo));
 436                     continue;
 437                 } else if (coordIdx < 0) {
 438                     FIXME("tex: %d - Coord index %d is less than zero, expect a crash.\n", textureNo, coordIdx);
 439                     continue;
 440                 }
 441
 442                 ptrToCoords = (float *)(texCoords[coordIdx] + (SkipnStrides * sd->u.s.texCoords[coordIdx].dwStride));
 443                 if (texCoords[coordIdx] == NULL) {
 444                     TRACE("tex: %d - Skipping tex coords, as no data supplied\n", textureNo);
 445                     if (GL_SUPPORT(ARB_MULTITEXTURE)) {
 446                         GL_EXTCALL(glMultiTexCoord4fARB(GL_TEXTURE0_ARB + texture_idx, 0, 0, 0, 1));
 447                     } else {
 448                         glTexCoord4f(0, 0, 0, 1);
 449                     }
 450                     continue;
 451                 } else {
 452                     int coordsToUse = sd->u.s.texCoords[coordIdx].dwType + 1; /* 0 == WINED3DDECLTYPE_FLOAT1 etc */
 453
 454                     if (texture_idx == -1) continue;
 455
 456                     /* The coords to supply depend completely on the fvf / vertex shader */
 457                     switch (coordsToUse) {
 458                     case 4: q = ptrToCoords[3]; /* drop through */
 459                     case 3: r = ptrToCoords[2]; /* drop through */
 460                     case 2: t = ptrToCoords[1]; /* drop through */
 461                     case 1: s = ptrToCoords[0];
 462                     }
 463
 464                     switch (coordsToUse) {   /* Supply the provided texture coords */
 465                     case WINED3DTTFF_COUNT1:
 466                         VTRACE(("tex:%d, s=%f\n", textureNo, s));
 467                         if (GL_SUPPORT(ARB_MULTITEXTURE)) {
 468                             GL_EXTCALL(glMultiTexCoord1fARB(GL_TEXTURE0_ARB + texture_idx, s));
 469                         } else {
 470                             glTexCoord1f(s);
 471                         }
 472                         break;
 473                     case WINED3DTTFF_COUNT2:
 474                         VTRACE(("tex:%d, s=%f, t=%f\n", textureNo, s, t));
 475                         if (GL_SUPPORT(ARB_MULTITEXTURE)) {
 476                             GL_EXTCALL(glMultiTexCoord2fARB(GL_TEXTURE0_ARB + texture_idx, s, t));
 477                         } else {
 478                             glTexCoord2f(s, t);
 479                         }
 480                         break;
 481                     case WINED3DTTFF_COUNT3:
 482                         VTRACE(("tex:%d, s=%f, t=%f, r=%f\n", textureNo, s, t, r));
 483                         if (GL_SUPPORT(ARB_MULTITEXTURE)) {
 484                             GL_EXTCALL(glMultiTexCoord3fARB(GL_TEXTURE0_ARB + texture_idx, s, t, r));
 485                         } else {
 486                             glTexCoord3f(s, t, r);
 487                         }
 488                         break;
 489                     case WINED3DTTFF_COUNT4:
 490                         VTRACE(("tex:%d, s=%f, t=%f, r=%f, q=%f\n", textureNo, s, t, r, q));
 491                         if (GL_SUPPORT(ARB_MULTITEXTURE)) {
 492                             GL_EXTCALL(glMultiTexCoord4fARB(GL_TEXTURE0_ARB + texture_idx, s, t, r, q));
 493                         } else {
 494                             glTexCoord4f(s, t, r, q);
 495                         }
 496                         break;
 497                     default:
 498                         FIXME("Should not get here as coordsToUse is two bits only (%x)!\n", coordsToUse);
 499                     }
 500                 }
 501             }
 502         } /* End of textures */
 503
 504         /* Diffuse -------------------------------- */
 505         if (diffuse) {
 506             DWORD *ptrToCoords = (DWORD *)(diffuse + (SkipnStrides * sd->u.s.diffuse.dwStride));
 507
 508             diffuse_funcs[sd->u.s.diffuse.dwType]((void *) ptrToCoords);
 509             if(This->activeContext->num_untracked_materials) {
 510                 unsigned char i;
 511                 float color[4];
 512
 513                 diffuseColor = ptrToCoords[0];
 514                 color[0] = D3DCOLOR_B_R(diffuseColor) / 255.0;
 515                 color[1] = D3DCOLOR_B_G(diffuseColor) / 255.0;
 516                 color[2] = D3DCOLOR_B_B(diffuseColor) / 255.0;
 517                 color[3] = D3DCOLOR_B_A(diffuseColor) / 255.0;
 518
 519                 for(i = 0; i < This->activeContext->num_untracked_materials; i++) {
 520                     glMaterialfv(GL_FRONT_AND_BACK, This->activeContext->untracked_materials[i], color);
 521                 }
 522             }
 523         }
 524
 525         /* Specular ------------------------------- */
 526         if (specular) {
 527             DWORD *ptrToCoords = (DWORD *)(specular + (SkipnStrides * sd->u.s.specular.dwStride));
 528
 529             /* special case where the fog density is stored in the specular alpha channel */
 530             if(This->stateBlock->renderState[WINED3DRS_FOGENABLE] &&
 531               (This->stateBlock->renderState[WINED3DRS_FOGVERTEXMODE] == WINED3DFOG_NONE || sd->u.s.position.dwType == WINED3DDECLTYPE_FLOAT4 )&&
 532               This->stateBlock->renderState[WINED3DRS_FOGTABLEMODE] == WINED3DFOG_NONE) {
 533                 if(GL_SUPPORT(EXT_FOG_COORD)) {
 534                     specularColor = ptrToCoords[0];
 535                     GL_EXTCALL(glFogCoordfEXT(specularColor >> 24));
 536                 } else {
 537                     static BOOL warned = FALSE;
 538                     if(!warned) {
 539                         /* TODO: Use the fog table code from old ddraw */
 540                         FIXME("Implement fog for transformed vertices in software\n");
 541                         warned = TRUE;
 542                     }
 543                 }
 544             }
 545
 546             specular_funcs[sd->u.s.specular.dwType]((void *) ptrToCoords);
 547         }
 548
 549         /* Normal -------------------------------- */
 550         if (normal != NULL) {
 551             float *ptrToCoords = (float *)(normal + (SkipnStrides * sd->u.s.normal.dwStride));
 552             normal_funcs[sd->u.s.normal.dwType](ptrToCoords);
 553         }
 554
 555         /* Position -------------------------------- */
 556         if (position) {
 557             float *ptrToCoords = (float *)(position + (SkipnStrides * sd->u.s.position.dwStride));
 558             position_funcs[sd->u.s.position.dwType](ptrToCoords);
 559         }
 560
 561         /* For non indexed mode, step onto next parts */
 562         if (idxData == NULL) {
 563             ++SkipnStrides;
 564         }
 565     }
 566
 567     glEnd();
 568     checkGLcall("glEnd and previous calls");
 569 }
 570
 571 static inline void send_attribute(IWineD3DDeviceImpl *This, const DWORD type, const UINT index, const void *ptr) {
 572     switch(type) {
 573         case WINED3DDECLTYPE_FLOAT1:
 574             GL_EXTCALL(glVertexAttrib1fvARB(index, (float *) ptr));
 575             break;
 576         case WINED3DDECLTYPE_FLOAT2:
 577             GL_EXTCALL(glVertexAttrib2fvARB(index, (float *) ptr));
 578             break;
 579         case WINED3DDECLTYPE_FLOAT3:
 580             GL_EXTCALL(glVertexAttrib3fvARB(index, (float *) ptr));
 581             break;
 582         case WINED3DDECLTYPE_FLOAT4:
 583             GL_EXTCALL(glVertexAttrib4fvARB(index, (float *) ptr));
 584             break;
 585
 586         case WINED3DDECLTYPE_UBYTE4:
 587             GL_EXTCALL(glVertexAttrib4ubvARB(index, ptr));
 588             break;
 589         case WINED3DDECLTYPE_UBYTE4N:
 590         case WINED3DDECLTYPE_D3DCOLOR:
 591             GL_EXTCALL(glVertexAttrib4NubvARB(index, ptr));
 592             break;
 593
 594         case WINED3DDECLTYPE_SHORT2:
 595             GL_EXTCALL(glVertexAttrib4svARB(index, (GLshort *) ptr));
 596             break;
 597         case WINED3DDECLTYPE_SHORT4:
 598             GL_EXTCALL(glVertexAttrib4svARB(index, (GLshort *) ptr));
 599             break;
 600
 601         case WINED3DDECLTYPE_SHORT2N:
 602         {
 603             GLshort s[4] = {((short *) ptr)[0], ((short *) ptr)[1], 0, 1};
 604             GL_EXTCALL(glVertexAttrib4NsvARB(index, s));
 605             break;
 606         }
 607         case WINED3DDECLTYPE_USHORT2N:
 608         {
 609             GLushort s[4] = {((unsigned short *) ptr)[0], ((unsigned short *) ptr)[1], 0, 1};
 610             GL_EXTCALL(glVertexAttrib4NusvARB(index, s));
 611             break;
 612         }
 613         case WINED3DDECLTYPE_SHORT4N:
 614             GL_EXTCALL(glVertexAttrib4NsvARB(index, (GLshort *) ptr));
 615             break;
 616         case WINED3DDECLTYPE_USHORT4N:
 617             GL_EXTCALL(glVertexAttrib4NusvARB(index, (GLushort *) ptr));
 618             break;
 619
 620         case WINED3DDECLTYPE_UDEC3:
 621             FIXME("Unsure about WINED3DDECLTYPE_UDEC3\n");
 622             /*glVertexAttrib3usvARB(instancedData[j], (GLushort *) ptr); Does not exist */
 623             break;
 624         case WINED3DDECLTYPE_DEC3N:
 625             FIXME("Unsure about WINED3DDECLTYPE_DEC3N\n");
 626             /*glVertexAttrib3NusvARB(instancedData[j], (GLushort *) ptr); Does not exist */
 627             break;
 628
 629         case WINED3DDECLTYPE_FLOAT16_2:
 630             /* Are those 16 bit floats. C doesn't have a 16 bit float type. I could read the single bits and calculate a 4
 631              * byte float according to the IEEE standard
 632              */
 633             if (GL_SUPPORT(NV_HALF_FLOAT)) {
 634                 GL_EXTCALL(glVertexAttrib2hvNV(index, (GLhalfNV *)ptr));
 635             } else {
 636                 float x = float_16_to_32(((unsigned short *) ptr) + 0);
 637                 float y = float_16_to_32(((unsigned short *) ptr) + 1);
 638                 GL_EXTCALL(glVertexAttrib2fARB(index, x, y));
 639             }
 640             break;
 641         case WINED3DDECLTYPE_FLOAT16_4:
 642             if (GL_SUPPORT(NV_HALF_FLOAT)) {
 643                 GL_EXTCALL(glVertexAttrib4hvNV(index, (GLhalfNV *)ptr));
 644             } else {
 645                 float x = float_16_to_32(((unsigned short *) ptr) + 0);
 646                 float y = float_16_to_32(((unsigned short *) ptr) + 1);
 647                 float z = float_16_to_32(((unsigned short *) ptr) + 2);
 648                 float w = float_16_to_32(((unsigned short *) ptr) + 3);
 649                 GL_EXTCALL(glVertexAttrib4fARB(index, x, y, z, w));
 650             }
 651             break;
 652
 653         case WINED3DDECLTYPE_UNUSED:
 654         default:
 655             ERR("Unexpected attribute declaration: %d\n", type);
 656             break;
 657     }
 658 }
 659
 660 static void drawStridedSlowVs(IWineD3DDevice *iface, WineDirect3DVertexStridedData *sd, UINT numberOfVertices,
 661                               GLenum glPrimitiveType, const void *idxData, short idxSize, ULONG minIndex, ULONG startIdx,
 662                               ULONG startVertex) {
 663
 664     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *) iface;
 665     long                      SkipnStrides = startVertex + This->stateBlock->loadBaseVertexIndex;
 666     const WORD                *pIdxBufS     = NULL;
 667     const DWORD               *pIdxBufL     = NULL;
 668     LONG                       vx_index;
 669     int i;
 670     IWineD3DStateBlockImpl *stateblock = This->stateBlock;
 671     BYTE *ptr;
 672
 673     if (idxSize != 0) {
 674         /* Immediate mode drawing can't make use of indices in a vbo - get the data from the index buffer.
 675          * If the index buffer has no vbo(not supported or other reason), or with user pointer drawing
 676          * idxData will be != NULL
 677          */
 678         if(idxData == NULL) {
 679             idxData = ((IWineD3DIndexBufferImpl *) stateblock->pIndexData)->resource.allocatedMemory;
 680         }
 681
 682         if (idxSize == 2) pIdxBufS = (const WORD *) idxData;
 683         else pIdxBufL = (const DWORD *) idxData;
 684     }
 685
 686     /* Start drawing in GL */
 687     VTRACE(("glBegin(%x)\n", glPrimType));
 688     glBegin(glPrimitiveType);
 689
 690     for (vx_index = 0; vx_index < numberOfVertices; ++vx_index) {
 691         if (idxData != NULL) {
 692
 693             /* Indexed so work out the number of strides to skip */
 694             if (idxSize == 2) {
 695                 VTRACE(("Idx for vertex %d = %d\n", vx_index, pIdxBufS[startIdx+vx_index]));
 696                 SkipnStrides = pIdxBufS[startIdx + vx_index] + stateblock->loadBaseVertexIndex;
 697             } else {
 698                 VTRACE(("Idx for vertex %d = %d\n", vx_index, pIdxBufL[startIdx+vx_index]));
 699                 SkipnStrides = pIdxBufL[startIdx + vx_index] + stateblock->loadBaseVertexIndex;
 700             }
 701         }
 702
 703         for(i = MAX_ATTRIBS - 1; i >= 0; i--) {
 704             if(!sd->u.input[i].lpData) continue;
 705
 706             ptr = sd->u.input[i].lpData +
 707                   sd->u.input[i].dwStride * SkipnStrides +
 708                   stateblock->streamOffset[sd->u.input[i].streamNo];
 709
 710             send_attribute(This, sd->u.input[i].dwType, i, ptr);
 711         }
 712         SkipnStrides++;
 713     }
 714
 715     glEnd();
 716 }
 717
 718 static void depth_blt(IWineD3DDevice *iface, GLuint texture) {
 719     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
 720     GLint old_binding = 0;
 721
 722     glPushAttrib(GL_ENABLE_BIT | GL_DEPTH_BUFFER_BIT | GL_COLOR_BUFFER_BIT);
 723
 724     glDisable(GL_CULL_FACE);
 725     glEnable(GL_BLEND);
 726     glDisable(GL_ALPHA_TEST);
 727     glDisable(GL_SCISSOR_TEST);
 728     glDisable(GL_STENCIL_TEST);
 729     glEnable(GL_DEPTH_TEST);
 730     glDepthFunc(GL_ALWAYS);
 731     glBlendFunc(GL_ZERO, GL_ONE);
 732
 733     GL_EXTCALL(glActiveTextureARB(GL_TEXTURE0_ARB));
 734     glGetIntegerv(GL_TEXTURE_BINDING_2D, &old_binding);
 735     glBindTexture(GL_TEXTURE_2D, texture);
 736     glEnable(GL_TEXTURE_2D);
 737
 738     This->shader_backend->shader_select_depth_blt(iface);
 739
 740     glBegin(GL_TRIANGLE_STRIP);
 741     glVertex2f(-1.0f, -1.0f);
 742     glVertex2f(1.0f, -1.0f);
 743     glVertex2f(-1.0f, 1.0f);
 744     glVertex2f(1.0f, 1.0f);
 745     glEnd();
 746
 747     glBindTexture(GL_TEXTURE_2D, old_binding);
 748
 749     glPopAttrib();
 750
 751     /* Reselect the old shaders. There doesn't seem to be any glPushAttrib bit for arb shaders,
 752      * and this seems easier and more efficient than providing the shader backend with a private
 753      * storage to read and restore the old shader settings
 754      */
 755     This->shader_backend->shader_select(iface, use_ps(This), use_vs(This));
 756 }
 757
 758 void depth_copy(IWineD3DDevice *iface) {
 759     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
 760     IWineD3DSurfaceImpl *depth_stencil = (IWineD3DSurfaceImpl *)This->auto_depth_stencil_buffer;
 761
 762     /* Only copy the depth buffer if there is one. */
 763     if (!depth_stencil) return;
 764
 765     /* TODO: Make this work for modes other than FBO */
 766     if (wined3d_settings.offscreen_rendering_mode != ORM_FBO) return;
 767
 768     if (depth_stencil->current_renderbuffer) {
 769         FIXME("Not supported with fixed up depth stencil\n");
 770         return;
 771     }
 772
 773     if (This->render_offscreen) {
 774         GLint old_binding = 0;
 775
 776         TRACE("Copying onscreen depth buffer to offscreen surface\n");
 777
 778         if (!This->depth_blt_texture) {
 779             glGenTextures(1, &This->depth_blt_texture);
 780         }
 781
 782         /* Note that we use depth_blt here as well, rather than glCopyTexImage2D
 783          * directly on the FBO texture. That's because we need to flip. */
 784         GL_EXTCALL(glBindFramebufferEXT(GL_FRAMEBUFFER_EXT, 0));
 785         glGetIntegerv(GL_TEXTURE_BINDING_2D, &old_binding);
 786         glBindTexture(GL_TEXTURE_2D, This->depth_blt_texture);
 787         glCopyTexImage2D(depth_stencil->glDescription.target,
 788                 depth_stencil->glDescription.level,
 789                 depth_stencil->glDescription.glFormatInternal,
 790                 0,
 791                 0,
 792                 depth_stencil->currentDesc.Width,
 793                 depth_stencil->currentDesc.Height,
 794                 0);
 795         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
 796         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
 797         glTexParameteri(GL_TEXTURE_2D, GL_DEPTH_TEXTURE_MODE_ARB, GL_LUMINANCE);
 798         glBindTexture(GL_TEXTURE_2D, old_binding);
 799
 800         GL_EXTCALL(glBindFramebufferEXT(GL_FRAMEBUFFER_EXT, This->fbo));
 801         checkGLcall("glBindFramebuffer()");
 802         depth_blt(iface, This->depth_blt_texture);
 803         checkGLcall("depth_blt");
 804     } else {
 805         TRACE("Copying offscreen surface to onscreen depth buffer\n");
 806
 807         GL_EXTCALL(glBindFramebufferEXT(GL_FRAMEBUFFER_EXT, 0));
 808         checkGLcall("glBindFramebuffer()");
 809         depth_blt(iface, depth_stencil->glDescription.textureName);
 810         checkGLcall("depth_blt");
 811     }
 812 }
 813
 814 static inline void drawStridedInstanced(IWineD3DDevice *iface, WineDirect3DVertexStridedData *sd, UINT numberOfVertices,
 815                                  GLenum glPrimitiveType, const void *idxData, short idxSize, ULONG minIndex,
 816                                  ULONG startIdx, ULONG startVertex) {
 817     UINT numInstances = 0;
 818     int numInstancedAttribs = 0, i, j;
 819     UINT instancedData[sizeof(sd->u.input) / sizeof(sd->u.input[0]) /* 16 */];
 820     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *) iface;
 821     IWineD3DStateBlockImpl *stateblock = This->stateBlock;
 822
 823     if (idxSize == 0) {
 824         /* This is a nasty thing. MSDN says no hardware supports that and apps have to use software vertex processing.
 825          * We don't support this for now
 826          *
 827          * Shouldn't be too hard to support with opengl, in theory just call glDrawArrays instead of drawElements.
 828          * But the StreamSourceFreq value has a different meaning in that situation.
 829          */
 830         FIXME("Non-indexed instanced drawing is not supported\n");
 831         return;
 832     }
 833
 834     TRACE("(%p) : glElements(%x, %d, %d, ...)\n", This, glPrimitiveType, numberOfVertices, minIndex);
 835     idxData = idxData == (void *)-1 ? NULL : idxData;
 836
 837     /* First, figure out how many instances we have to draw */
 838     for(i = 0; i < MAX_STREAMS; i++) {
 839         /* Look at the streams and take the first one which matches */
 840         if(((stateblock->streamFlags[i] & WINED3DSTREAMSOURCE_INSTANCEDATA) || (stateblock->streamFlags[i] & WINED3DSTREAMSOURCE_INDEXEDDATA)) && stateblock->streamSource[i]) {
 841             /* D3D9 could set streamFreq 0 with (INSTANCEDATA or INDEXEDDATA) and then it is handled as 1. See d3d9/tests/visual.c-> stream_test() */
 842             if(stateblock->streamFreq[i] == 0){
 843                 numInstances = 1;
 844             } else {
 845                 numInstances = stateblock->streamFreq[i]; /* use the specified number of instances from the first matched stream. See d3d9/tests/visual.c-> stream_test() */
 846             }
 847             break; /* break, because only the first suitable value is interesting */
 848         }
 849     }
 850
 851     for(i = 0; i < sizeof(sd->u.input) / sizeof(sd->u.input[0]); i++) {
 852         if(stateblock->streamFlags[sd->u.input[i].streamNo] & WINED3DSTREAMSOURCE_INSTANCEDATA) {
 853             instancedData[numInstancedAttribs] = i;
 854             numInstancedAttribs++;
 855         }
 856     }
 857
 858     /* now draw numInstances instances :-) */
 859     for(i = 0; i < numInstances; i++) {
 860         /* Specify the instanced attributes using immediate mode calls */
 861         for(j = 0; j < numInstancedAttribs; j++) {
 862             BYTE *ptr = sd->u.input[instancedData[j]].lpData +
 863                         sd->u.input[instancedData[j]].dwStride * i +
 864                         stateblock->streamOffset[sd->u.input[instancedData[j]].streamNo];
 865             if(sd->u.input[instancedData[j]].VBO) {
 866                 IWineD3DVertexBufferImpl *vb = (IWineD3DVertexBufferImpl *) stateblock->streamSource[sd->u.input[instancedData[j]].streamNo];
 867                 ptr += (long) vb->resource.allocatedMemory;
 868             }
 869
 870             send_attribute(This, sd->u.input[instancedData[j]].dwType, instancedData[j], ptr);
 871         }
 872
 873         glDrawElements(glPrimitiveType, numberOfVertices, idxSize == 2 ? GL_UNSIGNED_SHORT : GL_UNSIGNED_INT,
 874                     (const char *)idxData+(idxSize * startIdx));
 875         checkGLcall("glDrawElements");
 876     }
 877 }
 878
 879 static inline void remove_vbos(IWineD3DDeviceImpl *This, WineDirect3DVertexStridedData *s) {
 880     unsigned char i;
 881     IWineD3DVertexBufferImpl *vb;
 882
 883     if(s->u.s.position.VBO) {
 884         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.position.streamNo];
 885         s->u.s.position.VBO = 0;
 886         s->u.s.position.lpData = (BYTE *) ((unsigned long) s->u.s.position.lpData + (unsigned long) vb->resource.allocatedMemory);
 887     }
 888     if(s->u.s.blendWeights.VBO) {
 889         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.blendWeights.streamNo];
 890         s->u.s.blendWeights.VBO = 0;
 891         s->u.s.blendWeights.lpData = (BYTE *) ((unsigned long) s->u.s.blendWeights.lpData + (unsigned long) vb->resource.allocatedMemory);
 892     }
 893     if(s->u.s.blendMatrixIndices.VBO) {
 894         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.blendMatrixIndices.streamNo];
 895         s->u.s.blendMatrixIndices.VBO = 0;
 896         s->u.s.blendMatrixIndices.lpData = (BYTE *) ((unsigned long) s->u.s.blendMatrixIndices.lpData + (unsigned long) vb->resource.allocatedMemory);
 897     }
 898     if(s->u.s.normal.VBO) {
 899         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.normal.streamNo];
 900         s->u.s.normal.VBO = 0;
 901         s->u.s.normal.lpData = (BYTE *) ((unsigned long) s->u.s.normal.lpData + (unsigned long) vb->resource.allocatedMemory);
 902     }
 903     if(s->u.s.pSize.VBO) {
 904         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.pSize.streamNo];
 905         s->u.s.pSize.VBO = 0;
 906         s->u.s.pSize.lpData = (BYTE *) ((unsigned long) s->u.s.pSize.lpData + (unsigned long) vb->resource.allocatedMemory);
 907     }
 908     if(s->u.s.diffuse.VBO) {
 909         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.diffuse.streamNo];
 910         s->u.s.diffuse.VBO = 0;
 911         s->u.s.diffuse.lpData = (BYTE *) ((unsigned long) s->u.s.diffuse.lpData + (unsigned long) vb->resource.allocatedMemory);
 912     }
 913     if(s->u.s.specular.VBO) {
 914         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.specular.streamNo];
 915         s->u.s.specular.VBO = 0;
 916         s->u.s.specular.lpData = (BYTE *) ((unsigned long) s->u.s.specular.lpData + (unsigned long) vb->resource.allocatedMemory);
 917     }
 918     for(i = 0; i < WINED3DDP_MAXTEXCOORD; i++) {
 919         if(s->u.s.texCoords[i].VBO) {
 920             vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.texCoords[i].streamNo];
 921             s->u.s.texCoords[i].VBO = 0;
 922             s->u.s.texCoords[i].lpData = (BYTE *) ((unsigned long) s->u.s.texCoords[i].lpData + (unsigned long) vb->resource.allocatedMemory);
 923         }
 924     }
 925     if(s->u.s.position2.VBO) {
 926         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.position2.streamNo];
 927         s->u.s.position2.VBO = 0;
 928         s->u.s.position2.lpData = (BYTE *) ((unsigned long) s->u.s.position2.lpData + (unsigned long) vb->resource.allocatedMemory);
 929     }
 930     if(s->u.s.normal2.VBO) {
 931         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.normal2.streamNo];
 932         s->u.s.normal2.VBO = 0;
 933         s->u.s.normal2.lpData = (BYTE *) ((unsigned long) s->u.s.normal2.lpData + (unsigned long) vb->resource.allocatedMemory);
 934     }
 935     if(s->u.s.tangent.VBO) {
 936         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.tangent.streamNo];
 937         s->u.s.tangent.VBO = 0;
 938         s->u.s.tangent.lpData = (BYTE *) ((unsigned long) s->u.s.tangent.lpData + (unsigned long) vb->resource.allocatedMemory);
 939     }
 940     if(s->u.s.binormal.VBO) {
 941         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.binormal.streamNo];
 942         s->u.s.binormal.VBO = 0;
 943         s->u.s.binormal.lpData = (BYTE *) ((unsigned long) s->u.s.binormal.lpData + (unsigned long) vb->resource.allocatedMemory);
 944     }
 945     if(s->u.s.tessFactor.VBO) {
 946         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.tessFactor.streamNo];
 947         s->u.s.tessFactor.VBO = 0;
 948         s->u.s.tessFactor.lpData = (BYTE *) ((unsigned long) s->u.s.tessFactor.lpData + (unsigned long) vb->resource.allocatedMemory);
 949     }
 950     if(s->u.s.fog.VBO) {
 951         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.fog.streamNo];
 952         s->u.s.fog.VBO = 0;
 953         s->u.s.fog.lpData = (BYTE *) ((unsigned long) s->u.s.fog.lpData + (unsigned long) vb->resource.allocatedMemory);
 954     }
 955     if(s->u.s.depth.VBO) {
 956         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.depth.streamNo];
 957         s->u.s.depth.VBO = 0;
 958         s->u.s.depth.lpData = (BYTE *) ((unsigned long) s->u.s.depth.lpData + (unsigned long) vb->resource.allocatedMemory);
 959     }
 960     if(s->u.s.sample.VBO) {
 961         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.sample.streamNo];
 962         s->u.s.sample.VBO = 0;
 963         s->u.s.sample.lpData = (BYTE *) ((unsigned long) s->u.s.sample.lpData + (unsigned long) vb->resource.allocatedMemory);
 964     }
 965 }
 966
 967 /* Routine common to the draw primitive and draw indexed primitive routines */
 968 void drawPrimitive(IWineD3DDevice *iface,
 969                    int PrimitiveType,
 970                    long NumPrimitives,
 971                    /* for Indexed: */
 972                    long  StartVertexIndex,
 973                    UINT  numberOfVertices,
 974                    long  StartIdx,
 975                    short idxSize,
 976                    const void *idxData,
 977                    int   minIndex) {
 978
 979     IWineD3DDeviceImpl           *This = (IWineD3DDeviceImpl *)iface;
 980     IWineD3DSwapChain            *swapchain;
 981     IWineD3DBaseTexture          *texture = NULL;
 982     IWineD3DSurfaceImpl          *target;
 983     int i;
 984
 985     if (NumPrimitives == 0) return;
 986
 987     /* Invalidate the back buffer memory so LockRect will read it the next time */
 988     for(i = 0; i < GL_LIMITS(buffers); i++) {
 989         target = (IWineD3DSurfaceImpl *) This->render_targets[i];
 990
 991         /* TODO: Only do all that if we're going to change anything
 992          * Texture container dirtification does not work quite right yet
 993          */
 994         if(target /*&& target->Flags & (SFLAG_INTEXTURE | SFLAG_INSYSMEM)*/) {
 995             swapchain = NULL;
 996             texture = NULL;
 997
 998             if(i == 0) {
 999                 IWineD3DSurface_GetContainer((IWineD3DSurface *) target, &IID_IWineD3DSwapChain, (void **)&swapchain);
1000
1001                 /* Need the surface in the drawable! */
1002                 IWineD3DSurface_LoadLocation((IWineD3DSurface *) target, SFLAG_INDRAWABLE, NULL);
1003
1004                 /* TODO: Move fbo logic to ModifyLocation */
1005                 IWineD3DSurface_ModifyLocation((IWineD3DSurface *) target, SFLAG_INDRAWABLE, TRUE);
1006                 if(swapchain) {
1007                     /* Onscreen target. Invalidate system memory copy and texture copy */
1008                     IWineD3DSwapChain_Release(swapchain);
1009                 } else if(wined3d_settings.offscreen_rendering_mode != ORM_FBO) {
1010                     /* Non-FBO target: Invalidate system copy, texture copy and dirtify the container */
1011                     /* TODO: Move container dirtification to ModifyLocation */
1012                     IWineD3DSurface_GetContainer((IWineD3DSurface *) target, &IID_IWineD3DBaseTexture, (void **)&texture);
1013
1014                     if(texture) {
1015                         IWineD3DBaseTexture_SetDirty(texture, TRUE);
1016                         IWineD3DTexture_Release(texture);
1017                     }
1018                 } else {
1019                     /* FBO offscreen target. Texture == Drawable */
1020                     target->Flags |= SFLAG_INTEXTURE;
1021                 }
1022             } else {
1023                 /* Must be an fbo render target */
1024                 IWineD3DSurface_ModifyLocation((IWineD3DSurface *) target, SFLAG_INDRAWABLE, TRUE);
1025                 target->Flags |=  SFLAG_INTEXTURE;
1026             }
1027         }
1028     }
1029
1030     /* Signals other modules that a drawing is in progress and the stateblock finalized */
1031     This->isInDraw = TRUE;
1032
1033     /* Ok, we will be updating the screen from here onwards so grab the lock */
1034
1035     if (wined3d_settings.offscreen_rendering_mode == ORM_FBO) {
1036         ENTER_GL();
1037         apply_fbo_state(iface);
1038         LEAVE_GL();
1039     }
1040
1041     ActivateContext(This, This->render_targets[0], CTXUSAGE_DRAWPRIM);
1042     ENTER_GL();
1043
1044     if (This->depth_copy_state == WINED3D_DCS_COPY) {
1045         depth_copy(iface);
1046     }
1047     This->depth_copy_state = WINED3D_DCS_INITIAL;
1048
1049     {
1050         GLenum glPrimType;
1051         BOOL emulation = FALSE;
1052         WineDirect3DVertexStridedData *strided = &This->strided_streams;
1053         WineDirect3DVertexStridedData stridedlcl;
1054         /* Ok, Work out which primitive is requested and how many vertexes that
1055            will be                                                              */
1056         UINT calculatedNumberOfindices = primitiveToGl(PrimitiveType, NumPrimitives, &glPrimType);
1057         if (numberOfVertices == 0 )
1058             numberOfVertices = calculatedNumberOfindices;
1059
1060         if(!use_vs(This)) {
1061             if(!This->strided_streams.u.s.position_transformed && This->activeContext->num_untracked_materials &&
1062                 This->stateBlock->renderState[WINED3DRS_LIGHTING]) {
1063                 static BOOL first = TRUE;
1064                 if(first) {
1065                     FIXME("Using software emulation because not all material properties could be tracked\n");
1066                     first = FALSE;
1067                 } else {
1068                     TRACE("Using software emulation because not all material properties could be tracked\n");
1069                 }
1070                 emulation = TRUE;
1071             }
1072             else if(This->activeContext->fog_coord && This->stateBlock->renderState[WINED3DRS_FOGENABLE]) {
1073                 /* Either write a pipeline replacement shader or convert the specular alpha from unsigned byte
1074                  * to a float in the vertex buffer
1075                  */
1076                 static BOOL first = TRUE;
1077                 if(first) {
1078                     FIXME("Using software emulation because manual fog coordinates are provided\n");
1079                     first = FALSE;
1080                 } else {
1081                     TRACE("Using software emulation because manual fog coordinates are provided\n");
1082                 }
1083                 emulation = TRUE;
1084             }
1085
1086             if(emulation) {
1087                 strided = &stridedlcl;
1088                 memcpy(&stridedlcl, &This->strided_streams, sizeof(stridedlcl));
1089                 remove_vbos(This, &stridedlcl);
1090             }
1091         }
1092
1093         if (This->useDrawStridedSlow || emulation) {
1094             /* Immediate mode drawing */
1095             if(use_vs(This)) {
1096                 static BOOL first = TRUE;
1097                 if(first) {
1098                     FIXME("Using immediate mode with vertex shaders for half float emulation\n");
1099                     first = FALSE;
1100                 } else {
1101                     TRACE("Using immediate mode with vertex shaders for half float emulation\n");
1102                 }
1103                 drawStridedSlowVs(iface, strided, calculatedNumberOfindices, glPrimType,
1104                                   idxData, idxSize, minIndex, StartIdx, StartVertexIndex);
1105             } else {
1106                 drawStridedSlow(iface, strided, calculatedNumberOfindices,
1107                                 glPrimType, idxData, idxSize, minIndex, StartIdx, StartVertexIndex);
1108             }
1109         } else if(This->instancedDraw) {
1110             /* Instancing emulation with mixing immediate mode and arrays */
1111             drawStridedInstanced(iface, &This->strided_streams, calculatedNumberOfindices, glPrimType,
1112                             idxData, idxSize, minIndex, StartIdx, StartVertexIndex);
1113         } else {
1114             drawStridedFast(iface, calculatedNumberOfindices, glPrimType,
1115                             idxData, idxSize, minIndex, StartIdx, StartVertexIndex);
1116         }
1117     }
1118
1119     /* Finished updating the screen, restore lock */
1120     LEAVE_GL();
1121     TRACE("Done all gl drawing\n");
1122
1123     /* Diagnostics */
1124 #ifdef SHOW_FRAME_MAKEUP
1125     {
1126         static long int primCounter = 0;
1127         /* NOTE: set primCounter to the value reported by drawprim
1128            before you want to to write frame makeup to /tmp */
1129         if (primCounter >= 0) {
1130             WINED3DLOCKED_RECT r;
1131             char buffer[80];
1132             IWineD3DSurface_LockRect(This->renderTarget, &r, NULL, WINED3DLOCK_READONLY);
1133             sprintf(buffer, "/tmp/backbuffer_%d.tga", primCounter);
1134             TRACE("Saving screenshot %s\n", buffer);
1135             IWineD3DSurface_SaveSnapshot(This->renderTarget, buffer);
1136             IWineD3DSurface_UnlockRect(This->renderTarget);
1137
1138 #ifdef SHOW_TEXTURE_MAKEUP
1139            {
1140             IWineD3DSurface *pSur;
1141             int textureNo;
1142             for (textureNo = 0; textureNo < MAX_COMBINED_SAMPLERS; ++textureNo) {
1143                 if (This->stateBlock->textures[textureNo] != NULL) {
1144                     sprintf(buffer, "/tmp/texture_%p_%d_%d.tga", This->stateBlock->textures[textureNo], primCounter, textureNo);
1145                     TRACE("Saving texture %s\n", buffer);
1146                     if (IWineD3DBaseTexture_GetType(This->stateBlock->textures[textureNo]) == WINED3DRTYPE_TEXTURE) {
1147                             IWineD3DTexture_GetSurfaceLevel((IWineD3DTexture *)This->stateBlock->textures[textureNo], 0, &pSur);
1148                             IWineD3DSurface_SaveSnapshot(pSur, buffer);
1149                             IWineD3DSurface_Release(pSur);
1150                     } else  {
1151                         FIXME("base Texture isn't of type texture %d\n", IWineD3DBaseTexture_GetType(This->stateBlock->textures[textureNo]));
1152                     }
1153                 }
1154             }
1155            }
1156 #endif
1157         }
1158         TRACE("drawprim #%d\n", primCounter);
1159         ++primCounter;
1160     }
1161 #endif
1162
1163     /* Control goes back to the device, stateblock values may change again */
1164     This->isInDraw = FALSE;
1165 }
1166
1167 static void normalize_normal(float *n) {
1168     float length = n[0] * n[0] + n[1] * n[1] + n[2] * n[2];
1169     if(length == 0.0) return;
1170     length = sqrt(length);
1171     n[0] = n[0] / length;
1172     n[1] = n[1] / length;
1173     n[2] = n[2] / length;
1174 }
1175
1176 /* Tesselates a high order rectangular patch into single triangles using gl evaluators
1177  *
1178  * The problem is that OpenGL does not offer a direct way to return the tesselated primitives,
1179  * and they can't be sent off for rendering directly either. Tesselating is slow, so we want
1180  * to cache the patches in a vertex buffer. But more importantly, gl can't bind generated
1181  * attributes to numbered shader attributes, so we have to store them and rebind them as needed
1182  * in drawprim.
1183  *
1184  * To read back, the opengl feedback mode is used. This creates a proplem because we want
1185  * untransformed, unlit vertices, but feedback runs everything through transform and lighting.
1186  * Thus disable lighting and set identity matrices to get unmodified colors and positions.
1187  * To overcome clipping find the biggest x, y and z values of the vertices in the patch and scale
1188  * them to [-1.0;+1.0] and set the viewport up to scale them back.
1189  *
1190  * Normals are more tricky: Draw white vertices with 3 directional lights, and calculate the
1191  * resulting colors back to the normals.
1192  *
1193  * NOTE: This function activates a context for blitting, modifies matrices & viewport, but
1194  * does not restore it because normally a draw follows immediately afterwards. The caller is
1195  * responsible of taking care that either the gl states are restored, or the context activated
1196  * for drawing to reset the lastWasBlit flag.
1197  */
1198 HRESULT tesselate_rectpatch(IWineD3DDeviceImpl *This,
1199                             struct WineD3DRectPatch *patch) {
1200     unsigned int i, j, num_quads, out_vertex_size, buffer_size, d3d_out_vertex_size;
1201     float max_x = 0.0, max_y = 0.0, max_z = 0.0, neg_z = 0.0;
1202     WineDirect3DVertexStridedData strided;
1203     BYTE *data;
1204     WINED3DRECTPATCH_INFO *info = &patch->RectPatchInfo;
1205     DWORD vtxStride;
1206     GLenum feedback_type;
1207     GLfloat *feedbuffer;
1208
1209     /* First, locate the position data. This is provided in a vertex buffer in the stateblock.
1210      * Beware of vbos
1211      */
1212     memset(&strided, 0, sizeof(strided));
1213     primitiveDeclarationConvertToStridedData((IWineD3DDevice *) This, FALSE, &strided, NULL);
1214     if(strided.u.s.position.VBO) {
1215         IWineD3DVertexBufferImpl *vb;
1216         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[strided.u.s.position.streamNo];
1217         strided.u.s.position.lpData = (BYTE *) ((unsigned long) strided.u.s.position.lpData +
1218                                                 (unsigned long) vb->resource.allocatedMemory);
1219     }
1220     vtxStride = strided.u.s.position.dwStride;
1221     data = strided.u.s.position.lpData +
1222            vtxStride * info->Stride * info->StartVertexOffsetHeight +
1223            vtxStride * info->StartVertexOffsetWidth;
1224
1225     /* Not entirely sure about what happens with transformed vertices */
1226     if(strided.u.s.position_transformed) {
1227         FIXME("Transformed position in rectpatch generation\n");
1228     }
1229     if(vtxStride % sizeof(GLfloat)) {
1230         /* glMap2f reads vertex sizes in GLfloats, the d3d stride is in bytes.
1231          * I don't see how the stride could not be a multiple of 4, but make sure
1232          * to check it
1233          */
1234         ERR("Vertex stride is not a multiple of sizeof(GLfloat)\n");
1235     }
1236     if(info->Basis != WINED3DBASIS_BEZIER) {
1237         FIXME("Basis is %s, how to handle this?\n", debug_d3dbasis(info->Basis));
1238     }
1239     if(info->Degree != WINED3DDEGREE_CUBIC) {
1240         FIXME("Degree is %s, how to handle this?\n", debug_d3ddegree(info->Degree));
1241     }
1242
1243     /* First, get the boundary cube of the input data */
1244     for(j = 0; j < info->Height; j++) {
1245         for(i = 0; i < info->Width; i++) {
1246             float *v = (float *) (data + vtxStride * i + vtxStride * info->Stride * j);
1247             if(fabs(v[0]) > max_x) max_x = fabs(v[0]);
1248             if(fabs(v[1]) > max_y) max_y = fabs(v[1]);
1249             if(fabs(v[2]) > max_z) max_z = fabs(v[2]);
1250             if(v[2] < neg_z) neg_z = v[2];
1251         }
1252     }
1253
1254     /* This needs some improvements in the vertex decl code */
1255     FIXME("Cannot find data to generate. Only generating position and normals\n");
1256     patch->has_normals = TRUE;
1257     patch->has_texcoords = FALSE;
1258
1259     /* Simply activate the context for blitting. This disables all the things we don't want and
1260      * takes care of dirtifying. Dirtifying is preferred over pushing / popping, since drawing the
1261      * patch (as opposed to normal draws) will most likely need different changes anyway
1262      */
1263     ActivateContext(This, This->lastActiveRenderTarget, CTXUSAGE_BLIT);
1264     ENTER_GL();
1265
1266     glMatrixMode(GL_PROJECTION);
1267     checkGLcall("glMatrixMode(GL_PROJECTION)");
1268     glLoadIdentity();
1269     checkGLcall("glLoadIndentity()");
1270     glScalef(1 / (max_x) , 1 / (max_y), max_z == 0 ? 1 : 1 / ( 2 * max_z));
1271     glTranslatef(0, 0, 0.5);
1272     checkGLcall("glScalef");
1273     glViewport(-max_x, -max_y, 2 * (max_x), 2 * (max_y));
1274     checkGLcall("glViewport");
1275
1276     /* Some states to take care of. If we're in wireframe opengl will produce lines, and confuse
1277      * our feedback buffer parser
1278      */
1279     glPolygonMode(GL_FRONT_AND_BACK, GL_FILL);
1280     checkGLcall("glPolygonMode(GL_FRONT_AND_BACK, GL_FILL)");
1281     IWineD3DDeviceImpl_MarkStateDirty(This, STATE_RENDER(WINED3DRS_FILLMODE));
1282     if(patch->has_normals) {
1283         float black[4] = {0, 0, 0, 0};
1284         float red[4]   = {1, 0, 0, 0};
1285         float green[4] = {0, 1, 0, 0};
1286         float blue[4]  = {0, 0, 1, 0};
1287         float white[4] = {1, 1, 1, 1};
1288         glEnable(GL_LIGHTING);
1289         checkGLcall("glEnable(GL_LIGHTING)");
1290         glLightModelfv(GL_LIGHT_MODEL_AMBIENT, black);
1291         checkGLcall("glLightModel for MODEL_AMBIENT");
1292         IWineD3DDeviceImpl_MarkStateDirty(This, STATE_RENDER(WINED3DRS_AMBIENT));
1293
1294         for(i = 3; i < GL_LIMITS(lights); i++) {
1295             glDisable(GL_LIGHT0 + i);
1296             checkGLcall("glDisable(GL_LIGHT0 + i)");
1297             IWineD3DDeviceImpl_MarkStateDirty(This, STATE_ACTIVELIGHT(i));
1298         }
1299
1300         IWineD3DDeviceImpl_MarkStateDirty(This, STATE_ACTIVELIGHT(0));
1301         glLightfv(GL_LIGHT0, GL_DIFFUSE, red);
1302         glLightfv(GL_LIGHT0, GL_SPECULAR, black);
1303         glLightfv(GL_LIGHT0, GL_AMBIENT, black);
1304         glLightfv(GL_LIGHT0, GL_POSITION, red);
1305         glEnable(GL_LIGHT0);
1306         checkGLcall("Setting up light 1\n");
1307         IWineD3DDeviceImpl_MarkStateDirty(This, STATE_ACTIVELIGHT(1));
1308         glLightfv(GL_LIGHT1, GL_DIFFUSE, green);
1309         glLightfv(GL_LIGHT1, GL_SPECULAR, black);
1310         glLightfv(GL_LIGHT1, GL_AMBIENT, black);
1311         glLightfv(GL_LIGHT1, GL_POSITION, green);
1312         glEnable(GL_LIGHT1);
1313         checkGLcall("Setting up light 2\n");
1314         IWineD3DDeviceImpl_MarkStateDirty(This, STATE_ACTIVELIGHT(2));
1315         glLightfv(GL_LIGHT2, GL_DIFFUSE, blue);
1316         glLightfv(GL_LIGHT2, GL_SPECULAR, black);
1317         glLightfv(GL_LIGHT2, GL_AMBIENT, black);
1318         glLightfv(GL_LIGHT2, GL_POSITION, blue);
1319         glEnable(GL_LIGHT2);
1320         checkGLcall("Setting up light 3\n");
1321
1322         IWineD3DDeviceImpl_MarkStateDirty(This, STATE_MATERIAL);
1323         IWineD3DDeviceImpl_MarkStateDirty(This, STATE_RENDER(WINED3DRS_COLORVERTEX));
1324         glDisable(GL_COLOR_MATERIAL);
1325         glMaterialfv(GL_FRONT_AND_BACK, GL_EMISSION, black);
1326         glMaterialfv(GL_FRONT_AND_BACK, GL_SPECULAR, black);
1327         glMaterialfv(GL_FRONT_AND_BACK, GL_DIFFUSE, white);
1328         checkGLcall("Setting up materials\n");
1329     }
1330
1331     /* Enable the needed maps.
1332      * GL_MAP2_VERTEX_3 is needed for positional data.
1333      * GL_AUTO_NORMAL to generate normals from the position. Do not use GL_MAP2_NORMAL.
1334      * GL_MAP2_TEXTURE_COORD_4 for texture coords
1335      */
1336     num_quads = ceilf(patch->numSegs[0]) * ceilf(patch->numSegs[1]);
1337     out_vertex_size = 3 /* position */;
1338     d3d_out_vertex_size = 3;
1339     glEnable(GL_MAP2_VERTEX_3);
1340     if(patch->has_normals && patch->has_texcoords) {
1341         FIXME("Texcoords not handled yet\n");
1342         feedback_type = GL_3D_COLOR_TEXTURE;
1343         out_vertex_size += 8;
1344         d3d_out_vertex_size += 7;
1345         glEnable(GL_AUTO_NORMAL);
1346         glEnable(GL_MAP2_TEXTURE_COORD_4);
1347     } else if(patch->has_texcoords) {
1348         FIXME("Texcoords not handled yet\n");
1349         feedback_type = GL_3D_COLOR_TEXTURE;
1350         out_vertex_size += 7;
1351         d3d_out_vertex_size += 4;
1352         glEnable(GL_MAP2_TEXTURE_COORD_4);
1353     } else if(patch->has_normals) {
1354         feedback_type = GL_3D_COLOR;
1355         out_vertex_size += 4;
1356         d3d_out_vertex_size += 3;
1357         glEnable(GL_AUTO_NORMAL);
1358     } else {
1359         feedback_type = GL_3D;
1360     }
1361     checkGLcall("glEnable vertex attrib generation");
1362
1363     buffer_size = num_quads * out_vertex_size * 2 /* triangle list */ * 3 /* verts per tri */
1364                    + 4 * num_quads /* 2 triangle markers per quad + num verts in tri */;
1365     feedbuffer = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, buffer_size * sizeof(float) * 8);
1366
1367     glMap2f(GL_MAP2_VERTEX_3,
1368             0, 1, vtxStride / sizeof(float), info->Width,
1369             0, 1, info->Stride * vtxStride / sizeof(float), info->Height,
1370             (float *) data);
1371     checkGLcall("glMap2f");
1372     if(patch->has_texcoords) {
1373         glMap2f(GL_MAP2_TEXTURE_COORD_4,
1374                 0, 1, vtxStride / sizeof(float), info->Width,
1375                 0, 1, info->Stride * vtxStride / sizeof(float), info->Height,
1376                 (float *) data);
1377         checkGLcall("glMap2f");
1378     }
1379     glMapGrid2f(ceilf(patch->numSegs[0]), 0.0, 1.0, ceilf(patch->numSegs[1]), 0.0, 1.0);
1380     checkGLcall("glMapGrid2f");
1381
1382     glFeedbackBuffer(buffer_size * 2, feedback_type, feedbuffer);
1383     checkGLcall("glFeedbackBuffer");
1384     glRenderMode(GL_FEEDBACK);
1385
1386     glEvalMesh2(GL_FILL, 0, ceilf(patch->numSegs[0]), 0, ceilf(patch->numSegs[1]));
1387     checkGLcall("glEvalMesh2\n");
1388
1389     i = glRenderMode(GL_RENDER);
1390     if(i == -1) {
1391         LEAVE_GL();
1392         ERR("Feedback failed. Expected %d elements back\n", buffer_size);
1393         Sleep(10000);
1394         HeapFree(GetProcessHeap(), 0, feedbuffer);
1395         return WINED3DERR_DRIVERINTERNALERROR;
1396     } else if(i != buffer_size) {
1397         LEAVE_GL();
1398         ERR("Unexpected amount of elements returned. Expected %d, got %d\n", buffer_size, i);
1399         Sleep(10000);
1400         HeapFree(GetProcessHeap(), 0, feedbuffer);
1401         return WINED3DERR_DRIVERINTERNALERROR;
1402     } else {
1403         TRACE("Got %d elements as expected\n", i);
1404     }
1405
1406     HeapFree(GetProcessHeap(), 0, patch->mem);
1407     patch->mem = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, num_quads * 6 * d3d_out_vertex_size * sizeof(float) * 8);
1408     i = 0;
1409     for(j = 0; j < buffer_size; j += (3 /* num verts */ * out_vertex_size + 2 /* tri marker */)) {
1410         if(feedbuffer[j] != GL_POLYGON_TOKEN) {
1411             ERR("Unexpected token: %f\n", feedbuffer[j]);
1412             continue;
1413         }
1414         if(feedbuffer[j + 1] != 3) {
1415             ERR("Unexpected polygon: %f corners\n", feedbuffer[j + 1]);
1416             continue;
1417         }
1418         /* Somehow there are different ideas about back / front facing, so fix up the
1419          * vertex order
1420          */
1421         patch->mem[i + 0] =  feedbuffer[j + out_vertex_size * 2 + 2]; /* x, triangle 2 */
1422         patch->mem[i + 1] =  feedbuffer[j + out_vertex_size * 2 + 3]; /* y, triangle 2 */
1423         patch->mem[i + 2] = (feedbuffer[j + out_vertex_size * 2 + 4] - 0.5) * 4 * max_z; /* z, triangle 3 */
1424         if(patch->has_normals) {
1425             patch->mem[i + 3] = feedbuffer[j + out_vertex_size * 2 + 5];
1426             patch->mem[i + 4] = feedbuffer[j + out_vertex_size * 2 + 6];
1427             patch->mem[i + 5] = feedbuffer[j + out_vertex_size * 2 + 7];
1428         }
1429         i += d3d_out_vertex_size;
1430
1431         patch->mem[i + 0] =  feedbuffer[j + out_vertex_size * 1 + 2]; /* x, triangle 2 */
1432         patch->mem[i + 1] =  feedbuffer[j + out_vertex_size * 1 + 3]; /* y, triangle 2 */
1433         patch->mem[i + 2] = (feedbuffer[j + out_vertex_size * 1 + 4] - 0.5) * 4 * max_z; /* z, triangle 2 */
1434         if(patch->has_normals) {
1435             patch->mem[i + 3] = feedbuffer[j + out_vertex_size * 1 + 5];
1436             patch->mem[i + 4] = feedbuffer[j + out_vertex_size * 1 + 6];
1437             patch->mem[i + 5] = feedbuffer[j + out_vertex_size * 1 + 7];
1438         }
1439         i += d3d_out_vertex_size;
1440
1441         patch->mem[i + 0] =  feedbuffer[j + out_vertex_size * 0 + 2]; /* x, triangle 1 */
1442         patch->mem[i + 1] =  feedbuffer[j + out_vertex_size * 0 + 3]; /* y, triangle 1 */
1443         patch->mem[i + 2] = (feedbuffer[j + out_vertex_size * 0 + 4] - 0.5) * 4 * max_z; /* z, triangle 1 */
1444         if(patch->has_normals) {
1445             patch->mem[i + 3] = feedbuffer[j + out_vertex_size * 0 + 5];
1446             patch->mem[i + 4] = feedbuffer[j + out_vertex_size * 0 + 6];
1447             patch->mem[i + 5] = feedbuffer[j + out_vertex_size * 0 + 7];
1448         }
1449         i += d3d_out_vertex_size;
1450     }
1451
1452     if(patch->has_normals) {
1453         /* Now do the same with reverse light directions */
1454         float x[4] = {-1,  0,  0, 0};
1455         float y[4] = { 0, -1,  0, 0};
1456         float z[4] = { 0,  0, -1, 0};
1457         glLightfv(GL_LIGHT0, GL_POSITION, x);
1458         glLightfv(GL_LIGHT1, GL_POSITION, y);
1459         glLightfv(GL_LIGHT2, GL_POSITION, z);
1460         checkGLcall("Setting up reverse light directions\n");
1461
1462         glRenderMode(GL_FEEDBACK);
1463         checkGLcall("glRenderMode(GL_FEEDBACK)");
1464         glEvalMesh2(GL_FILL, 0, ceilf(patch->numSegs[0]), 0, ceilf(patch->numSegs[1]));
1465         checkGLcall("glEvalMesh2\n");
1466         i = glRenderMode(GL_RENDER);
1467         checkGLcall("glRenderMode(GL_RENDER)");
1468
1469         i = 0;
1470         for(j = 0; j < buffer_size; j += (3 /* num verts */ * out_vertex_size + 2 /* tri marker */)) {
1471             if(feedbuffer[j] != GL_POLYGON_TOKEN) {
1472                 ERR("Unexpected token: %f\n", feedbuffer[j]);
1473                 continue;
1474             }
1475             if(feedbuffer[j + 1] != 3) {
1476                 ERR("Unexpected polygon: %f corners\n", feedbuffer[j + 1]);
1477                 continue;
1478             }
1479             if(patch->mem[i + 3] == 0.0)
1480                 patch->mem[i + 3] = -feedbuffer[j + out_vertex_size * 2 + 5];
1481             if(patch->mem[i + 4] == 0.0)
1482                 patch->mem[i + 4] = -feedbuffer[j + out_vertex_size * 2 + 6];
1483             if(patch->mem[i + 5] == 0.0)
1484                 patch->mem[i + 5] = -feedbuffer[j + out_vertex_size * 2 + 7];
1485             normalize_normal(patch->mem + i + 3);
1486             i += d3d_out_vertex_size;
1487
1488             if(patch->mem[i + 3] == 0.0)
1489                 patch->mem[i + 3] = -feedbuffer[j + out_vertex_size * 1 + 5];
1490             if(patch->mem[i + 4] == 0.0)
1491                 patch->mem[i + 4] = -feedbuffer[j + out_vertex_size * 1 + 6];
1492             if(patch->mem[i + 5] == 0.0)
1493                 patch->mem[i + 5] = -feedbuffer[j + out_vertex_size * 1 + 7];
1494             normalize_normal(patch->mem + i + 3);
1495             i += d3d_out_vertex_size;
1496
1497             if(patch->mem[i + 3] == 0.0)
1498                 patch->mem[i + 3] = -feedbuffer[j + out_vertex_size * 0 + 5];
1499             if(patch->mem[i + 4] == 0.0)
1500                 patch->mem[i + 4] = -feedbuffer[j + out_vertex_size * 0 + 6];
1501             if(patch->mem[i + 5] == 0.0)
1502                 patch->mem[i + 5] = -feedbuffer[j + out_vertex_size * 0 + 7];
1503             normalize_normal(patch->mem + i + 3);
1504             i += d3d_out_vertex_size;
1505         }
1506     }
1507
1508     glDisable(GL_MAP2_VERTEX_3);
1509     glDisable(GL_AUTO_NORMAL);
1510     glDisable(GL_MAP2_NORMAL);
1511     glDisable(GL_MAP2_TEXTURE_COORD_4);
1512     checkGLcall("glDisable vertex attrib generation");
1513     LEAVE_GL();
1514
1515     HeapFree(GetProcessHeap(), 0, feedbuffer);
1516
1517     vtxStride = 3 * sizeof(float);
1518     if(patch->has_normals) {
1519         vtxStride += 3 * sizeof(float);
1520     }
1521     if(patch->has_texcoords) {
1522         vtxStride += 4 * sizeof(float);
1523     }
1524     memset(&patch->strided, 0, sizeof(&patch->strided));
1525     patch->strided.u.s.position.lpData = (BYTE *) patch->mem;
1526     patch->strided.u.s.position.dwStride = vtxStride;
1527     patch->strided.u.s.position.dwType = WINED3DDECLTYPE_FLOAT3;
1528     patch->strided.u.s.position.streamNo = 255;
1529
1530     if(patch->has_normals) {
1531         patch->strided.u.s.normal.lpData = (BYTE *) patch->mem + 3 * sizeof(float) /* pos */;
1532         patch->strided.u.s.normal.dwStride = vtxStride;
1533         patch->strided.u.s.normal.dwType = WINED3DDECLTYPE_FLOAT3;
1534         patch->strided.u.s.normal.streamNo = 255;
1535     }
1536     if(patch->has_texcoords) {
1537         patch->strided.u.s.texCoords[0].lpData = (BYTE *) patch->mem + 3 * sizeof(float) /* pos */;
1538         if(patch->has_normals) {
1539             patch->strided.u.s.texCoords[0].lpData += 3 * sizeof(float);
1540         }
1541         patch->strided.u.s.texCoords[0].dwStride = vtxStride;
1542         patch->strided.u.s.texCoords[0].dwType = WINED3DDECLTYPE_FLOAT4;
1543         /* MAX_STREAMS index points to an unused element in stateblock->streamOffsets which
1544          * always remains set to 0. Windows uses stream 255 here, but this is not visible to the
1545          * application.
1546          */
1547         patch->strided.u.s.texCoords[0].streamNo = MAX_STREAMS;
1548     }
1549
1550     return WINED3D_OK;
1551 }