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, 2008 Henri Verbeet
   9  * Copyright 2007-2008 Stefan Dösinger for CodeWeavers
  10  *
  11  * This library is free software; you can redistribute it and/or
  12  * modify it under the terms of the GNU Lesser General Public
  13  * License as published by the Free Software Foundation; either
  14  * version 2.1 of the License, or (at your option) any later version.
  15  *
  16  * This library is distributed in the hope that it will be useful,
  17  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  18  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  19  * Lesser General Public License for more details.
  20  *
  21  * You should have received a copy of the GNU Lesser General Public
  22  * License along with this library; if not, write to the Free Software
  23  * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA
  24  */
  25
  26 #include "config.h"
  27 #include "wined3d_private.h"
  28
  29 WINE_DEFAULT_DEBUG_CHANNEL(d3d_draw);
  30 #define GLINFO_LOCATION This->adapter->gl_info
  31
  32 #include <stdio.h>
  33 #include <math.h>
  34
  35 /* Issues the glBegin call for gl given the primitive type and count */
  36 static DWORD primitiveToGl(WINED3DPRIMITIVETYPE PrimitiveType,
  37                     DWORD            NumPrimitives,
  38                     GLenum          *primType)
  39 {
  40     DWORD   NumVertexes = NumPrimitives;
  41
  42     switch (PrimitiveType) {
  43     case WINED3DPT_POINTLIST:
  44         TRACE("POINTS\n");
  45         *primType   = GL_POINTS;
  46         NumVertexes = NumPrimitives;
  47         break;
  48
  49     case WINED3DPT_LINELIST:
  50         TRACE("LINES\n");
  51         *primType   = GL_LINES;
  52         NumVertexes = NumPrimitives * 2;
  53         break;
  54
  55     case WINED3DPT_LINESTRIP:
  56         TRACE("LINE_STRIP\n");
  57         *primType   = GL_LINE_STRIP;
  58         NumVertexes = NumPrimitives + 1;
  59         break;
  60
  61     case WINED3DPT_TRIANGLELIST:
  62         TRACE("TRIANGLES\n");
  63         *primType   = GL_TRIANGLES;
  64         NumVertexes = NumPrimitives * 3;
  65         break;
  66
  67     case WINED3DPT_TRIANGLESTRIP:
  68         TRACE("TRIANGLE_STRIP\n");
  69         *primType   = GL_TRIANGLE_STRIP;
  70         NumVertexes = NumPrimitives + 2;
  71         break;
  72
  73     case WINED3DPT_TRIANGLEFAN:
  74         TRACE("TRIANGLE_FAN\n");
  75         *primType   = GL_TRIANGLE_FAN;
  76         NumVertexes = NumPrimitives + 2;
  77         break;
  78
  79     default:
  80         FIXME("Unhandled primitive\n");
  81         *primType    = GL_POINTS;
  82         break;
  83     }
  84     return NumVertexes;
  85 }
  86
  87 static BOOL fixed_get_input(
  88     BYTE usage, BYTE usage_idx,
  89     unsigned int* regnum) {
  90
  91     *regnum = -1;
  92
  93     /* Those positions must have the order in the
  94      * named part of the strided data */
  95
  96     if ((usage == WINED3DDECLUSAGE_POSITION || usage == WINED3DDECLUSAGE_POSITIONT) && usage_idx == 0)
  97         *regnum = 0;
  98     else if (usage == WINED3DDECLUSAGE_BLENDWEIGHT && usage_idx == 0)
  99         *regnum = 1;
 100     else if (usage == WINED3DDECLUSAGE_BLENDINDICES && usage_idx == 0)
 101         *regnum = 2;
 102     else if (usage == WINED3DDECLUSAGE_NORMAL && usage_idx == 0)
 103         *regnum = 3;
 104     else if (usage == WINED3DDECLUSAGE_PSIZE && usage_idx == 0)
 105         *regnum = 4;
 106     else if (usage == WINED3DDECLUSAGE_COLOR && usage_idx == 0)
 107         *regnum = 5;
 108     else if (usage == WINED3DDECLUSAGE_COLOR && usage_idx == 1)
 109         *regnum = 6;
 110     else if (usage == WINED3DDECLUSAGE_TEXCOORD && usage_idx < WINED3DDP_MAXTEXCOORD)
 111         *regnum = 7 + usage_idx;
 112     else if ((usage == WINED3DDECLUSAGE_POSITION || usage == WINED3DDECLUSAGE_POSITIONT) && usage_idx == 1)
 113         *regnum = 7 + WINED3DDP_MAXTEXCOORD;
 114     else if (usage == WINED3DDECLUSAGE_NORMAL && usage_idx == 1)
 115         *regnum = 8 + WINED3DDP_MAXTEXCOORD;
 116     else if (usage == WINED3DDECLUSAGE_TANGENT && usage_idx == 0)
 117         *regnum = 9 + WINED3DDP_MAXTEXCOORD;
 118     else if (usage == WINED3DDECLUSAGE_BINORMAL && usage_idx == 0)
 119         *regnum = 10 + WINED3DDP_MAXTEXCOORD;
 120     else if (usage == WINED3DDECLUSAGE_TESSFACTOR && usage_idx == 0)
 121         *regnum = 11 + WINED3DDP_MAXTEXCOORD;
 122     else if (usage == WINED3DDECLUSAGE_FOG && usage_idx == 0)
 123         *regnum = 12 + WINED3DDP_MAXTEXCOORD;
 124     else if (usage == WINED3DDECLUSAGE_DEPTH && usage_idx == 0)
 125         *regnum = 13 + WINED3DDP_MAXTEXCOORD;
 126     else if (usage == WINED3DDECLUSAGE_SAMPLE && usage_idx == 0)
 127         *regnum = 14 + WINED3DDP_MAXTEXCOORD;
 128
 129     if (*regnum == -1) {
 130         FIXME("Unsupported input stream [usage=%s, usage_idx=%u]\n",
 131             debug_d3ddeclusage(usage), usage_idx);
 132         return FALSE;
 133     }
 134     return TRUE;
 135 }
 136
 137 void primitiveDeclarationConvertToStridedData(
 138      IWineD3DDevice *iface,
 139      BOOL useVertexShaderFunction,
 140      WineDirect3DVertexStridedData *strided,
 141      BOOL *fixup) {
 142
 143      /* We need to deal with frequency data!*/
 144
 145     BYTE  *data    = NULL;
 146     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
 147     IWineD3DVertexDeclarationImpl* vertexDeclaration = (IWineD3DVertexDeclarationImpl *)This->stateBlock->vertexDecl;
 148     int i;
 149     WINED3DVERTEXELEMENT *element;
 150     DWORD stride;
 151     DWORD numPreloadStreams = This->stateBlock->streamIsUP ? 0 : vertexDeclaration->num_streams;
 152     DWORD *streams = vertexDeclaration->streams;
 153
 154     /* Check for transformed vertices, disable vertex shader if present */
 155     strided->u.s.position_transformed = vertexDeclaration->position_transformed;
 156     if(vertexDeclaration->position_transformed) {
 157         useVertexShaderFunction = FALSE;
 158     }
 159
 160     /* Translate the declaration into strided data */
 161     for (i = 0 ; i < vertexDeclaration->declarationWNumElements - 1; ++i) {
 162         GLint streamVBO = 0;
 163         BOOL stride_used;
 164         unsigned int idx;
 165
 166         element = vertexDeclaration->pDeclarationWine + i;
 167         TRACE("%p Element %p (%d of %d)\n", vertexDeclaration->pDeclarationWine,
 168             element,  i + 1, vertexDeclaration->declarationWNumElements - 1);
 169
 170         if (This->stateBlock->streamSource[element->Stream] == NULL)
 171             continue;
 172
 173         stride  = This->stateBlock->streamStride[element->Stream];
 174         if (This->stateBlock->streamIsUP) {
 175             TRACE("Stream is up %d, %p\n", element->Stream, This->stateBlock->streamSource[element->Stream]);
 176             streamVBO = 0;
 177             data    = (BYTE *)This->stateBlock->streamSource[element->Stream];
 178         } else {
 179             TRACE("Stream isn't up %d, %p\n", element->Stream, This->stateBlock->streamSource[element->Stream]);
 180             data    = IWineD3DVertexBufferImpl_GetMemory(This->stateBlock->streamSource[element->Stream], 0, &streamVBO);
 181
 182             /* Can't use vbo's if the base vertex index is negative. OpenGL doesn't accept negative offsets
 183              * (or rather offsets bigger than the vbo, because the pointer is unsigned), so use system memory
 184              * sources. In most sane cases the pointer - offset will still be > 0, otherwise it will wrap
 185              * around to some big value. Hope that with the indices, the driver wraps it back internally. If
 186              * not, drawStridedSlow is needed, including a vertex buffer path.
 187              */
 188             if(This->stateBlock->loadBaseVertexIndex < 0) {
 189                 WARN("loadBaseVertexIndex is < 0 (%d), not using vbos\n", This->stateBlock->loadBaseVertexIndex);
 190                 streamVBO = 0;
 191                 data = ((IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[element->Stream])->resource.allocatedMemory;
 192                 if((UINT_PTR)data < -This->stateBlock->loadBaseVertexIndex * stride) {
 193                     FIXME("System memory vertex data load offset is negative!\n");
 194                 }
 195             }
 196
 197             if(fixup) {
 198                 if( streamVBO != 0) *fixup = TRUE;
 199                 else if(*fixup && !useVertexShaderFunction &&
 200                        (element->Usage == WINED3DDECLUSAGE_COLOR ||
 201                         element->Usage == WINED3DDECLUSAGE_POSITIONT)) {
 202                     static BOOL warned = FALSE;
 203                     if(!warned) {
 204                         /* This may be bad with the fixed function pipeline */
 205                         FIXME("Missing vbo streams with unfixed colors or transformed position, expect problems\n");
 206                         warned = TRUE;
 207                     }
 208                 }
 209             }
 210         }
 211         data += element->Offset;
 212
 213         TRACE("Offset %d Stream %d UsageIndex %d\n", element->Offset, element->Stream, element->UsageIndex);
 214
 215         if (useVertexShaderFunction)
 216             stride_used = vshader_get_input(This->stateBlock->vertexShader,
 217                 element->Usage, element->UsageIndex, &idx);
 218         else
 219             stride_used = fixed_get_input(element->Usage, element->UsageIndex, &idx);
 220
 221         if (stride_used) {
 222             TRACE("Load %s array %u [usage=%s, usage_idx=%u, "
 223                     "stream=%u, offset=%u, stride=%u, type=%s, VBO=%u]\n",
 224                     useVertexShaderFunction? "shader": "fixed function", idx,
 225                     debug_d3ddeclusage(element->Usage), element->UsageIndex,
 226                     element->Stream, element->Offset, stride, debug_d3ddecltype(element->Type), streamVBO);
 227
 228             if (!useVertexShaderFunction && !vertexDeclaration->ffp_valid[i]) {
 229                 WARN("Skipping unsupported fixed function element of type %s and usage %s\n",
 230                     debug_d3ddecltype(element->Type), debug_d3ddeclusage(element->Usage));
 231             } else {
 232                 strided->u.input[idx].lpData = data;
 233                 strided->u.input[idx].dwType = element->Type;
 234                 strided->u.input[idx].dwStride = stride;
 235                 strided->u.input[idx].VBO = streamVBO;
 236                 strided->u.input[idx].streamNo = element->Stream;
 237             }
 238         }
 239     }
 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         TRACE("(%p) : glDrawArrays(%#x, %d, %d)\n", This, glPrimitiveType, startVertex, numberOfVertices);
 277         glDrawArrays(glPrimitiveType, startVertex, numberOfVertices);
 278         checkGLcall("glDrawArrays");
 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     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
 298     UINT *streamOffset = This->stateBlock->streamOffset;
 299     long                      SkipnStrides = startVertex + This->stateBlock->loadBaseVertexIndex;
 300     BOOL                      pixelShader = use_ps(This);
 301
 302     BYTE *texCoords[WINED3DDP_MAXTEXCOORD];
 303     BYTE *diffuse = NULL, *specular = NULL, *normal = NULL, *position = NULL;
 304
 305     TRACE("Using slow vertex array code\n");
 306
 307     /* Variable Initialization */
 308     if (idxSize != 0) {
 309         /* Immediate mode drawing can't make use of indices in a vbo - get the data from the index buffer.
 310          * If the index buffer has no vbo(not supported or other reason), or with user pointer drawing
 311          * idxData will be != NULL
 312          */
 313         if(idxData == NULL) {
 314             idxData = ((IWineD3DIndexBufferImpl *) This->stateBlock->pIndexData)->resource.allocatedMemory;
 315         }
 316
 317         if (idxSize == 2) pIdxBufS = (const WORD *) idxData;
 318         else pIdxBufL = (const DWORD *) idxData;
 319     } else if (idxData) {
 320         ERR("non-NULL idxData with 0 idxSize, this should never happen\n");
 321         return;
 322     }
 323
 324     /* Adding the stream offset once is cheaper than doing it every iteration. Do not modify the strided data, it is a pointer
 325      * to the strided Data in the device and might be needed intact on the next draw
 326      */
 327     for (textureNo = 0; textureNo < GL_LIMITS(texture_stages); ++textureNo) {
 328         if(sd->u.s.texCoords[textureNo].lpData) {
 329             texCoords[textureNo] = sd->u.s.texCoords[textureNo].lpData + streamOffset[sd->u.s.texCoords[textureNo].streamNo];
 330         } else {
 331             texCoords[textureNo] = NULL;
 332         }
 333     }
 334     if(sd->u.s.diffuse.lpData) {
 335         diffuse = sd->u.s.diffuse.lpData + streamOffset[sd->u.s.diffuse.streamNo];
 336     }
 337     if(sd->u.s.specular.lpData) {
 338         specular = sd->u.s.specular.lpData + streamOffset[sd->u.s.specular.streamNo];
 339     }
 340     if(sd->u.s.normal.lpData) {
 341         normal = sd->u.s.normal.lpData + streamOffset[sd->u.s.normal.streamNo];
 342     }
 343     if(sd->u.s.position.lpData) {
 344         position = sd->u.s.position.lpData + streamOffset[sd->u.s.position.streamNo];
 345     }
 346
 347     if(FIXME_ON(d3d_draw)) {
 348         if(specular && This->stateBlock->renderState[WINED3DRS_FOGENABLE] &&
 349            (This->stateBlock->renderState[WINED3DRS_FOGVERTEXMODE] == WINED3DFOG_NONE || sd->u.s.position_transformed )&&
 350            This->stateBlock->renderState[WINED3DRS_FOGTABLEMODE] == WINED3DFOG_NONE) {
 351             if(GL_SUPPORT(EXT_FOG_COORD) && sd->u.s.specular.dwType != WINED3DDECLTYPE_D3DCOLOR) {
 352                 FIXME("Implement fog coordinates from %s\n", debug_d3ddecltype(sd->u.s.specular.dwType));
 353             }
 354         }
 355         if(This->activeContext->num_untracked_materials && sd->u.s.diffuse.dwType != WINED3DDECLTYPE_D3DCOLOR) {
 356             FIXME("Implement diffuse color tracking from %s\n", debug_d3ddecltype(sd->u.s.diffuse.dwType));
 357         }
 358     }
 359
 360     /* Start drawing in GL */
 361     VTRACE(("glBegin(%x)\n", glPrimType));
 362     glBegin(glPrimType);
 363
 364     /* Default settings for data that is not passed */
 365     if (sd->u.s.normal.lpData == NULL) {
 366         glNormal3f(0, 0, 0);
 367     }
 368     if(sd->u.s.diffuse.lpData == NULL) {
 369         glColor4f(1.0f, 1.0f, 1.0f, 1.0f);
 370     }
 371     if(sd->u.s.specular.lpData == NULL) {
 372         if (GL_SUPPORT(EXT_SECONDARY_COLOR)) {
 373             GL_EXTCALL(glSecondaryColor3fEXT)(0, 0, 0);
 374         }
 375     }
 376
 377     /* We shouldn't start this function if any VBO is involved. Should I put a safety check here?
 378      * Guess it's not necessary(we crash then anyway) and would only eat CPU time
 379      */
 380
 381     /* For each primitive */
 382     for (vx_index = 0; vx_index < NumVertexes; ++vx_index) {
 383         /* Blending data and Point sizes are not supported by this function. They are not supported by the fixed
 384          * function pipeline at all. A Fixme for them is printed after decoding the vertex declaration
 385          */
 386
 387         /* For indexed data, we need to go a few more strides in */
 388         if (idxData != NULL) {
 389
 390             /* Indexed so work out the number of strides to skip */
 391             if (idxSize == 2) {
 392                 VTRACE(("Idx for vertex %d = %d\n", vx_index, pIdxBufS[startIdx+vx_index]));
 393                 SkipnStrides = pIdxBufS[startIdx + vx_index] + This->stateBlock->loadBaseVertexIndex;
 394             } else {
 395                 VTRACE(("Idx for vertex %d = %d\n", vx_index, pIdxBufL[startIdx+vx_index]));
 396                 SkipnStrides = pIdxBufL[startIdx + vx_index] + This->stateBlock->loadBaseVertexIndex;
 397             }
 398         }
 399
 400         /* Texture coords --------------------------- */
 401         for (textureNo = 0; textureNo < GL_LIMITS(texture_stages); ++textureNo) {
 402
 403             if (!GL_SUPPORT(ARB_MULTITEXTURE) && textureNo > 0) {
 404                 FIXME("Program using multiple concurrent textures which this opengl implementation doesn't support\n");
 405                 continue ;
 406             }
 407
 408             /* Query tex coords */
 409             if (This->stateBlock->textures[textureNo] != NULL || pixelShader) {
 410                 int    coordIdx = This->stateBlock->textureState[textureNo][WINED3DTSS_TEXCOORDINDEX];
 411                 int texture_idx = This->texUnitMap[textureNo];
 412                 void *ptrToCoords;
 413
 414                 if (coordIdx > 7) {
 415                     VTRACE(("tex: %d - Skip tex coords, as being system generated\n", textureNo));
 416                     continue;
 417                 } else if (coordIdx < 0) {
 418                     FIXME("tex: %d - Coord index %d is less than zero, expect a crash.\n", textureNo, coordIdx);
 419                     continue;
 420                 }
 421
 422                 if (texture_idx == -1) continue;
 423
 424                 if (texCoords[coordIdx] == NULL) {
 425                     TRACE("tex: %d - Skipping tex coords, as no data supplied\n", textureNo);
 426                     if (GL_SUPPORT(ARB_MULTITEXTURE)) {
 427                         GL_EXTCALL(glMultiTexCoord4fARB(GL_TEXTURE0_ARB + texture_idx, 0, 0, 0, 1));
 428                     } else {
 429                         glTexCoord4f(0, 0, 0, 1);
 430                     }
 431                     continue;
 432                 }
 433
 434                 ptrToCoords = texCoords[coordIdx] + (SkipnStrides * sd->u.s.texCoords[coordIdx].dwStride);
 435                 if (GL_SUPPORT(ARB_MULTITEXTURE))
 436                     multi_texcoord_funcs[sd->u.s.texCoords[coordIdx].dwType](GL_TEXTURE0_ARB + texture_idx, ptrToCoords);
 437                 else
 438                     texcoord_funcs[sd->u.s.texCoords[coordIdx].dwType](ptrToCoords);
 439             }
 440         } /* End of textures */
 441
 442         /* Diffuse -------------------------------- */
 443         if (diffuse) {
 444             DWORD *ptrToCoords = (DWORD *)(diffuse + (SkipnStrides * sd->u.s.diffuse.dwStride));
 445
 446             diffuse_funcs[sd->u.s.diffuse.dwType]((void *) ptrToCoords);
 447             if(This->activeContext->num_untracked_materials) {
 448                 DWORD diffuseColor = ptrToCoords[0];
 449                 unsigned char i;
 450                 float color[4];
 451
 452                 color[0] = D3DCOLOR_B_R(diffuseColor) / 255.0;
 453                 color[1] = D3DCOLOR_B_G(diffuseColor) / 255.0;
 454                 color[2] = D3DCOLOR_B_B(diffuseColor) / 255.0;
 455                 color[3] = D3DCOLOR_B_A(diffuseColor) / 255.0;
 456
 457                 for(i = 0; i < This->activeContext->num_untracked_materials; i++) {
 458                     glMaterialfv(GL_FRONT_AND_BACK, This->activeContext->untracked_materials[i], color);
 459                 }
 460             }
 461         }
 462
 463         /* Specular ------------------------------- */
 464         if (specular) {
 465             DWORD *ptrToCoords = (DWORD *)(specular + (SkipnStrides * sd->u.s.specular.dwStride));
 466
 467             /* special case where the fog density is stored in the specular alpha channel */
 468             if(This->stateBlock->renderState[WINED3DRS_FOGENABLE] &&
 469               (This->stateBlock->renderState[WINED3DRS_FOGVERTEXMODE] == WINED3DFOG_NONE || sd->u.s.position.dwType == WINED3DDECLTYPE_FLOAT4 )&&
 470               This->stateBlock->renderState[WINED3DRS_FOGTABLEMODE] == WINED3DFOG_NONE) {
 471                 if(GL_SUPPORT(EXT_FOG_COORD)) {
 472                     DWORD specularColor = ptrToCoords[0];
 473                     GL_EXTCALL(glFogCoordfEXT(specularColor >> 24));
 474                 } else {
 475                     static BOOL warned = FALSE;
 476                     if(!warned) {
 477                         /* TODO: Use the fog table code from old ddraw */
 478                         FIXME("Implement fog for transformed vertices in software\n");
 479                         warned = TRUE;
 480                     }
 481                 }
 482             }
 483
 484             specular_funcs[sd->u.s.specular.dwType]((void *) ptrToCoords);
 485         }
 486
 487         /* Normal -------------------------------- */
 488         if (normal != NULL) {
 489             float *ptrToCoords = (float *)(normal + (SkipnStrides * sd->u.s.normal.dwStride));
 490             normal_funcs[sd->u.s.normal.dwType](ptrToCoords);
 491         }
 492
 493         /* Position -------------------------------- */
 494         if (position) {
 495             float *ptrToCoords = (float *)(position + (SkipnStrides * sd->u.s.position.dwStride));
 496             position_funcs[sd->u.s.position.dwType](ptrToCoords);
 497         }
 498
 499         /* For non indexed mode, step onto next parts */
 500         if (idxData == NULL) {
 501             ++SkipnStrides;
 502         }
 503     }
 504
 505     glEnd();
 506     checkGLcall("glEnd and previous calls");
 507 }
 508
 509 static inline void send_attribute(IWineD3DDeviceImpl *This, const DWORD type, const UINT index, const void *ptr) {
 510     switch(type) {
 511         case WINED3DDECLTYPE_FLOAT1:
 512             GL_EXTCALL(glVertexAttrib1fvARB(index, (float *) ptr));
 513             break;
 514         case WINED3DDECLTYPE_FLOAT2:
 515             GL_EXTCALL(glVertexAttrib2fvARB(index, (float *) ptr));
 516             break;
 517         case WINED3DDECLTYPE_FLOAT3:
 518             GL_EXTCALL(glVertexAttrib3fvARB(index, (float *) ptr));
 519             break;
 520         case WINED3DDECLTYPE_FLOAT4:
 521             GL_EXTCALL(glVertexAttrib4fvARB(index, (float *) ptr));
 522             break;
 523
 524         case WINED3DDECLTYPE_UBYTE4:
 525             GL_EXTCALL(glVertexAttrib4ubvARB(index, ptr));
 526             break;
 527         case WINED3DDECLTYPE_UBYTE4N:
 528         case WINED3DDECLTYPE_D3DCOLOR:
 529             GL_EXTCALL(glVertexAttrib4NubvARB(index, ptr));
 530             break;
 531
 532         case WINED3DDECLTYPE_SHORT2:
 533             GL_EXTCALL(glVertexAttrib4svARB(index, (GLshort *) ptr));
 534             break;
 535         case WINED3DDECLTYPE_SHORT4:
 536             GL_EXTCALL(glVertexAttrib4svARB(index, (GLshort *) ptr));
 537             break;
 538
 539         case WINED3DDECLTYPE_SHORT2N:
 540         {
 541             GLshort s[4] = {((short *) ptr)[0], ((short *) ptr)[1], 0, 1};
 542             GL_EXTCALL(glVertexAttrib4NsvARB(index, s));
 543             break;
 544         }
 545         case WINED3DDECLTYPE_USHORT2N:
 546         {
 547             GLushort s[4] = {((unsigned short *) ptr)[0], ((unsigned short *) ptr)[1], 0, 1};
 548             GL_EXTCALL(glVertexAttrib4NusvARB(index, s));
 549             break;
 550         }
 551         case WINED3DDECLTYPE_SHORT4N:
 552             GL_EXTCALL(glVertexAttrib4NsvARB(index, (GLshort *) ptr));
 553             break;
 554         case WINED3DDECLTYPE_USHORT4N:
 555             GL_EXTCALL(glVertexAttrib4NusvARB(index, (GLushort *) ptr));
 556             break;
 557
 558         case WINED3DDECLTYPE_UDEC3:
 559             FIXME("Unsure about WINED3DDECLTYPE_UDEC3\n");
 560             /*glVertexAttrib3usvARB(instancedData[j], (GLushort *) ptr); Does not exist */
 561             break;
 562         case WINED3DDECLTYPE_DEC3N:
 563             FIXME("Unsure about WINED3DDECLTYPE_DEC3N\n");
 564             /*glVertexAttrib3NusvARB(instancedData[j], (GLushort *) ptr); Does not exist */
 565             break;
 566
 567         case WINED3DDECLTYPE_FLOAT16_2:
 568             /* Are those 16 bit floats. C doesn't have a 16 bit float type. I could read the single bits and calculate a 4
 569              * byte float according to the IEEE standard
 570              */
 571             if (GL_SUPPORT(NV_HALF_FLOAT)) {
 572                 GL_EXTCALL(glVertexAttrib2hvNV(index, (GLhalfNV *)ptr));
 573             } else {
 574                 float x = float_16_to_32(((unsigned short *) ptr) + 0);
 575                 float y = float_16_to_32(((unsigned short *) ptr) + 1);
 576                 GL_EXTCALL(glVertexAttrib2fARB(index, x, y));
 577             }
 578             break;
 579         case WINED3DDECLTYPE_FLOAT16_4:
 580             if (GL_SUPPORT(NV_HALF_FLOAT)) {
 581                 GL_EXTCALL(glVertexAttrib4hvNV(index, (GLhalfNV *)ptr));
 582             } else {
 583                 float x = float_16_to_32(((unsigned short *) ptr) + 0);
 584                 float y = float_16_to_32(((unsigned short *) ptr) + 1);
 585                 float z = float_16_to_32(((unsigned short *) ptr) + 2);
 586                 float w = float_16_to_32(((unsigned short *) ptr) + 3);
 587                 GL_EXTCALL(glVertexAttrib4fARB(index, x, y, z, w));
 588             }
 589             break;
 590
 591         case WINED3DDECLTYPE_UNUSED:
 592         default:
 593             ERR("Unexpected attribute declaration: %d\n", type);
 594             break;
 595     }
 596 }
 597
 598 static void drawStridedSlowVs(IWineD3DDevice *iface, WineDirect3DVertexStridedData *sd, UINT numberOfVertices,
 599                               GLenum glPrimitiveType, const void *idxData, short idxSize, ULONG minIndex, ULONG startIdx,
 600                               ULONG startVertex) {
 601
 602     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *) iface;
 603     long                      SkipnStrides = startVertex + This->stateBlock->loadBaseVertexIndex;
 604     const WORD                *pIdxBufS     = NULL;
 605     const DWORD               *pIdxBufL     = NULL;
 606     LONG                       vx_index;
 607     int i;
 608     IWineD3DStateBlockImpl *stateblock = This->stateBlock;
 609     BYTE *ptr;
 610
 611     if (idxSize != 0) {
 612         /* Immediate mode drawing can't make use of indices in a vbo - get the data from the index buffer.
 613          * If the index buffer has no vbo(not supported or other reason), or with user pointer drawing
 614          * idxData will be != NULL
 615          */
 616         if(idxData == NULL) {
 617             idxData = ((IWineD3DIndexBufferImpl *) stateblock->pIndexData)->resource.allocatedMemory;
 618         }
 619
 620         if (idxSize == 2) pIdxBufS = (const WORD *) idxData;
 621         else pIdxBufL = (const DWORD *) idxData;
 622     } else if (idxData) {
 623         ERR("non-NULL idxData with 0 idxSize, this should never happen\n");
 624         return;
 625     }
 626
 627     /* Start drawing in GL */
 628     VTRACE(("glBegin(%x)\n", glPrimitiveType));
 629     glBegin(glPrimitiveType);
 630
 631     for (vx_index = 0; vx_index < numberOfVertices; ++vx_index) {
 632         if (idxData != NULL) {
 633
 634             /* Indexed so work out the number of strides to skip */
 635             if (idxSize == 2) {
 636                 VTRACE(("Idx for vertex %d = %d\n", vx_index, pIdxBufS[startIdx+vx_index]));
 637                 SkipnStrides = pIdxBufS[startIdx + vx_index] + stateblock->loadBaseVertexIndex;
 638             } else {
 639                 VTRACE(("Idx for vertex %d = %d\n", vx_index, pIdxBufL[startIdx+vx_index]));
 640                 SkipnStrides = pIdxBufL[startIdx + vx_index] + stateblock->loadBaseVertexIndex;
 641             }
 642         }
 643
 644         for(i = MAX_ATTRIBS - 1; i >= 0; i--) {
 645             if(!sd->u.input[i].lpData) continue;
 646
 647             ptr = sd->u.input[i].lpData +
 648                   sd->u.input[i].dwStride * SkipnStrides +
 649                   stateblock->streamOffset[sd->u.input[i].streamNo];
 650
 651             send_attribute(This, sd->u.input[i].dwType, i, ptr);
 652         }
 653         SkipnStrides++;
 654     }
 655
 656     glEnd();
 657 }
 658
 659 static inline void drawStridedInstanced(IWineD3DDevice *iface, WineDirect3DVertexStridedData *sd, UINT numberOfVertices,
 660                                  GLenum glPrimitiveType, const void *idxData, short idxSize, ULONG minIndex,
 661                                  ULONG startIdx, ULONG startVertex) {
 662     UINT numInstances = 0;
 663     int numInstancedAttribs = 0, i, j;
 664     UINT instancedData[sizeof(sd->u.input) / sizeof(sd->u.input[0]) /* 16 */];
 665     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *) iface;
 666     IWineD3DStateBlockImpl *stateblock = This->stateBlock;
 667
 668     if (idxSize == 0) {
 669         /* This is a nasty thing. MSDN says no hardware supports that and apps have to use software vertex processing.
 670          * We don't support this for now
 671          *
 672          * Shouldn't be too hard to support with opengl, in theory just call glDrawArrays instead of drawElements.
 673          * But the StreamSourceFreq value has a different meaning in that situation.
 674          */
 675         FIXME("Non-indexed instanced drawing is not supported\n");
 676         return;
 677     }
 678
 679     TRACE("(%p) : glElements(%x, %d, %d, ...)\n", This, glPrimitiveType, numberOfVertices, minIndex);
 680     idxData = idxData == (void *)-1 ? NULL : idxData;
 681
 682     /* First, figure out how many instances we have to draw */
 683     for(i = 0; i < MAX_STREAMS; i++) {
 684         /* Look at the streams and take the first one which matches */
 685         if(((stateblock->streamFlags[i] & WINED3DSTREAMSOURCE_INSTANCEDATA) || (stateblock->streamFlags[i] & WINED3DSTREAMSOURCE_INDEXEDDATA)) && stateblock->streamSource[i]) {
 686             /* D3D9 could set streamFreq 0 with (INSTANCEDATA or INDEXEDDATA) and then it is handled as 1. See d3d9/tests/visual.c-> stream_test() */
 687             if(stateblock->streamFreq[i] == 0){
 688                 numInstances = 1;
 689             } else {
 690                 numInstances = stateblock->streamFreq[i]; /* use the specified number of instances from the first matched stream. See d3d9/tests/visual.c-> stream_test() */
 691             }
 692             break; /* break, because only the first suitable value is interesting */
 693         }
 694     }
 695
 696     for(i = 0; i < sizeof(sd->u.input) / sizeof(sd->u.input[0]); i++) {
 697         if(stateblock->streamFlags[sd->u.input[i].streamNo] & WINED3DSTREAMSOURCE_INSTANCEDATA) {
 698             instancedData[numInstancedAttribs] = i;
 699             numInstancedAttribs++;
 700         }
 701     }
 702
 703     /* now draw numInstances instances :-) */
 704     for(i = 0; i < numInstances; i++) {
 705         /* Specify the instanced attributes using immediate mode calls */
 706         for(j = 0; j < numInstancedAttribs; j++) {
 707             BYTE *ptr = sd->u.input[instancedData[j]].lpData +
 708                         sd->u.input[instancedData[j]].dwStride * i +
 709                         stateblock->streamOffset[sd->u.input[instancedData[j]].streamNo];
 710             if(sd->u.input[instancedData[j]].VBO) {
 711                 IWineD3DVertexBufferImpl *vb = (IWineD3DVertexBufferImpl *) stateblock->streamSource[sd->u.input[instancedData[j]].streamNo];
 712                 ptr += (long) vb->resource.allocatedMemory;
 713             }
 714
 715             send_attribute(This, sd->u.input[instancedData[j]].dwType, instancedData[j], ptr);
 716         }
 717
 718         glDrawElements(glPrimitiveType, numberOfVertices, idxSize == 2 ? GL_UNSIGNED_SHORT : GL_UNSIGNED_INT,
 719                     (const char *)idxData+(idxSize * startIdx));
 720         checkGLcall("glDrawElements");
 721     }
 722 }
 723
 724 static inline void remove_vbos(IWineD3DDeviceImpl *This, WineDirect3DVertexStridedData *s) {
 725     unsigned char i;
 726     IWineD3DVertexBufferImpl *vb;
 727
 728     if(s->u.s.position.VBO) {
 729         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.position.streamNo];
 730         s->u.s.position.VBO = 0;
 731         s->u.s.position.lpData = (BYTE *) ((unsigned long) s->u.s.position.lpData + (unsigned long) vb->resource.allocatedMemory);
 732     }
 733     if(s->u.s.blendWeights.VBO) {
 734         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.blendWeights.streamNo];
 735         s->u.s.blendWeights.VBO = 0;
 736         s->u.s.blendWeights.lpData = (BYTE *) ((unsigned long) s->u.s.blendWeights.lpData + (unsigned long) vb->resource.allocatedMemory);
 737     }
 738     if(s->u.s.blendMatrixIndices.VBO) {
 739         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.blendMatrixIndices.streamNo];
 740         s->u.s.blendMatrixIndices.VBO = 0;
 741         s->u.s.blendMatrixIndices.lpData = (BYTE *) ((unsigned long) s->u.s.blendMatrixIndices.lpData + (unsigned long) vb->resource.allocatedMemory);
 742     }
 743     if(s->u.s.normal.VBO) {
 744         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.normal.streamNo];
 745         s->u.s.normal.VBO = 0;
 746         s->u.s.normal.lpData = (BYTE *) ((unsigned long) s->u.s.normal.lpData + (unsigned long) vb->resource.allocatedMemory);
 747     }
 748     if(s->u.s.pSize.VBO) {
 749         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.pSize.streamNo];
 750         s->u.s.pSize.VBO = 0;
 751         s->u.s.pSize.lpData = (BYTE *) ((unsigned long) s->u.s.pSize.lpData + (unsigned long) vb->resource.allocatedMemory);
 752     }
 753     if(s->u.s.diffuse.VBO) {
 754         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.diffuse.streamNo];
 755         s->u.s.diffuse.VBO = 0;
 756         s->u.s.diffuse.lpData = (BYTE *) ((unsigned long) s->u.s.diffuse.lpData + (unsigned long) vb->resource.allocatedMemory);
 757     }
 758     if(s->u.s.specular.VBO) {
 759         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.specular.streamNo];
 760         s->u.s.specular.VBO = 0;
 761         s->u.s.specular.lpData = (BYTE *) ((unsigned long) s->u.s.specular.lpData + (unsigned long) vb->resource.allocatedMemory);
 762     }
 763     for(i = 0; i < WINED3DDP_MAXTEXCOORD; i++) {
 764         if(s->u.s.texCoords[i].VBO) {
 765             vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.texCoords[i].streamNo];
 766             s->u.s.texCoords[i].VBO = 0;
 767             s->u.s.texCoords[i].lpData = (BYTE *) ((unsigned long) s->u.s.texCoords[i].lpData + (unsigned long) vb->resource.allocatedMemory);
 768         }
 769     }
 770     if(s->u.s.position2.VBO) {
 771         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.position2.streamNo];
 772         s->u.s.position2.VBO = 0;
 773         s->u.s.position2.lpData = (BYTE *) ((unsigned long) s->u.s.position2.lpData + (unsigned long) vb->resource.allocatedMemory);
 774     }
 775     if(s->u.s.normal2.VBO) {
 776         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.normal2.streamNo];
 777         s->u.s.normal2.VBO = 0;
 778         s->u.s.normal2.lpData = (BYTE *) ((unsigned long) s->u.s.normal2.lpData + (unsigned long) vb->resource.allocatedMemory);
 779     }
 780     if(s->u.s.tangent.VBO) {
 781         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.tangent.streamNo];
 782         s->u.s.tangent.VBO = 0;
 783         s->u.s.tangent.lpData = (BYTE *) ((unsigned long) s->u.s.tangent.lpData + (unsigned long) vb->resource.allocatedMemory);
 784     }
 785     if(s->u.s.binormal.VBO) {
 786         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.binormal.streamNo];
 787         s->u.s.binormal.VBO = 0;
 788         s->u.s.binormal.lpData = (BYTE *) ((unsigned long) s->u.s.binormal.lpData + (unsigned long) vb->resource.allocatedMemory);
 789     }
 790     if(s->u.s.tessFactor.VBO) {
 791         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.tessFactor.streamNo];
 792         s->u.s.tessFactor.VBO = 0;
 793         s->u.s.tessFactor.lpData = (BYTE *) ((unsigned long) s->u.s.tessFactor.lpData + (unsigned long) vb->resource.allocatedMemory);
 794     }
 795     if(s->u.s.fog.VBO) {
 796         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.fog.streamNo];
 797         s->u.s.fog.VBO = 0;
 798         s->u.s.fog.lpData = (BYTE *) ((unsigned long) s->u.s.fog.lpData + (unsigned long) vb->resource.allocatedMemory);
 799     }
 800     if(s->u.s.depth.VBO) {
 801         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.depth.streamNo];
 802         s->u.s.depth.VBO = 0;
 803         s->u.s.depth.lpData = (BYTE *) ((unsigned long) s->u.s.depth.lpData + (unsigned long) vb->resource.allocatedMemory);
 804     }
 805     if(s->u.s.sample.VBO) {
 806         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.sample.streamNo];
 807         s->u.s.sample.VBO = 0;
 808         s->u.s.sample.lpData = (BYTE *) ((unsigned long) s->u.s.sample.lpData + (unsigned long) vb->resource.allocatedMemory);
 809     }
 810 }
 811
 812 /* Routine common to the draw primitive and draw indexed primitive routines */
 813 void drawPrimitive(IWineD3DDevice *iface,
 814                    int PrimitiveType,
 815                    long NumPrimitives,
 816                    /* for Indexed: */
 817                    long  StartVertexIndex,
 818                    UINT  numberOfVertices,
 819                    long  StartIdx,
 820                    short idxSize,
 821                    const void *idxData,
 822                    int   minIndex) {
 823
 824     IWineD3DDeviceImpl           *This = (IWineD3DDeviceImpl *)iface;
 825     IWineD3DSurfaceImpl          *target;
 826     int i;
 827
 828     if (NumPrimitives == 0) return;
 829
 830     /* Invalidate the back buffer memory so LockRect will read it the next time */
 831     for(i = 0; i < GL_LIMITS(buffers); i++) {
 832         target = (IWineD3DSurfaceImpl *) This->render_targets[i];
 833         if (target) {
 834             IWineD3DSurface_LoadLocation((IWineD3DSurface *) target, SFLAG_INDRAWABLE, NULL);
 835             IWineD3DSurface_ModifyLocation((IWineD3DSurface *) target, SFLAG_INDRAWABLE, TRUE);
 836         }
 837     }
 838
 839     /* Signals other modules that a drawing is in progress and the stateblock finalized */
 840     This->isInDraw = TRUE;
 841
 842     ActivateContext(This, This->render_targets[0], CTXUSAGE_DRAWPRIM);
 843
 844     if (This->stencilBufferTarget) {
 845         /* Note that this depends on the ActivateContext call above to set
 846          * This->render_offscreen properly */
 847         DWORD location = This->render_offscreen ? SFLAG_DS_OFFSCREEN : SFLAG_DS_ONSCREEN;
 848         surface_load_ds_location(This->stencilBufferTarget, location);
 849         surface_modify_ds_location(This->stencilBufferTarget, location);
 850     }
 851
 852     /* Ok, we will be updating the screen from here onwards so grab the lock */
 853     ENTER_GL();
 854     {
 855         GLenum glPrimType;
 856         BOOL emulation = FALSE;
 857         WineDirect3DVertexStridedData *strided = &This->strided_streams;
 858         WineDirect3DVertexStridedData stridedlcl;
 859         /* Ok, Work out which primitive is requested and how many vertexes that
 860            will be                                                              */
 861         UINT calculatedNumberOfindices = primitiveToGl(PrimitiveType, NumPrimitives, &glPrimType);
 862         if (numberOfVertices == 0 )
 863             numberOfVertices = calculatedNumberOfindices;
 864
 865         if(!use_vs(This)) {
 866             if(!This->strided_streams.u.s.position_transformed && This->activeContext->num_untracked_materials &&
 867                 This->stateBlock->renderState[WINED3DRS_LIGHTING]) {
 868                 static BOOL first = TRUE;
 869                 if(first) {
 870                     FIXME("Using software emulation because not all material properties could be tracked\n");
 871                     first = FALSE;
 872                 } else {
 873                     TRACE("Using software emulation because not all material properties could be tracked\n");
 874                 }
 875                 emulation = TRUE;
 876             }
 877             else if(This->activeContext->fog_coord && This->stateBlock->renderState[WINED3DRS_FOGENABLE]) {
 878                 /* Either write a pipeline replacement shader or convert the specular alpha from unsigned byte
 879                  * to a float in the vertex buffer
 880                  */
 881                 static BOOL first = TRUE;
 882                 if(first) {
 883                     FIXME("Using software emulation because manual fog coordinates are provided\n");
 884                     first = FALSE;
 885                 } else {
 886                     TRACE("Using software emulation because manual fog coordinates are provided\n");
 887                 }
 888                 emulation = TRUE;
 889             }
 890
 891             if(emulation) {
 892                 strided = &stridedlcl;
 893                 memcpy(&stridedlcl, &This->strided_streams, sizeof(stridedlcl));
 894                 remove_vbos(This, &stridedlcl);
 895             }
 896         }
 897
 898         if (This->useDrawStridedSlow || emulation) {
 899             /* Immediate mode drawing */
 900             if(use_vs(This)) {
 901                 static BOOL first = TRUE;
 902                 if(first) {
 903                     FIXME("Using immediate mode with vertex shaders for half float emulation\n");
 904                     first = FALSE;
 905                 } else {
 906                     TRACE("Using immediate mode with vertex shaders for half float emulation\n");
 907                 }
 908                 drawStridedSlowVs(iface, strided, calculatedNumberOfindices, glPrimType,
 909                                   idxData, idxSize, minIndex, StartIdx, StartVertexIndex);
 910             } else {
 911                 drawStridedSlow(iface, strided, calculatedNumberOfindices,
 912                                 glPrimType, idxData, idxSize, minIndex, StartIdx, StartVertexIndex);
 913             }
 914         } else if(This->instancedDraw) {
 915             /* Instancing emulation with mixing immediate mode and arrays */
 916             drawStridedInstanced(iface, &This->strided_streams, calculatedNumberOfindices, glPrimType,
 917                             idxData, idxSize, minIndex, StartIdx, StartVertexIndex);
 918         } else {
 919             drawStridedFast(iface, calculatedNumberOfindices, glPrimType,
 920                             idxData, idxSize, minIndex, StartIdx, StartVertexIndex);
 921         }
 922     }
 923
 924     /* Finished updating the screen, restore lock */
 925     LEAVE_GL();
 926     TRACE("Done all gl drawing\n");
 927
 928     /* Diagnostics */
 929 #ifdef SHOW_FRAME_MAKEUP
 930     {
 931         static long int primCounter = 0;
 932         /* NOTE: set primCounter to the value reported by drawprim
 933            before you want to to write frame makeup to /tmp */
 934         if (primCounter >= 0) {
 935             WINED3DLOCKED_RECT r;
 936             char buffer[80];
 937             IWineD3DSurface_LockRect(This->render_targets[0], &r, NULL, WINED3DLOCK_READONLY);
 938             sprintf(buffer, "/tmp/backbuffer_%ld.tga", primCounter);
 939             TRACE("Saving screenshot %s\n", buffer);
 940             IWineD3DSurface_SaveSnapshot(This->render_targets[0], buffer);
 941             IWineD3DSurface_UnlockRect(This->render_targets[0]);
 942
 943 #ifdef SHOW_TEXTURE_MAKEUP
 944            {
 945             IWineD3DSurface *pSur;
 946             int textureNo;
 947             for (textureNo = 0; textureNo < MAX_COMBINED_SAMPLERS; ++textureNo) {
 948                 if (This->stateBlock->textures[textureNo] != NULL) {
 949                     sprintf(buffer, "/tmp/texture_%p_%ld_%d.tga", This->stateBlock->textures[textureNo], primCounter, textureNo);
 950                     TRACE("Saving texture %s\n", buffer);
 951                     if (IWineD3DBaseTexture_GetType(This->stateBlock->textures[textureNo]) == WINED3DRTYPE_TEXTURE) {
 952                             IWineD3DTexture_GetSurfaceLevel((IWineD3DTexture *)This->stateBlock->textures[textureNo], 0, &pSur);
 953                             IWineD3DSurface_SaveSnapshot(pSur, buffer);
 954                             IWineD3DSurface_Release(pSur);
 955                     } else  {
 956                         FIXME("base Texture isn't of type texture %d\n", IWineD3DBaseTexture_GetType(This->stateBlock->textures[textureNo]));
 957                     }
 958                 }
 959             }
 960            }
 961 #endif
 962         }
 963         TRACE("drawprim #%ld\n", primCounter);
 964         ++primCounter;
 965     }
 966 #endif
 967
 968     /* Control goes back to the device, stateblock values may change again */
 969     This->isInDraw = FALSE;
 970 }
 971
 972 static void normalize_normal(float *n) {
 973     float length = n[0] * n[0] + n[1] * n[1] + n[2] * n[2];
 974     if(length == 0.0) return;
 975     length = sqrt(length);
 976     n[0] = n[0] / length;
 977     n[1] = n[1] / length;
 978     n[2] = n[2] / length;
 979 }
 980
 981 /* Tesselates a high order rectangular patch into single triangles using gl evaluators
 982  *
 983  * The problem is that OpenGL does not offer a direct way to return the tesselated primitives,
 984  * and they can't be sent off for rendering directly either. Tesselating is slow, so we want
 985  * to cache the patches in a vertex buffer. But more importantly, gl can't bind generated
 986  * attributes to numbered shader attributes, so we have to store them and rebind them as needed
 987  * in drawprim.
 988  *
 989  * To read back, the opengl feedback mode is used. This creates a problem because we want
 990  * untransformed, unlit vertices, but feedback runs everything through transform and lighting.
 991  * Thus disable lighting and set identity matrices to get unmodified colors and positions.
 992  * To overcome clipping find the biggest x, y and z values of the vertices in the patch and scale
 993  * them to [-1.0;+1.0] and set the viewport up to scale them back.
 994  *
 995  * Normals are more tricky: Draw white vertices with 3 directional lights, and calculate the
 996  * resulting colors back to the normals.
 997  *
 998  * NOTE: This function activates a context for blitting, modifies matrices & viewport, but
 999  * does not restore it because normally a draw follows immediately afterwards. The caller is
1000  * responsible of taking care that either the gl states are restored, or the context activated
1001  * for drawing to reset the lastWasBlit flag.
1002  */
1003 HRESULT tesselate_rectpatch(IWineD3DDeviceImpl *This,
1004                             struct WineD3DRectPatch *patch) {
1005     unsigned int i, j, num_quads, out_vertex_size, buffer_size, d3d_out_vertex_size;
1006     float max_x = 0.0, max_y = 0.0, max_z = 0.0, neg_z = 0.0;
1007     WineDirect3DVertexStridedData strided;
1008     BYTE *data;
1009     WINED3DRECTPATCH_INFO *info = &patch->RectPatchInfo;
1010     DWORD vtxStride;
1011     GLenum feedback_type;
1012     GLfloat *feedbuffer;
1013
1014     /* First, locate the position data. This is provided in a vertex buffer in the stateblock.
1015      * Beware of vbos
1016      */
1017     memset(&strided, 0, sizeof(strided));
1018     primitiveDeclarationConvertToStridedData((IWineD3DDevice *) This, FALSE, &strided, NULL);
1019     if(strided.u.s.position.VBO) {
1020         IWineD3DVertexBufferImpl *vb;
1021         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[strided.u.s.position.streamNo];
1022         strided.u.s.position.lpData = (BYTE *) ((unsigned long) strided.u.s.position.lpData +
1023                                                 (unsigned long) vb->resource.allocatedMemory);
1024     }
1025     vtxStride = strided.u.s.position.dwStride;
1026     data = strided.u.s.position.lpData +
1027            vtxStride * info->Stride * info->StartVertexOffsetHeight +
1028            vtxStride * info->StartVertexOffsetWidth;
1029
1030     /* Not entirely sure about what happens with transformed vertices */
1031     if(strided.u.s.position_transformed) {
1032         FIXME("Transformed position in rectpatch generation\n");
1033     }
1034     if(vtxStride % sizeof(GLfloat)) {
1035         /* glMap2f reads vertex sizes in GLfloats, the d3d stride is in bytes.
1036          * I don't see how the stride could not be a multiple of 4, but make sure
1037          * to check it
1038          */
1039         ERR("Vertex stride is not a multiple of sizeof(GLfloat)\n");
1040     }
1041     if(info->Basis != WINED3DBASIS_BEZIER) {
1042         FIXME("Basis is %s, how to handle this?\n", debug_d3dbasis(info->Basis));
1043     }
1044     if(info->Degree != WINED3DDEGREE_CUBIC) {
1045         FIXME("Degree is %s, how to handle this?\n", debug_d3ddegree(info->Degree));
1046     }
1047
1048     /* First, get the boundary cube of the input data */
1049     for(j = 0; j < info->Height; j++) {
1050         for(i = 0; i < info->Width; i++) {
1051             float *v = (float *) (data + vtxStride * i + vtxStride * info->Stride * j);
1052             if(fabs(v[0]) > max_x) max_x = fabs(v[0]);
1053             if(fabs(v[1]) > max_y) max_y = fabs(v[1]);
1054             if(fabs(v[2]) > max_z) max_z = fabs(v[2]);
1055             if(v[2] < neg_z) neg_z = v[2];
1056         }
1057     }
1058
1059     /* This needs some improvements in the vertex decl code */
1060     FIXME("Cannot find data to generate. Only generating position and normals\n");
1061     patch->has_normals = TRUE;
1062     patch->has_texcoords = FALSE;
1063
1064     /* Simply activate the context for blitting. This disables all the things we don't want and
1065      * takes care of dirtifying. Dirtifying is preferred over pushing / popping, since drawing the
1066      * patch (as opposed to normal draws) will most likely need different changes anyway
1067      */
1068     ActivateContext(This, This->lastActiveRenderTarget, CTXUSAGE_BLIT);
1069     ENTER_GL();
1070
1071     glMatrixMode(GL_PROJECTION);
1072     checkGLcall("glMatrixMode(GL_PROJECTION)");
1073     glLoadIdentity();
1074     checkGLcall("glLoadIndentity()");
1075     glScalef(1 / (max_x) , 1 / (max_y), max_z == 0 ? 1 : 1 / ( 2 * max_z));
1076     glTranslatef(0, 0, 0.5);
1077     checkGLcall("glScalef");
1078     glViewport(-max_x, -max_y, 2 * (max_x), 2 * (max_y));
1079     checkGLcall("glViewport");
1080
1081     /* Some states to take care of. If we're in wireframe opengl will produce lines, and confuse
1082      * our feedback buffer parser
1083      */
1084     glPolygonMode(GL_FRONT_AND_BACK, GL_FILL);
1085     checkGLcall("glPolygonMode(GL_FRONT_AND_BACK, GL_FILL)");
1086     IWineD3DDeviceImpl_MarkStateDirty(This, STATE_RENDER(WINED3DRS_FILLMODE));
1087     if(patch->has_normals) {
1088         float black[4] = {0, 0, 0, 0};
1089         float red[4]   = {1, 0, 0, 0};
1090         float green[4] = {0, 1, 0, 0};
1091         float blue[4]  = {0, 0, 1, 0};
1092         float white[4] = {1, 1, 1, 1};
1093         glEnable(GL_LIGHTING);
1094         checkGLcall("glEnable(GL_LIGHTING)");
1095         glLightModelfv(GL_LIGHT_MODEL_AMBIENT, black);
1096         checkGLcall("glLightModel for MODEL_AMBIENT");
1097         IWineD3DDeviceImpl_MarkStateDirty(This, STATE_RENDER(WINED3DRS_AMBIENT));
1098
1099         for(i = 3; i < GL_LIMITS(lights); i++) {
1100             glDisable(GL_LIGHT0 + i);
1101             checkGLcall("glDisable(GL_LIGHT0 + i)");
1102             IWineD3DDeviceImpl_MarkStateDirty(This, STATE_ACTIVELIGHT(i));
1103         }
1104
1105         IWineD3DDeviceImpl_MarkStateDirty(This, STATE_ACTIVELIGHT(0));
1106         glLightfv(GL_LIGHT0, GL_DIFFUSE, red);
1107         glLightfv(GL_LIGHT0, GL_SPECULAR, black);
1108         glLightfv(GL_LIGHT0, GL_AMBIENT, black);
1109         glLightfv(GL_LIGHT0, GL_POSITION, red);
1110         glEnable(GL_LIGHT0);
1111         checkGLcall("Setting up light 1\n");
1112         IWineD3DDeviceImpl_MarkStateDirty(This, STATE_ACTIVELIGHT(1));
1113         glLightfv(GL_LIGHT1, GL_DIFFUSE, green);
1114         glLightfv(GL_LIGHT1, GL_SPECULAR, black);
1115         glLightfv(GL_LIGHT1, GL_AMBIENT, black);
1116         glLightfv(GL_LIGHT1, GL_POSITION, green);
1117         glEnable(GL_LIGHT1);
1118         checkGLcall("Setting up light 2\n");
1119         IWineD3DDeviceImpl_MarkStateDirty(This, STATE_ACTIVELIGHT(2));
1120         glLightfv(GL_LIGHT2, GL_DIFFUSE, blue);
1121         glLightfv(GL_LIGHT2, GL_SPECULAR, black);
1122         glLightfv(GL_LIGHT2, GL_AMBIENT, black);
1123         glLightfv(GL_LIGHT2, GL_POSITION, blue);
1124         glEnable(GL_LIGHT2);
1125         checkGLcall("Setting up light 3\n");
1126
1127         IWineD3DDeviceImpl_MarkStateDirty(This, STATE_MATERIAL);
1128         IWineD3DDeviceImpl_MarkStateDirty(This, STATE_RENDER(WINED3DRS_COLORVERTEX));
1129         glDisable(GL_COLOR_MATERIAL);
1130         glMaterialfv(GL_FRONT_AND_BACK, GL_EMISSION, black);
1131         glMaterialfv(GL_FRONT_AND_BACK, GL_SPECULAR, black);
1132         glMaterialfv(GL_FRONT_AND_BACK, GL_DIFFUSE, white);
1133         checkGLcall("Setting up materials\n");
1134     }
1135
1136     /* Enable the needed maps.
1137      * GL_MAP2_VERTEX_3 is needed for positional data.
1138      * GL_AUTO_NORMAL to generate normals from the position. Do not use GL_MAP2_NORMAL.
1139      * GL_MAP2_TEXTURE_COORD_4 for texture coords
1140      */
1141     num_quads = ceilf(patch->numSegs[0]) * ceilf(patch->numSegs[1]);
1142     out_vertex_size = 3 /* position */;
1143     d3d_out_vertex_size = 3;
1144     glEnable(GL_MAP2_VERTEX_3);
1145     if(patch->has_normals && patch->has_texcoords) {
1146         FIXME("Texcoords not handled yet\n");
1147         feedback_type = GL_3D_COLOR_TEXTURE;
1148         out_vertex_size += 8;
1149         d3d_out_vertex_size += 7;
1150         glEnable(GL_AUTO_NORMAL);
1151         glEnable(GL_MAP2_TEXTURE_COORD_4);
1152     } else if(patch->has_texcoords) {
1153         FIXME("Texcoords not handled yet\n");
1154         feedback_type = GL_3D_COLOR_TEXTURE;
1155         out_vertex_size += 7;
1156         d3d_out_vertex_size += 4;
1157         glEnable(GL_MAP2_TEXTURE_COORD_4);
1158     } else if(patch->has_normals) {
1159         feedback_type = GL_3D_COLOR;
1160         out_vertex_size += 4;
1161         d3d_out_vertex_size += 3;
1162         glEnable(GL_AUTO_NORMAL);
1163     } else {
1164         feedback_type = GL_3D;
1165     }
1166     checkGLcall("glEnable vertex attrib generation");
1167
1168     buffer_size = num_quads * out_vertex_size * 2 /* triangle list */ * 3 /* verts per tri */
1169                    + 4 * num_quads /* 2 triangle markers per quad + num verts in tri */;
1170     feedbuffer = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, buffer_size * sizeof(float) * 8);
1171
1172     glMap2f(GL_MAP2_VERTEX_3,
1173             0, 1, vtxStride / sizeof(float), info->Width,
1174             0, 1, info->Stride * vtxStride / sizeof(float), info->Height,
1175             (float *) data);
1176     checkGLcall("glMap2f");
1177     if(patch->has_texcoords) {
1178         glMap2f(GL_MAP2_TEXTURE_COORD_4,
1179                 0, 1, vtxStride / sizeof(float), info->Width,
1180                 0, 1, info->Stride * vtxStride / sizeof(float), info->Height,
1181                 (float *) data);
1182         checkGLcall("glMap2f");
1183     }
1184     glMapGrid2f(ceilf(patch->numSegs[0]), 0.0, 1.0, ceilf(patch->numSegs[1]), 0.0, 1.0);
1185     checkGLcall("glMapGrid2f");
1186
1187     glFeedbackBuffer(buffer_size * 2, feedback_type, feedbuffer);
1188     checkGLcall("glFeedbackBuffer");
1189     glRenderMode(GL_FEEDBACK);
1190
1191     glEvalMesh2(GL_FILL, 0, ceilf(patch->numSegs[0]), 0, ceilf(patch->numSegs[1]));
1192     checkGLcall("glEvalMesh2\n");
1193
1194     i = glRenderMode(GL_RENDER);
1195     if(i == -1) {
1196         LEAVE_GL();
1197         ERR("Feedback failed. Expected %d elements back\n", buffer_size);
1198         Sleep(10000);
1199         HeapFree(GetProcessHeap(), 0, feedbuffer);
1200         return WINED3DERR_DRIVERINTERNALERROR;
1201     } else if(i != buffer_size) {
1202         LEAVE_GL();
1203         ERR("Unexpected amount of elements returned. Expected %d, got %d\n", buffer_size, i);
1204         Sleep(10000);
1205         HeapFree(GetProcessHeap(), 0, feedbuffer);
1206         return WINED3DERR_DRIVERINTERNALERROR;
1207     } else {
1208         TRACE("Got %d elements as expected\n", i);
1209     }
1210
1211     HeapFree(GetProcessHeap(), 0, patch->mem);
1212     patch->mem = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, num_quads * 6 * d3d_out_vertex_size * sizeof(float) * 8);
1213     i = 0;
1214     for(j = 0; j < buffer_size; j += (3 /* num verts */ * out_vertex_size + 2 /* tri marker */)) {
1215         if(feedbuffer[j] != GL_POLYGON_TOKEN) {
1216             ERR("Unexpected token: %f\n", feedbuffer[j]);
1217             continue;
1218         }
1219         if(feedbuffer[j + 1] != 3) {
1220             ERR("Unexpected polygon: %f corners\n", feedbuffer[j + 1]);
1221             continue;
1222         }
1223         /* Somehow there are different ideas about back / front facing, so fix up the
1224          * vertex order
1225          */
1226         patch->mem[i + 0] =  feedbuffer[j + out_vertex_size * 2 + 2]; /* x, triangle 2 */
1227         patch->mem[i + 1] =  feedbuffer[j + out_vertex_size * 2 + 3]; /* y, triangle 2 */
1228         patch->mem[i + 2] = (feedbuffer[j + out_vertex_size * 2 + 4] - 0.5) * 4 * max_z; /* z, triangle 3 */
1229         if(patch->has_normals) {
1230             patch->mem[i + 3] = feedbuffer[j + out_vertex_size * 2 + 5];
1231             patch->mem[i + 4] = feedbuffer[j + out_vertex_size * 2 + 6];
1232             patch->mem[i + 5] = feedbuffer[j + out_vertex_size * 2 + 7];
1233         }
1234         i += d3d_out_vertex_size;
1235
1236         patch->mem[i + 0] =  feedbuffer[j + out_vertex_size * 1 + 2]; /* x, triangle 2 */
1237         patch->mem[i + 1] =  feedbuffer[j + out_vertex_size * 1 + 3]; /* y, triangle 2 */
1238         patch->mem[i + 2] = (feedbuffer[j + out_vertex_size * 1 + 4] - 0.5) * 4 * max_z; /* z, triangle 2 */
1239         if(patch->has_normals) {
1240             patch->mem[i + 3] = feedbuffer[j + out_vertex_size * 1 + 5];
1241             patch->mem[i + 4] = feedbuffer[j + out_vertex_size * 1 + 6];
1242             patch->mem[i + 5] = feedbuffer[j + out_vertex_size * 1 + 7];
1243         }
1244         i += d3d_out_vertex_size;
1245
1246         patch->mem[i + 0] =  feedbuffer[j + out_vertex_size * 0 + 2]; /* x, triangle 1 */
1247         patch->mem[i + 1] =  feedbuffer[j + out_vertex_size * 0 + 3]; /* y, triangle 1 */
1248         patch->mem[i + 2] = (feedbuffer[j + out_vertex_size * 0 + 4] - 0.5) * 4 * max_z; /* z, triangle 1 */
1249         if(patch->has_normals) {
1250             patch->mem[i + 3] = feedbuffer[j + out_vertex_size * 0 + 5];
1251             patch->mem[i + 4] = feedbuffer[j + out_vertex_size * 0 + 6];
1252             patch->mem[i + 5] = feedbuffer[j + out_vertex_size * 0 + 7];
1253         }
1254         i += d3d_out_vertex_size;
1255     }
1256
1257     if(patch->has_normals) {
1258         /* Now do the same with reverse light directions */
1259         float x[4] = {-1,  0,  0, 0};
1260         float y[4] = { 0, -1,  0, 0};
1261         float z[4] = { 0,  0, -1, 0};
1262         glLightfv(GL_LIGHT0, GL_POSITION, x);
1263         glLightfv(GL_LIGHT1, GL_POSITION, y);
1264         glLightfv(GL_LIGHT2, GL_POSITION, z);
1265         checkGLcall("Setting up reverse light directions\n");
1266
1267         glRenderMode(GL_FEEDBACK);
1268         checkGLcall("glRenderMode(GL_FEEDBACK)");
1269         glEvalMesh2(GL_FILL, 0, ceilf(patch->numSegs[0]), 0, ceilf(patch->numSegs[1]));
1270         checkGLcall("glEvalMesh2\n");
1271         i = glRenderMode(GL_RENDER);
1272         checkGLcall("glRenderMode(GL_RENDER)");
1273
1274         i = 0;
1275         for(j = 0; j < buffer_size; j += (3 /* num verts */ * out_vertex_size + 2 /* tri marker */)) {
1276             if(feedbuffer[j] != GL_POLYGON_TOKEN) {
1277                 ERR("Unexpected token: %f\n", feedbuffer[j]);
1278                 continue;
1279             }
1280             if(feedbuffer[j + 1] != 3) {
1281                 ERR("Unexpected polygon: %f corners\n", feedbuffer[j + 1]);
1282                 continue;
1283             }
1284             if(patch->mem[i + 3] == 0.0)
1285                 patch->mem[i + 3] = -feedbuffer[j + out_vertex_size * 2 + 5];
1286             if(patch->mem[i + 4] == 0.0)
1287                 patch->mem[i + 4] = -feedbuffer[j + out_vertex_size * 2 + 6];
1288             if(patch->mem[i + 5] == 0.0)
1289                 patch->mem[i + 5] = -feedbuffer[j + out_vertex_size * 2 + 7];
1290             normalize_normal(patch->mem + i + 3);
1291             i += d3d_out_vertex_size;
1292
1293             if(patch->mem[i + 3] == 0.0)
1294                 patch->mem[i + 3] = -feedbuffer[j + out_vertex_size * 1 + 5];
1295             if(patch->mem[i + 4] == 0.0)
1296                 patch->mem[i + 4] = -feedbuffer[j + out_vertex_size * 1 + 6];
1297             if(patch->mem[i + 5] == 0.0)
1298                 patch->mem[i + 5] = -feedbuffer[j + out_vertex_size * 1 + 7];
1299             normalize_normal(patch->mem + i + 3);
1300             i += d3d_out_vertex_size;
1301
1302             if(patch->mem[i + 3] == 0.0)
1303                 patch->mem[i + 3] = -feedbuffer[j + out_vertex_size * 0 + 5];
1304             if(patch->mem[i + 4] == 0.0)
1305                 patch->mem[i + 4] = -feedbuffer[j + out_vertex_size * 0 + 6];
1306             if(patch->mem[i + 5] == 0.0)
1307                 patch->mem[i + 5] = -feedbuffer[j + out_vertex_size * 0 + 7];
1308             normalize_normal(patch->mem + i + 3);
1309             i += d3d_out_vertex_size;
1310         }
1311     }
1312
1313     glDisable(GL_MAP2_VERTEX_3);
1314     glDisable(GL_AUTO_NORMAL);
1315     glDisable(GL_MAP2_NORMAL);
1316     glDisable(GL_MAP2_TEXTURE_COORD_4);
1317     checkGLcall("glDisable vertex attrib generation");
1318     LEAVE_GL();
1319
1320     HeapFree(GetProcessHeap(), 0, feedbuffer);
1321
1322     vtxStride = 3 * sizeof(float);
1323     if(patch->has_normals) {
1324         vtxStride += 3 * sizeof(float);
1325     }
1326     if(patch->has_texcoords) {
1327         vtxStride += 4 * sizeof(float);
1328     }
1329     memset(&patch->strided, 0, sizeof(&patch->strided));
1330     patch->strided.u.s.position.lpData = (BYTE *) patch->mem;
1331     patch->strided.u.s.position.dwStride = vtxStride;
1332     patch->strided.u.s.position.dwType = WINED3DDECLTYPE_FLOAT3;
1333     patch->strided.u.s.position.streamNo = 255;
1334
1335     if(patch->has_normals) {
1336         patch->strided.u.s.normal.lpData = (BYTE *) patch->mem + 3 * sizeof(float) /* pos */;
1337         patch->strided.u.s.normal.dwStride = vtxStride;
1338         patch->strided.u.s.normal.dwType = WINED3DDECLTYPE_FLOAT3;
1339         patch->strided.u.s.normal.streamNo = 255;
1340     }
1341     if(patch->has_texcoords) {
1342         patch->strided.u.s.texCoords[0].lpData = (BYTE *) patch->mem + 3 * sizeof(float) /* pos */;
1343         if(patch->has_normals) {
1344             patch->strided.u.s.texCoords[0].lpData += 3 * sizeof(float);
1345         }
1346         patch->strided.u.s.texCoords[0].dwStride = vtxStride;
1347         patch->strided.u.s.texCoords[0].dwType = WINED3DDECLTYPE_FLOAT4;
1348         /* MAX_STREAMS index points to an unused element in stateblock->streamOffsets which
1349          * always remains set to 0. Windows uses stream 255 here, but this is not visible to the
1350          * application.
1351          */
1352         patch->strided.u.s.texCoords[0].streamNo = MAX_STREAMS;
1353     }
1354
1355     return WINED3D_OK;
1356 }