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 *vb = This->stateBlock->streamSource[streams[i]];
 250         if(vb) {
 251             IWineD3DVertexBuffer_PreLoad(vb);
 252         }
 253     }
 254 }
 255
 256 static void drawStridedFast(IWineD3DDevice *iface,UINT numberOfVertices, GLenum glPrimitiveType,
 257                      const void *idxData, short idxSize, ULONG minIndex, ULONG startIdx, ULONG startVertex) {
 258     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
 259
 260     if (idxSize != 0 /* This crashes sometimes!*/) {
 261         TRACE("(%p) : glElements(%x, %d, %d, ...)\n", This, glPrimitiveType, numberOfVertices, minIndex);
 262         idxData = idxData == (void *)-1 ? NULL : idxData;
 263 #if 1
 264         glDrawElements(glPrimitiveType, numberOfVertices, idxSize == 2 ? GL_UNSIGNED_SHORT : GL_UNSIGNED_INT,
 265                      (const char *)idxData+(idxSize * startIdx));
 266         checkGLcall("glDrawElements");
 267 #else /* using drawRangeElements may be faster */
 268
 269         glDrawRangeElements(glPrimitiveType, minIndex, minIndex + numberOfVertices - 1, numberOfVertices,
 270                       idxSize == 2 ? GL_UNSIGNED_SHORT : GL_UNSIGNED_INT,
 271                       (const char *)idxData+(idxSize * startIdx));
 272         checkGLcall("glDrawRangeElements");
 273 #endif
 274
 275     } else {
 276
 277         /* Note first is now zero as we shuffled along earlier */
 278         TRACE("(%p) : glDrawArrays(%x, 0, %d)\n", This, glPrimitiveType, numberOfVertices);
 279         glDrawArrays(glPrimitiveType, startVertex, numberOfVertices);
 280         checkGLcall("glDrawArrays");
 281
 282     }
 283
 284     return;
 285 }
 286
 287 /*
 288  * Actually draw using the supplied information.
 289  * Slower GL version which extracts info about each vertex in turn
 290  */
 291
 292 static void drawStridedSlow(IWineD3DDevice *iface, WineDirect3DVertexStridedData *sd,
 293                      UINT NumVertexes, GLenum glPrimType,
 294                      const void *idxData, short idxSize, ULONG minIndex, ULONG startIdx, ULONG startVertex) {
 295
 296     unsigned int               textureNo    = 0;
 297     const WORD                *pIdxBufS     = NULL;
 298     const DWORD               *pIdxBufL     = NULL;
 299     LONG                       vx_index;
 300     float x  = 0.0f, y  = 0.0f, z = 0.0f;  /* x,y,z coordinates          */
 301     float rhw = 0.0f;                      /* rhw                        */
 302     DWORD diffuseColor = 0xFFFFFFFF;       /* Diffuse Color              */
 303     DWORD specularColor = 0;               /* Specular Color             */
 304     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
 305     UINT *streamOffset = This->stateBlock->streamOffset;
 306     long                      SkipnStrides = startVertex + This->stateBlock->loadBaseVertexIndex;
 307     BOOL                      pixelShader = use_ps(This);
 308
 309     BYTE *texCoords[WINED3DDP_MAXTEXCOORD];
 310     BYTE *diffuse = NULL, *specular = NULL, *normal = NULL, *position = NULL;
 311
 312     TRACE("Using slow vertex array code\n");
 313
 314     /* Variable Initialization */
 315     if (idxSize != 0) {
 316         /* Immediate mode drawing can't make use of indices in a vbo - get the data from the index buffer.
 317          * If the index buffer has no vbo(not supported or other reason), or with user pointer drawing
 318          * idxData will be != NULL
 319          */
 320         if(idxData == NULL) {
 321             idxData = ((IWineD3DIndexBufferImpl *) This->stateBlock->pIndexData)->resource.allocatedMemory;
 322         }
 323
 324         if (idxSize == 2) pIdxBufS = (const WORD *) idxData;
 325         else pIdxBufL = (const DWORD *) idxData;
 326     }
 327
 328     /* Adding the stream offset once is cheaper than doing it every iteration. Do not modify the strided data, it is a pointer
 329      * to the strided Data in the device and might be needed intact on the next draw
 330      */
 331     for (textureNo = 0; textureNo < GL_LIMITS(texture_stages); ++textureNo) {
 332         if(sd->u.s.texCoords[textureNo].lpData) {
 333             texCoords[textureNo] = sd->u.s.texCoords[textureNo].lpData + streamOffset[sd->u.s.texCoords[textureNo].streamNo];
 334         } else {
 335             texCoords[textureNo] = NULL;
 336         }
 337     }
 338     if(sd->u.s.diffuse.lpData) {
 339         diffuse = sd->u.s.diffuse.lpData + streamOffset[sd->u.s.diffuse.streamNo];
 340     }
 341     if(sd->u.s.specular.lpData) {
 342         specular = sd->u.s.specular.lpData + streamOffset[sd->u.s.specular.streamNo];
 343     }
 344     if(sd->u.s.normal.lpData) {
 345         normal = sd->u.s.normal.lpData + streamOffset[sd->u.s.normal.streamNo];
 346     }
 347     if(sd->u.s.position.lpData) {
 348         position = sd->u.s.position.lpData + streamOffset[sd->u.s.position.streamNo];
 349     }
 350
 351     /* Start drawing in GL */
 352     VTRACE(("glBegin(%x)\n", glPrimType));
 353     glBegin(glPrimType);
 354
 355     /* Default settings for data that is not passed */
 356     if (sd->u.s.normal.lpData == NULL) {
 357         glNormal3f(0, 0, 0);
 358     }
 359     if(sd->u.s.diffuse.lpData == NULL) {
 360         glColor4f(1.0f, 1.0f, 1.0f, 1.0f);
 361     }
 362     if(sd->u.s.specular.lpData == NULL) {
 363         if (GL_SUPPORT(EXT_SECONDARY_COLOR)) {
 364             GL_EXTCALL(glSecondaryColor3fEXT)(0, 0, 0);
 365         }
 366     }
 367
 368     /* We shouldn't start this function if any VBO is involved. Should I put a safety check here?
 369      * Guess it's not necessary(we crash then anyway) and would only eat CPU time
 370      */
 371
 372     /* For each primitive */
 373     for (vx_index = 0; vx_index < NumVertexes; ++vx_index) {
 374
 375         /* Initialize diffuse color */
 376         diffuseColor = 0xFFFFFFFF;
 377
 378         /* Blending data and Point sizes are not supported by this function. They are not supported by the fixed
 379          * function pipeline at all. A Fixme for them is printed after decoding the vertex declaration
 380          */
 381
 382         /* For indexed data, we need to go a few more strides in */
 383         if (idxData != NULL) {
 384
 385             /* Indexed so work out the number of strides to skip */
 386             if (idxSize == 2) {
 387                 VTRACE(("Idx for vertex %d = %d\n", vx_index, pIdxBufS[startIdx+vx_index]));
 388                 SkipnStrides = pIdxBufS[startIdx + vx_index] + This->stateBlock->loadBaseVertexIndex;
 389             } else {
 390                 VTRACE(("Idx for vertex %d = %d\n", vx_index, pIdxBufL[startIdx+vx_index]));
 391                 SkipnStrides = pIdxBufL[startIdx + vx_index] + This->stateBlock->loadBaseVertexIndex;
 392             }
 393         }
 394
 395         /* Texture coords --------------------------- */
 396         for (textureNo = 0; textureNo < GL_LIMITS(texture_stages); ++textureNo) {
 397
 398             if (!GL_SUPPORT(ARB_MULTITEXTURE) && textureNo > 0) {
 399                 FIXME("Program using multiple concurrent textures which this opengl implementation doesn't support\n");
 400                 continue ;
 401             }
 402
 403             /* Query tex coords */
 404             if (This->stateBlock->textures[textureNo] != NULL || pixelShader) {
 405
 406                 int    coordIdx = This->stateBlock->textureState[textureNo][WINED3DTSS_TEXCOORDINDEX];
 407                 int texture_idx = This->texUnitMap[textureNo];
 408                 float *ptrToCoords = NULL;
 409                 float  s = 0.0, t = 0.0, r = 0.0, q = 0.0;
 410
 411                 if (coordIdx > 7) {
 412                     VTRACE(("tex: %d - Skip tex coords, as being system generated\n", textureNo));
 413                     continue;
 414                 } else if (coordIdx < 0) {
 415                     FIXME("tex: %d - Coord index %d is less than zero, expect a crash.\n", textureNo, coordIdx);
 416                     continue;
 417                 }
 418
 419                 ptrToCoords = (float *)(texCoords[coordIdx] + (SkipnStrides * sd->u.s.texCoords[coordIdx].dwStride));
 420                 if (texCoords[coordIdx] == NULL) {
 421                     TRACE("tex: %d - Skipping tex coords, as no data supplied\n", textureNo);
 422                     if (GL_SUPPORT(ARB_MULTITEXTURE)) {
 423                         GL_EXTCALL(glMultiTexCoord4fARB(GL_TEXTURE0_ARB + texture_idx, 0, 0, 0, 1));
 424                     } else {
 425                         glTexCoord4f(0, 0, 0, 1);
 426                     }
 427                     continue;
 428                 } else {
 429                     int coordsToUse = sd->u.s.texCoords[coordIdx].dwType + 1; /* 0 == WINED3DDECLTYPE_FLOAT1 etc */
 430
 431                     if (texture_idx == -1) continue;
 432
 433                     /* The coords to supply depend completely on the fvf / vertex shader */
 434                     switch (coordsToUse) {
 435                     case 4: q = ptrToCoords[3]; /* drop through */
 436                     case 3: r = ptrToCoords[2]; /* drop through */
 437                     case 2: t = ptrToCoords[1]; /* drop through */
 438                     case 1: s = ptrToCoords[0];
 439                     }
 440
 441                     switch (coordsToUse) {   /* Supply the provided texture coords */
 442                     case WINED3DTTFF_COUNT1:
 443                         VTRACE(("tex:%d, s=%f\n", textureNo, s));
 444                         if (GL_SUPPORT(ARB_MULTITEXTURE)) {
 445                             GL_EXTCALL(glMultiTexCoord1fARB(GL_TEXTURE0_ARB + texture_idx, s));
 446                         } else {
 447                             glTexCoord1f(s);
 448                         }
 449                         break;
 450                     case WINED3DTTFF_COUNT2:
 451                         VTRACE(("tex:%d, s=%f, t=%f\n", textureNo, s, t));
 452                         if (GL_SUPPORT(ARB_MULTITEXTURE)) {
 453                             GL_EXTCALL(glMultiTexCoord2fARB(GL_TEXTURE0_ARB + texture_idx, s, t));
 454                         } else {
 455                             glTexCoord2f(s, t);
 456                         }
 457                         break;
 458                     case WINED3DTTFF_COUNT3:
 459                         VTRACE(("tex:%d, s=%f, t=%f, r=%f\n", textureNo, s, t, r));
 460                         if (GL_SUPPORT(ARB_MULTITEXTURE)) {
 461                             GL_EXTCALL(glMultiTexCoord3fARB(GL_TEXTURE0_ARB + texture_idx, s, t, r));
 462                         } else {
 463                             glTexCoord3f(s, t, r);
 464                         }
 465                         break;
 466                     case WINED3DTTFF_COUNT4:
 467                         VTRACE(("tex:%d, s=%f, t=%f, r=%f, q=%f\n", textureNo, s, t, r, q));
 468                         if (GL_SUPPORT(ARB_MULTITEXTURE)) {
 469                             GL_EXTCALL(glMultiTexCoord4fARB(GL_TEXTURE0_ARB + texture_idx, s, t, r, q));
 470                         } else {
 471                             glTexCoord4f(s, t, r, q);
 472                         }
 473                         break;
 474                     default:
 475                         FIXME("Should not get here as coordsToUse is two bits only (%x)!\n", coordsToUse);
 476                     }
 477                 }
 478             }
 479         } /* End of textures */
 480
 481         /* Diffuse -------------------------------- */
 482         if (diffuse) {
 483             DWORD *ptrToCoords = (DWORD *)(diffuse + (SkipnStrides * sd->u.s.diffuse.dwStride));
 484             diffuseColor = ptrToCoords[0];
 485             VTRACE(("diffuseColor=%lx\n", diffuseColor));
 486
 487             glColor4ub(D3DCOLOR_B_R(diffuseColor),
 488                      D3DCOLOR_B_G(diffuseColor),
 489                      D3DCOLOR_B_B(diffuseColor),
 490                      D3DCOLOR_B_A(diffuseColor));
 491             VTRACE(("glColor4ub: r,g,b,a=%lu,%lu,%lu,%lu\n",
 492                     D3DCOLOR_B_R(diffuseColor),
 493                     D3DCOLOR_B_G(diffuseColor),
 494                     D3DCOLOR_B_B(diffuseColor),
 495                     D3DCOLOR_B_A(diffuseColor)));
 496
 497             if(This->activeContext->num_untracked_materials) {
 498                 unsigned char i;
 499                 float color[4];
 500                 color[0] = D3DCOLOR_B_R(diffuseColor) / 255.0;
 501                 color[1] = D3DCOLOR_B_G(diffuseColor) / 255.0;
 502                 color[2] = D3DCOLOR_B_B(diffuseColor) / 255.0;
 503                 color[3] = D3DCOLOR_B_A(diffuseColor) / 255.0;
 504
 505                 for(i = 0; i < This->activeContext->num_untracked_materials; i++) {
 506                     glMaterialfv(GL_FRONT_AND_BACK, This->activeContext->untracked_materials[i], color);
 507                 }
 508             }
 509         }
 510
 511         /* Specular ------------------------------- */
 512         if (specular) {
 513             DWORD *ptrToCoords = (DWORD *)(specular + (SkipnStrides * sd->u.s.specular.dwStride));
 514             specularColor = ptrToCoords[0];
 515             VTRACE(("specularColor=%lx\n", specularColor));
 516
 517             /* special case where the fog density is stored in the specular alpha channel */
 518             if(This->stateBlock->renderState[WINED3DRS_FOGENABLE] &&
 519               (This->stateBlock->renderState[WINED3DRS_FOGVERTEXMODE] == WINED3DFOG_NONE || sd->u.s.position.dwType == WINED3DDECLTYPE_FLOAT4 )&&
 520               This->stateBlock->renderState[WINED3DRS_FOGTABLEMODE] == WINED3DFOG_NONE) {
 521                 if(GL_SUPPORT(EXT_FOG_COORD)) {
 522                     GL_EXTCALL(glFogCoordfEXT(specularColor >> 24));
 523                 } else {
 524                     static BOOL warned = FALSE;
 525                     if(!warned) {
 526                         /* TODO: Use the fog table code from old ddraw */
 527                         FIXME("Implement fog for transformed vertices in software\n");
 528                         warned = TRUE;
 529                     }
 530                 }
 531             }
 532
 533             VTRACE(("glSecondaryColor4ub: r,g,b=%lu,%lu,%lu\n",
 534                     D3DCOLOR_B_R(specularColor),
 535                     D3DCOLOR_B_G(specularColor),
 536                     D3DCOLOR_B_B(specularColor)));
 537             if (GL_SUPPORT(EXT_SECONDARY_COLOR)) {
 538                 GL_EXTCALL(glSecondaryColor3ubEXT)(
 539                            D3DCOLOR_B_R(specularColor),
 540                            D3DCOLOR_B_G(specularColor),
 541                            D3DCOLOR_B_B(specularColor));
 542             } else {
 543                 /* Do not worry if specular colour missing and disable request */
 544                 VTRACE(("Specular color extensions not supplied\n"));
 545             }
 546         }
 547
 548         /* Normal -------------------------------- */
 549         if (normal != NULL) {
 550             float *ptrToCoords = (float *)(normal + (SkipnStrides * sd->u.s.normal.dwStride));
 551
 552             VTRACE(("glNormal:nx,ny,nz=%f,%f,%f\n", ptrToCoords[0], ptrToCoords[1], ptrToCoords[2]));
 553             glNormal3f(ptrToCoords[0], ptrToCoords[1], ptrToCoords[2]);
 554         }
 555
 556         /* Position -------------------------------- */
 557         if (position) {
 558             float *ptrToCoords = (float *)(position + (SkipnStrides * sd->u.s.position.dwStride));
 559             x = ptrToCoords[0];
 560             y = ptrToCoords[1];
 561             z = ptrToCoords[2];
 562             rhw = 1.0;
 563             VTRACE(("x,y,z=%f,%f,%f\n", x,y,z));
 564
 565             /* RHW follows, only if transformed, ie 4 floats were provided */
 566             if (sd->u.s.position_transformed) {
 567                 rhw = ptrToCoords[3];
 568                 VTRACE(("rhw=%f\n", rhw));
 569             }
 570
 571             if (1.0f == rhw || ((rhw < eps) && (rhw > -eps))) {
 572                 VTRACE(("Vertex: glVertex:x,y,z=%f,%f,%f\n", x,y,z));
 573                 glVertex3f(x, y, z);
 574             } else {
 575                 GLfloat w = 1.0 / rhw;
 576                 VTRACE(("Vertex: glVertex:x,y,z=%f,%f,%f / rhw=%f\n", x,y,z,rhw));
 577                 glVertex4f(x*w, y*w, z*w, w);
 578             }
 579         }
 580
 581         /* For non indexed mode, step onto next parts */
 582         if (idxData == NULL) {
 583             ++SkipnStrides;
 584         }
 585     }
 586
 587     glEnd();
 588     checkGLcall("glEnd and previous calls");
 589 }
 590
 591 static void depth_blt(IWineD3DDevice *iface, GLuint texture) {
 592     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
 593     GLint old_binding = 0;
 594
 595     glPushAttrib(GL_ENABLE_BIT | GL_DEPTH_BUFFER_BIT | GL_COLOR_BUFFER_BIT);
 596
 597     glDisable(GL_CULL_FACE);
 598     glEnable(GL_BLEND);
 599     glDisable(GL_ALPHA_TEST);
 600     glDisable(GL_SCISSOR_TEST);
 601     glDisable(GL_STENCIL_TEST);
 602     glEnable(GL_DEPTH_TEST);
 603     glDepthFunc(GL_ALWAYS);
 604     glBlendFunc(GL_ZERO, GL_ONE);
 605
 606     GL_EXTCALL(glActiveTextureARB(GL_TEXTURE0_ARB));
 607     glGetIntegerv(GL_TEXTURE_BINDING_2D, &old_binding);
 608     glBindTexture(GL_TEXTURE_2D, texture);
 609     glEnable(GL_TEXTURE_2D);
 610
 611     This->shader_backend->shader_select_depth_blt(iface);
 612
 613     glBegin(GL_TRIANGLE_STRIP);
 614     glVertex2f(-1.0f, -1.0f);
 615     glVertex2f(1.0f, -1.0f);
 616     glVertex2f(-1.0f, 1.0f);
 617     glVertex2f(1.0f, 1.0f);
 618     glEnd();
 619
 620     glBindTexture(GL_TEXTURE_2D, old_binding);
 621
 622     glPopAttrib();
 623
 624     /* Reselect the old shaders. There doesn't seem to be any glPushAttrib bit for arb shaders,
 625      * and this seems easier and more efficient than providing the shader backend with a private
 626      * storage to read and restore the old shader settings
 627      */
 628     This->shader_backend->shader_select(iface, use_ps(This), use_vs(This));
 629 }
 630
 631 static void depth_copy(IWineD3DDevice *iface) {
 632     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
 633     IWineD3DSurfaceImpl *depth_stencil = (IWineD3DSurfaceImpl *)This->depthStencilBuffer;
 634
 635     /* Only copy the depth buffer if there is one. */
 636     if (!depth_stencil) return;
 637
 638     /* TODO: Make this work for modes other than FBO */
 639     if (wined3d_settings.offscreen_rendering_mode != ORM_FBO) return;
 640
 641     if (depth_stencil->current_renderbuffer) {
 642         FIXME("Not supported with fixed up depth stencil\n");
 643         return;
 644     }
 645
 646     if (This->render_offscreen) {
 647         static GLuint tmp_texture = 0;
 648         GLint old_binding = 0;
 649
 650         TRACE("Copying onscreen depth buffer to offscreen surface\n");
 651
 652         if (!tmp_texture) {
 653             glGenTextures(1, &tmp_texture);
 654         }
 655
 656         /* Note that we use depth_blt here as well, rather than glCopyTexImage2D
 657          * directly on the FBO texture. That's because we need to flip. */
 658         GL_EXTCALL(glBindFramebufferEXT(GL_FRAMEBUFFER_EXT, 0));
 659         glGetIntegerv(GL_TEXTURE_BINDING_2D, &old_binding);
 660         glBindTexture(GL_TEXTURE_2D, tmp_texture);
 661         glCopyTexImage2D(depth_stencil->glDescription.target,
 662                 depth_stencil->glDescription.level,
 663                 depth_stencil->glDescription.glFormatInternal,
 664                 0,
 665                 0,
 666                 depth_stencil->currentDesc.Width,
 667                 depth_stencil->currentDesc.Height,
 668                 0);
 669         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
 670         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
 671         glTexParameteri(GL_TEXTURE_2D, GL_DEPTH_TEXTURE_MODE_ARB, GL_LUMINANCE);
 672         glBindTexture(GL_TEXTURE_2D, old_binding);
 673
 674         GL_EXTCALL(glBindFramebufferEXT(GL_FRAMEBUFFER_EXT, This->fbo));
 675         checkGLcall("glBindFramebuffer()");
 676         depth_blt(iface, tmp_texture);
 677         checkGLcall("depth_blt");
 678     } else {
 679         TRACE("Copying offscreen surface to onscreen depth buffer\n");
 680
 681         GL_EXTCALL(glBindFramebufferEXT(GL_FRAMEBUFFER_EXT, 0));
 682         checkGLcall("glBindFramebuffer()");
 683         depth_blt(iface, depth_stencil->glDescription.textureName);
 684         checkGLcall("depth_blt");
 685     }
 686 }
 687
 688 static inline void drawStridedInstanced(IWineD3DDevice *iface, WineDirect3DVertexStridedData *sd, UINT numberOfVertices,
 689                                  GLenum glPrimitiveType, const void *idxData, short idxSize, ULONG minIndex,
 690                                  ULONG startIdx, ULONG startVertex) {
 691     UINT numInstances = 0;
 692     int numInstancedAttribs = 0, i, j;
 693     UINT instancedData[sizeof(sd->u.input) / sizeof(sd->u.input[0]) /* 16 */];
 694     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *) iface;
 695     IWineD3DStateBlockImpl *stateblock = This->stateBlock;
 696
 697     if (idxSize == 0) {
 698         /* This is a nasty thing. MSDN says no hardware supports that and apps have to use software vertex processing.
 699          * We don't support this for now
 700          *
 701          * Shouldn't be too hard to support with opengl, in theory just call glDrawArrays instead of drawElements.
 702          * But the StreamSourceFreq value has a different meaning in that situation.
 703          */
 704         FIXME("Non-indexed instanced drawing is not supported\n");
 705         return;
 706     }
 707
 708     TRACE("(%p) : glElements(%x, %d, %d, ...)\n", This, glPrimitiveType, numberOfVertices, minIndex);
 709     idxData = idxData == (void *)-1 ? NULL : idxData;
 710
 711     /* First, figure out how many instances we have to draw */
 712     for(i = 0; i < MAX_STREAMS; i++) {
 713         /* Look at all non-instanced streams */
 714         if(!(stateblock->streamFlags[i] & WINED3DSTREAMSOURCE_INSTANCEDATA) &&
 715            stateblock->streamSource[i]) {
 716             int inst = stateblock->streamFreq[i];
 717
 718             if(numInstances && inst != numInstances) {
 719                 ERR("Two streams specify a different number of instances. Got %d, new is %d\n", numInstances, inst);
 720             }
 721             numInstances = inst;
 722         }
 723     }
 724
 725     for(i = 0; i < sizeof(sd->u.input) / sizeof(sd->u.input[0]); i++) {
 726         if(stateblock->streamFlags[sd->u.input[i].streamNo] & WINED3DSTREAMSOURCE_INSTANCEDATA) {
 727             instancedData[numInstancedAttribs] = i;
 728             numInstancedAttribs++;
 729         }
 730     }
 731
 732     /* now draw numInstances instances :-) */
 733     for(i = 0; i < numInstances; i++) {
 734         /* Specify the instanced attributes using immediate mode calls */
 735         for(j = 0; j < numInstancedAttribs; j++) {
 736             BYTE *ptr = sd->u.input[instancedData[j]].lpData +
 737                         sd->u.input[instancedData[j]].dwStride * i +
 738                         stateblock->streamOffset[sd->u.input[instancedData[j]].streamNo];
 739             if(sd->u.input[instancedData[j]].VBO) {
 740                 IWineD3DVertexBufferImpl *vb = (IWineD3DVertexBufferImpl *) stateblock->streamSource[sd->u.input[instancedData[j]].streamNo];
 741                 ptr += (long) vb->resource.allocatedMemory;
 742             }
 743
 744             switch(sd->u.input[instancedData[j]].dwType) {
 745                 case WINED3DDECLTYPE_FLOAT1:
 746                     GL_EXTCALL(glVertexAttrib1fvARB(instancedData[j], (float *) ptr));
 747                     break;
 748                 case WINED3DDECLTYPE_FLOAT2:
 749                     GL_EXTCALL(glVertexAttrib2fvARB(instancedData[j], (float *) ptr));
 750                     break;
 751                 case WINED3DDECLTYPE_FLOAT3:
 752                     GL_EXTCALL(glVertexAttrib3fvARB(instancedData[j], (float *) ptr));
 753                     break;
 754                 case WINED3DDECLTYPE_FLOAT4:
 755                     GL_EXTCALL(glVertexAttrib4fvARB(instancedData[j], (float *) ptr));
 756                     break;
 757
 758                 case WINED3DDECLTYPE_UBYTE4:
 759                     GL_EXTCALL(glVertexAttrib4ubvARB(instancedData[j], ptr));
 760                     break;
 761                 case WINED3DDECLTYPE_UBYTE4N:
 762                 case WINED3DDECLTYPE_D3DCOLOR:
 763                     GL_EXTCALL(glVertexAttrib4NubvARB(instancedData[j], ptr));
 764                     break;
 765
 766                 case WINED3DDECLTYPE_SHORT2:
 767                     GL_EXTCALL(glVertexAttrib4svARB(instancedData[j], (GLshort *) ptr));
 768                     break;
 769                 case WINED3DDECLTYPE_SHORT4:
 770                     GL_EXTCALL(glVertexAttrib4svARB(instancedData[j], (GLshort *) ptr));
 771                     break;
 772
 773                 case WINED3DDECLTYPE_SHORT2N:
 774                 {
 775                     GLshort s[4] = {((short *) ptr)[0], ((short *) ptr)[1], 0, 1};
 776                     GL_EXTCALL(glVertexAttrib4NsvARB(instancedData[j], s));
 777                     break;
 778                 }
 779                 case WINED3DDECLTYPE_USHORT2N:
 780                 {
 781                     GLushort s[4] = {((unsigned short *) ptr)[0], ((unsigned short *) ptr)[1], 0, 1};
 782                     GL_EXTCALL(glVertexAttrib4NusvARB(instancedData[j], s));
 783                     break;
 784                 }
 785                 case WINED3DDECLTYPE_SHORT4N:
 786                     GL_EXTCALL(glVertexAttrib4NsvARB(instancedData[j], (GLshort *) ptr));
 787                     break;
 788                 case WINED3DDECLTYPE_USHORT4N:
 789                     GL_EXTCALL(glVertexAttrib4NusvARB(instancedData[j], (GLushort *) ptr));
 790                     break;
 791
 792                 case WINED3DDECLTYPE_UDEC3:
 793                     FIXME("Unsure about WINED3DDECLTYPE_UDEC3\n");
 794                     /*glVertexAttrib3usvARB(instancedData[j], (GLushort *) ptr); Does not exist */
 795                     break;
 796                 case WINED3DDECLTYPE_DEC3N:
 797                     FIXME("Unsure about WINED3DDECLTYPE_DEC3N\n");
 798                     /*glVertexAttrib3NusvARB(instancedData[j], (GLushort *) ptr); Does not exist */
 799                     break;
 800
 801                 case WINED3DDECLTYPE_FLOAT16_2:
 802                     /* Are those 16 bit floats. C doesn't have a 16 bit float type. I could read the single bits and calculate a 4
 803                      * byte float according to the IEEE standard
 804                      */
 805                     if (GL_SUPPORT(NV_HALF_FLOAT)) {
 806                         GL_EXTCALL(glVertexAttrib2hvNV(instancedData[j], (GLhalfNV *)ptr));
 807                     } else {
 808                         FIXME("Unsupported WINED3DDECLTYPE_FLOAT16_2\n");
 809                     }
 810                     break;
 811                 case WINED3DDECLTYPE_FLOAT16_4:
 812                     if (GL_SUPPORT(NV_HALF_FLOAT)) {
 813                         GL_EXTCALL(glVertexAttrib4hvNV(instancedData[j], (GLhalfNV *)ptr));
 814                     } else {
 815                         FIXME("Unsupported WINED3DDECLTYPE_FLOAT16_4\n");
 816                     }
 817                     break;
 818
 819                 case WINED3DDECLTYPE_UNUSED:
 820                 default:
 821                     ERR("Unexpected declaration in instanced attributes\n");
 822                     break;
 823             }
 824         }
 825
 826         glDrawElements(glPrimitiveType, numberOfVertices, idxSize == 2 ? GL_UNSIGNED_SHORT : GL_UNSIGNED_INT,
 827                     (const char *)idxData+(idxSize * startIdx));
 828         checkGLcall("glDrawElements");
 829     }
 830 }
 831
 832 struct coords {
 833     int x, y, z;
 834 };
 835
 836 void blt_to_drawable(IWineD3DDeviceImpl *This, IWineD3DSurfaceImpl *surface) {
 837     struct coords coords[4];
 838     int low_coord;
 839
 840     /* TODO: This could be supported for lazy unlocking */
 841     if(!(surface->Flags & SFLAG_INTEXTURE)) {
 842         /* It is ok at init to be nowhere */
 843         if(!(surface->Flags & SFLAG_INSYSMEM)) {
 844             ERR("Blitting surfaces from sysmem not supported yet\n");
 845         }
 846         return;
 847     }
 848
 849     ActivateContext(This, This->render_targets[0], CTXUSAGE_BLIT);
 850     ENTER_GL();
 851
 852     if(surface->glDescription.target == GL_TEXTURE_2D) {
 853         glBindTexture(GL_TEXTURE_2D, surface->glDescription.textureName);
 854         checkGLcall("GL_TEXTURE_2D, This->glDescription.textureName)");
 855
 856         coords[0].x = 0;    coords[0].y = 0;    coords[0].z = 0;
 857         coords[1].x = 0;    coords[1].y = 1;    coords[1].z = 0;
 858         coords[2].x = 1;    coords[2].y = 1;    coords[2].z = 0;
 859         coords[3].x = 1;    coords[3].y = 0;    coords[3].z = 0;
 860
 861         low_coord = 0;
 862     } else {
 863         /* Must be a cube map */
 864         glDisable(GL_TEXTURE_2D);
 865         checkGLcall("glDisable(GL_TEXTURE_2D)");
 866         glEnable(GL_TEXTURE_CUBE_MAP_ARB);
 867         checkGLcall("glEnable(surface->glDescription.target)");
 868         glBindTexture(GL_TEXTURE_CUBE_MAP_ARB, surface->glDescription.textureName);
 869         checkGLcall("GL_TEXTURE_CUBE_MAP_ARB, This->glDescription.textureName)");
 870
 871         switch(surface->glDescription.target) {
 872             case GL_TEXTURE_CUBE_MAP_POSITIVE_X:
 873                 coords[0].x =  1;   coords[0].y = -1;   coords[0].z =  1;
 874                 coords[1].x =  1;   coords[1].y =  1;   coords[1].z =  1;
 875                 coords[2].x =  1;   coords[2].y =  1;   coords[2].z = -1;
 876                 coords[3].x =  1;   coords[3].y = -1;   coords[3].z = -1;
 877                 break;
 878
 879             case GL_TEXTURE_CUBE_MAP_NEGATIVE_X:
 880                 coords[0].x = -1;   coords[0].y = -1;   coords[0].z =  1;
 881                 coords[1].x = -1;   coords[1].y =  1;   coords[1].z =  1;
 882                 coords[2].x = -1;   coords[2].y =  1;   coords[2].z = -1;
 883                 coords[3].x = -1;   coords[3].y = -1;   coords[3].z = -1;
 884                 break;
 885
 886             case GL_TEXTURE_CUBE_MAP_POSITIVE_Y:
 887                 coords[0].x = -1;   coords[0].y =  1;   coords[0].z =  1;
 888                 coords[1].x =  1;   coords[1].y =  1;   coords[1].z =  1;
 889                 coords[2].x =  1;   coords[2].y =  1;   coords[2].z = -1;
 890                 coords[3].x = -1;   coords[3].y =  1;   coords[3].z = -1;
 891                 break;
 892
 893             case GL_TEXTURE_CUBE_MAP_NEGATIVE_Y:
 894                 coords[0].x = -1;   coords[0].y = -1;   coords[0].z =  1;
 895                 coords[1].x =  1;   coords[1].y = -1;   coords[1].z =  1;
 896                 coords[2].x =  1;   coords[2].y = -1;   coords[2].z = -1;
 897                 coords[3].x = -1;   coords[3].y = -1;   coords[3].z = -1;
 898                 break;
 899
 900             case GL_TEXTURE_CUBE_MAP_POSITIVE_Z:
 901                 coords[0].x = -1;   coords[0].y = -1;   coords[0].z =  1;
 902                 coords[1].x =  1;   coords[1].y = -1;   coords[1].z =  1;
 903                 coords[2].x =  1;   coords[2].y = -1;   coords[2].z =  1;
 904                 coords[3].x = -1;   coords[3].y = -1;   coords[3].z =  1;
 905                 break;
 906
 907             case GL_TEXTURE_CUBE_MAP_NEGATIVE_Z:
 908                 coords[0].x = -1;   coords[0].y = -1;   coords[0].z = -1;
 909                 coords[1].x =  1;   coords[1].y = -1;   coords[1].z = -1;
 910                 coords[2].x =  1;   coords[2].y = -1;   coords[2].z = -1;
 911                 coords[3].x = -1;   coords[3].y = -1;   coords[3].z = -1;
 912
 913             default:
 914                 ERR("Unexpected texture target\n");
 915                 LEAVE_GL();
 916                 return;
 917         }
 918
 919         low_coord = -1;
 920     }
 921
 922     if(This->render_offscreen) {
 923         coords[0].y = coords[0].y == 1 ? low_coord : 1;
 924         coords[1].y = coords[1].y == 1 ? low_coord : 1;
 925         coords[2].y = coords[2].y == 1 ? low_coord : 1;
 926         coords[3].y = coords[3].y == 1 ? low_coord : 1;
 927     }
 928
 929     glBegin(GL_QUADS);
 930         glTexCoord3iv((GLint *) &coords[0]);
 931         glVertex2i(0, 0);
 932
 933         glTexCoord3iv((GLint *) &coords[1]);
 934         glVertex2i(0, surface->pow2Height);
 935
 936         glTexCoord3iv((GLint *) &coords[2]);
 937         glVertex2i(surface->pow2Width, surface->pow2Height);
 938
 939         glTexCoord3iv((GLint *) &coords[3]);
 940         glVertex2i(surface->pow2Width, 0);
 941     glEnd();
 942     checkGLcall("glEnd");
 943
 944     if(surface->glDescription.target != GL_TEXTURE_2D) {
 945         glEnable(GL_TEXTURE_2D);
 946         checkGLcall("glEnable(GL_TEXTURE_2D)");
 947         glDisable(GL_TEXTURE_CUBE_MAP_ARB);
 948         checkGLcall("glDisable(GL_TEXTURE_CUBE_MAP_ARB)");
 949     }
 950     LEAVE_GL();
 951 }
 952
 953 static inline void remove_vbos(IWineD3DDeviceImpl *This, WineDirect3DVertexStridedData *s) {
 954     unsigned char i;
 955     IWineD3DVertexBufferImpl *vb;
 956
 957     if(s->u.s.position.VBO) {
 958         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.position.streamNo];
 959         s->u.s.position.VBO = 0;
 960         s->u.s.position.lpData = (BYTE *) ((unsigned long) s->u.s.position.lpData + (unsigned long) vb->resource.allocatedMemory);
 961     }
 962     if(s->u.s.blendWeights.VBO) {
 963         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.blendWeights.streamNo];
 964         s->u.s.blendWeights.VBO = 0;
 965         s->u.s.blendWeights.lpData = (BYTE *) ((unsigned long) s->u.s.blendWeights.lpData + (unsigned long) vb->resource.allocatedMemory);
 966     }
 967     if(s->u.s.blendMatrixIndices.VBO) {
 968         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.blendMatrixIndices.streamNo];
 969         s->u.s.blendMatrixIndices.VBO = 0;
 970         s->u.s.blendMatrixIndices.lpData = (BYTE *) ((unsigned long) s->u.s.blendMatrixIndices.lpData + (unsigned long) vb->resource.allocatedMemory);
 971     }
 972     if(s->u.s.normal.VBO) {
 973         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.normal.streamNo];
 974         s->u.s.normal.VBO = 0;
 975         s->u.s.normal.lpData = (BYTE *) ((unsigned long) s->u.s.normal.lpData + (unsigned long) vb->resource.allocatedMemory);
 976     }
 977     if(s->u.s.pSize.VBO) {
 978         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.pSize.streamNo];
 979         s->u.s.pSize.VBO = 0;
 980         s->u.s.pSize.lpData = (BYTE *) ((unsigned long) s->u.s.pSize.lpData + (unsigned long) vb->resource.allocatedMemory);
 981     }
 982     if(s->u.s.diffuse.VBO) {
 983         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.diffuse.streamNo];
 984         s->u.s.diffuse.VBO = 0;
 985         s->u.s.diffuse.lpData = (BYTE *) ((unsigned long) s->u.s.diffuse.lpData + (unsigned long) vb->resource.allocatedMemory);
 986     }
 987     if(s->u.s.specular.VBO) {
 988         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.specular.streamNo];
 989         s->u.s.specular.VBO = 0;
 990         s->u.s.specular.lpData = (BYTE *) ((unsigned long) s->u.s.specular.lpData + (unsigned long) vb->resource.allocatedMemory);
 991     }
 992     for(i = 0; i < WINED3DDP_MAXTEXCOORD; i++) {
 993         if(s->u.s.texCoords[i].VBO) {
 994             vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.texCoords[i].streamNo];
 995             s->u.s.texCoords[i].VBO = 0;
 996             s->u.s.texCoords[i].lpData = (BYTE *) ((unsigned long) s->u.s.texCoords[i].lpData + (unsigned long) vb->resource.allocatedMemory);
 997         }
 998     }
 999     if(s->u.s.position2.VBO) {
1000         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.position2.streamNo];
1001         s->u.s.position2.VBO = 0;
1002         s->u.s.position2.lpData = (BYTE *) ((unsigned long) s->u.s.position2.lpData + (unsigned long) vb->resource.allocatedMemory);
1003     }
1004     if(s->u.s.normal2.VBO) {
1005         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.normal2.streamNo];
1006         s->u.s.normal2.VBO = 0;
1007         s->u.s.normal2.lpData = (BYTE *) ((unsigned long) s->u.s.normal2.lpData + (unsigned long) vb->resource.allocatedMemory);
1008     }
1009     if(s->u.s.tangent.VBO) {
1010         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.tangent.streamNo];
1011         s->u.s.tangent.VBO = 0;
1012         s->u.s.tangent.lpData = (BYTE *) ((unsigned long) s->u.s.tangent.lpData + (unsigned long) vb->resource.allocatedMemory);
1013     }
1014     if(s->u.s.binormal.VBO) {
1015         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.binormal.streamNo];
1016         s->u.s.binormal.VBO = 0;
1017         s->u.s.binormal.lpData = (BYTE *) ((unsigned long) s->u.s.binormal.lpData + (unsigned long) vb->resource.allocatedMemory);
1018     }
1019     if(s->u.s.tessFactor.VBO) {
1020         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.tessFactor.streamNo];
1021         s->u.s.tessFactor.VBO = 0;
1022         s->u.s.tessFactor.lpData = (BYTE *) ((unsigned long) s->u.s.tessFactor.lpData + (unsigned long) vb->resource.allocatedMemory);
1023     }
1024     if(s->u.s.fog.VBO) {
1025         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.fog.streamNo];
1026         s->u.s.fog.VBO = 0;
1027         s->u.s.fog.lpData = (BYTE *) ((unsigned long) s->u.s.fog.lpData + (unsigned long) vb->resource.allocatedMemory);
1028     }
1029     if(s->u.s.depth.VBO) {
1030         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.depth.streamNo];
1031         s->u.s.depth.VBO = 0;
1032         s->u.s.depth.lpData = (BYTE *) ((unsigned long) s->u.s.depth.lpData + (unsigned long) vb->resource.allocatedMemory);
1033     }
1034     if(s->u.s.sample.VBO) {
1035         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.sample.streamNo];
1036         s->u.s.sample.VBO = 0;
1037         s->u.s.sample.lpData = (BYTE *) ((unsigned long) s->u.s.sample.lpData + (unsigned long) vb->resource.allocatedMemory);
1038     }
1039 }
1040
1041 /* Routine common to the draw primitive and draw indexed primitive routines */
1042 void drawPrimitive(IWineD3DDevice *iface,
1043                    int PrimitiveType,
1044                    long NumPrimitives,
1045                    /* for Indexed: */
1046                    long  StartVertexIndex,
1047                    UINT  numberOfVertices,
1048                    long  StartIdx,
1049                    short idxSize,
1050                    const void *idxData,
1051                    int   minIndex) {
1052
1053     IWineD3DDeviceImpl           *This = (IWineD3DDeviceImpl *)iface;
1054     IWineD3DSwapChain            *swapchain;
1055     IWineD3DBaseTexture          *texture = NULL;
1056     IWineD3DSurfaceImpl          *target;
1057     int i;
1058
1059     /* Signals other modules that a drawing is in progress and the stateblock finalized */
1060     This->isInDraw = TRUE;
1061
1062     /* Invalidate the back buffer memory so LockRect will read it the next time */
1063     for(i = 0; i < GL_LIMITS(buffers); i++) {
1064         target = (IWineD3DSurfaceImpl *) This->render_targets[i];
1065
1066         /* TODO: Only do all that if we're going to change anything
1067          * Texture container dirtification does not work quite right yet
1068          */
1069         if(target /*&& target->Flags & (SFLAG_INTEXTURE | SFLAG_INSYSMEM)*/) {
1070             swapchain = NULL;
1071             texture = NULL;
1072
1073             if(i == 0) {
1074                 IWineD3DSurface_GetContainer((IWineD3DSurface *) target, &IID_IWineD3DSwapChain, (void **)&swapchain);
1075
1076                 /* Need the surface in the drawable! */
1077                 if(!(target->Flags & SFLAG_INDRAWABLE) && (swapchain || wined3d_settings.offscreen_rendering_mode != ORM_FBO)) {
1078                     blt_to_drawable(This, target);
1079                 }
1080
1081                 if(swapchain) {
1082                     /* Onscreen target. Invalidate system memory copy and texture copy */
1083                     target->Flags &= ~(SFLAG_INSYSMEM | SFLAG_INTEXTURE);
1084                     target->Flags |= SFLAG_INDRAWABLE;
1085                     IWineD3DSwapChain_Release(swapchain);
1086                 } else if(wined3d_settings.offscreen_rendering_mode != ORM_FBO) {
1087                     /* Non-FBO target: Invalidate system copy, texture copy and dirtify the container */
1088                     IWineD3DSurface_GetContainer((IWineD3DSurface *) target, &IID_IWineD3DBaseTexture, (void **)&texture);
1089
1090                     if(texture) {
1091                         IWineD3DBaseTexture_SetDirty(texture, TRUE);
1092                         IWineD3DTexture_Release(texture);
1093                     }
1094
1095                     target->Flags &= ~(SFLAG_INSYSMEM | SFLAG_INTEXTURE);
1096                     target->Flags |= SFLAG_INDRAWABLE;
1097                 } else {
1098                     /* FBO offscreen target. Invalidate system memory copy */
1099                     target->Flags &= ~SFLAG_INSYSMEM;
1100                 }
1101             } else {
1102                 /* Must be an fbo render target */
1103                 target->Flags &= ~SFLAG_INSYSMEM;
1104                 target->Flags |=  SFLAG_INTEXTURE;
1105             }
1106         }
1107     }
1108
1109     /* Ok, we will be updating the screen from here onwards so grab the lock */
1110
1111     if (wined3d_settings.offscreen_rendering_mode == ORM_FBO) {
1112         ENTER_GL();
1113         apply_fbo_state(iface);
1114         LEAVE_GL();
1115     }
1116
1117     ActivateContext(This, This->render_targets[0], CTXUSAGE_DRAWPRIM);
1118     ENTER_GL();
1119
1120     if (This->depth_copy_state == WINED3D_DCS_COPY) {
1121         depth_copy(iface);
1122     }
1123     This->depth_copy_state = WINED3D_DCS_INITIAL;
1124
1125     {
1126         GLenum glPrimType;
1127         BOOL emulation = FALSE;
1128         WineDirect3DVertexStridedData *strided = &This->strided_streams;
1129         WineDirect3DVertexStridedData stridedlcl;
1130         /* Ok, Work out which primitive is requested and how many vertexes that
1131            will be                                                              */
1132         UINT calculatedNumberOfindices = primitiveToGl(PrimitiveType, NumPrimitives, &glPrimType);
1133         if (numberOfVertices == 0 )
1134             numberOfVertices = calculatedNumberOfindices;
1135
1136         if(!use_vs(This)) {
1137             if(!This->strided_streams.u.s.position_transformed && This->activeContext->num_untracked_materials &&
1138                 This->stateBlock->renderState[WINED3DRS_LIGHTING]) {
1139                 FIXME("Using software emulation because not all material properties could be tracked\n");
1140                 emulation = TRUE;
1141             }
1142             else if(This->activeContext->fog_coord && This->stateBlock->renderState[WINED3DRS_FOGENABLE]) {
1143                 /* Either write a pipeline replacement shader or convert the specular alpha from unsigned byte
1144                  * to a float in the vertex buffer
1145                  */
1146                 FIXME("Using software emulation because manual fog coordinates are provided\n");
1147                 emulation = TRUE;
1148             }
1149
1150             if(emulation) {
1151                 strided = &stridedlcl;
1152                 memcpy(&stridedlcl, &This->strided_streams, sizeof(stridedlcl));
1153                 remove_vbos(This, &stridedlcl);
1154             }
1155         }
1156
1157         if (This->useDrawStridedSlow || emulation) {
1158             /* Immediate mode drawing */
1159             drawStridedSlow(iface, strided, calculatedNumberOfindices,
1160                             glPrimType, idxData, idxSize, minIndex, StartIdx, StartVertexIndex);
1161         } else if(This->instancedDraw) {
1162             /* Instancing emulation with mixing immediate mode and arrays */
1163             drawStridedInstanced(iface, &This->strided_streams, calculatedNumberOfindices, glPrimType,
1164                             idxData, idxSize, minIndex, StartIdx, StartVertexIndex);
1165         } else {
1166             /* Simple array draw call */
1167             drawStridedFast(iface, calculatedNumberOfindices, glPrimType,
1168                             idxData, idxSize, minIndex, StartIdx, StartVertexIndex);
1169         }
1170     }
1171
1172     /* Finshed updating the screen, restore lock */
1173     LEAVE_GL();
1174     TRACE("Done all gl drawing\n");
1175
1176     /* Diagnostics */
1177 #ifdef SHOW_FRAME_MAKEUP
1178     {
1179         static long int primCounter = 0;
1180         /* NOTE: set primCounter to the value reported by drawprim
1181            before you want to to write frame makeup to /tmp */
1182         if (primCounter >= 0) {
1183             WINED3DLOCKED_RECT r;
1184             char buffer[80];
1185             IWineD3DSurface_LockRect(This->renderTarget, &r, NULL, WINED3DLOCK_READONLY);
1186             sprintf(buffer, "/tmp/backbuffer_%d.tga", primCounter);
1187             TRACE("Saving screenshot %s\n", buffer);
1188             IWineD3DSurface_SaveSnapshot(This->renderTarget, buffer);
1189             IWineD3DSurface_UnlockRect(This->renderTarget);
1190
1191 #ifdef SHOW_TEXTURE_MAKEUP
1192            {
1193             IWineD3DSurface *pSur;
1194             int textureNo;
1195             for (textureNo = 0; textureNo < MAX_COMBINED_SAMPLERS; ++textureNo) {
1196                 if (This->stateBlock->textures[textureNo] != NULL) {
1197                     sprintf(buffer, "/tmp/texture_%p_%d_%d.tga", This->stateBlock->textures[textureNo], primCounter, textureNo);
1198                     TRACE("Saving texture %s\n", buffer);
1199                     if (IWineD3DBaseTexture_GetType(This->stateBlock->textures[textureNo]) == WINED3DRTYPE_TEXTURE) {
1200                             IWineD3DTexture_GetSurfaceLevel((IWineD3DTexture *)This->stateBlock->textures[textureNo], 0, &pSur);
1201                             IWineD3DSurface_SaveSnapshot(pSur, buffer);
1202                             IWineD3DSurface_Release(pSur);
1203                     } else  {
1204                         FIXME("base Texture isn't of type texture %d\n", IWineD3DBaseTexture_GetType(This->stateBlock->textures[textureNo]));
1205                     }
1206                 }
1207             }
1208            }
1209 #endif
1210         }
1211         TRACE("drawprim #%d\n", primCounter);
1212         ++primCounter;
1213     }
1214 #endif
1215
1216     /* Control goes back to the device, stateblock values may change again */
1217     This->isInDraw = FALSE;
1218 }
1219
1220 static void normalize_normal(float *n) {
1221     float length = n[0] * n[0] + n[1] * n[1] + n[2] * n[2];
1222     if(length == 0.0) return;
1223     length = sqrt(length);
1224     n[0] = n[0] / length;
1225     n[1] = n[1] / length;
1226     n[2] = n[2] / length;
1227 }
1228
1229 /* Tesselates a high order rectangular patch into single triangles using gl evaluators
1230  *
1231  * The problem is that OpenGL does not offer a direct way to return the tesselated primitives,
1232  * and they can't be sent off for rendering directly either. Tesselating is slow, so we want
1233  * to chache the patches in a vertex buffer. But more importantly, gl can't bind generated
1234  * attributes to numbered shader attributes, so we have to store them and rebind them as needed
1235  * in drawprim.
1236  *
1237  * To read back, the opengl feedback mode is used. This creates a proplem because we want
1238  * untransformed, unlit vertices, but feedback runs everything through transform and lighting.
1239  * Thus disable lighting and set identity matrices to get unmodified colors and positions.
1240  * To overcome clipping find the biggest x, y and z values of the vertices in the patch and scale
1241  * them to [-1.0;+1.0] and set the viewport up to scale them back.
1242  *
1243  * Normals are more tricky: Draw white vertices with 3 directional lights, and calculate the
1244  * resulting colors back to the normals.
1245  *
1246  * NOTE: This function activates a context for blitting, modifies matrices & viewport, but
1247  * does not restore it because normally a draw follows immediately afterwards. The caller is
1248  * responsible of taking care that either the gl states are restored, or the context activated
1249  * for drawing to reset the lastWasBlit flag.
1250  */
1251 HRESULT tesselate_rectpatch(IWineD3DDeviceImpl *This,
1252                             struct WineD3DRectPatch *patch) {
1253     unsigned int i, j, num_quads, out_vertex_size, buffer_size, d3d_out_vertex_size;
1254     float max_x = 0.0, max_y = 0.0, max_z = 0.0, neg_z = 0.0;
1255     WineDirect3DVertexStridedData strided;
1256     BYTE *data;
1257     WINED3DRECTPATCH_INFO *info = &patch->RectPatchInfo;
1258     DWORD vtxStride;
1259     GLenum feedback_type;
1260     GLfloat *feedbuffer;
1261
1262     /* First, locate the position data. This is provided in a vertex buffer in the stateblock.
1263      * Beware of vbos
1264      */
1265     memset(&strided, 0, sizeof(strided));
1266     primitiveDeclarationConvertToStridedData((IWineD3DDevice *) This, FALSE, &strided, NULL);
1267     if(strided.u.s.position.VBO) {
1268         IWineD3DVertexBufferImpl *vb;
1269         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[strided.u.s.position.streamNo];
1270         strided.u.s.position.lpData = (BYTE *) ((unsigned long) strided.u.s.position.lpData +
1271                                                 (unsigned long) vb->resource.allocatedMemory);
1272     }
1273     vtxStride = strided.u.s.position.dwStride;
1274     data = strided.u.s.position.lpData +
1275            vtxStride * info->Stride * info->StartVertexOffsetHeight +
1276            vtxStride * info->StartVertexOffsetWidth;
1277
1278     /* Not entirely sure about what happens with transformed vertices */
1279     if(strided.u.s.position_transformed) {
1280         FIXME("Transformed position in rectpatch generation\n");
1281     }
1282     if(vtxStride % sizeof(GLfloat)) {
1283         /* glMap2f reads vertex sizes in GLfloats, the d3d stride is in bytes.
1284          * I don't see how the stride could not be a multiple of 4, but make sure
1285          * to check it
1286          */
1287         ERR("Vertex stride is not a multiple of sizeof(GLfloat)\n");
1288     }
1289     if(info->Basis != WINED3DBASIS_BEZIER) {
1290         FIXME("Basis is %s, how to handle this?\n", debug_d3dbasis(info->Basis));
1291     }
1292     if(info->Degree != WINED3DDEGREE_CUBIC) {
1293         FIXME("Degree is %s, how to handle this?\n", debug_d3ddegree(info->Degree));
1294     }
1295
1296     /* First, get the boundary cube of the input data */
1297     for(j = 0; j < info->Height; j++) {
1298         for(i = 0; i < info->Width; i++) {
1299             float *v = (float *) (data + vtxStride * i + vtxStride * info->Stride * j);
1300             if(fabs(v[0]) > max_x) max_x = fabs(v[0]);
1301             if(fabs(v[1]) > max_y) max_y = fabs(v[1]);
1302             if(fabs(v[2]) > max_z) max_z = fabs(v[2]);
1303             if(v[2] < neg_z) neg_z = v[2];
1304         }
1305     }
1306
1307     /* This needs some improvements in the vertex decl code */
1308     FIXME("Cannot find data to generate. Only generating position and normals\n");
1309     patch->has_normals = TRUE;
1310     patch->has_texcoords = FALSE;
1311
1312     /* Simply activate the context for blitting. This disables all the things we don't want and
1313      * takes care of dirtifying. Dirtifying is preferred over pushing / popping, since drawing the
1314      * patch (as opposed to normal draws) will most likely need different changes anyway
1315      */
1316     ActivateContext(This, This->lastActiveRenderTarget, CTXUSAGE_BLIT);
1317     ENTER_GL();
1318
1319     glMatrixMode(GL_PROJECTION);
1320     checkGLcall("glMatrixMode(GL_PROJECTION)");
1321     glLoadIdentity();
1322     checkGLcall("glLoadIndentity()");
1323     glScalef(1 / (max_x) , 1 / (max_y), max_z == 0 ? 1 : 1 / ( 2 * max_z));
1324     glTranslatef(0, 0, 0.5);
1325     checkGLcall("glScalef");
1326     glViewport(-max_x, -max_y, 2 * (max_x), 2 * (max_y));
1327     checkGLcall("glViewport");
1328
1329     /* Some states to take care of. If we're in wireframe opengl will produce lines, and confuse
1330      * our feedback buffer parser
1331      */
1332     glPolygonMode(GL_FRONT_AND_BACK, GL_FILL);
1333     checkGLcall("glPolygonMode(GL_FRONT_AND_BACK, GL_FILL)");
1334     IWineD3DDeviceImpl_MarkStateDirty(This, STATE_RENDER(WINED3DRS_FILLMODE));
1335     if(patch->has_normals) {
1336         float black[4] = {0, 0, 0, 0};
1337         float red[4]   = {1, 0, 0, 0};
1338         float green[4] = {0, 1, 0, 0};
1339         float blue[4]  = {0, 0, 1, 0};
1340         float white[4] = {1, 1, 1, 1};
1341         glEnable(GL_LIGHTING);
1342         checkGLcall("glEnable(GL_LIGHTING)");
1343         glLightModelfv(GL_LIGHT_MODEL_AMBIENT, black);
1344         checkGLcall("glLightModel for MODEL_AMBIENT");
1345         IWineD3DDeviceImpl_MarkStateDirty(This, STATE_RENDER(WINED3DRS_AMBIENT));
1346
1347         for(i = 3; i < GL_LIMITS(lights); i++) {
1348             glDisable(GL_LIGHT0 + i);
1349             checkGLcall("glDisable(GL_LIGHT0 + i)");
1350             IWineD3DDeviceImpl_MarkStateDirty(This, STATE_ACTIVELIGHT(i));
1351         }
1352
1353         IWineD3DDeviceImpl_MarkStateDirty(This, STATE_ACTIVELIGHT(0));
1354         glLightfv(GL_LIGHT0, GL_DIFFUSE, red);
1355         glLightfv(GL_LIGHT0, GL_SPECULAR, black);
1356         glLightfv(GL_LIGHT0, GL_AMBIENT, black);
1357         glLightfv(GL_LIGHT0, GL_POSITION, red);
1358         glEnable(GL_LIGHT0);
1359         checkGLcall("Setting up light 1\n");
1360         IWineD3DDeviceImpl_MarkStateDirty(This, STATE_ACTIVELIGHT(1));
1361         glLightfv(GL_LIGHT1, GL_DIFFUSE, green);
1362         glLightfv(GL_LIGHT1, GL_SPECULAR, black);
1363         glLightfv(GL_LIGHT1, GL_AMBIENT, black);
1364         glLightfv(GL_LIGHT1, GL_POSITION, green);
1365         glEnable(GL_LIGHT1);
1366         checkGLcall("Setting up light 2\n");
1367         IWineD3DDeviceImpl_MarkStateDirty(This, STATE_ACTIVELIGHT(2));
1368         glLightfv(GL_LIGHT2, GL_DIFFUSE, blue);
1369         glLightfv(GL_LIGHT2, GL_SPECULAR, black);
1370         glLightfv(GL_LIGHT2, GL_AMBIENT, black);
1371         glLightfv(GL_LIGHT2, GL_POSITION, blue);
1372         glEnable(GL_LIGHT2);
1373         checkGLcall("Setting up light 3\n");
1374
1375         IWineD3DDeviceImpl_MarkStateDirty(This, STATE_MATERIAL);
1376         IWineD3DDeviceImpl_MarkStateDirty(This, STATE_RENDER(WINED3DRS_COLORVERTEX));
1377         glDisable(GL_COLOR_MATERIAL);
1378         glMaterialfv(GL_FRONT_AND_BACK, GL_EMISSION, black);
1379         glMaterialfv(GL_FRONT_AND_BACK, GL_SPECULAR, black);
1380         glMaterialfv(GL_FRONT_AND_BACK, GL_DIFFUSE, white);
1381         checkGLcall("Setting up materials\n");
1382     }
1383
1384     /* Enable the needed maps.
1385      * GL_MAP2_VERTEX_3 is needed for positional data.
1386      * GL_AUTO_NORMAL to generate normals from the position. Do not use GL_MAP2_NORMAL.
1387      * GL_MAP2_TEXTURE_COORD_4 for texture coords
1388      */
1389     num_quads = ceilf(patch->numSegs[0]) * ceilf(patch->numSegs[1]);
1390     out_vertex_size = 3 /* position */;
1391     d3d_out_vertex_size = 3;
1392     glEnable(GL_MAP2_VERTEX_3);
1393     if(patch->has_normals && patch->has_texcoords) {
1394         FIXME("Texcoords not handled yet\n");
1395         feedback_type = GL_3D_COLOR_TEXTURE;
1396         out_vertex_size += 8;
1397         d3d_out_vertex_size += 7;
1398         glEnable(GL_AUTO_NORMAL);
1399         glEnable(GL_MAP2_TEXTURE_COORD_4);
1400     } else if(patch->has_texcoords) {
1401         FIXME("Texcoords not handled yet\n");
1402         feedback_type = GL_3D_COLOR_TEXTURE;
1403         out_vertex_size += 7;
1404         d3d_out_vertex_size += 4;
1405         glEnable(GL_MAP2_TEXTURE_COORD_4);
1406     } else if(patch->has_normals) {
1407         feedback_type = GL_3D_COLOR;
1408         out_vertex_size += 4;
1409         d3d_out_vertex_size += 3;
1410         glEnable(GL_AUTO_NORMAL);
1411     } else {
1412         feedback_type = GL_3D;
1413     }
1414     checkGLcall("glEnable vertex attrib generation");
1415
1416     buffer_size = num_quads * out_vertex_size * 2 /* triangle list */ * 3 /* verts per tri */
1417                    + 4 * num_quads /* 2 triangle markers per quad + num verts in tri */;
1418     feedbuffer = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, buffer_size * sizeof(float) * 8);
1419
1420     glMap2f(GL_MAP2_VERTEX_3,
1421             0, 1, vtxStride / sizeof(float), info->Width,
1422             0, 1, info->Stride * vtxStride / sizeof(float), info->Height,
1423             (float *) data);
1424     checkGLcall("glMap2f");
1425     if(patch->has_texcoords) {
1426         glMap2f(GL_MAP2_TEXTURE_COORD_4,
1427                 0, 1, vtxStride / sizeof(float), info->Width,
1428                 0, 1, info->Stride * vtxStride / sizeof(float), info->Height,
1429                 (float *) data);
1430         checkGLcall("glMap2f");
1431     }
1432     glMapGrid2f(ceilf(patch->numSegs[0]), 0.0, 1.0, ceilf(patch->numSegs[1]), 0.0, 1.0);
1433     checkGLcall("glMapGrid2f");
1434
1435     glFeedbackBuffer(buffer_size * 2, feedback_type, feedbuffer);
1436     checkGLcall("glFeedbackBuffer");
1437     glRenderMode(GL_FEEDBACK);
1438
1439     glEvalMesh2(GL_FILL, 0, ceilf(patch->numSegs[0]), 0, ceilf(patch->numSegs[1]));
1440     checkGLcall("glEvalMesh2\n");
1441
1442     i = glRenderMode(GL_RENDER);
1443     if(i == -1) {
1444         ERR("Feedback failed. Expected %d elements back\n", buffer_size);
1445         Sleep(10000);
1446         HeapFree(GetProcessHeap(), 0, feedbuffer);
1447         return WINED3DERR_DRIVERINTERNALERROR;
1448     } else if(i != buffer_size) {
1449         ERR("Unexpected amount of elements returned. Expected %d, got %d\n", buffer_size, i);
1450         Sleep(10000);
1451         HeapFree(GetProcessHeap(), 0, feedbuffer);
1452         return WINED3DERR_DRIVERINTERNALERROR;
1453     } else {
1454         TRACE("Got %d elements as expected\n", i);
1455     }
1456
1457     HeapFree(GetProcessHeap(), 0, patch->mem);
1458     patch->mem = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, num_quads * 6 * d3d_out_vertex_size * sizeof(float) * 8);
1459     i = 0;
1460     for(j = 0; j < buffer_size; j += (3 /* num verts */ * out_vertex_size + 2 /* tri marker */)) {
1461         if(feedbuffer[j] != GL_POLYGON_TOKEN) {
1462             ERR("Unexpected token: %f\n", feedbuffer[j]);
1463             continue;
1464         }
1465         if(feedbuffer[j + 1] != 3) {
1466             ERR("Unexpected polygon: %f corners\n", feedbuffer[j + 1]);
1467             continue;
1468         }
1469         /* Somehow there are different ideas about back / front facing, so fix up the
1470          * vertex order
1471          */
1472         patch->mem[i + 0] =  feedbuffer[j + out_vertex_size * 2 + 2]; /* x, triangle 2 */
1473         patch->mem[i + 1] =  feedbuffer[j + out_vertex_size * 2 + 3]; /* y, triangle 2 */
1474         patch->mem[i + 2] = (feedbuffer[j + out_vertex_size * 2 + 4] - 0.5) * 4 * max_z; /* z, triangle 3 */
1475         if(patch->has_normals) {
1476             patch->mem[i + 3] = feedbuffer[j + out_vertex_size * 2 + 5];
1477             patch->mem[i + 4] = feedbuffer[j + out_vertex_size * 2 + 6];
1478             patch->mem[i + 5] = feedbuffer[j + out_vertex_size * 2 + 7];
1479         }
1480         i += d3d_out_vertex_size;
1481
1482         patch->mem[i + 0] =  feedbuffer[j + out_vertex_size * 1 + 2]; /* x, triangle 2 */
1483         patch->mem[i + 1] =  feedbuffer[j + out_vertex_size * 1 + 3]; /* y, triangle 2 */
1484         patch->mem[i + 2] = (feedbuffer[j + out_vertex_size * 1 + 4] - 0.5) * 4 * max_z; /* z, triangle 2 */
1485         if(patch->has_normals) {
1486             patch->mem[i + 3] = feedbuffer[j + out_vertex_size * 1 + 5];
1487             patch->mem[i + 4] = feedbuffer[j + out_vertex_size * 1 + 6];
1488             patch->mem[i + 5] = feedbuffer[j + out_vertex_size * 1 + 7];
1489         }
1490         i += d3d_out_vertex_size;
1491
1492         patch->mem[i + 0] =  feedbuffer[j + out_vertex_size * 0 + 2]; /* x, triangle 1 */
1493         patch->mem[i + 1] =  feedbuffer[j + out_vertex_size * 0 + 3]; /* y, triangle 1 */
1494         patch->mem[i + 2] = (feedbuffer[j + out_vertex_size * 0 + 4] - 0.5) * 4 * max_z; /* z, triangle 1 */
1495         if(patch->has_normals) {
1496             patch->mem[i + 3] = feedbuffer[j + out_vertex_size * 0 + 5];
1497             patch->mem[i + 4] = feedbuffer[j + out_vertex_size * 0 + 6];
1498             patch->mem[i + 5] = feedbuffer[j + out_vertex_size * 0 + 7];
1499         }
1500         i += d3d_out_vertex_size;
1501     }
1502
1503     if(patch->has_normals) {
1504         /* Now do the same with reverse light directions */
1505         float x[4] = {-1,  0,  0, 0};
1506         float y[4] = { 0, -1,  0, 0};
1507         float z[4] = { 0,  0, -1, 0};
1508         glLightfv(GL_LIGHT0, GL_POSITION, x);
1509         glLightfv(GL_LIGHT1, GL_POSITION, y);
1510         glLightfv(GL_LIGHT2, GL_POSITION, z);
1511         checkGLcall("Setting up reverse light directions\n");
1512
1513         glRenderMode(GL_FEEDBACK);
1514         checkGLcall("glRenderMode(GL_FEEDBACK)");
1515         glEvalMesh2(GL_FILL, 0, ceilf(patch->numSegs[0]), 0, ceilf(patch->numSegs[1]));
1516         checkGLcall("glEvalMesh2\n");
1517         i = glRenderMode(GL_RENDER);
1518         checkGLcall("glRenderMode(GL_RENDER)");
1519
1520         i = 0;
1521         for(j = 0; j < buffer_size; j += (3 /* num verts */ * out_vertex_size + 2 /* tri marker */)) {
1522             if(feedbuffer[j] != GL_POLYGON_TOKEN) {
1523                 ERR("Unexpected token: %f\n", feedbuffer[j]);
1524                 continue;
1525             }
1526             if(feedbuffer[j + 1] != 3) {
1527                 ERR("Unexpected polygon: %f corners\n", feedbuffer[j + 1]);
1528                 continue;
1529             }
1530             if(patch->mem[i + 3] == 0.0)
1531                 patch->mem[i + 3] = -feedbuffer[j + out_vertex_size * 2 + 5];
1532             if(patch->mem[i + 4] == 0.0)
1533                 patch->mem[i + 4] = -feedbuffer[j + out_vertex_size * 2 + 6];
1534             if(patch->mem[i + 5] == 0.0)
1535                 patch->mem[i + 5] = -feedbuffer[j + out_vertex_size * 2 + 7];
1536             normalize_normal(patch->mem + i + 3);
1537             i += d3d_out_vertex_size;
1538
1539             if(patch->mem[i + 3] == 0.0)
1540                 patch->mem[i + 3] = -feedbuffer[j + out_vertex_size * 1 + 5];
1541             if(patch->mem[i + 4] == 0.0)
1542                 patch->mem[i + 4] = -feedbuffer[j + out_vertex_size * 1 + 6];
1543             if(patch->mem[i + 5] == 0.0)
1544                 patch->mem[i + 5] = -feedbuffer[j + out_vertex_size * 1 + 7];
1545             normalize_normal(patch->mem + i + 3);
1546             i += d3d_out_vertex_size;
1547
1548             if(patch->mem[i + 3] == 0.0)
1549                 patch->mem[i + 3] = -feedbuffer[j + out_vertex_size * 0 + 5];
1550             if(patch->mem[i + 4] == 0.0)
1551                 patch->mem[i + 4] = -feedbuffer[j + out_vertex_size * 0 + 6];
1552             if(patch->mem[i + 5] == 0.0)
1553                 patch->mem[i + 5] = -feedbuffer[j + out_vertex_size * 0 + 7];
1554             normalize_normal(patch->mem + i + 3);
1555             i += d3d_out_vertex_size;
1556         }
1557     }
1558
1559     glDisable(GL_MAP2_VERTEX_3);
1560     glDisable(GL_AUTO_NORMAL);
1561     glDisable(GL_MAP2_NORMAL);
1562     glDisable(GL_MAP2_TEXTURE_COORD_4);
1563     checkGLcall("glDisable vertex attrib generation");
1564     LEAVE_GL();
1565
1566     HeapFree(GetProcessHeap(), 0, feedbuffer);
1567
1568     vtxStride = 3 * sizeof(float);
1569     if(patch->has_normals) {
1570         vtxStride += 3 * sizeof(float);
1571     }
1572     if(patch->has_texcoords) {
1573         vtxStride += 4 * sizeof(float);
1574     }
1575     memset(&patch->strided, 0, sizeof(&patch->strided));
1576     patch->strided.u.s.position.lpData = (BYTE *) patch->mem;
1577     patch->strided.u.s.position.dwStride = vtxStride;
1578     patch->strided.u.s.position.dwType = WINED3DDECLTYPE_FLOAT3;
1579     patch->strided.u.s.position.streamNo = 255;
1580
1581     if(patch->has_normals) {
1582         patch->strided.u.s.normal.lpData = (BYTE *) patch->mem + 3 * sizeof(float) /* pos */;
1583         patch->strided.u.s.normal.dwStride = vtxStride;
1584         patch->strided.u.s.normal.dwType = WINED3DDECLTYPE_FLOAT3;
1585         patch->strided.u.s.normal.streamNo = 255;
1586     }
1587     if(patch->has_texcoords) {
1588         patch->strided.u.s.texCoords[0].lpData = (BYTE *) patch->mem + 3 * sizeof(float) /* pos */;
1589         if(patch->has_normals) {
1590             patch->strided.u.s.texCoords[0].lpData += 3 * sizeof(float);
1591         }
1592         patch->strided.u.s.texCoords[0].dwStride = vtxStride;
1593         patch->strided.u.s.texCoords[0].dwType = WINED3DDECLTYPE_FLOAT4;
1594         /* MAX_STREAMS index points to an unused element in stateblock->streamOffsets which
1595          * always remains set to 0. Windows uses stream 255 here, but this is not visible to the
1596          * application.
1597          */
1598         patch->strided.u.s.texCoords[0].streamNo = MAX_STREAMS;
1599     }
1600
1601     return WINED3D_OK;
1602 }