reactos/dll/directx/wine/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     const BYTE *data = NULL;
 146     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
 147     IWineD3DVertexDeclarationImpl* vertexDeclaration = (IWineD3DVertexDeclarationImpl *)This->stateBlock->vertexDecl;
 148     unsigned int i;
 149     const WINED3DVERTEXELEMENT *element;
 150     DWORD stride;
 151     DWORD numPreloadStreams = This->stateBlock->streamIsUP ? 0 : vertexDeclaration->num_streams;
 152     const 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 (%u of %u)\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, const WineDirect3DVertexStridedData *sd, UINT NumVertexes,
 290         GLenum glPrimType, const void *idxData, short idxSize, ULONG minIndex, ULONG startIdx, ULONG startVertex)
 291 {
 292     unsigned int               textureNo    = 0;
 293     const WORD                *pIdxBufS     = NULL;
 294     const DWORD               *pIdxBufL     = NULL;
 295     ULONG                      vx_index;
 296     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
 297     const UINT *streamOffset = This->stateBlock->streamOffset;
 298     long                      SkipnStrides = startVertex + This->stateBlock->loadBaseVertexIndex;
 299     BOOL                      pixelShader = use_ps(This);
 300     BOOL specular_fog = FALSE;
 301     UINT texture_stages = GL_LIMITS(texture_stages);
 302     const BYTE *texCoords[WINED3DDP_MAXTEXCOORD];
 303     const BYTE *diffuse = NULL, *specular = NULL, *normal = NULL, *position = NULL;
 304     DWORD tex_mask = 0;
 305
 306     TRACE("Using slow vertex array code\n");
 307
 308     /* Variable Initialization */
 309     if (idxSize != 0) {
 310         /* Immediate mode drawing can't make use of indices in a vbo - get the data from the index buffer.
 311          * If the index buffer has no vbo(not supported or other reason), or with user pointer drawing
 312          * idxData will be != NULL
 313          */
 314         if(idxData == NULL) {
 315             idxData = ((IWineD3DIndexBufferImpl *) This->stateBlock->pIndexData)->resource.allocatedMemory;
 316         }
 317
 318         if (idxSize == 2) pIdxBufS = (const WORD *) idxData;
 319         else pIdxBufL = (const DWORD *) idxData;
 320     } else if (idxData) {
 321         ERR("non-NULL idxData with 0 idxSize, this should never happen\n");
 322         return;
 323     }
 324
 325     /* Start drawing in GL */
 326     VTRACE(("glBegin(%x)\n", glPrimType));
 327     glBegin(glPrimType);
 328
 329     if (sd->u.s.position.lpData) position = sd->u.s.position.lpData + streamOffset[sd->u.s.position.streamNo];
 330
 331     if (sd->u.s.normal.lpData) normal = sd->u.s.normal.lpData + streamOffset[sd->u.s.normal.streamNo];
 332     else glNormal3f(0, 0, 0);
 333
 334     if (sd->u.s.diffuse.lpData) diffuse = sd->u.s.diffuse.lpData + streamOffset[sd->u.s.diffuse.streamNo];
 335     else glColor4f(1.0f, 1.0f, 1.0f, 1.0f);
 336     if (This->activeContext->num_untracked_materials && sd->u.s.diffuse.dwType != WINED3DDECLTYPE_D3DCOLOR)
 337         FIXME("Implement diffuse color tracking from %s\n", debug_d3ddecltype(sd->u.s.diffuse.dwType));
 338
 339     if (sd->u.s.specular.lpData)
 340     {
 341         specular = sd->u.s.specular.lpData + streamOffset[sd->u.s.specular.streamNo];
 342
 343         /* special case where the fog density is stored in the specular alpha channel */
 344         if (This->stateBlock->renderState[WINED3DRS_FOGENABLE]
 345                 && (This->stateBlock->renderState[WINED3DRS_FOGVERTEXMODE] == WINED3DFOG_NONE
 346                     || sd->u.s.position.dwType == WINED3DDECLTYPE_FLOAT4)
 347                 && This->stateBlock->renderState[WINED3DRS_FOGTABLEMODE] == WINED3DFOG_NONE)
 348         {
 349             if (GL_SUPPORT(EXT_FOG_COORD))
 350             {
 351                 if (sd->u.s.specular.dwType == WINED3DDECLTYPE_D3DCOLOR) specular_fog = TRUE;
 352                 else FIXME("Implement fog coordinates from %s\n", debug_d3ddecltype(sd->u.s.specular.dwType));
 353             }
 354             else
 355             {
 356                 static BOOL warned;
 357
 358                 if (!warned)
 359                 {
 360                     /* TODO: Use the fog table code from old ddraw */
 361                     FIXME("Implement fog for transformed vertices in software\n");
 362                     warned = TRUE;
 363                 }
 364             }
 365         }
 366     }
 367     else if (GL_SUPPORT(EXT_SECONDARY_COLOR))
 368     {
 369         GL_EXTCALL(glSecondaryColor3fEXT)(0, 0, 0);
 370     }
 371
 372     for (textureNo = 0; textureNo < texture_stages; ++textureNo)
 373     {
 374         int coordIdx = This->stateBlock->textureState[textureNo][WINED3DTSS_TEXCOORDINDEX];
 375         int texture_idx = This->texUnitMap[textureNo];
 376
 377         if (!GL_SUPPORT(ARB_MULTITEXTURE) && textureNo > 0)
 378         {
 379             FIXME("Program using multiple concurrent textures which this opengl implementation doesn't support\n");
 380             continue;
 381         }
 382
 383         if (!pixelShader && !This->stateBlock->textures[textureNo]) continue;
 384
 385         if (texture_idx == -1) continue;
 386
 387         if (coordIdx > 7)
 388         {
 389             TRACE("tex: %d - Skip tex coords, as being system generated\n", textureNo);
 390             continue;
 391         }
 392         else if (coordIdx < 0)
 393         {
 394             FIXME("tex: %d - Coord index %d is less than zero, expect a crash.\n", textureNo, coordIdx);
 395             continue;
 396         }
 397
 398         if(sd->u.s.texCoords[coordIdx].lpData)
 399         {
 400             texCoords[coordIdx] =
 401                     sd->u.s.texCoords[coordIdx].lpData + streamOffset[sd->u.s.texCoords[coordIdx].streamNo];
 402             tex_mask |= (1 << textureNo);
 403         }
 404         else
 405         {
 406             TRACE("tex: %d - Skipping tex coords, as no data supplied\n", textureNo);
 407             if (GL_SUPPORT(ARB_MULTITEXTURE))
 408                 GL_EXTCALL(glMultiTexCoord4fARB(GL_TEXTURE0_ARB + texture_idx, 0, 0, 0, 1));
 409             else
 410                 glTexCoord4f(0, 0, 0, 1);
 411         }
 412     }
 413
 414     /* We shouldn't start this function if any VBO is involved. Should I put a safety check here?
 415      * Guess it's not necessary(we crash then anyway) and would only eat CPU time
 416      */
 417
 418     /* For each primitive */
 419     for (vx_index = 0; vx_index < NumVertexes; ++vx_index) {
 420         UINT texture, tmp_tex_mask;
 421         /* Blending data and Point sizes are not supported by this function. They are not supported by the fixed
 422          * function pipeline at all. A Fixme for them is printed after decoding the vertex declaration
 423          */
 424
 425         /* For indexed data, we need to go a few more strides in */
 426         if (idxData != NULL) {
 427
 428             /* Indexed so work out the number of strides to skip */
 429             if (idxSize == 2) {
 430                 VTRACE(("Idx for vertex %u = %u\n", vx_index, pIdxBufS[startIdx+vx_index]));
 431                 SkipnStrides = pIdxBufS[startIdx + vx_index] + This->stateBlock->loadBaseVertexIndex;
 432             } else {
 433                 VTRACE(("Idx for vertex %u = %u\n", vx_index, pIdxBufL[startIdx+vx_index]));
 434                 SkipnStrides = pIdxBufL[startIdx + vx_index] + This->stateBlock->loadBaseVertexIndex;
 435             }
 436         }
 437
 438         tmp_tex_mask = tex_mask;
 439         for (texture = 0; tmp_tex_mask; tmp_tex_mask >>= 1, ++texture)
 440         {
 441             int coord_idx;
 442             const void *ptr;
 443
 444             if (!(tmp_tex_mask & 1)) continue;
 445
 446             coord_idx = This->stateBlock->textureState[texture][WINED3DTSS_TEXCOORDINDEX];
 447             ptr = texCoords[coord_idx] + (SkipnStrides * sd->u.s.texCoords[coord_idx].dwStride);
 448
 449             if (GL_SUPPORT(ARB_MULTITEXTURE))
 450             {
 451                 int texture_idx = This->texUnitMap[texture];
 452                 multi_texcoord_funcs[sd->u.s.texCoords[coord_idx].dwType](GL_TEXTURE0_ARB + texture_idx, ptr);
 453             }
 454             else
 455             {
 456                 texcoord_funcs[sd->u.s.texCoords[coord_idx].dwType](ptr);
 457             }
 458         }
 459
 460         /* Diffuse -------------------------------- */
 461         if (diffuse) {
 462             const void *ptrToCoords = diffuse + SkipnStrides * sd->u.s.diffuse.dwStride;
 463
 464             diffuse_funcs[sd->u.s.diffuse.dwType](ptrToCoords);
 465             if(This->activeContext->num_untracked_materials) {
 466                 DWORD diffuseColor = ((const DWORD *)ptrToCoords)[0];
 467                 unsigned char i;
 468                 float color[4];
 469
 470                 color[0] = D3DCOLOR_B_R(diffuseColor) / 255.0;
 471                 color[1] = D3DCOLOR_B_G(diffuseColor) / 255.0;
 472                 color[2] = D3DCOLOR_B_B(diffuseColor) / 255.0;
 473                 color[3] = D3DCOLOR_B_A(diffuseColor) / 255.0;
 474
 475                 for(i = 0; i < This->activeContext->num_untracked_materials; i++) {
 476                     glMaterialfv(GL_FRONT_AND_BACK, This->activeContext->untracked_materials[i], color);
 477                 }
 478             }
 479         }
 480
 481         /* Specular ------------------------------- */
 482         if (specular) {
 483             const void *ptrToCoords = specular + SkipnStrides * sd->u.s.specular.dwStride;
 484
 485             specular_funcs[sd->u.s.specular.dwType](ptrToCoords);
 486
 487             if (specular_fog)
 488             {
 489                 DWORD specularColor = *(const DWORD *)ptrToCoords;
 490                 GL_EXTCALL(glFogCoordfEXT(specularColor >> 24));
 491             }
 492         }
 493
 494         /* Normal -------------------------------- */
 495         if (normal != NULL) {
 496             const void *ptrToCoords = normal + SkipnStrides * sd->u.s.normal.dwStride;
 497             normal_funcs[sd->u.s.normal.dwType](ptrToCoords);
 498         }
 499
 500         /* Position -------------------------------- */
 501         if (position) {
 502             const void *ptrToCoords = position + SkipnStrides * sd->u.s.position.dwStride;
 503             position_funcs[sd->u.s.position.dwType](ptrToCoords);
 504         }
 505
 506         /* For non indexed mode, step onto next parts */
 507         if (idxData == NULL) {
 508             ++SkipnStrides;
 509         }
 510     }
 511
 512     glEnd();
 513     checkGLcall("glEnd and previous calls");
 514 }
 515
 516 static inline void send_attribute(IWineD3DDeviceImpl *This, const DWORD type, const UINT index, const void *ptr) {
 517     switch(type) {
 518         case WINED3DDECLTYPE_FLOAT1:
 519             GL_EXTCALL(glVertexAttrib1fvARB(index, (const float *)ptr));
 520             break;
 521         case WINED3DDECLTYPE_FLOAT2:
 522             GL_EXTCALL(glVertexAttrib2fvARB(index, (const float *)ptr));
 523             break;
 524         case WINED3DDECLTYPE_FLOAT3:
 525             GL_EXTCALL(glVertexAttrib3fvARB(index, (const float *)ptr));
 526             break;
 527         case WINED3DDECLTYPE_FLOAT4:
 528             GL_EXTCALL(glVertexAttrib4fvARB(index, (const float *)ptr));
 529             break;
 530
 531         case WINED3DDECLTYPE_UBYTE4:
 532             GL_EXTCALL(glVertexAttrib4ubvARB(index, ptr));
 533             break;
 534         case WINED3DDECLTYPE_UBYTE4N:
 535         case WINED3DDECLTYPE_D3DCOLOR:
 536             GL_EXTCALL(glVertexAttrib4NubvARB(index, ptr));
 537             break;
 538
 539         case WINED3DDECLTYPE_SHORT2:
 540             GL_EXTCALL(glVertexAttrib4svARB(index, (const GLshort *)ptr));
 541             break;
 542         case WINED3DDECLTYPE_SHORT4:
 543             GL_EXTCALL(glVertexAttrib4svARB(index, (const GLshort *)ptr));
 544             break;
 545
 546         case WINED3DDECLTYPE_SHORT2N:
 547         {
 548             GLshort s[4] = {((const GLshort *)ptr)[0], ((const GLshort *)ptr)[1], 0, 1};
 549             GL_EXTCALL(glVertexAttrib4NsvARB(index, s));
 550             break;
 551         }
 552         case WINED3DDECLTYPE_USHORT2N:
 553         {
 554             GLushort s[4] = {((const GLushort *)ptr)[0], ((const GLushort *)ptr)[1], 0, 1};
 555             GL_EXTCALL(glVertexAttrib4NusvARB(index, s));
 556             break;
 557         }
 558         case WINED3DDECLTYPE_SHORT4N:
 559             GL_EXTCALL(glVertexAttrib4NsvARB(index, (const GLshort *)ptr));
 560             break;
 561         case WINED3DDECLTYPE_USHORT4N:
 562             GL_EXTCALL(glVertexAttrib4NusvARB(index, (const GLushort *)ptr));
 563             break;
 564
 565         case WINED3DDECLTYPE_UDEC3:
 566             FIXME("Unsure about WINED3DDECLTYPE_UDEC3\n");
 567             /*glVertexAttrib3usvARB(instancedData[j], (GLushort *) ptr); Does not exist */
 568             break;
 569         case WINED3DDECLTYPE_DEC3N:
 570             FIXME("Unsure about WINED3DDECLTYPE_DEC3N\n");
 571             /*glVertexAttrib3NusvARB(instancedData[j], (GLushort *) ptr); Does not exist */
 572             break;
 573
 574         case WINED3DDECLTYPE_FLOAT16_2:
 575             /* Are those 16 bit floats. C doesn't have a 16 bit float type. I could read the single bits and calculate a 4
 576              * byte float according to the IEEE standard
 577              */
 578             if (GL_SUPPORT(NV_HALF_FLOAT)) {
 579                 GL_EXTCALL(glVertexAttrib2hvNV(index, (const GLhalfNV *)ptr));
 580             } else {
 581                 float x = float_16_to_32(((const unsigned short *)ptr) + 0);
 582                 float y = float_16_to_32(((const unsigned short *)ptr) + 1);
 583                 GL_EXTCALL(glVertexAttrib2fARB(index, x, y));
 584             }
 585             break;
 586         case WINED3DDECLTYPE_FLOAT16_4:
 587             if (GL_SUPPORT(NV_HALF_FLOAT)) {
 588                 GL_EXTCALL(glVertexAttrib4hvNV(index, (const GLhalfNV *)ptr));
 589             } else {
 590                 float x = float_16_to_32(((const unsigned short *)ptr) + 0);
 591                 float y = float_16_to_32(((const unsigned short *)ptr) + 1);
 592                 float z = float_16_to_32(((const unsigned short *)ptr) + 2);
 593                 float w = float_16_to_32(((const unsigned short *)ptr) + 3);
 594                 GL_EXTCALL(glVertexAttrib4fARB(index, x, y, z, w));
 595             }
 596             break;
 597
 598         case WINED3DDECLTYPE_UNUSED:
 599         default:
 600             ERR("Unexpected attribute declaration: %d\n", type);
 601             break;
 602     }
 603 }
 604
 605 static void drawStridedSlowVs(IWineD3DDevice *iface, const WineDirect3DVertexStridedData *sd, UINT numberOfVertices,
 606         GLenum glPrimitiveType, const void *idxData, short idxSize, ULONG minIndex, ULONG startIdx, ULONG startVertex)
 607 {
 608     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *) iface;
 609     long                      SkipnStrides = startVertex + This->stateBlock->loadBaseVertexIndex;
 610     const WORD                *pIdxBufS     = NULL;
 611     const DWORD               *pIdxBufL     = NULL;
 612     ULONG                      vx_index;
 613     int i;
 614     IWineD3DStateBlockImpl *stateblock = This->stateBlock;
 615     const BYTE *ptr;
 616
 617     if (idxSize != 0) {
 618         /* Immediate mode drawing can't make use of indices in a vbo - get the data from the index buffer.
 619          * If the index buffer has no vbo(not supported or other reason), or with user pointer drawing
 620          * idxData will be != NULL
 621          */
 622         if(idxData == NULL) {
 623             idxData = ((IWineD3DIndexBufferImpl *) stateblock->pIndexData)->resource.allocatedMemory;
 624         }
 625
 626         if (idxSize == 2) pIdxBufS = (const WORD *) idxData;
 627         else pIdxBufL = (const DWORD *) idxData;
 628     } else if (idxData) {
 629         ERR("non-NULL idxData with 0 idxSize, this should never happen\n");
 630         return;
 631     }
 632
 633     /* Start drawing in GL */
 634     VTRACE(("glBegin(%x)\n", glPrimitiveType));
 635     glBegin(glPrimitiveType);
 636
 637     for (vx_index = 0; vx_index < numberOfVertices; ++vx_index) {
 638         if (idxData != NULL) {
 639
 640             /* Indexed so work out the number of strides to skip */
 641             if (idxSize == 2) {
 642                 VTRACE(("Idx for vertex %d = %d\n", vx_index, pIdxBufS[startIdx+vx_index]));
 643                 SkipnStrides = pIdxBufS[startIdx + vx_index] + stateblock->loadBaseVertexIndex;
 644             } else {
 645                 VTRACE(("Idx for vertex %d = %d\n", vx_index, pIdxBufL[startIdx+vx_index]));
 646                 SkipnStrides = pIdxBufL[startIdx + vx_index] + stateblock->loadBaseVertexIndex;
 647             }
 648         }
 649
 650         for(i = MAX_ATTRIBS - 1; i >= 0; i--) {
 651             if(!sd->u.input[i].lpData) continue;
 652
 653             ptr = sd->u.input[i].lpData +
 654                   sd->u.input[i].dwStride * SkipnStrides +
 655                   stateblock->streamOffset[sd->u.input[i].streamNo];
 656
 657             send_attribute(This, sd->u.input[i].dwType, i, ptr);
 658         }
 659         SkipnStrides++;
 660     }
 661
 662     glEnd();
 663 }
 664
 665 static inline void drawStridedInstanced(IWineD3DDevice *iface, const WineDirect3DVertexStridedData *sd,
 666         UINT numberOfVertices, GLenum glPrimitiveType, const void *idxData, short idxSize, ULONG minIndex,
 667         ULONG startIdx, ULONG startVertex)
 668 {
 669     UINT numInstances = 0, i;
 670     int numInstancedAttribs = 0, j;
 671     UINT instancedData[sizeof(sd->u.input) / sizeof(sd->u.input[0]) /* 16 */];
 672     IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *) iface;
 673     IWineD3DStateBlockImpl *stateblock = This->stateBlock;
 674
 675     if (idxSize == 0) {
 676         /* This is a nasty thing. MSDN says no hardware supports that and apps have to use software vertex processing.
 677          * We don't support this for now
 678          *
 679          * Shouldn't be too hard to support with opengl, in theory just call glDrawArrays instead of drawElements.
 680          * But the StreamSourceFreq value has a different meaning in that situation.
 681          */
 682         FIXME("Non-indexed instanced drawing is not supported\n");
 683         return;
 684     }
 685
 686     TRACE("(%p) : glElements(%x, %d, %d, ...)\n", This, glPrimitiveType, numberOfVertices, minIndex);
 687     idxData = idxData == (void *)-1 ? NULL : idxData;
 688
 689     /* First, figure out how many instances we have to draw */
 690     for(i = 0; i < MAX_STREAMS; i++) {
 691         /* Look at the streams and take the first one which matches */
 692         if(((stateblock->streamFlags[i] & WINED3DSTREAMSOURCE_INSTANCEDATA) || (stateblock->streamFlags[i] & WINED3DSTREAMSOURCE_INDEXEDDATA)) && stateblock->streamSource[i]) {
 693             /* D3D9 could set streamFreq 0 with (INSTANCEDATA or INDEXEDDATA) and then it is handled as 1. See d3d9/tests/visual.c-> stream_test() */
 694             if(stateblock->streamFreq[i] == 0){
 695                 numInstances = 1;
 696             } else {
 697                 numInstances = stateblock->streamFreq[i]; /* use the specified number of instances from the first matched stream. See d3d9/tests/visual.c-> stream_test() */
 698             }
 699             break; /* break, because only the first suitable value is interesting */
 700         }
 701     }
 702
 703     for(i = 0; i < sizeof(sd->u.input) / sizeof(sd->u.input[0]); i++) {
 704         if(stateblock->streamFlags[sd->u.input[i].streamNo] & WINED3DSTREAMSOURCE_INSTANCEDATA) {
 705             instancedData[numInstancedAttribs] = i;
 706             numInstancedAttribs++;
 707         }
 708     }
 709
 710     /* now draw numInstances instances :-) */
 711     for(i = 0; i < numInstances; i++) {
 712         /* Specify the instanced attributes using immediate mode calls */
 713         for(j = 0; j < numInstancedAttribs; j++) {
 714             const BYTE *ptr = sd->u.input[instancedData[j]].lpData +
 715                         sd->u.input[instancedData[j]].dwStride * i +
 716                         stateblock->streamOffset[sd->u.input[instancedData[j]].streamNo];
 717             if(sd->u.input[instancedData[j]].VBO) {
 718                 IWineD3DVertexBufferImpl *vb = (IWineD3DVertexBufferImpl *) stateblock->streamSource[sd->u.input[instancedData[j]].streamNo];
 719                 ptr += (long) vb->resource.allocatedMemory;
 720             }
 721
 722             send_attribute(This, sd->u.input[instancedData[j]].dwType, instancedData[j], ptr);
 723         }
 724
 725         glDrawElements(glPrimitiveType, numberOfVertices, idxSize == 2 ? GL_UNSIGNED_SHORT : GL_UNSIGNED_INT,
 726                     (const char *)idxData+(idxSize * startIdx));
 727         checkGLcall("glDrawElements");
 728     }
 729 }
 730
 731 static inline void remove_vbos(IWineD3DDeviceImpl *This, WineDirect3DVertexStridedData *s) {
 732     unsigned char i;
 733     IWineD3DVertexBufferImpl *vb;
 734
 735     if(s->u.s.position.VBO) {
 736         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.position.streamNo];
 737         s->u.s.position.VBO = 0;
 738         s->u.s.position.lpData = (BYTE *) ((unsigned long) s->u.s.position.lpData + (unsigned long) vb->resource.allocatedMemory);
 739     }
 740     if(s->u.s.blendWeights.VBO) {
 741         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.blendWeights.streamNo];
 742         s->u.s.blendWeights.VBO = 0;
 743         s->u.s.blendWeights.lpData = (BYTE *) ((unsigned long) s->u.s.blendWeights.lpData + (unsigned long) vb->resource.allocatedMemory);
 744     }
 745     if(s->u.s.blendMatrixIndices.VBO) {
 746         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.blendMatrixIndices.streamNo];
 747         s->u.s.blendMatrixIndices.VBO = 0;
 748         s->u.s.blendMatrixIndices.lpData = (BYTE *) ((unsigned long) s->u.s.blendMatrixIndices.lpData + (unsigned long) vb->resource.allocatedMemory);
 749     }
 750     if(s->u.s.normal.VBO) {
 751         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.normal.streamNo];
 752         s->u.s.normal.VBO = 0;
 753         s->u.s.normal.lpData = (BYTE *) ((unsigned long) s->u.s.normal.lpData + (unsigned long) vb->resource.allocatedMemory);
 754     }
 755     if(s->u.s.pSize.VBO) {
 756         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.pSize.streamNo];
 757         s->u.s.pSize.VBO = 0;
 758         s->u.s.pSize.lpData = (BYTE *) ((unsigned long) s->u.s.pSize.lpData + (unsigned long) vb->resource.allocatedMemory);
 759     }
 760     if(s->u.s.diffuse.VBO) {
 761         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.diffuse.streamNo];
 762         s->u.s.diffuse.VBO = 0;
 763         s->u.s.diffuse.lpData = (BYTE *) ((unsigned long) s->u.s.diffuse.lpData + (unsigned long) vb->resource.allocatedMemory);
 764     }
 765     if(s->u.s.specular.VBO) {
 766         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.specular.streamNo];
 767         s->u.s.specular.VBO = 0;
 768         s->u.s.specular.lpData = (BYTE *) ((unsigned long) s->u.s.specular.lpData + (unsigned long) vb->resource.allocatedMemory);
 769     }
 770     for(i = 0; i < WINED3DDP_MAXTEXCOORD; i++) {
 771         if(s->u.s.texCoords[i].VBO) {
 772             vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.texCoords[i].streamNo];
 773             s->u.s.texCoords[i].VBO = 0;
 774             s->u.s.texCoords[i].lpData = (BYTE *) ((unsigned long) s->u.s.texCoords[i].lpData + (unsigned long) vb->resource.allocatedMemory);
 775         }
 776     }
 777     if(s->u.s.position2.VBO) {
 778         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.position2.streamNo];
 779         s->u.s.position2.VBO = 0;
 780         s->u.s.position2.lpData = (BYTE *) ((unsigned long) s->u.s.position2.lpData + (unsigned long) vb->resource.allocatedMemory);
 781     }
 782     if(s->u.s.normal2.VBO) {
 783         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.normal2.streamNo];
 784         s->u.s.normal2.VBO = 0;
 785         s->u.s.normal2.lpData = (BYTE *) ((unsigned long) s->u.s.normal2.lpData + (unsigned long) vb->resource.allocatedMemory);
 786     }
 787     if(s->u.s.tangent.VBO) {
 788         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.tangent.streamNo];
 789         s->u.s.tangent.VBO = 0;
 790         s->u.s.tangent.lpData = (BYTE *) ((unsigned long) s->u.s.tangent.lpData + (unsigned long) vb->resource.allocatedMemory);
 791     }
 792     if(s->u.s.binormal.VBO) {
 793         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.binormal.streamNo];
 794         s->u.s.binormal.VBO = 0;
 795         s->u.s.binormal.lpData = (BYTE *) ((unsigned long) s->u.s.binormal.lpData + (unsigned long) vb->resource.allocatedMemory);
 796     }
 797     if(s->u.s.tessFactor.VBO) {
 798         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.tessFactor.streamNo];
 799         s->u.s.tessFactor.VBO = 0;
 800         s->u.s.tessFactor.lpData = (BYTE *) ((unsigned long) s->u.s.tessFactor.lpData + (unsigned long) vb->resource.allocatedMemory);
 801     }
 802     if(s->u.s.fog.VBO) {
 803         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.fog.streamNo];
 804         s->u.s.fog.VBO = 0;
 805         s->u.s.fog.lpData = (BYTE *) ((unsigned long) s->u.s.fog.lpData + (unsigned long) vb->resource.allocatedMemory);
 806     }
 807     if(s->u.s.depth.VBO) {
 808         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.depth.streamNo];
 809         s->u.s.depth.VBO = 0;
 810         s->u.s.depth.lpData = (BYTE *) ((unsigned long) s->u.s.depth.lpData + (unsigned long) vb->resource.allocatedMemory);
 811     }
 812     if(s->u.s.sample.VBO) {
 813         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[s->u.s.sample.streamNo];
 814         s->u.s.sample.VBO = 0;
 815         s->u.s.sample.lpData = (BYTE *) ((unsigned long) s->u.s.sample.lpData + (unsigned long) vb->resource.allocatedMemory);
 816     }
 817 }
 818
 819 /* Routine common to the draw primitive and draw indexed primitive routines */
 820 void drawPrimitive(IWineD3DDevice *iface,
 821                    int PrimitiveType,
 822                    long NumPrimitives,
 823                    /* for Indexed: */
 824                    long  StartVertexIndex,
 825                    UINT  numberOfVertices,
 826                    long  StartIdx,
 827                    short idxSize,
 828                    const void *idxData,
 829                    int   minIndex) {
 830
 831     IWineD3DDeviceImpl           *This = (IWineD3DDeviceImpl *)iface;
 832     IWineD3DSurfaceImpl          *target;
 833     unsigned int i;
 834
 835     if (NumPrimitives == 0) return;
 836
 837     /* Invalidate the back buffer memory so LockRect will read it the next time */
 838     for(i = 0; i < GL_LIMITS(buffers); i++) {
 839         target = (IWineD3DSurfaceImpl *) This->render_targets[i];
 840         if (target) {
 841             IWineD3DSurface_LoadLocation((IWineD3DSurface *) target, SFLAG_INDRAWABLE, NULL);
 842             IWineD3DSurface_ModifyLocation((IWineD3DSurface *) target, SFLAG_INDRAWABLE, TRUE);
 843         }
 844     }
 845
 846     /* Signals other modules that a drawing is in progress and the stateblock finalized */
 847     This->isInDraw = TRUE;
 848
 849     ActivateContext(This, This->render_targets[0], CTXUSAGE_DRAWPRIM);
 850
 851     if (This->stencilBufferTarget) {
 852         /* Note that this depends on the ActivateContext call above to set
 853          * This->render_offscreen properly */
 854         DWORD location = This->render_offscreen ? SFLAG_DS_OFFSCREEN : SFLAG_DS_ONSCREEN;
 855         surface_load_ds_location(This->stencilBufferTarget, location);
 856         surface_modify_ds_location(This->stencilBufferTarget, location);
 857     }
 858
 859     /* Ok, we will be updating the screen from here onwards so grab the lock */
 860     ENTER_GL();
 861     {
 862         GLenum glPrimType;
 863         BOOL emulation = FALSE;
 864         const WineDirect3DVertexStridedData *strided = &This->strided_streams;
 865         WineDirect3DVertexStridedData stridedlcl;
 866         /* Ok, Work out which primitive is requested and how many vertexes that
 867            will be                                                              */
 868         UINT calculatedNumberOfindices = primitiveToGl(PrimitiveType, NumPrimitives, &glPrimType);
 869         if (numberOfVertices == 0 )
 870             numberOfVertices = calculatedNumberOfindices;
 871
 872         if(!use_vs(This)) {
 873             if(!This->strided_streams.u.s.position_transformed && This->activeContext->num_untracked_materials &&
 874                 This->stateBlock->renderState[WINED3DRS_LIGHTING]) {
 875                 static BOOL warned;
 876                 if (!warned) {
 877                     FIXME("Using software emulation because not all material properties could be tracked\n");
 878                     warned = TRUE;
 879                 } else {
 880                     TRACE("Using software emulation because not all material properties could be tracked\n");
 881                 }
 882                 emulation = TRUE;
 883             }
 884             else if(This->activeContext->fog_coord && This->stateBlock->renderState[WINED3DRS_FOGENABLE]) {
 885                 /* Either write a pipeline replacement shader or convert the specular alpha from unsigned byte
 886                  * to a float in the vertex buffer
 887                  */
 888                 static BOOL warned;
 889                 if (!warned) {
 890                     FIXME("Using software emulation because manual fog coordinates are provided\n");
 891                     warned = TRUE;
 892                 } else {
 893                     TRACE("Using software emulation because manual fog coordinates are provided\n");
 894                 }
 895                 emulation = TRUE;
 896             }
 897
 898             if(emulation) {
 899                 strided = &stridedlcl;
 900                 memcpy(&stridedlcl, &This->strided_streams, sizeof(stridedlcl));
 901                 remove_vbos(This, &stridedlcl);
 902             }
 903         }
 904
 905         if (This->useDrawStridedSlow || emulation) {
 906             /* Immediate mode drawing */
 907             if(use_vs(This)) {
 908                 static BOOL warned;
 909                 if (!warned) {
 910                     FIXME("Using immediate mode with vertex shaders for half float emulation\n");
 911                     warned = TRUE;
 912                 } else {
 913                     TRACE("Using immediate mode with vertex shaders for half float emulation\n");
 914                 }
 915                 drawStridedSlowVs(iface, strided, calculatedNumberOfindices, glPrimType,
 916                                   idxData, idxSize, minIndex, StartIdx, StartVertexIndex);
 917             } else {
 918                 drawStridedSlow(iface, strided, calculatedNumberOfindices,
 919                                 glPrimType, idxData, idxSize, minIndex, StartIdx, StartVertexIndex);
 920             }
 921         } else if(This->instancedDraw) {
 922             /* Instancing emulation with mixing immediate mode and arrays */
 923             drawStridedInstanced(iface, &This->strided_streams, calculatedNumberOfindices, glPrimType,
 924                             idxData, idxSize, minIndex, StartIdx, StartVertexIndex);
 925         } else {
 926             drawStridedFast(iface, calculatedNumberOfindices, glPrimType,
 927                             idxData, idxSize, minIndex, StartIdx, StartVertexIndex);
 928         }
 929     }
 930
 931     /* Finished updating the screen, restore lock */
 932     LEAVE_GL();
 933     TRACE("Done all gl drawing\n");
 934
 935     /* Diagnostics */
 936 #ifdef SHOW_FRAME_MAKEUP
 937     {
 938         static long int primCounter = 0;
 939         /* NOTE: set primCounter to the value reported by drawprim
 940            before you want to to write frame makeup to /tmp */
 941         if (primCounter >= 0) {
 942             WINED3DLOCKED_RECT r;
 943             char buffer[80];
 944             IWineD3DSurface_LockRect(This->render_targets[0], &r, NULL, WINED3DLOCK_READONLY);
 945             sprintf(buffer, "/tmp/backbuffer_%ld.tga", primCounter);
 946             TRACE("Saving screenshot %s\n", buffer);
 947             IWineD3DSurface_SaveSnapshot(This->render_targets[0], buffer);
 948             IWineD3DSurface_UnlockRect(This->render_targets[0]);
 949
 950 #ifdef SHOW_TEXTURE_MAKEUP
 951            {
 952             IWineD3DSurface *pSur;
 953             int textureNo;
 954             for (textureNo = 0; textureNo < MAX_COMBINED_SAMPLERS; ++textureNo) {
 955                 if (This->stateBlock->textures[textureNo] != NULL) {
 956                     sprintf(buffer, "/tmp/texture_%p_%ld_%d.tga", This->stateBlock->textures[textureNo], primCounter, textureNo);
 957                     TRACE("Saving texture %s\n", buffer);
 958                     if (IWineD3DBaseTexture_GetType(This->stateBlock->textures[textureNo]) == WINED3DRTYPE_TEXTURE) {
 959                             IWineD3DTexture_GetSurfaceLevel((IWineD3DTexture *)This->stateBlock->textures[textureNo], 0, &pSur);
 960                             IWineD3DSurface_SaveSnapshot(pSur, buffer);
 961                             IWineD3DSurface_Release(pSur);
 962                     } else  {
 963                         FIXME("base Texture isn't of type texture %d\n", IWineD3DBaseTexture_GetType(This->stateBlock->textures[textureNo]));
 964                     }
 965                 }
 966             }
 967            }
 968 #endif
 969         }
 970         TRACE("drawprim #%ld\n", primCounter);
 971         ++primCounter;
 972     }
 973 #endif
 974
 975     /* Control goes back to the device, stateblock values may change again */
 976     This->isInDraw = FALSE;
 977 }
 978
 979 static void normalize_normal(float *n) {
 980     float length = n[0] * n[0] + n[1] * n[1] + n[2] * n[2];
 981     if(length == 0.0) return;
 982     length = sqrt(length);
 983     n[0] = n[0] / length;
 984     n[1] = n[1] / length;
 985     n[2] = n[2] / length;
 986 }
 987
 988 /* Tesselates a high order rectangular patch into single triangles using gl evaluators
 989  *
 990  * The problem is that OpenGL does not offer a direct way to return the tesselated primitives,
 991  * and they can't be sent off for rendering directly either. Tesselating is slow, so we want
 992  * to cache the patches in a vertex buffer. But more importantly, gl can't bind generated
 993  * attributes to numbered shader attributes, so we have to store them and rebind them as needed
 994  * in drawprim.
 995  *
 996  * To read back, the opengl feedback mode is used. This creates a problem because we want
 997  * untransformed, unlit vertices, but feedback runs everything through transform and lighting.
 998  * Thus disable lighting and set identity matrices to get unmodified colors and positions.
 999  * To overcome clipping find the biggest x, y and z values of the vertices in the patch and scale
1000  * them to [-1.0;+1.0] and set the viewport up to scale them back.
1001  *
1002  * Normals are more tricky: Draw white vertices with 3 directional lights, and calculate the
1003  * resulting colors back to the normals.
1004  *
1005  * NOTE: This function activates a context for blitting, modifies matrices & viewport, but
1006  * does not restore it because normally a draw follows immediately afterwards. The caller is
1007  * responsible of taking care that either the gl states are restored, or the context activated
1008  * for drawing to reset the lastWasBlit flag.
1009  */
1010 HRESULT tesselate_rectpatch(IWineD3DDeviceImpl *This,
1011                             struct WineD3DRectPatch *patch) {
1012     unsigned int i, j, num_quads, out_vertex_size, buffer_size, d3d_out_vertex_size;
1013     float max_x = 0.0, max_y = 0.0, max_z = 0.0, neg_z = 0.0;
1014     WineDirect3DVertexStridedData strided;
1015     const BYTE *data;
1016     const WINED3DRECTPATCH_INFO *info = &patch->RectPatchInfo;
1017     DWORD vtxStride;
1018     GLenum feedback_type;
1019     GLfloat *feedbuffer;
1020
1021     /* First, locate the position data. This is provided in a vertex buffer in the stateblock.
1022      * Beware of vbos
1023      */
1024     memset(&strided, 0, sizeof(strided));
1025     primitiveDeclarationConvertToStridedData((IWineD3DDevice *) This, FALSE, &strided, NULL);
1026     if(strided.u.s.position.VBO) {
1027         IWineD3DVertexBufferImpl *vb;
1028         vb = (IWineD3DVertexBufferImpl *) This->stateBlock->streamSource[strided.u.s.position.streamNo];
1029         strided.u.s.position.lpData = (BYTE *) ((unsigned long) strided.u.s.position.lpData +
1030                                                 (unsigned long) vb->resource.allocatedMemory);
1031     }
1032     vtxStride = strided.u.s.position.dwStride;
1033     data = strided.u.s.position.lpData +
1034            vtxStride * info->Stride * info->StartVertexOffsetHeight +
1035            vtxStride * info->StartVertexOffsetWidth;
1036
1037     /* Not entirely sure about what happens with transformed vertices */
1038     if(strided.u.s.position_transformed) {
1039         FIXME("Transformed position in rectpatch generation\n");
1040     }
1041     if(vtxStride % sizeof(GLfloat)) {
1042         /* glMap2f reads vertex sizes in GLfloats, the d3d stride is in bytes.
1043          * I don't see how the stride could not be a multiple of 4, but make sure
1044          * to check it
1045          */
1046         ERR("Vertex stride is not a multiple of sizeof(GLfloat)\n");
1047     }
1048     if(info->Basis != WINED3DBASIS_BEZIER) {
1049         FIXME("Basis is %s, how to handle this?\n", debug_d3dbasis(info->Basis));
1050     }
1051     if(info->Degree != WINED3DDEGREE_CUBIC) {
1052         FIXME("Degree is %s, how to handle this?\n", debug_d3ddegree(info->Degree));
1053     }
1054
1055     /* First, get the boundary cube of the input data */
1056     for(j = 0; j < info->Height; j++) {
1057         for(i = 0; i < info->Width; i++) {
1058             const float *v = (const float *)(data + vtxStride * i + vtxStride * info->Stride * j);
1059             if(fabs(v[0]) > max_x) max_x = fabs(v[0]);
1060             if(fabs(v[1]) > max_y) max_y = fabs(v[1]);
1061             if(fabs(v[2]) > max_z) max_z = fabs(v[2]);
1062             if(v[2] < neg_z) neg_z = v[2];
1063         }
1064     }
1065
1066     /* This needs some improvements in the vertex decl code */
1067     FIXME("Cannot find data to generate. Only generating position and normals\n");
1068     patch->has_normals = TRUE;
1069     patch->has_texcoords = FALSE;
1070
1071     /* Simply activate the context for blitting. This disables all the things we don't want and
1072      * takes care of dirtifying. Dirtifying is preferred over pushing / popping, since drawing the
1073      * patch (as opposed to normal draws) will most likely need different changes anyway
1074      */
1075     ActivateContext(This, This->lastActiveRenderTarget, CTXUSAGE_BLIT);
1076     ENTER_GL();
1077
1078     glMatrixMode(GL_PROJECTION);
1079     checkGLcall("glMatrixMode(GL_PROJECTION)");
1080     glLoadIdentity();
1081     checkGLcall("glLoadIndentity()");
1082     glScalef(1 / (max_x) , 1 / (max_y), max_z == 0 ? 1 : 1 / ( 2 * max_z));
1083     glTranslatef(0, 0, 0.5);
1084     checkGLcall("glScalef");
1085     glViewport(-max_x, -max_y, 2 * (max_x), 2 * (max_y));
1086     checkGLcall("glViewport");
1087
1088     /* Some states to take care of. If we're in wireframe opengl will produce lines, and confuse
1089      * our feedback buffer parser
1090      */
1091     glPolygonMode(GL_FRONT_AND_BACK, GL_FILL);
1092     checkGLcall("glPolygonMode(GL_FRONT_AND_BACK, GL_FILL)");
1093     IWineD3DDeviceImpl_MarkStateDirty(This, STATE_RENDER(WINED3DRS_FILLMODE));
1094     if(patch->has_normals) {
1095         const GLfloat black[4] = {0, 0, 0, 0};
1096         const GLfloat red[4]   = {1, 0, 0, 0};
1097         const GLfloat green[4] = {0, 1, 0, 0};
1098         const GLfloat blue[4]  = {0, 0, 1, 0};
1099         const GLfloat white[4] = {1, 1, 1, 1};
1100         glEnable(GL_LIGHTING);
1101         checkGLcall("glEnable(GL_LIGHTING)");
1102         glLightModelfv(GL_LIGHT_MODEL_AMBIENT, black);
1103         checkGLcall("glLightModel for MODEL_AMBIENT");
1104         IWineD3DDeviceImpl_MarkStateDirty(This, STATE_RENDER(WINED3DRS_AMBIENT));
1105
1106         for(i = 3; i < GL_LIMITS(lights); i++) {
1107             glDisable(GL_LIGHT0 + i);
1108             checkGLcall("glDisable(GL_LIGHT0 + i)");
1109             IWineD3DDeviceImpl_MarkStateDirty(This, STATE_ACTIVELIGHT(i));
1110         }
1111
1112         IWineD3DDeviceImpl_MarkStateDirty(This, STATE_ACTIVELIGHT(0));
1113         glLightfv(GL_LIGHT0, GL_DIFFUSE, red);
1114         glLightfv(GL_LIGHT0, GL_SPECULAR, black);
1115         glLightfv(GL_LIGHT0, GL_AMBIENT, black);
1116         glLightfv(GL_LIGHT0, GL_POSITION, red);
1117         glEnable(GL_LIGHT0);
1118         checkGLcall("Setting up light 1\n");
1119         IWineD3DDeviceImpl_MarkStateDirty(This, STATE_ACTIVELIGHT(1));
1120         glLightfv(GL_LIGHT1, GL_DIFFUSE, green);
1121         glLightfv(GL_LIGHT1, GL_SPECULAR, black);
1122         glLightfv(GL_LIGHT1, GL_AMBIENT, black);
1123         glLightfv(GL_LIGHT1, GL_POSITION, green);
1124         glEnable(GL_LIGHT1);
1125         checkGLcall("Setting up light 2\n");
1126         IWineD3DDeviceImpl_MarkStateDirty(This, STATE_ACTIVELIGHT(2));
1127         glLightfv(GL_LIGHT2, GL_DIFFUSE, blue);
1128         glLightfv(GL_LIGHT2, GL_SPECULAR, black);
1129         glLightfv(GL_LIGHT2, GL_AMBIENT, black);
1130         glLightfv(GL_LIGHT2, GL_POSITION, blue);
1131         glEnable(GL_LIGHT2);
1132         checkGLcall("Setting up light 3\n");
1133
1134         IWineD3DDeviceImpl_MarkStateDirty(This, STATE_MATERIAL);
1135         IWineD3DDeviceImpl_MarkStateDirty(This, STATE_RENDER(WINED3DRS_COLORVERTEX));
1136         glDisable(GL_COLOR_MATERIAL);
1137         glMaterialfv(GL_FRONT_AND_BACK, GL_EMISSION, black);
1138         glMaterialfv(GL_FRONT_AND_BACK, GL_SPECULAR, black);
1139         glMaterialfv(GL_FRONT_AND_BACK, GL_DIFFUSE, white);
1140         checkGLcall("Setting up materials\n");
1141     }
1142
1143     /* Enable the needed maps.
1144      * GL_MAP2_VERTEX_3 is needed for positional data.
1145      * GL_AUTO_NORMAL to generate normals from the position. Do not use GL_MAP2_NORMAL.
1146      * GL_MAP2_TEXTURE_COORD_4 for texture coords
1147      */
1148     num_quads = ceilf(patch->numSegs[0]) * ceilf(patch->numSegs[1]);
1149     out_vertex_size = 3 /* position */;
1150     d3d_out_vertex_size = 3;
1151     glEnable(GL_MAP2_VERTEX_3);
1152     if(patch->has_normals && patch->has_texcoords) {
1153         FIXME("Texcoords not handled yet\n");
1154         feedback_type = GL_3D_COLOR_TEXTURE;
1155         out_vertex_size += 8;
1156         d3d_out_vertex_size += 7;
1157         glEnable(GL_AUTO_NORMAL);
1158         glEnable(GL_MAP2_TEXTURE_COORD_4);
1159     } else if(patch->has_texcoords) {
1160         FIXME("Texcoords not handled yet\n");
1161         feedback_type = GL_3D_COLOR_TEXTURE;
1162         out_vertex_size += 7;
1163         d3d_out_vertex_size += 4;
1164         glEnable(GL_MAP2_TEXTURE_COORD_4);
1165     } else if(patch->has_normals) {
1166         feedback_type = GL_3D_COLOR;
1167         out_vertex_size += 4;
1168         d3d_out_vertex_size += 3;
1169         glEnable(GL_AUTO_NORMAL);
1170     } else {
1171         feedback_type = GL_3D;
1172     }
1173     checkGLcall("glEnable vertex attrib generation");
1174
1175     buffer_size = num_quads * out_vertex_size * 2 /* triangle list */ * 3 /* verts per tri */
1176                    + 4 * num_quads /* 2 triangle markers per quad + num verts in tri */;
1177     feedbuffer = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, buffer_size * sizeof(float) * 8);
1178
1179     glMap2f(GL_MAP2_VERTEX_3,
1180             0, 1, vtxStride / sizeof(float), info->Width,
1181             0, 1, info->Stride * vtxStride / sizeof(float), info->Height,
1182             (const GLfloat *)data);
1183     checkGLcall("glMap2f");
1184     if(patch->has_texcoords) {
1185         glMap2f(GL_MAP2_TEXTURE_COORD_4,
1186                 0, 1, vtxStride / sizeof(float), info->Width,
1187                 0, 1, info->Stride * vtxStride / sizeof(float), info->Height,
1188                 (const GLfloat *)data);
1189         checkGLcall("glMap2f");
1190     }
1191     glMapGrid2f(ceilf(patch->numSegs[0]), 0.0, 1.0, ceilf(patch->numSegs[1]), 0.0, 1.0);
1192     checkGLcall("glMapGrid2f");
1193
1194     glFeedbackBuffer(buffer_size * 2, feedback_type, feedbuffer);
1195     checkGLcall("glFeedbackBuffer");
1196     glRenderMode(GL_FEEDBACK);
1197
1198     glEvalMesh2(GL_FILL, 0, ceilf(patch->numSegs[0]), 0, ceilf(patch->numSegs[1]));
1199     checkGLcall("glEvalMesh2\n");
1200
1201     i = glRenderMode(GL_RENDER);
1202     if(i == -1) {
1203         LEAVE_GL();
1204         ERR("Feedback failed. Expected %d elements back\n", buffer_size);
1205         Sleep(10000);
1206         HeapFree(GetProcessHeap(), 0, feedbuffer);
1207         return WINED3DERR_DRIVERINTERNALERROR;
1208     } else if(i != buffer_size) {
1209         LEAVE_GL();
1210         ERR("Unexpected amount of elements returned. Expected %d, got %d\n", buffer_size, i);
1211         Sleep(10000);
1212         HeapFree(GetProcessHeap(), 0, feedbuffer);
1213         return WINED3DERR_DRIVERINTERNALERROR;
1214     } else {
1215         TRACE("Got %d elements as expected\n", i);
1216     }
1217
1218     HeapFree(GetProcessHeap(), 0, patch->mem);
1219     patch->mem = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, num_quads * 6 * d3d_out_vertex_size * sizeof(float) * 8);
1220     i = 0;
1221     for(j = 0; j < buffer_size; j += (3 /* num verts */ * out_vertex_size + 2 /* tri marker */)) {
1222         if(feedbuffer[j] != GL_POLYGON_TOKEN) {
1223             ERR("Unexpected token: %f\n", feedbuffer[j]);
1224             continue;
1225         }
1226         if(feedbuffer[j + 1] != 3) {
1227             ERR("Unexpected polygon: %f corners\n", feedbuffer[j + 1]);
1228             continue;
1229         }
1230         /* Somehow there are different ideas about back / front facing, so fix up the
1231          * vertex order
1232          */
1233         patch->mem[i + 0] =  feedbuffer[j + out_vertex_size * 2 + 2]; /* x, triangle 2 */
1234         patch->mem[i + 1] =  feedbuffer[j + out_vertex_size * 2 + 3]; /* y, triangle 2 */
1235         patch->mem[i + 2] = (feedbuffer[j + out_vertex_size * 2 + 4] - 0.5) * 4 * max_z; /* z, triangle 3 */
1236         if(patch->has_normals) {
1237             patch->mem[i + 3] = feedbuffer[j + out_vertex_size * 2 + 5];
1238             patch->mem[i + 4] = feedbuffer[j + out_vertex_size * 2 + 6];
1239             patch->mem[i + 5] = feedbuffer[j + out_vertex_size * 2 + 7];
1240         }
1241         i += d3d_out_vertex_size;
1242
1243         patch->mem[i + 0] =  feedbuffer[j + out_vertex_size * 1 + 2]; /* x, triangle 2 */
1244         patch->mem[i + 1] =  feedbuffer[j + out_vertex_size * 1 + 3]; /* y, triangle 2 */
1245         patch->mem[i + 2] = (feedbuffer[j + out_vertex_size * 1 + 4] - 0.5) * 4 * max_z; /* z, triangle 2 */
1246         if(patch->has_normals) {
1247             patch->mem[i + 3] = feedbuffer[j + out_vertex_size * 1 + 5];
1248             patch->mem[i + 4] = feedbuffer[j + out_vertex_size * 1 + 6];
1249             patch->mem[i + 5] = feedbuffer[j + out_vertex_size * 1 + 7];
1250         }
1251         i += d3d_out_vertex_size;
1252
1253         patch->mem[i + 0] =  feedbuffer[j + out_vertex_size * 0 + 2]; /* x, triangle 1 */
1254         patch->mem[i + 1] =  feedbuffer[j + out_vertex_size * 0 + 3]; /* y, triangle 1 */
1255         patch->mem[i + 2] = (feedbuffer[j + out_vertex_size * 0 + 4] - 0.5) * 4 * max_z; /* z, triangle 1 */
1256         if(patch->has_normals) {
1257             patch->mem[i + 3] = feedbuffer[j + out_vertex_size * 0 + 5];
1258             patch->mem[i + 4] = feedbuffer[j + out_vertex_size * 0 + 6];
1259             patch->mem[i + 5] = feedbuffer[j + out_vertex_size * 0 + 7];
1260         }
1261         i += d3d_out_vertex_size;
1262     }
1263
1264     if(patch->has_normals) {
1265         /* Now do the same with reverse light directions */
1266         const GLfloat x[4] = {-1,  0,  0, 0};
1267         const GLfloat y[4] = { 0, -1,  0, 0};
1268         const GLfloat z[4] = { 0,  0, -1, 0};
1269         glLightfv(GL_LIGHT0, GL_POSITION, x);
1270         glLightfv(GL_LIGHT1, GL_POSITION, y);
1271         glLightfv(GL_LIGHT2, GL_POSITION, z);
1272         checkGLcall("Setting up reverse light directions\n");
1273
1274         glRenderMode(GL_FEEDBACK);
1275         checkGLcall("glRenderMode(GL_FEEDBACK)");
1276         glEvalMesh2(GL_FILL, 0, ceilf(patch->numSegs[0]), 0, ceilf(patch->numSegs[1]));
1277         checkGLcall("glEvalMesh2\n");
1278         i = glRenderMode(GL_RENDER);
1279         checkGLcall("glRenderMode(GL_RENDER)");
1280
1281         i = 0;
1282         for(j = 0; j < buffer_size; j += (3 /* num verts */ * out_vertex_size + 2 /* tri marker */)) {
1283             if(feedbuffer[j] != GL_POLYGON_TOKEN) {
1284                 ERR("Unexpected token: %f\n", feedbuffer[j]);
1285                 continue;
1286             }
1287             if(feedbuffer[j + 1] != 3) {
1288                 ERR("Unexpected polygon: %f corners\n", feedbuffer[j + 1]);
1289                 continue;
1290             }
1291             if(patch->mem[i + 3] == 0.0)
1292                 patch->mem[i + 3] = -feedbuffer[j + out_vertex_size * 2 + 5];
1293             if(patch->mem[i + 4] == 0.0)
1294                 patch->mem[i + 4] = -feedbuffer[j + out_vertex_size * 2 + 6];
1295             if(patch->mem[i + 5] == 0.0)
1296                 patch->mem[i + 5] = -feedbuffer[j + out_vertex_size * 2 + 7];
1297             normalize_normal(patch->mem + i + 3);
1298             i += d3d_out_vertex_size;
1299
1300             if(patch->mem[i + 3] == 0.0)
1301                 patch->mem[i + 3] = -feedbuffer[j + out_vertex_size * 1 + 5];
1302             if(patch->mem[i + 4] == 0.0)
1303                 patch->mem[i + 4] = -feedbuffer[j + out_vertex_size * 1 + 6];
1304             if(patch->mem[i + 5] == 0.0)
1305                 patch->mem[i + 5] = -feedbuffer[j + out_vertex_size * 1 + 7];
1306             normalize_normal(patch->mem + i + 3);
1307             i += d3d_out_vertex_size;
1308
1309             if(patch->mem[i + 3] == 0.0)
1310                 patch->mem[i + 3] = -feedbuffer[j + out_vertex_size * 0 + 5];
1311             if(patch->mem[i + 4] == 0.0)
1312                 patch->mem[i + 4] = -feedbuffer[j + out_vertex_size * 0 + 6];
1313             if(patch->mem[i + 5] == 0.0)
1314                 patch->mem[i + 5] = -feedbuffer[j + out_vertex_size * 0 + 7];
1315             normalize_normal(patch->mem + i + 3);
1316             i += d3d_out_vertex_size;
1317         }
1318     }
1319
1320     glDisable(GL_MAP2_VERTEX_3);
1321     glDisable(GL_AUTO_NORMAL);
1322     glDisable(GL_MAP2_NORMAL);
1323     glDisable(GL_MAP2_TEXTURE_COORD_4);
1324     checkGLcall("glDisable vertex attrib generation");
1325     LEAVE_GL();
1326
1327     HeapFree(GetProcessHeap(), 0, feedbuffer);
1328
1329     vtxStride = 3 * sizeof(float);
1330     if(patch->has_normals) {
1331         vtxStride += 3 * sizeof(float);
1332     }
1333     if(patch->has_texcoords) {
1334         vtxStride += 4 * sizeof(float);
1335     }
1336     memset(&patch->strided, 0, sizeof(&patch->strided));
1337     patch->strided.u.s.position.lpData = (BYTE *) patch->mem;
1338     patch->strided.u.s.position.dwStride = vtxStride;
1339     patch->strided.u.s.position.dwType = WINED3DDECLTYPE_FLOAT3;
1340     patch->strided.u.s.position.streamNo = 255;
1341
1342     if(patch->has_normals) {
1343         patch->strided.u.s.normal.lpData = (BYTE *) patch->mem + 3 * sizeof(float) /* pos */;
1344         patch->strided.u.s.normal.dwStride = vtxStride;
1345         patch->strided.u.s.normal.dwType = WINED3DDECLTYPE_FLOAT3;
1346         patch->strided.u.s.normal.streamNo = 255;
1347     }
1348     if(patch->has_texcoords) {
1349         patch->strided.u.s.texCoords[0].lpData = (BYTE *) patch->mem + 3 * sizeof(float) /* pos */;
1350         if(patch->has_normals) {
1351             patch->strided.u.s.texCoords[0].lpData += 3 * sizeof(float);
1352         }
1353         patch->strided.u.s.texCoords[0].dwStride = vtxStride;
1354         patch->strided.u.s.texCoords[0].dwType = WINED3DDECLTYPE_FLOAT4;
1355         /* MAX_STREAMS index points to an unused element in stateblock->streamOffsets which
1356          * always remains set to 0. Windows uses stream 255 here, but this is not visible to the
1357          * application.
1358          */
1359         patch->strided.u.s.texCoords[0].streamNo = MAX_STREAMS;
1360     }
1361
1362     return WINED3D_OK;
1363 }