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
  34 #if 0 /* TODO */
  35 extern IWineD3DVertexShaderImpl*            VertexShaders[64];
  36 extern IWineD3DVertexDeclarationImpl*       VertexShaderDeclarations[64];
  37 extern IWineD3DPixelShaderImpl*             PixelShaders[64];
  38
  39 #undef GL_VERSION_1_4 /* To be fixed, caused by mesa headers */
  40 #endif
  41
  42 /* Issues the glBegin call for gl given the primitive type and count */
  43 static DWORD primitiveToGl(WINED3DPRIMITIVETYPE PrimitiveType,
  44                     DWORD            NumPrimitives,
  45                     GLenum          *primType)
  46 {
  47     DWORD   NumVertexes = NumPrimitives;
  48
  49     switch (PrimitiveType) {
  50     case WINED3DPT_POINTLIST:
  51         TRACE("POINTS\n");
  52         *primType   = GL_POINTS;
  53         NumVertexes = NumPrimitives;
  54         break;
  55
  56     case WINED3DPT_LINELIST:
  57         TRACE("LINES\n");
  58         *primType   = GL_LINES;
  59         NumVertexes = NumPrimitives * 2;
  60         break;
  61
  62     case WINED3DPT_LINESTRIP:
  63         TRACE("LINE_STRIP\n");
  64         *primType   = GL_LINE_STRIP;
  65         NumVertexes = NumPrimitives + 1;
  66         break;
  67
  68     case WINED3DPT_TRIANGLELIST:
  69         TRACE("TRIANGLES\n");
  70         *primType   = GL_TRIANGLES;
  71         NumVertexes = NumPrimitives * 3;
  72         break;
  73
  74     case WINED3DPT_TRIANGLESTRIP:
  75         TRACE("TRIANGLE_STRIP\n");
  76         *primType   = GL_TRIANGLE_STRIP;
  77         NumVertexes = NumPrimitives + 2;
  78         break;
  79
  80     case WINED3DPT_TRIANGLEFAN:
  81         TRACE("TRIANGLE_FAN\n");
  82         *primType   = GL_TRIANGLE_FAN;
  83         NumVertexes = NumPrimitives + 2;
  84         break;
  85
  86     default:
  87         FIXME("Unhandled primitive\n");
  88         *primType    = GL_POINTS;
  89         break;
  90     }
  91     return NumVertexes;
  92 }
  93
  94 static BOOL fixed_get_input(
  95     BYTE usage, BYTE usage_idx,
  96     unsigned int* regnum) {
  97
  98     *regnum = -1;
  99
 100     /* Those positions must have the order in the
 101      * named part of the strided data */
 102
 103     if ((usage == WINED3DDECLUSAGE_POSITION || usage == WINED3DDECLUSAGE_POSITIONT) && usage_idx == 0)
 104         *regnum = 0;
 105     else if (usage == WINED3DDECLUSAGE_BLENDWEIGHT && usage_idx == 0)
 106         *regnum = 1;
 107     else if (usage == WINED3DDECLUSAGE_BLENDINDICES && usage_idx == 0)
 108         *regnum = 2;
 109     else if (usage == WINED3DDECLUSAGE_NORMAL && usage_idx == 0)
 110         *regnum = 3;
 111     else if (usage == WINED3DDECLUSAGE_PSIZE && usage_idx == 0)
 112         *regnum = 4;
 113     else if (usage == WINED3DDECLUSAGE_COLOR && usage_idx == 0)
 114         *regnum = 5;
 115     else if (usage == WINED3DDECLUSAGE_COLOR && usage_idx == 1)
 116         *regnum = 6;
 117     else if (usage == WINED3DDECLUSAGE_TEXCOORD && usage_idx < WINED3DDP_MAXTEXCOORD)
 118         *regnum = 7 + usage_idx;
 119     else if ((usage == WINED3DDECLUSAGE_POSITION || usage == WINED3DDECLUSAGE_POSITIONT) && usage_idx == 1)
 120         *regnum = 7 + WINED3DDP_MAXTEXCOORD;
 121     else if (usage == WINED3DDECLUSAGE_NORMAL && usage_idx == 1)
 122         *regnum = 8 + WINED3DDP_MAXTEXCOORD;
 123     else if (usage == WINED3DDECLUSAGE_TANGENT && usage_idx == 0)
 124         *regnum = 9 + WINED3DDP_MAXTEXCOORD;
 125     else if (usage == WINED3DDECLUSAGE_BINORMAL && usage_idx == 0)
 126         *regnum = 10 + WINED3DDP_MAXTEXCOORD;
 127     else if (usage == WINED3DDECLUSAGE_TESSFACTOR && usage_idx == 0)
 128         *regnum = 11 + WINED3DDP_MAXTEXCOORD;
 129     else if (usage == WINED3DDECLUSAGE_FOG && usage_idx == 0)
 130         *regnum = 12 + WINED3DDP_MAXTEXCOORD;
 131     else if (usage == WINED3DDECLUSAGE_DEPTH && usage_idx == 0)
 132         *regnum = 13 + WINED3DDP_MAXTEXCOORD;
 133     else if (usage == WINED3DDECLUSAGE_SAMPLE && usage_idx == 0)
 134         *regnum = 14 + WINED3DDP_MAXTEXCOORD;
 135
 136     if (*regnum < 0) {
 137         FIXME("Unsupported input stream [usage=%s, usage_idx=%u]\n",
 138             debug_d3ddeclusage(usage), usage_idx);
 139         return FALSE;
 140     }
 141     return TRUE;
 142 }
 143
 144 void primitiveDeclarationConvertToStridedData(
 145      IWineD3DDevice *iface,
 146      BOOL useVertexShaderFunction,
 147      WineDirect3DVertexStridedData *strided,
 148      BOOL *fixup) {
 149
 150      /* We need to deal with frequency data!*/
 151
 152     BYTE  *data    = NULL;
 153     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
 154     IWineD3DVertexDeclarationImpl* vertexDeclaration = (IWineD3DVertexDeclarationImpl *)This->stateBlock->vertexDecl;
 155     int i;
 156     WINED3DVERTEXELEMENT *element;
 157     DWORD stride;
 158     int reg;
 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(data + This->stateBlock->loadBaseVertexIndex * stride < 0) {
 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         reg = element->Reg;
 217
 218         TRACE("Offset %d Stream %d UsageIndex %d\n", element->Offset, element->Stream, element->UsageIndex);
 219
 220         if (useVertexShaderFunction)
 221             stride_used = vshader_get_input(This->stateBlock->vertexShader,
 222                 element->Usage, element->UsageIndex, &idx);
 223         else
 224             stride_used = fixed_get_input(element->Usage, element->UsageIndex, &idx);
 225
 226         if (stride_used) {
 227             TRACE("Loaded %s array %u [usage=%s, usage_idx=%u, "
 228                     "stream=%u, offset=%u, stride=%u, type=%s, VBO=%u]\n",
 229                     useVertexShaderFunction? "shader": "fixed function", idx,
 230                     debug_d3ddeclusage(element->Usage), element->UsageIndex,
 231                     element->Stream, element->Offset, stride, debug_d3ddecltype(element->Type), streamVBO);
 232
 233             strided->u.input[idx].lpData = data;
 234             strided->u.input[idx].dwType = element->Type;
 235             strided->u.input[idx].dwStride = stride;
 236             strided->u.input[idx].VBO = streamVBO;
 237             strided->u.input[idx].streamNo = element->Stream;
 238         }
 239     }
 240     /* Now call PreLoad on all the vertex buffers. In the very rare case
 241      * that the buffers stopps converting PreLoad will dirtify the VDECL again.
 242      * The vertex buffer can now use the strided structure in the device instead of finding its
 243      * own again.
 244      *
 245      * NULL streams won't be recorded in the array, UP streams won't be either. A stream is only
 246      * once in there.
 247      */
 248     for(i=0; i < numPreloadStreams; i++) {
 249         IWineD3DVertexBuffer_PreLoad(This->stateBlock->streamSource[streams[i]]);
 250     }
 251 }
 252
 253 static void drawStridedFast(IWineD3DDevice *iface,UINT numberOfVertices, GLenum glPrimitiveType,
 254                      const void *idxData, short idxSize, ULONG minIndex, ULONG startIdx, ULONG startVertex) {
 255     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
 256
 257     if (idxSize != 0 /* This crashes sometimes!*/) {
 258         TRACE("(%p) : glElements(%x, %d, %d, ...)\n", This, glPrimitiveType, numberOfVertices, minIndex);
 259         idxData = idxData == (void *)-1 ? NULL : idxData;
 260 #if 1
 261         glDrawElements(glPrimitiveType, numberOfVertices, idxSize == 2 ? GL_UNSIGNED_SHORT : GL_UNSIGNED_INT,
 262                      (const char *)idxData+(idxSize * startIdx));
 263         checkGLcall("glDrawElements");
 264 #else /* using drawRangeElements may be faster */
 265
 266         glDrawRangeElements(glPrimitiveType, minIndex, minIndex + numberOfVertices - 1, numberOfVertices,
 267                       idxSize == 2 ? GL_UNSIGNED_SHORT : GL_UNSIGNED_INT,
 268                       (const char *)idxData+(idxSize * startIdx));
 269         checkGLcall("glDrawRangeElements");
 270 #endif
 271
 272     } else {
 273
 274         /* Note first is now zero as we shuffled along earlier */
 275         TRACE("(%p) : glDrawArrays(%x, 0, %d)\n", This, glPrimitiveType, numberOfVertices);
 276         glDrawArrays(glPrimitiveType, startVertex, numberOfVertices);
 277         checkGLcall("glDrawArrays");
 278
 279     }
 280
 281     return;
 282 }
 283
 284 /*
 285  * Actually draw using the supplied information.
 286  * Slower GL version which extracts info about each vertex in turn
 287  */
 288
 289 static void drawStridedSlow(IWineD3DDevice *iface, WineDirect3DVertexStridedData *sd,
 290                      UINT NumVertexes, GLenum glPrimType,
 291                      const void *idxData, short idxSize, ULONG minIndex, ULONG startIdx, ULONG startVertex) {
 292
 293     unsigned int               textureNo    = 0;
 294     const WORD                *pIdxBufS     = NULL;
 295     const DWORD               *pIdxBufL     = NULL;
 296     LONG                       vx_index;
 297     float x  = 0.0f, y  = 0.0f, z = 0.0f;  /* x,y,z coordinates          */
 298     float rhw = 0.0f;                      /* rhw                        */
 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     /* Start drawing in GL */
 349     VTRACE(("glBegin(%x)\n", glPrimType));
 350     glBegin(glPrimType);
 351
 352     /* Default settings for data that is not passed */
 353     if (sd->u.s.normal.lpData == NULL) {
 354         glNormal3f(0, 0, 0);
 355     }
 356     if(sd->u.s.diffuse.lpData == NULL) {
 357         glColor4f(1.0f, 1.0f, 1.0f, 1.0f);
 358     }
 359     if(sd->u.s.specular.lpData == NULL) {
 360         if (GL_SUPPORT(EXT_SECONDARY_COLOR)) {
 361             GL_EXTCALL(glSecondaryColor3fEXT)(0, 0, 0);
 362         }
 363     }
 364
 365     /* We shouldn't start this function if any VBO is involved. Should I put a safety check here?
 366      * Guess it's not necessary(we crash then anyway) and would only eat CPU time
 367      */
 368
 369     /* For each primitive */
 370     for (vx_index = 0; vx_index < NumVertexes; ++vx_index) {
 371
 372         /* Initialize diffuse color */
 373         diffuseColor = 0xFFFFFFFF;
 374
 375         /* Blending data and Point sizes are not supported by this function. They are not supported by the fixed
 376          * function pipeline at all. A Fixme for them is printed after decoding the vertex declaration
 377          */
 378
 379         /* For indexed data, we need to go a few more strides in */
 380         if (idxData != NULL) {
 381
 382             /* Indexed so work out the number of strides to skip */
 383             if (idxSize == 2) {
 384                 VTRACE(("Idx for vertex %d = %d\n", vx_index, pIdxBufS[startIdx+vx_index]));
 385                 SkipnStrides = pIdxBufS[startIdx + vx_index] + This->stateBlock->loadBaseVertexIndex;
 386             } else {
 387                 VTRACE(("Idx for vertex %d = %d\n", vx_index, pIdxBufL[startIdx+vx_index]));
 388                 SkipnStrides = pIdxBufL[startIdx + vx_index] + This->stateBlock->loadBaseVertexIndex;
 389             }
 390         }
 391
 392         /* Texture coords --------------------------- */
 393         for (textureNo = 0; textureNo < GL_LIMITS(texture_stages); ++textureNo) {
 394
 395             if (!GL_SUPPORT(ARB_MULTITEXTURE) && textureNo > 0) {
 396                 FIXME("Program using multiple concurrent textures which this opengl implementation doesn't support\n");
 397                 continue ;
 398             }
 399
 400             /* Query tex coords */
 401             if (This->stateBlock->textures[textureNo] != NULL || pixelShader) {
 402
 403                 int    coordIdx = This->stateBlock->textureState[textureNo][WINED3DTSS_TEXCOORDINDEX];
 404                 int texture_idx = This->texUnitMap[textureNo];
 405                 float *ptrToCoords = NULL;
 406                 float  s = 0.0, t = 0.0, r = 0.0, q = 0.0;
 407
 408                 if (coordIdx > 7) {
 409                     VTRACE(("tex: %d - Skip tex coords, as being system generated\n", textureNo));
 410                     continue;
 411                 } else if (coordIdx < 0) {
 412                     FIXME("tex: %d - Coord index %d is less than zero, expect a crash.\n", textureNo, coordIdx);
 413                     continue;
 414                 }
 415
 416                 ptrToCoords = (float *)(texCoords[coordIdx] + (SkipnStrides * sd->u.s.texCoords[coordIdx].dwStride));
 417                 if (texCoords[coordIdx] == NULL) {
 418                     TRACE("tex: %d - Skipping tex coords, as no data supplied\n", textureNo);
 419                     if (GL_SUPPORT(ARB_MULTITEXTURE)) {
 420                         GL_EXTCALL(glMultiTexCoord4fARB(GL_TEXTURE0_ARB + texture_idx, 0, 0, 0, 1));
 421                     } else {
 422                         glTexCoord4f(0, 0, 0, 1);
 423                     }
 424                     continue;
 425                 } else {
 426                     int coordsToUse = sd->u.s.texCoords[coordIdx].dwType + 1; /* 0 == WINED3DDECLTYPE_FLOAT1 etc */
 427
 428                     if (texture_idx == -1) continue;
 429
 430                     /* The coords to supply depend completely on the fvf / vertex shader */
 431                     switch (coordsToUse) {
 432                     case 4: q = ptrToCoords[3]; /* drop through */
 433                     case 3: r = ptrToCoords[2]; /* drop through */
 434                     case 2: t = ptrToCoords[1]; /* drop through */
 435                     case 1: s = ptrToCoords[0];
 436                     }
 437
 438                     switch (coordsToUse) {   /* Supply the provided texture coords */
 439                     case WINED3DTTFF_COUNT1:
 440                         VTRACE(("tex:%d, s=%f\n", textureNo, s));
 441                         if (GL_SUPPORT(ARB_MULTITEXTURE)) {
 442                             GL_EXTCALL(glMultiTexCoord1fARB(GL_TEXTURE0_ARB + texture_idx, s));
 443                         } else {
 444                             glTexCoord1f(s);
 445                         }
 446                         break;
 447                     case WINED3DTTFF_COUNT2:
 448                         VTRACE(("tex:%d, s=%f, t=%f\n", textureNo, s, t));
 449                         if (GL_SUPPORT(ARB_MULTITEXTURE)) {
 450                             GL_EXTCALL(glMultiTexCoord2fARB(GL_TEXTURE0_ARB + texture_idx, s, t));
 451                         } else {
 452                             glTexCoord2f(s, t);
 453                         }
 454                         break;
 455                     case WINED3DTTFF_COUNT3:
 456                         VTRACE(("tex:%d, s=%f, t=%f, r=%f\n", textureNo, s, t, r));
 457                         if (GL_SUPPORT(ARB_MULTITEXTURE)) {
 458                             GL_EXTCALL(glMultiTexCoord3fARB(GL_TEXTURE0_ARB + texture_idx, s, t, r));
 459                         } else {
 460                             glTexCoord3f(s, t, r);
 461                         }
 462                         break;
 463                     case WINED3DTTFF_COUNT4:
 464                         VTRACE(("tex:%d, s=%f, t=%f, r=%f, q=%f\n", textureNo, s, t, r, q));
 465                         if (GL_SUPPORT(ARB_MULTITEXTURE)) {
 466                             GL_EXTCALL(glMultiTexCoord4fARB(GL_TEXTURE0_ARB + texture_idx, s, t, r, q));
 467                         } else {
 468                             glTexCoord4f(s, t, r, q);
 469                         }
 470                         break;
 471                     default:
 472                         FIXME("Should not get here as coordsToUse is two bits only (%x)!\n", coordsToUse);
 473                     }
 474                 }
 475             }
 476         } /* End of textures */
 477
 478         /* Diffuse -------------------------------- */
 479         if (diffuse) {
 480             DWORD *ptrToCoords = (DWORD *)(diffuse + (SkipnStrides * sd->u.s.diffuse.dwStride));
 481             diffuseColor = ptrToCoords[0];
 482             VTRACE(("diffuseColor=%lx\n", diffuseColor));
 483
 484             glColor4ub(D3DCOLOR_B_R(diffuseColor),
 485                      D3DCOLOR_B_G(diffuseColor),
 486                      D3DCOLOR_B_B(diffuseColor),
 487                      D3DCOLOR_B_A(diffuseColor));
 488             VTRACE(("glColor4ub: r,g,b,a=%lu,%lu,%lu,%lu\n",
 489                     D3DCOLOR_B_R(diffuseColor),
 490                     D3DCOLOR_B_G(diffuseColor),
 491                     D3DCOLOR_B_B(diffuseColor),
 492                     D3DCOLOR_B_A(diffuseColor)));
 493
 494             if(This->activeContext->num_untracked_materials) {
 495                 unsigned char i;
 496                 float color[4];
 497                 color[0] = D3DCOLOR_B_R(diffuseColor) / 255.0;
 498                 color[1] = D3DCOLOR_B_G(diffuseColor) / 255.0;
 499                 color[2] = D3DCOLOR_B_B(diffuseColor) / 255.0;
 500                 color[3] = D3DCOLOR_B_A(diffuseColor) / 255.0;
 501
 502                 for(i = 0; i < This->activeContext->num_untracked_materials; i++) {
 503                     glMaterialfv(GL_FRONT_AND_BACK, This->activeContext->untracked_materials[i], color);
 504                 }
 505             }
 506         }
 507
 508         /* Specular ------------------------------- */
 509         if (specular) {
 510             DWORD *ptrToCoords = (DWORD *)(specular + (SkipnStrides * sd->u.s.specular.dwStride));
 511             specularColor = ptrToCoords[0];
 512             VTRACE(("specularColor=%lx\n", specularColor));
 513
 514             /* special case where the fog density is stored in the specular alpha channel */
 515             if(This->stateBlock->renderState[WINED3DRS_FOGENABLE] &&
 516               (This->stateBlock->renderState[WINED3DRS_FOGVERTEXMODE] == WINED3DFOG_NONE || sd->u.s.position.dwType == WINED3DDECLTYPE_FLOAT4 )&&
 517               This->stateBlock->renderState[WINED3DRS_FOGTABLEMODE] == WINED3DFOG_NONE) {
 518                 if(GL_SUPPORT(EXT_FOG_COORD)) {
 519                     GL_EXTCALL(glFogCoordfEXT(specularColor >> 24));
 520                 } else {
 521                     static BOOL warned = FALSE;
 522                     if(!warned) {
 523                         /* TODO: Use the fog table code from old ddraw */
 524                         FIXME("Implement fog for transformed vertices in software\n");
 525                         warned = TRUE;
 526                     }
 527                 }
 528             }
 529
 530             VTRACE(("glSecondaryColor4ub: r,g,b=%lu,%lu,%lu\n",
 531                     D3DCOLOR_B_R(specularColor),
 532                     D3DCOLOR_B_G(specularColor),
 533                     D3DCOLOR_B_B(specularColor)));
 534             if (GL_SUPPORT(EXT_SECONDARY_COLOR)) {
 535                 GL_EXTCALL(glSecondaryColor3ubEXT)(
 536                            D3DCOLOR_B_R(specularColor),
 537                            D3DCOLOR_B_G(specularColor),
 538                            D3DCOLOR_B_B(specularColor));
 539             } else {
 540                 /* Do not worry if specular colour missing and disable request */
 541                 VTRACE(("Specular color extensions not supplied\n"));
 542             }
 543         }
 544
 545         /* Normal -------------------------------- */
 546         if (normal != NULL) {
 547             float *ptrToCoords = (float *)(normal + (SkipnStrides * sd->u.s.normal.dwStride));
 548
 549             VTRACE(("glNormal:nx,ny,nz=%f,%f,%f\n", ptrToCoords[0], ptrToCoords[1], ptrToCoords[2]));
 550             glNormal3f(ptrToCoords[0], ptrToCoords[1], ptrToCoords[2]);
 551         }
 552
 553         /* Position -------------------------------- */
 554         if (position) {
 555             float *ptrToCoords = (float *)(position + (SkipnStrides * sd->u.s.position.dwStride));
 556             x = ptrToCoords[0];
 557             y = ptrToCoords[1];
 558             z = ptrToCoords[2];
 559             rhw = 1.0;
 560             VTRACE(("x,y,z=%f,%f,%f\n", x,y,z));
 561
 562             /* RHW follows, only if transformed, ie 4 floats were provided */
 563             if (sd->u.s.position_transformed) {
 564                 rhw = ptrToCoords[3];
 565                 VTRACE(("rhw=%f\n", rhw));
 566             }
 567
 568             if (1.0f == rhw || ((rhw < eps) && (rhw > -eps))) {
 569                 VTRACE(("Vertex: glVertex:x,y,z=%f,%f,%f\n", x,y,z));
 570                 glVertex3f(x, y, z);
 571             } else {
 572                 GLfloat w = 1.0 / rhw;
 573                 VTRACE(("Vertex: glVertex:x,y,z=%f,%f,%f / rhw=%f\n", x,y,z,rhw));
 574                 glVertex4f(x*w, y*w, z*w, w);
 575             }
 576         }
 577
 578         /* For non indexed mode, step onto next parts */
 579         if (idxData == NULL) {
 580             ++SkipnStrides;
 581         }
 582     }
 583
 584     glEnd();
 585     checkGLcall("glEnd and previous calls");
 586 }
 587
 588 static void depth_blt(IWineD3DDevice *iface, GLuint texture) {
 589     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
 590     GLint old_binding = 0;
 591
 592     glPushAttrib(GL_ENABLE_BIT | GL_DEPTH_BUFFER_BIT | GL_COLOR_BUFFER_BIT);
 593
 594     glDisable(GL_CULL_FACE);
 595     glEnable(GL_BLEND);
 596     glDisable(GL_ALPHA_TEST);
 597     glDisable(GL_SCISSOR_TEST);
 598     glDisable(GL_STENCIL_TEST);
 599     glEnable(GL_DEPTH_TEST);
 600     glDepthFunc(GL_ALWAYS);
 601     glBlendFunc(GL_ZERO, GL_ONE);
 602
 603     GL_EXTCALL(glActiveTextureARB(GL_TEXTURE0_ARB));
 604     glGetIntegerv(GL_TEXTURE_BINDING_2D, &old_binding);
 605     glBindTexture(GL_TEXTURE_2D, texture);
 606     glEnable(GL_TEXTURE_2D);
 607
 608     This->shader_backend->shader_select_depth_blt(iface);
 609
 610     glBegin(GL_TRIANGLE_STRIP);
 611     glVertex2f(-1.0f, -1.0f);
 612     glVertex2f(1.0f, -1.0f);
 613     glVertex2f(-1.0f, 1.0f);
 614     glVertex2f(1.0f, 1.0f);
 615     glEnd();
 616
 617     glBindTexture(GL_TEXTURE_2D, old_binding);
 618
 619     glPopAttrib();
 620
 621     /* Reselect the old shaders. There doesn't seem to be any glPushAttrib bit for arb shaders,
 622      * and this seems easier and more efficient than providing the shader backend with a private
 623      * storage to read and restore the old shader settings
 624      */
 625     This->shader_backend->shader_select(iface, use_ps(This), use_vs(This));
 626 }
 627
 628 static void depth_copy(IWineD3DDevice *iface) {
 629     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
 630     IWineD3DSurfaceImpl *depth_stencil = (IWineD3DSurfaceImpl *)This->depthStencilBuffer;
 631
 632     /* Only copy the depth buffer if there is one. */
 633     if (!depth_stencil) return;
 634
 635     /* TODO: Make this work for modes other than FBO */
 636     if (wined3d_settings.offscreen_rendering_mode != ORM_FBO) return;
 637
 638     if (depth_stencil->current_renderbuffer) {
 639         FIXME("Not supported with fixed up depth stencil\n");
 640         return;
 641     }
 642
 643     if (This->render_offscreen) {
 644         static GLuint tmp_texture = 0;
 645         GLint old_binding = 0;
 646
 647         TRACE("Copying onscreen depth buffer to offscreen surface\n");
 648
 649         if (!tmp_texture) {
 650             glGenTextures(1, &tmp_texture);
 651         }
 652
 653         /* Note that we use depth_blt here as well, rather than glCopyTexImage2D
 654          * directly on the FBO texture. That's because we need to flip. */
 655         GL_EXTCALL(glBindFramebufferEXT(GL_FRAMEBUFFER_EXT, 0));
 656         glGetIntegerv(GL_TEXTURE_BINDING_2D, &old_binding);
 657         glBindTexture(GL_TEXTURE_2D, tmp_texture);
 658         glCopyTexImage2D(depth_stencil->glDescription.target,
 659                 depth_stencil->glDescription.level,
 660                 depth_stencil->glDescription.glFormatInternal,
 661                 0,
 662                 0,
 663                 depth_stencil->currentDesc.Width,
 664                 depth_stencil->currentDesc.Height,
 665                 0);
 666         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
 667         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
 668         glTexParameteri(GL_TEXTURE_2D, GL_DEPTH_TEXTURE_MODE_ARB, GL_LUMINANCE);
 669         glBindTexture(GL_TEXTURE_2D, old_binding);
 670
 671         GL_EXTCALL(glBindFramebufferEXT(GL_FRAMEBUFFER_EXT, This->fbo));
 672         checkGLcall("glBindFramebuffer()");
 673         depth_blt(iface, tmp_texture);
 674         checkGLcall("depth_blt");
 675     } else {
 676         TRACE("Copying offscreen surface to onscreen depth buffer\n");
 677
 678         GL_EXTCALL(glBindFramebufferEXT(GL_FRAMEBUFFER_EXT, 0));
 679         checkGLcall("glBindFramebuffer()");
 680         depth_blt(iface, depth_stencil->glDescription.textureName);
 681         checkGLcall("depth_blt");
 682     }
 683 }
 684
 685 static inline void drawStridedInstanced(IWineD3DDevice *iface, WineDirect3DVertexStridedData *sd, UINT numberOfVertices,
 686                                  GLenum glPrimitiveType, const void *idxData, short idxSize, ULONG minIndex,
 687                                  ULONG startIdx, ULONG startVertex) {
 688     UINT numInstances = 0;
 689     int numInstancedAttribs = 0, i, j;
 690     UINT instancedData[sizeof(sd->u.input) / sizeof(sd->u.input[0]) /* 16 */];
 691     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *) iface;
 692     IWineD3DStateBlockImpl *stateblock = This->stateBlock;
 693
 694     if (idxSize == 0) {
 695         /* This is a nasty thing. MSDN says no hardware supports that and apps have to use software vertex processing.
 696          * We don't support this for now
 697          *
 698          * Shouldn't be too hard to support with opengl, in theory just call glDrawArrays instead of drawElements.
 699          * But the StreamSourceFreq value has a different meaning in that situation.
 700          */
 701         FIXME("Non-indexed instanced drawing is not supported\n");
 702         return;
 703     }
 704
 705     TRACE("(%p) : glElements(%x, %d, %d, ...)\n", This, glPrimitiveType, numberOfVertices, minIndex);
 706     idxData = idxData == (void *)-1 ? NULL : idxData;
 707
 708     /* First, figure out how many instances we have to draw */
 709     for(i = 0; i < MAX_STREAMS; i++) {
 710         /* Look at all non-instanced streams */
 711         if(!(stateblock->streamFlags[i] & WINED3DSTREAMSOURCE_INSTANCEDATA) &&
 712            stateblock->streamSource[i]) {
 713             int inst = stateblock->streamFreq[i];
 714
 715             if(numInstances && inst != numInstances) {
 716                 ERR("Two streams specify a different number of instances. Got %d, new is %d\n", numInstances, inst);
 717             }
 718             numInstances = inst;
 719         }
 720     }
 721
 722     for(i = 0; i < sizeof(sd->u.input) / sizeof(sd->u.input[0]); i++) {
 723         if(stateblock->streamFlags[sd->u.input[i].streamNo] & WINED3DSTREAMSOURCE_INSTANCEDATA) {
 724             instancedData[numInstancedAttribs] = i;
 725             numInstancedAttribs++;
 726         }
 727     }
 728
 729     /* now draw numInstances instances :-) */
 730     for(i = 0; i < numInstances; i++) {
 731         /* Specify the instanced attributes using immediate mode calls */
 732         for(j = 0; j < numInstancedAttribs; j++) {
 733             BYTE *ptr = sd->u.input[instancedData[j]].lpData +
 734                         sd->u.input[instancedData[j]].dwStride * i +
 735                         stateblock->streamOffset[sd->u.input[instancedData[j]].streamNo];
 736             if(sd->u.input[instancedData[j]].VBO) {
 737                 IWineD3DVertexBufferImpl *vb = (IWineD3DVertexBufferImpl *) stateblock->streamSource[sd->u.input[instancedData[j]].streamNo];
 738                 ptr += (long) vb->resource.allocatedMemory;
 739             }
 740
 741             switch(sd->u.input[instancedData[j]].dwType) {
 742                 case WINED3DDECLTYPE_FLOAT1:
 743                     GL_EXTCALL(glVertexAttrib1fvARB(instancedData[j], (float *) ptr));
 744                     break;
 745                 case WINED3DDECLTYPE_FLOAT2:
 746                     GL_EXTCALL(glVertexAttrib2fvARB(instancedData[j], (float *) ptr));
 747                     break;
 748                 case WINED3DDECLTYPE_FLOAT3:
 749                     GL_EXTCALL(glVertexAttrib3fvARB(instancedData[j], (float *) ptr));
 750                     break;
 751                 case WINED3DDECLTYPE_FLOAT4:
 752                     GL_EXTCALL(glVertexAttrib4fvARB(instancedData[j], (float *) ptr));
 753                     break;
 754
 755                 case WINED3DDECLTYPE_UBYTE4:
 756                     GL_EXTCALL(glVertexAttrib4ubvARB(instancedData[j], ptr));
 757                     break;
 758                 case WINED3DDECLTYPE_UBYTE4N:
 759                 case WINED3DDECLTYPE_D3DCOLOR:
 760                     GL_EXTCALL(glVertexAttrib4NubvARB(instancedData[j], ptr));
 761                     break;
 762
 763                 case WINED3DDECLTYPE_SHORT2:
 764                     GL_EXTCALL(glVertexAttrib4svARB(instancedData[j], (GLshort *) ptr));
 765                     break;
 766                 case WINED3DDECLTYPE_SHORT4:
 767                     GL_EXTCALL(glVertexAttrib4svARB(instancedData[j], (GLshort *) ptr));
 768                     break;
 769
 770                 case WINED3DDECLTYPE_SHORT2N:
 771                 {
 772                     GLshort s[4] = {((short *) ptr)[0], ((short *) ptr)[1], 0, 1};
 773                     GL_EXTCALL(glVertexAttrib4NsvARB(instancedData[j], s));
 774                     break;
 775                 }
 776                 case WINED3DDECLTYPE_USHORT2N:
 777                 {
 778                     GLushort s[4] = {((unsigned short *) ptr)[0], ((unsigned short *) ptr)[1], 0, 1};
 779                     GL_EXTCALL(glVertexAttrib4NusvARB(instancedData[j], s));
 780                     break;
 781                 }
 782                 case WINED3DDECLTYPE_SHORT4N:
 783                     GL_EXTCALL(glVertexAttrib4NsvARB(instancedData[j], (GLshort *) ptr));
 784                     break;
 785                 case WINED3DDECLTYPE_USHORT4N:
 786                     GL_EXTCALL(glVertexAttrib4NusvARB(instancedData[j], (GLushort *) ptr));
 787                     break;
 788
 789                 case WINED3DDECLTYPE_UDEC3:
 790                     FIXME("Unsure about WINED3DDECLTYPE_UDEC3\n");
 791                     /*glVertexAttrib3usvARB(instancedData[j], (GLushort *) ptr); Does not exist */
 792                     break;
 793                 case WINED3DDECLTYPE_DEC3N:
 794                     FIXME("Unsure about WINED3DDECLTYPE_DEC3N\n");
 795                     /*glVertexAttrib3NusvARB(instancedData[j], (GLushort *) ptr); Does not exist */
 796                     break;
 797
 798                 case WINED3DDECLTYPE_FLOAT16_2:
 799                     /* Are those 16 bit floats. C doesn't have a 16 bit float type. I could read the single bits and calculate a 4
 800                      * byte float according to the IEEE standard
 801                      */
 802                     if (GL_SUPPORT(NV_HALF_FLOAT)) {
 803                         GL_EXTCALL(glVertexAttrib2hvNV(instancedData[j], (GLhalfNV *)ptr));
 804                     } else {
 805                         FIXME("Unsupported WINED3DDECLTYPE_FLOAT16_2\n");
 806                     }
 807                     break;
 808                 case WINED3DDECLTYPE_FLOAT16_4:
 809                     if (GL_SUPPORT(NV_HALF_FLOAT)) {
 810                         GL_EXTCALL(glVertexAttrib4hvNV(instancedData[j], (GLhalfNV *)ptr));
 811                     } else {
 812                         FIXME("Unsupported WINED3DDECLTYPE_FLOAT16_4\n");
 813                     }
 814                     break;
 815
 816                 case WINED3DDECLTYPE_UNUSED:
 817                 default:
 818                     ERR("Unexpected declaration in instanced attributes\n");
 819                     break;
 820             }
 821         }
 822
 823         glDrawElements(glPrimitiveType, numberOfVertices, idxSize == 2 ? GL_UNSIGNED_SHORT : GL_UNSIGNED_INT,
 824                     (const char *)idxData+(idxSize * startIdx));
 825         checkGLcall("glDrawElements");
 826     }
 827 }
 828
 829 struct coords {
 830     int x, y, z;
 831 };
 832
 833 void blt_to_drawable(IWineD3DDeviceImpl *This, IWineD3DSurfaceImpl *surface) {
 834     struct coords coords[4];
 835     int low_coord;
 836
 837     /* TODO: This could be supported for lazy unlocking */
 838     if(!(surface->Flags & SFLAG_INTEXTURE)) {
 839         /* It is ok at init to be nowhere */
 840         if(!(surface->Flags & SFLAG_INSYSMEM)) {
 841             ERR("Blitting surfaces from sysmem not supported yet\n");
 842         }
 843         return;
 844     }
 845
 846     ActivateContext(This, This->render_targets[0], CTXUSAGE_BLIT);
 847     ENTER_GL();
 848
 849     if(surface->glDescription.target == GL_TEXTURE_2D) {
 850         glBindTexture(GL_TEXTURE_2D, surface->glDescription.textureName);
 851         checkGLcall("GL_TEXTURE_2D, This->glDescription.textureName)");
 852
 853         coords[0].x = 0;    coords[0].y = 0;    coords[0].z = 0;
 854         coords[1].x = 0;    coords[1].y = 1;    coords[1].z = 0;
 855         coords[2].x = 1;    coords[2].y = 1;    coords[2].z = 0;
 856         coords[3].x = 1;    coords[3].y = 0;    coords[3].z = 0;
 857
 858         low_coord = 0;
 859     } else {
 860         /* Must be a cube map */
 861         glDisable(GL_TEXTURE_2D);
 862         checkGLcall("glDisable(GL_TEXTURE_2D)");
 863         glEnable(GL_TEXTURE_CUBE_MAP_ARB);
 864         checkGLcall("glEnable(surface->glDescription.target)");
 865         glBindTexture(GL_TEXTURE_CUBE_MAP_ARB, surface->glDescription.textureName);
 866         checkGLcall("GL_TEXTURE_CUBE_MAP_ARB, This->glDescription.textureName)");
 867
 868         switch(surface->glDescription.target) {
 869             case GL_TEXTURE_CUBE_MAP_POSITIVE_X:
 870                 coords[0].x =  1;   coords[0].y = -1;   coords[0].z =  1;
 871                 coords[1].x =  1;   coords[1].y =  1;   coords[1].z =  1;
 872                 coords[2].x =  1;   coords[2].y =  1;   coords[2].z = -1;
 873                 coords[3].x =  1;   coords[3].y = -1;   coords[3].z = -1;
 874                 break;
 875
 876             case GL_TEXTURE_CUBE_MAP_NEGATIVE_X:
 877                 coords[0].x = -1;   coords[0].y = -1;   coords[0].z =  1;
 878                 coords[1].x = -1;   coords[1].y =  1;   coords[1].z =  1;
 879                 coords[2].x = -1;   coords[2].y =  1;   coords[2].z = -1;
 880                 coords[3].x = -1;   coords[3].y = -1;   coords[3].z = -1;
 881                 break;
 882
 883             case GL_TEXTURE_CUBE_MAP_POSITIVE_Y:
 884                 coords[0].x = -1;   coords[0].y =  1;   coords[0].z =  1;
 885                 coords[1].x =  1;   coords[1].y =  1;   coords[1].z =  1;
 886                 coords[2].x =  1;   coords[2].y =  1;   coords[2].z = -1;
 887                 coords[3].x = -1;   coords[3].y =  1;   coords[3].z = -1;
 888                 break;
 889
 890             case GL_TEXTURE_CUBE_MAP_NEGATIVE_Y:
 891                 coords[0].x = -1;   coords[0].y = -1;   coords[0].z =  1;
 892                 coords[1].x =  1;   coords[1].y = -1;   coords[1].z =  1;
 893                 coords[2].x =  1;   coords[2].y = -1;   coords[2].z = -1;
 894                 coords[3].x = -1;   coords[3].y = -1;   coords[3].z = -1;
 895                 break;
 896
 897             case GL_TEXTURE_CUBE_MAP_POSITIVE_Z:
 898                 coords[0].x = -1;   coords[0].y = -1;   coords[0].z =  1;
 899                 coords[1].x =  1;   coords[1].y = -1;   coords[1].z =  1;
 900                 coords[2].x =  1;   coords[2].y = -1;   coords[2].z =  1;
 901                 coords[3].x = -1;   coords[3].y = -1;   coords[3].z =  1;
 902                 break;
 903
 904             case GL_TEXTURE_CUBE_MAP_NEGATIVE_Z:
 905                 coords[0].x = -1;   coords[0].y = -1;   coords[0].z = -1;
 906                 coords[1].x =  1;   coords[1].y = -1;   coords[1].z = -1;
 907                 coords[2].x =  1;   coords[2].y = -1;   coords[2].z = -1;
 908                 coords[3].x = -1;   coords[3].y = -1;   coords[3].z = -1;
 909
 910             default:
 911                 ERR("Unexpected texture target\n");
 912                 LEAVE_GL();
 913                 return;
 914         }
 915
 916         low_coord = -1;
 917     }
 918
 919     if(This->render_offscreen) {
 920         coords[0].y = coords[0].y == 1 ? low_coord : 1;
 921         coords[1].y = coords[1].y == 1 ? low_coord : 1;
 922         coords[2].y = coords[2].y == 1 ? low_coord : 1;
 923         coords[3].y = coords[3].y == 1 ? low_coord : 1;
 924     }
 925
 926     glBegin(GL_QUADS);
 927         glTexCoord3iv((GLint *) &coords[0]);
 928         glVertex2i(0, 0);
 929
 930         glTexCoord3iv((GLint *) &coords[1]);
 931         glVertex2i(0, surface->pow2Height);
 932
 933         glTexCoord3iv((GLint *) &coords[2]);
 934         glVertex2i(surface->pow2Width, surface->pow2Height);
 935
 936         glTexCoord3iv((GLint *) &coords[3]);
 937         glVertex2i(surface->pow2Width, 0);
 938     glEnd();
 939     checkGLcall("glEnd");
 940
 941     if(surface->glDescription.target != GL_TEXTURE_2D) {
 942         glEnable(GL_TEXTURE_2D);
 943         checkGLcall("glEnable(GL_TEXTURE_2D)");
 944         glDisable(GL_TEXTURE_CUBE_MAP_ARB);
 945         checkGLcall("glDisable(GL_TEXTURE_CUBE_MAP_ARB)");
 946     }
 947     LEAVE_GL();
 948 }
 949
 950 static inline void remove_vbos(IWineD3DDeviceImpl *This, WineDirect3DVertexStridedData *s) {
 951     unsigned char i;
 952     IWineD3DVertexBufferImpl *vb;
 953
 954     if(s->u.s.position.VBO) {
 955         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.position.streamNo];
 956         s->u.s.position.VBO = 0;
 957         s->u.s.position.lpData = (BYTE *) ((unsigned long) s->u.s.position.lpData + (unsigned long) vb->resource.allocatedMemory);
 958     }
 959     if(s->u.s.blendWeights.VBO) {
 960         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.blendWeights.streamNo];
 961         s->u.s.blendWeights.VBO = 0;
 962         s->u.s.blendWeights.lpData = (BYTE *) ((unsigned long) s->u.s.blendWeights.lpData + (unsigned long) vb->resource.allocatedMemory);
 963     }
 964     if(s->u.s.blendMatrixIndices.VBO) {
 965         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.blendMatrixIndices.streamNo];
 966         s->u.s.blendMatrixIndices.VBO = 0;
 967         s->u.s.blendMatrixIndices.lpData = (BYTE *) ((unsigned long) s->u.s.blendMatrixIndices.lpData + (unsigned long) vb->resource.allocatedMemory);
 968     }
 969     if(s->u.s.normal.VBO) {
 970         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.normal.streamNo];
 971         s->u.s.normal.VBO = 0;
 972         s->u.s.normal.lpData = (BYTE *) ((unsigned long) s->u.s.normal.lpData + (unsigned long) vb->resource.allocatedMemory);
 973     }
 974     if(s->u.s.pSize.VBO) {
 975         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.pSize.streamNo];
 976         s->u.s.pSize.VBO = 0;
 977         s->u.s.pSize.lpData = (BYTE *) ((unsigned long) s->u.s.pSize.lpData + (unsigned long) vb->resource.allocatedMemory);
 978     }
 979     if(s->u.s.diffuse.VBO) {
 980         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.diffuse.streamNo];
 981         s->u.s.diffuse.VBO = 0;
 982         s->u.s.diffuse.lpData = (BYTE *) ((unsigned long) s->u.s.diffuse.lpData + (unsigned long) vb->resource.allocatedMemory);
 983     }
 984     if(s->u.s.specular.VBO) {
 985         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.specular.streamNo];
 986         s->u.s.specular.VBO = 0;
 987         s->u.s.specular.lpData = (BYTE *) ((unsigned long) s->u.s.specular.lpData + (unsigned long) vb->resource.allocatedMemory);
 988     }
 989     for(i = 0; i < WINED3DDP_MAXTEXCOORD; i++) {
 990         if(s->u.s.texCoords[i].VBO) {
 991             vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.texCoords[i].streamNo];
 992             s->u.s.texCoords[i].VBO = 0;
 993             s->u.s.texCoords[i].lpData = (BYTE *) ((unsigned long) s->u.s.texCoords[i].lpData + (unsigned long) vb->resource.allocatedMemory);
 994         }
 995     }
 996     if(s->u.s.position2.VBO) {
 997         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.position2.streamNo];
 998         s->u.s.position2.VBO = 0;
 999         s->u.s.position2.lpData = (BYTE *) ((unsigned long) s->u.s.position2.lpData + (unsigned long) vb->resource.allocatedMemory);
1000     }
1001     if(s->u.s.normal2.VBO) {
1002         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.normal2.streamNo];
1003         s->u.s.normal2.VBO = 0;
1004         s->u.s.normal2.lpData = (BYTE *) ((unsigned long) s->u.s.normal2.lpData + (unsigned long) vb->resource.allocatedMemory);
1005     }
1006     if(s->u.s.tangent.VBO) {
1007         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.tangent.streamNo];
1008         s->u.s.tangent.VBO = 0;
1009         s->u.s.tangent.lpData = (BYTE *) ((unsigned long) s->u.s.tangent.lpData + (unsigned long) vb->resource.allocatedMemory);
1010     }
1011     if(s->u.s.binormal.VBO) {
1012         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.binormal.streamNo];
1013         s->u.s.binormal.VBO = 0;
1014         s->u.s.binormal.lpData = (BYTE *) ((unsigned long) s->u.s.binormal.lpData + (unsigned long) vb->resource.allocatedMemory);
1015     }
1016     if(s->u.s.tessFactor.VBO) {
1017         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.tessFactor.streamNo];
1018         s->u.s.tessFactor.VBO = 0;
1019         s->u.s.tessFactor.lpData = (BYTE *) ((unsigned long) s->u.s.tessFactor.lpData + (unsigned long) vb->resource.allocatedMemory);
1020     }
1021     if(s->u.s.fog.VBO) {
1022         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.fog.streamNo];
1023         s->u.s.fog.VBO = 0;
1024         s->u.s.fog.lpData = (BYTE *) ((unsigned long) s->u.s.fog.lpData + (unsigned long) vb->resource.allocatedMemory);
1025     }
1026     if(s->u.s.depth.VBO) {
1027         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.depth.streamNo];
1028         s->u.s.depth.VBO = 0;
1029         s->u.s.depth.lpData = (BYTE *) ((unsigned long) s->u.s.depth.lpData + (unsigned long) vb->resource.allocatedMemory);
1030     }
1031     if(s->u.s.sample.VBO) {
1032         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.sample.streamNo];
1033         s->u.s.sample.VBO = 0;
1034         s->u.s.sample.lpData = (BYTE *) ((unsigned long) s->u.s.sample.lpData + (unsigned long) vb->resource.allocatedMemory);
1035     }
1036 }
1037
1038 /* Routine common to the draw primitive and draw indexed primitive routines */
1039 void drawPrimitive(IWineD3DDevice *iface,
1040                    int PrimitiveType,
1041                    long NumPrimitives,
1042                    /* for Indexed: */
1043                    long  StartVertexIndex,
1044                    UINT  numberOfVertices,
1045                    long  StartIdx,
1046                    short idxSize,
1047                    const void *idxData,
1048                    int   minIndex) {
1049
1050     IWineD3DDeviceImpl           *This = (IWineD3DDeviceImpl *)iface;
1051     IWineD3DSwapChain            *swapchain;
1052     IWineD3DBaseTexture          *texture = NULL;
1053     IWineD3DSurfaceImpl          *target;
1054     int i;
1055
1056     /* Signals other modules that a drawing is in progress and the stateblock finalized */
1057     This->isInDraw = TRUE;
1058
1059     /* Invalidate the back buffer memory so LockRect will read it the next time */
1060     for(i = 0; i < GL_LIMITS(buffers); i++) {
1061         target = (IWineD3DSurfaceImpl *) This->render_targets[i];
1062
1063         /* TODO: Only do all that if we're going to change anything
1064          * Texture container dirtification does not work quite right yet
1065          */
1066         if(target /*&& target->Flags & (SFLAG_INTEXTURE | SFLAG_INSYSMEM)*/) {
1067             swapchain = NULL;
1068             texture = NULL;
1069
1070             if(i == 0) {
1071                 IWineD3DSurface_GetContainer((IWineD3DSurface *) target, &IID_IWineD3DSwapChain, (void **)&swapchain);
1072
1073                 /* Need the surface in the drawable! */
1074                 if(!(target->Flags & SFLAG_INDRAWABLE) && (swapchain || wined3d_settings.offscreen_rendering_mode != ORM_FBO)) {
1075                     blt_to_drawable(This, target);
1076                 }
1077
1078                 if(swapchain) {
1079                     /* Onscreen target. Invalidate system memory copy and texture copy */
1080                     target->Flags &= ~(SFLAG_INSYSMEM | SFLAG_INTEXTURE);
1081                     target->Flags |= SFLAG_INDRAWABLE;
1082                     IWineD3DSwapChain_Release(swapchain);
1083                 } else if(wined3d_settings.offscreen_rendering_mode != ORM_FBO) {
1084                     /* Non-FBO target: Invalidate system copy, texture copy and dirtify the container */
1085                     IWineD3DSurface_GetContainer((IWineD3DSurface *) target, &IID_IWineD3DBaseTexture, (void **)&texture);
1086
1087                     if(texture) {
1088                         IWineD3DBaseTexture_SetDirty(texture, TRUE);
1089                         IWineD3DTexture_Release(texture);
1090                     }
1091
1092                     target->Flags &= ~(SFLAG_INSYSMEM | SFLAG_INTEXTURE);
1093                     target->Flags |= SFLAG_INDRAWABLE;
1094                 } else {
1095                     /* FBO offscreen target. Invalidate system memory copy */
1096                     target->Flags &= ~SFLAG_INSYSMEM;
1097                 }
1098             } else {
1099                 /* Must be an fbo render target */
1100                 target->Flags &= ~SFLAG_INSYSMEM;
1101                 target->Flags |=  SFLAG_INTEXTURE;
1102             }
1103         }
1104     }
1105
1106     /* Ok, we will be updating the screen from here onwards so grab the lock */
1107
1108     if (wined3d_settings.offscreen_rendering_mode == ORM_FBO) {
1109         ENTER_GL();
1110         apply_fbo_state(iface);
1111         LEAVE_GL();
1112     }
1113
1114     ActivateContext(This, This->render_targets[0], CTXUSAGE_DRAWPRIM);
1115     ENTER_GL();
1116
1117     if (This->depth_copy_state == WINED3D_DCS_COPY) {
1118         depth_copy(iface);
1119     }
1120     This->depth_copy_state = WINED3D_DCS_INITIAL;
1121
1122     {
1123         GLenum glPrimType;
1124         BOOL emulation = FALSE;
1125         WineDirect3DVertexStridedData *strided = &This->strided_streams;
1126         WineDirect3DVertexStridedData stridedlcl;
1127         /* Ok, Work out which primitive is requested and how many vertexes that
1128            will be                                                              */
1129         UINT calculatedNumberOfindices = primitiveToGl(PrimitiveType, NumPrimitives, &glPrimType);
1130         if (numberOfVertices == 0 )
1131             numberOfVertices = calculatedNumberOfindices;
1132
1133         if(!use_vs(This)) {
1134             if(!This->strided_streams.u.s.position_transformed && This->activeContext->num_untracked_materials &&
1135                 This->stateBlock->renderState[WINED3DRS_LIGHTING]) {
1136                 FIXME("Using software emulation because not all material properties could be tracked\n");
1137                 emulation = TRUE;
1138             }
1139             else if(This->activeContext->fog_coord && This->stateBlock->renderState[WINED3DRS_FOGENABLE]) {
1140                 /* Either write a pipeline replacement shader or convert the specular alpha from unsigned byte
1141                  * to a float in the vertex buffer
1142                  */
1143                 FIXME("Using software emulation because manual fog coordinates are provided\n");
1144                 emulation = TRUE;
1145             }
1146
1147             if(emulation) {
1148                 strided = &stridedlcl;
1149                 memcpy(&stridedlcl, &This->strided_streams, sizeof(stridedlcl));
1150                 remove_vbos(This, &stridedlcl);
1151             }
1152         }
1153
1154         if (This->useDrawStridedSlow || emulation) {
1155             /* Immediate mode drawing */
1156             drawStridedSlow(iface, strided, calculatedNumberOfindices,
1157                             glPrimType, idxData, idxSize, minIndex, StartIdx, StartVertexIndex);
1158         } else if(This->instancedDraw) {
1159             /* Instancing emulation with mixing immediate mode and arrays */
1160             drawStridedInstanced(iface, &This->strided_streams, calculatedNumberOfindices, glPrimType,
1161                             idxData, idxSize, minIndex, StartIdx, StartVertexIndex);
1162         } else {
1163             /* Simple array draw call */
1164             drawStridedFast(iface, calculatedNumberOfindices, glPrimType,
1165                             idxData, idxSize, minIndex, StartIdx, StartVertexIndex);
1166         }
1167     }
1168
1169     /* Finshed updating the screen, restore lock */
1170     LEAVE_GL();
1171     TRACE("Done all gl drawing\n");
1172
1173     /* Diagnostics */
1174 #ifdef SHOW_FRAME_MAKEUP
1175     {
1176         static long int primCounter = 0;
1177         /* NOTE: set primCounter to the value reported by drawprim
1178            before you want to to write frame makeup to /tmp */
1179         if (primCounter >= 0) {
1180             WINED3DLOCKED_RECT r;
1181             char buffer[80];
1182             IWineD3DSurface_LockRect(This->renderTarget, &r, NULL, WINED3DLOCK_READONLY);
1183             sprintf(buffer, "/tmp/backbuffer_%d.tga", primCounter);
1184             TRACE("Saving screenshot %s\n", buffer);
1185             IWineD3DSurface_SaveSnapshot(This->renderTarget, buffer);
1186             IWineD3DSurface_UnlockRect(This->renderTarget);
1187
1188 #ifdef SHOW_TEXTURE_MAKEUP
1189            {
1190             IWineD3DSurface *pSur;
1191             int textureNo;
1192             for (textureNo = 0; textureNo < MAX_COMBINED_SAMPLERS; ++textureNo) {
1193                 if (This->stateBlock->textures[textureNo] != NULL) {
1194                     sprintf(buffer, "/tmp/texture_%p_%d_%d.tga", This->stateBlock->textures[textureNo], primCounter, textureNo);
1195                     TRACE("Saving texture %s\n", buffer);
1196                     if (IWineD3DBaseTexture_GetType(This->stateBlock->textures[textureNo]) == WINED3DRTYPE_TEXTURE) {
1197                             IWineD3DTexture_GetSurfaceLevel((IWineD3DTexture *)This->stateBlock->textures[textureNo], 0, &pSur);
1198                             IWineD3DSurface_SaveSnapshot(pSur, buffer);
1199                             IWineD3DSurface_Release(pSur);
1200                     } else  {
1201                         FIXME("base Texture isn't of type texture %d\n", IWineD3DBaseTexture_GetType(This->stateBlock->textures[textureNo]));
1202                     }
1203                 }
1204             }
1205            }
1206 #endif
1207         }
1208         TRACE("drawprim #%d\n", primCounter);
1209         ++primCounter;
1210     }
1211 #endif
1212
1213     /* Control goes back to the device, stateblock values may change again */
1214     This->isInDraw = FALSE;
1215 }
1216
1217 static void normalize_normal(float *n) {
1218     float length = n[0] * n[0] + n[1] * n[1] + n[2] * n[2];
1219     if(length == 0.0) return;
1220     length = sqrt(length);
1221     n[0] = n[0] / length;
1222     n[1] = n[1] / length;
1223     n[2] = n[2] / length;
1224 }
1225
1226 /* Tesselates a high order rectangular patch into single triangles using gl evaluators
1227  *
1228  * The problem is that OpenGL does not offer a direct way to return the tesselated primitives,
1229  * and they can't be sent off for rendering directly either. Tesselating is slow, so we want
1230  * to chache the patches in a vertex buffer. But more importantly, gl can't bind generated
1231  * attributes to numbered shader attributes, so we have to store them and rebind them as needed
1232  * in drawprim.
1233  *
1234  * To read back, the opengl feedback mode is used. This creates a proplem because we want
1235  * untransformed, unlit vertices, but feedback runs everything through transform and lighting.
1236  * Thus disable lighting and set identity matrices to get unmodified colors and positions.
1237  * To overcome clipping find the biggest x, y and z values of the vertices in the patch and scale
1238  * them to [-1.0;+1.0] and set the viewport up to scale them back.
1239  *
1240  * Normals are more tricky: Draw white vertices with 3 directional lights, and calculate the
1241  * resulting colors back to the normals.
1242  *
1243  * NOTE: This function activates a context for blitting, modifies matrices & viewport, but
1244  * does not restore it because normally a draw follows immediately afterwards. The caller is
1245  * responsible of taking care that either the gl states are restored, or the context activated
1246  * for drawing to reset the lastWasBlit flag.
1247  */
1248 HRESULT tesselate_rectpatch(IWineD3DDeviceImpl *This,
1249                             struct WineD3DRectPatch *patch) {
1250     unsigned int i, j, num_quads, out_vertex_size, buffer_size, d3d_out_vertex_size;
1251     float max_x = 0.0, max_y = 0.0, max_z = 0.0, neg_z = 0.0;
1252     WineDirect3DVertexStridedData strided;
1253     BYTE *data;
1254     WINED3DRECTPATCH_INFO *info = &patch->RectPatchInfo;
1255     DWORD vtxStride;
1256     GLenum feedback_type;
1257     GLfloat *feedbuffer;
1258
1259     /* First, locate the position data. This is provided in a vertex buffer in the stateblock.
1260      * Beware of vbos
1261      */
1262     memset(&strided, 0, sizeof(strided));
1263     primitiveDeclarationConvertToStridedData((IWineD3DDevice *) This, FALSE, &strided, NULL);
1264     if(strided.u.s.position.VBO) {
1265         IWineD3DVertexBufferImpl *vb;
1266         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[strided.u.s.position.streamNo];
1267         strided.u.s.position.lpData = (BYTE *) ((unsigned long) strided.u.s.position.lpData +
1268                                                 (unsigned long) vb->resource.allocatedMemory);
1269     }
1270     vtxStride = strided.u.s.position.dwStride;
1271     data = strided.u.s.position.lpData +
1272            vtxStride * info->Stride * info->StartVertexOffsetHeight +
1273            vtxStride * info->StartVertexOffsetWidth;
1274
1275     /* Not entirely sure about what happens with transformed vertices */
1276     if(strided.u.s.position_transformed) {
1277         FIXME("Transformed position in rectpatch generation\n");
1278     }
1279     if(vtxStride % sizeof(GLfloat)) {
1280         /* glMap2f reads vertex sizes in GLfloats, the d3d stride is in bytes.
1281          * I don't see how the stride could not be a multiple of 4, but make sure
1282          * to check it
1283          */
1284         ERR("Vertex stride is not a multiple of sizeof(GLfloat)\n");
1285     }
1286     if(info->Basis != WINED3DBASIS_BEZIER) {
1287         FIXME("Basis is %s, how to handle this?\n", debug_d3dbasis(info->Basis));
1288     }
1289     if(info->Degree != WINED3DDEGREE_CUBIC) {
1290         FIXME("Degree is %s, how to handle this?\n", debug_d3ddegree(info->Degree));
1291     }
1292
1293     /* First, get the boundary cube of the input data */
1294     for(j = 0; j < info->Height; j++) {
1295         for(i = 0; i < info->Width; i++) {
1296             float *v = (float *) (data + vtxStride * i + vtxStride * info->Stride * j);
1297             if(fabs(v[0]) > max_x) max_x = fabs(v[0]);
1298             if(fabs(v[1]) > max_y) max_y = fabs(v[1]);
1299             if(fabs(v[2]) > max_z) max_z = fabs(v[2]);
1300             if(v[2] < neg_z) neg_z = v[2];
1301         }
1302     }
1303
1304     /* This needs some improvements in the vertex decl code */
1305     FIXME("Cannot find data to generate. Only generating position and normals\n");
1306     patch->has_normals = TRUE;
1307     patch->has_texcoords = FALSE;
1308
1309     /* Simply activate the context for blitting. This disables all the things we don't want and
1310      * takes care of dirtifying. Dirtifying is preferred over pushing / popping, since drawing the
1311      * patch (as opposed to normal draws) will most likely need different changes anyway
1312      */
1313     ActivateContext(This, This->lastActiveRenderTarget, CTXUSAGE_BLIT);
1314     ENTER_GL();
1315
1316     glMatrixMode(GL_PROJECTION);
1317     checkGLcall("glMatrixMode(GL_PROJECTION)");
1318     glLoadIdentity();
1319     checkGLcall("glLoadIndentity()");
1320     glScalef(1 / (max_x) , 1 / (max_y), max_z == 0 ? 1 : 1 / ( 2 * max_z));
1321     glTranslatef(0, 0, 0.5);
1322     checkGLcall("glScalef");
1323     glViewport(-max_x, -max_y, 2 * (max_x), 2 * (max_y));
1324     checkGLcall("glViewport");
1325
1326     /* Some states to take care of. If we're in wireframe opengl will produce lines, and confuse
1327      * our feedback buffer parser
1328      */
1329     glPolygonMode(GL_FRONT_AND_BACK, GL_FILL);
1330     checkGLcall("glPolygonMode(GL_FRONT_AND_BACK, GL_FILL)");
1331     IWineD3DDeviceImpl_MarkStateDirty(This, STATE_RENDER(WINED3DRS_FILLMODE));
1332     if(patch->has_normals) {
1333         float black[4] = {0, 0, 0, 0};
1334         float red[4]   = {1, 0, 0, 0};
1335         float green[4] = {0, 1, 0, 0};
1336         float blue[4]  = {0, 0, 1, 0};
1337         float white[4] = {1, 1, 1, 1};
1338         glEnable(GL_LIGHTING);
1339         checkGLcall("glEnable(GL_LIGHTING)");
1340         glLightModelfv(GL_LIGHT_MODEL_AMBIENT, black);
1341         checkGLcall("glLightModel for MODEL_AMBIENT");
1342         IWineD3DDeviceImpl_MarkStateDirty(This, STATE_RENDER(WINED3DRS_AMBIENT));
1343
1344         for(i = 3; i < GL_LIMITS(lights); i++) {
1345             glDisable(GL_LIGHT0 + i);
1346             checkGLcall("glDisable(GL_LIGHT0 + i)");
1347             IWineD3DDeviceImpl_MarkStateDirty(This, STATE_ACTIVELIGHT(i));
1348         }
1349
1350         IWineD3DDeviceImpl_MarkStateDirty(This, STATE_ACTIVELIGHT(0));
1351         glLightfv(GL_LIGHT0, GL_DIFFUSE, red);
1352         glLightfv(GL_LIGHT0, GL_SPECULAR, black);
1353         glLightfv(GL_LIGHT0, GL_AMBIENT, black);
1354         glLightfv(GL_LIGHT0, GL_POSITION, red);
1355         glEnable(GL_LIGHT0);
1356         checkGLcall("Setting up light 1\n");
1357         IWineD3DDeviceImpl_MarkStateDirty(This, STATE_ACTIVELIGHT(1));
1358         glLightfv(GL_LIGHT1, GL_DIFFUSE, green);
1359         glLightfv(GL_LIGHT1, GL_SPECULAR, black);
1360         glLightfv(GL_LIGHT1, GL_AMBIENT, black);
1361         glLightfv(GL_LIGHT1, GL_POSITION, green);
1362         glEnable(GL_LIGHT1);
1363         checkGLcall("Setting up light 2\n");
1364         IWineD3DDeviceImpl_MarkStateDirty(This, STATE_ACTIVELIGHT(2));
1365         glLightfv(GL_LIGHT2, GL_DIFFUSE, blue);
1366         glLightfv(GL_LIGHT2, GL_SPECULAR, black);
1367         glLightfv(GL_LIGHT2, GL_AMBIENT, black);
1368         glLightfv(GL_LIGHT2, GL_POSITION, blue);
1369         glEnable(GL_LIGHT2);
1370         checkGLcall("Setting up light 3\n");
1371
1372         IWineD3DDeviceImpl_MarkStateDirty(This, STATE_MATERIAL);
1373         IWineD3DDeviceImpl_MarkStateDirty(This, STATE_RENDER(WINED3DRS_COLORVERTEX));
1374         glDisable(GL_COLOR_MATERIAL);
1375         glMaterialfv(GL_FRONT_AND_BACK, GL_EMISSION, black);
1376         glMaterialfv(GL_FRONT_AND_BACK, GL_SPECULAR, black);
1377         glMaterialfv(GL_FRONT_AND_BACK, GL_DIFFUSE, white);
1378         checkGLcall("Setting up materials\n");
1379     }
1380
1381     /* Enable the needed maps.
1382      * GL_MAP2_VERTEX_3 is needed for positional data.
1383      * GL_AUTO_NORMAL to generate normals from the position. Do not use GL_MAP2_NORMAL.
1384      * GL_MAP2_TEXTURE_COORD_4 for texture coords
1385      */
1386     num_quads = ceilf(patch->numSegs[0]) * ceilf(patch->numSegs[1]);
1387     out_vertex_size = 3 /* position */;
1388     d3d_out_vertex_size = 3;
1389     glEnable(GL_MAP2_VERTEX_3);
1390     if(patch->has_normals && patch->has_texcoords) {
1391         FIXME("Texcoords not handled yet\n");
1392         feedback_type = GL_3D_COLOR_TEXTURE;
1393         out_vertex_size += 8;
1394         d3d_out_vertex_size += 7;
1395         glEnable(GL_AUTO_NORMAL);
1396         glEnable(GL_MAP2_TEXTURE_COORD_4);
1397     } else if(patch->has_texcoords) {
1398         FIXME("Texcoords not handled yet\n");
1399         feedback_type = GL_3D_COLOR_TEXTURE;
1400         out_vertex_size += 7;
1401         d3d_out_vertex_size += 4;
1402         glEnable(GL_MAP2_TEXTURE_COORD_4);
1403     } else if(patch->has_normals) {
1404         feedback_type = GL_3D_COLOR;
1405         out_vertex_size += 4;
1406         d3d_out_vertex_size += 3;
1407         glEnable(GL_AUTO_NORMAL);
1408     } else {
1409         feedback_type = GL_3D;
1410     }
1411     checkGLcall("glEnable vertex attrib generation");
1412
1413     buffer_size = num_quads * out_vertex_size * 2 /* triangle list */ * 3 /* verts per tri */
1414                    + 4 * num_quads /* 2 triangle markers per quad + num verts in tri */;
1415     feedbuffer = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, buffer_size * sizeof(float) * 8);
1416
1417     glMap2f(GL_MAP2_VERTEX_3,
1418             0, 1, vtxStride / sizeof(float), info->Width,
1419             0, 1, info->Stride * vtxStride / sizeof(float), info->Height,
1420             (float *) data);
1421     checkGLcall("glMap2f");
1422     if(patch->has_texcoords) {
1423         glMap2f(GL_MAP2_TEXTURE_COORD_4,
1424                 0, 1, vtxStride / sizeof(float), info->Width,
1425                 0, 1, info->Stride * vtxStride / sizeof(float), info->Height,
1426                 (float *) data);
1427         checkGLcall("glMap2f");
1428     }
1429     glMapGrid2f(ceilf(patch->numSegs[0]), 0.0, 1.0, ceilf(patch->numSegs[1]), 0.0, 1.0);
1430     checkGLcall("glMapGrid2f");
1431
1432     glFeedbackBuffer(buffer_size * 2, feedback_type, feedbuffer);
1433     checkGLcall("glFeedbackBuffer");
1434     glRenderMode(GL_FEEDBACK);
1435
1436     glEvalMesh2(GL_FILL, 0, ceilf(patch->numSegs[0]), 0, ceilf(patch->numSegs[1]));
1437     checkGLcall("glEvalMesh2\n");
1438
1439     i = glRenderMode(GL_RENDER);
1440     if(i == -1) {
1441         ERR("Feedback failed. Expected %d elements back\n", buffer_size);
1442         Sleep(10000);
1443         HeapFree(GetProcessHeap(), 0, feedbuffer);
1444         return WINED3DERR_DRIVERINTERNALERROR;
1445     } else if(i != buffer_size) {
1446         ERR("Unexpected amount of elements returned. Expected %d, got %d\n", buffer_size, i);
1447         Sleep(10000);
1448         HeapFree(GetProcessHeap(), 0, feedbuffer);
1449         return WINED3DERR_DRIVERINTERNALERROR;
1450     } else {
1451         TRACE("Got %d elements as expected\n", i);
1452     }
1453
1454     HeapFree(GetProcessHeap(), 0, patch->mem);
1455     patch->mem = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, num_quads * 6 * d3d_out_vertex_size * sizeof(float) * 8);
1456     i = 0;
1457     for(j = 0; j < buffer_size; j += (3 /* num verts */ * out_vertex_size + 2 /* tri marker */)) {
1458         if(feedbuffer[j] != GL_POLYGON_TOKEN) {
1459             ERR("Unexpected token: %f\n", feedbuffer[j]);
1460             continue;
1461         }
1462         if(feedbuffer[j + 1] != 3) {
1463             ERR("Unexpected polygon: %f corners\n", feedbuffer[j + 1]);
1464             continue;
1465         }
1466         /* Somehow there are different ideas about back / front facing, so fix up the
1467          * vertex order
1468          */
1469         patch->mem[i + 0] =  feedbuffer[j + out_vertex_size * 2 + 2]; /* x, triangle 2 */
1470         patch->mem[i + 1] =  feedbuffer[j + out_vertex_size * 2 + 3]; /* y, triangle 2 */
1471         patch->mem[i + 2] = (feedbuffer[j + out_vertex_size * 2 + 4] - 0.5) * 4 * max_z; /* z, triangle 3 */
1472         if(patch->has_normals) {
1473             patch->mem[i + 3] = feedbuffer[j + out_vertex_size * 2 + 5];
1474             patch->mem[i + 4] = feedbuffer[j + out_vertex_size * 2 + 6];
1475             patch->mem[i + 5] = feedbuffer[j + out_vertex_size * 2 + 7];
1476         }
1477         i += d3d_out_vertex_size;
1478
1479         patch->mem[i + 0] =  feedbuffer[j + out_vertex_size * 1 + 2]; /* x, triangle 2 */
1480         patch->mem[i + 1] =  feedbuffer[j + out_vertex_size * 1 + 3]; /* y, triangle 2 */
1481         patch->mem[i + 2] = (feedbuffer[j + out_vertex_size * 1 + 4] - 0.5) * 4 * max_z; /* z, triangle 2 */
1482         if(patch->has_normals) {
1483             patch->mem[i + 3] = feedbuffer[j + out_vertex_size * 1 + 5];
1484             patch->mem[i + 4] = feedbuffer[j + out_vertex_size * 1 + 6];
1485             patch->mem[i + 5] = feedbuffer[j + out_vertex_size * 1 + 7];
1486         }
1487         i += d3d_out_vertex_size;
1488
1489         patch->mem[i + 0] =  feedbuffer[j + out_vertex_size * 0 + 2]; /* x, triangle 1 */
1490         patch->mem[i + 1] =  feedbuffer[j + out_vertex_size * 0 + 3]; /* y, triangle 1 */
1491         patch->mem[i + 2] = (feedbuffer[j + out_vertex_size * 0 + 4] - 0.5) * 4 * max_z; /* z, triangle 1 */
1492         if(patch->has_normals) {
1493             patch->mem[i + 3] = feedbuffer[j + out_vertex_size * 0 + 5];
1494             patch->mem[i + 4] = feedbuffer[j + out_vertex_size * 0 + 6];
1495             patch->mem[i + 5] = feedbuffer[j + out_vertex_size * 0 + 7];
1496         }
1497         i += d3d_out_vertex_size;
1498     }
1499
1500     if(patch->has_normals) {
1501         /* Now do the same with reverse light directions */
1502         float x[4] = {-1,  0,  0, 0};
1503         float y[4] = { 0, -1,  0, 0};
1504         float z[4] = { 0,  0, -1, 0};
1505         glLightfv(GL_LIGHT0, GL_POSITION, x);
1506         glLightfv(GL_LIGHT1, GL_POSITION, y);
1507         glLightfv(GL_LIGHT2, GL_POSITION, z);
1508         checkGLcall("Setting up reverse light directions\n");
1509
1510         glRenderMode(GL_FEEDBACK);
1511         checkGLcall("glRenderMode(GL_FEEDBACK)");
1512         glEvalMesh2(GL_FILL, 0, ceilf(patch->numSegs[0]), 0, ceilf(patch->numSegs[1]));
1513         checkGLcall("glEvalMesh2\n");
1514         i = glRenderMode(GL_RENDER);
1515         checkGLcall("glRenderMode(GL_RENDER)");
1516
1517         i = 0;
1518         for(j = 0; j < buffer_size; j += (3 /* num verts */ * out_vertex_size + 2 /* tri marker */)) {
1519             if(feedbuffer[j] != GL_POLYGON_TOKEN) {
1520                 ERR("Unexpected token: %f\n", feedbuffer[j]);
1521                 continue;
1522             }
1523             if(feedbuffer[j + 1] != 3) {
1524                 ERR("Unexpected polygon: %f corners\n", feedbuffer[j + 1]);
1525                 continue;
1526             }
1527             if(patch->mem[i + 3] == 0.0)
1528                 patch->mem[i + 3] = -feedbuffer[j + out_vertex_size * 2 + 5];
1529             if(patch->mem[i + 4] == 0.0)
1530                 patch->mem[i + 4] = -feedbuffer[j + out_vertex_size * 2 + 6];
1531             if(patch->mem[i + 5] == 0.0)
1532                 patch->mem[i + 5] = -feedbuffer[j + out_vertex_size * 2 + 7];
1533             normalize_normal(patch->mem + i + 3);
1534             i += d3d_out_vertex_size;
1535
1536             if(patch->mem[i + 3] == 0.0)
1537                 patch->mem[i + 3] = -feedbuffer[j + out_vertex_size * 1 + 5];
1538             if(patch->mem[i + 4] == 0.0)
1539                 patch->mem[i + 4] = -feedbuffer[j + out_vertex_size * 1 + 6];
1540             if(patch->mem[i + 5] == 0.0)
1541                 patch->mem[i + 5] = -feedbuffer[j + out_vertex_size * 1 + 7];
1542             normalize_normal(patch->mem + i + 3);
1543             i += d3d_out_vertex_size;
1544
1545             if(patch->mem[i + 3] == 0.0)
1546                 patch->mem[i + 3] = -feedbuffer[j + out_vertex_size * 0 + 5];
1547             if(patch->mem[i + 4] == 0.0)
1548                 patch->mem[i + 4] = -feedbuffer[j + out_vertex_size * 0 + 6];
1549             if(patch->mem[i + 5] == 0.0)
1550                 patch->mem[i + 5] = -feedbuffer[j + out_vertex_size * 0 + 7];
1551             normalize_normal(patch->mem + i + 3);
1552             i += d3d_out_vertex_size;
1553         }
1554     }
1555
1556     glDisable(GL_MAP2_VERTEX_3);
1557     glDisable(GL_AUTO_NORMAL);
1558     glDisable(GL_MAP2_NORMAL);
1559     glDisable(GL_MAP2_TEXTURE_COORD_4);
1560     checkGLcall("glDisable vertex attrib generation");
1561     LEAVE_GL();
1562
1563     HeapFree(GetProcessHeap(), 0, feedbuffer);
1564
1565     vtxStride = 3 * sizeof(float);
1566     if(patch->has_normals) {
1567         vtxStride += 3 * sizeof(float);
1568     }
1569     if(patch->has_texcoords) {
1570         vtxStride += 4 * sizeof(float);
1571     }
1572     memset(&patch->strided, 0, sizeof(&patch->strided));
1573     patch->strided.u.s.position.lpData = (BYTE *) patch->mem;
1574     patch->strided.u.s.position.dwStride = vtxStride;
1575     patch->strided.u.s.position.dwType = WINED3DDECLTYPE_FLOAT3;
1576     patch->strided.u.s.position.streamNo = 255;
1577
1578     if(patch->has_normals) {
1579         patch->strided.u.s.normal.lpData = (BYTE *) patch->mem + 3 * sizeof(float) /* pos */;
1580         patch->strided.u.s.normal.dwStride = vtxStride;
1581         patch->strided.u.s.normal.dwType = WINED3DDECLTYPE_FLOAT3;
1582         patch->strided.u.s.normal.streamNo = 255;
1583     }
1584     if(patch->has_texcoords) {
1585         patch->strided.u.s.texCoords[0].lpData = (BYTE *) patch->mem + 3 * sizeof(float) /* pos */;
1586         if(patch->has_normals) {
1587             patch->strided.u.s.texCoords[0].lpData += 3 * sizeof(float);
1588         }
1589         patch->strided.u.s.texCoords[0].dwStride = vtxStride;
1590         patch->strided.u.s.texCoords[0].dwType = WINED3DDECLTYPE_FLOAT4;
1591         /* MAX_STREAMS index points to an unused element in stateblock->streamOffsets which
1592          * always remains set to 0. Windows uses stream 255 here, but this is not visible to the
1593          * application.
1594          */
1595         patch->strided.u.s.texCoords[0].streamNo = MAX_STREAMS;
1596     }
1597
1598     return WINED3D_OK;
1599 }