dlls/wined3d/vertexbuffer.c

   1 /*
   2  * IWineD3DVertexBuffer Implementation
   3  *
   4  * Copyright 2002-2005 Jason Edmeades
   5  *                     Raphael Junqueira
   6  * Copyright 2004 Christian Costa
   7  * Copyright 2007 Stefan Dösinger for CodeWeavers
   8  *
   9  * This library is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public
  11  * License as published by the Free Software Foundation; either
  12  * version 2.1 of the License, or (at your option) any later version.
  13  *
  14  * This library is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with this library; if not, write to the Free Software
  21  * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA
  22  */
  23
  24 #include "config.h"
  25 #include "wined3d_private.h"
  26
  27 WINE_DEFAULT_DEBUG_CHANNEL(d3d);
  28 #define GLINFO_LOCATION This->resource.wineD3DDevice->adapter->gl_info
  29
  30 #define VB_MAXDECLCHANGES     100     /* After that number we stop converting */
  31 #define VB_RESETDECLCHANGE    1000    /* Reset the changecount after that number of draws */
  32
  33 /* *******************************************
  34    IWineD3DVertexBuffer IUnknown parts follow
  35    ******************************************* */
  36 static HRESULT WINAPI IWineD3DVertexBufferImpl_QueryInterface(IWineD3DVertexBuffer *iface, REFIID riid, LPVOID *ppobj)
  37 {
  38     IWineD3DVertexBufferImpl *This = (IWineD3DVertexBufferImpl *)iface;
  39     TRACE("(%p)->(%s,%p)\n",This,debugstr_guid(riid),ppobj);
  40     if (IsEqualGUID(riid, &IID_IUnknown)
  41         || IsEqualGUID(riid, &IID_IWineD3DBase)
  42         || IsEqualGUID(riid, &IID_IWineD3DResource)
  43         || IsEqualGUID(riid, &IID_IWineD3DVertexBuffer)){
  44         IUnknown_AddRef(iface);
  45         *ppobj = This;
  46         return S_OK;
  47     }
  48     *ppobj = NULL;
  49     return E_NOINTERFACE;
  50 }
  51
  52 static ULONG WINAPI IWineD3DVertexBufferImpl_AddRef(IWineD3DVertexBuffer *iface) {
  53     IWineD3DVertexBufferImpl *This = (IWineD3DVertexBufferImpl *)iface;
  54     ULONG ref = InterlockedIncrement(&This->resource.ref);
  55     TRACE("(%p) : AddRef increasing from %d\n", This, ref - 1);
  56     return ref;
  57 }
  58
  59 static ULONG WINAPI IWineD3DVertexBufferImpl_Release(IWineD3DVertexBuffer *iface) {
  60     IWineD3DVertexBufferImpl *This = (IWineD3DVertexBufferImpl *)iface;
  61     ULONG ref = InterlockedDecrement(&This->resource.ref);
  62     TRACE("(%p) : Releasing from %d\n", This, ref + 1);
  63     if (ref == 0) {
  64
  65         if(This->vbo) {
  66             IWineD3DDeviceImpl *device = This->resource.wineD3DDevice;
  67
  68             ActivateContext(device, device->lastActiveRenderTarget, CTXUSAGE_RESOURCELOAD);
  69             ENTER_GL();
  70             GL_EXTCALL(glDeleteBuffersARB(1, &This->vbo));
  71             checkGLcall("glDeleteBuffersARB");
  72             LEAVE_GL();
  73         }
  74
  75         IWineD3DResourceImpl_CleanUp((IWineD3DResource *)iface);
  76         HeapFree(GetProcessHeap(), 0, This);
  77     }
  78     return ref;
  79 }
  80
  81 /* ****************************************************
  82    IWineD3DVertexBuffer IWineD3DResource parts follow
  83    **************************************************** */
  84 static HRESULT WINAPI IWineD3DVertexBufferImpl_GetDevice(IWineD3DVertexBuffer *iface, IWineD3DDevice** ppDevice) {
  85     return IWineD3DResourceImpl_GetDevice((IWineD3DResource *)iface, ppDevice);
  86 }
  87
  88 static HRESULT WINAPI IWineD3DVertexBufferImpl_SetPrivateData(IWineD3DVertexBuffer *iface, REFGUID refguid, CONST void* pData, DWORD SizeOfData, DWORD Flags) {
  89     return IWineD3DResourceImpl_SetPrivateData((IWineD3DResource *)iface, refguid, pData, SizeOfData, Flags);
  90 }
  91
  92 static HRESULT WINAPI IWineD3DVertexBufferImpl_GetPrivateData(IWineD3DVertexBuffer *iface, REFGUID refguid, void* pData, DWORD* pSizeOfData) {
  93     return IWineD3DResourceImpl_GetPrivateData((IWineD3DResource *)iface, refguid, pData, pSizeOfData);
  94 }
  95
  96 static HRESULT WINAPI IWineD3DVertexBufferImpl_FreePrivateData(IWineD3DVertexBuffer *iface, REFGUID refguid) {
  97     return IWineD3DResourceImpl_FreePrivateData((IWineD3DResource *)iface, refguid);
  98 }
  99
 100 static DWORD    WINAPI IWineD3DVertexBufferImpl_SetPriority(IWineD3DVertexBuffer *iface, DWORD PriorityNew) {
 101     return IWineD3DResourceImpl_SetPriority((IWineD3DResource *)iface, PriorityNew);
 102 }
 103
 104 static DWORD    WINAPI IWineD3DVertexBufferImpl_GetPriority(IWineD3DVertexBuffer *iface) {
 105     return IWineD3DResourceImpl_GetPriority((IWineD3DResource *)iface);
 106 }
 107
 108 static inline void fixup_d3dcolor(DWORD *pos) {
 109     DWORD srcColor = *pos;
 110
 111     /* Color conversion like in drawStridedSlow. watch out for little endianity
 112      * If we want that stuff to work on big endian machines too we have to consider more things
 113      *
 114      * 0xff000000: Alpha mask
 115      * 0x00ff0000: Blue mask
 116      * 0x0000ff00: Green mask
 117      * 0x000000ff: Red mask
 118      */
 119     *pos = 0;
 120     *pos |= (srcColor & 0xff00ff00)      ;   /* Alpha Green */
 121     *pos |= (srcColor & 0x00ff0000) >> 16;   /* Red */
 122     *pos |= (srcColor & 0x000000ff) << 16;   /* Blue */
 123 }
 124
 125 static inline void fixup_transformed_pos(float *p) {
 126     float x, y, z, w;
 127
 128     /* rhw conversion like in drawStridedSlow */
 129     if(p[3] == 1.0 || ((p[3] < eps) && (p[3] > -eps))) {
 130         x = p[0];
 131         y = p[1];
 132         z = p[2];
 133         w = 1.0;
 134     } else {
 135         w = 1.0 / p[3];
 136         x = p[0] * w;
 137         y = p[1] * w;
 138         z = p[2] * w;
 139     }
 140     p[0] = x;
 141     p[1] = y;
 142     p[2] = z;
 143     p[3] = w;
 144 }
 145
 146 DWORD *find_conversion_shift(IWineD3DVertexBufferImpl *This, WineDirect3DVertexStridedData *strided, DWORD stride) {
 147     DWORD *ret, i, shift, j, type;
 148     DWORD orig_type_size;
 149
 150     if(!stride) {
 151         TRACE("No shift\n");
 152         return NULL;
 153     }
 154
 155     This->conv_stride = stride;
 156     ret = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(DWORD) * stride);
 157     for(i = 0; i < MAX_ATTRIBS; i++) {
 158         if(strided->u.input[i].VBO != This->vbo) continue;
 159
 160         type = strided->u.input[i].dwType;
 161         if(type == WINED3DDECLTYPE_FLOAT16_2) {
 162             shift = 4;
 163         } else if(type == WINED3DDECLTYPE_FLOAT16_4) {
 164             shift = 8;
 165             /* Pre-shift the last 4 bytes in the FLOAT16_4 by 4 bytes - this makes FLOAT16_2 and FLOAT16_4 conversions
 166              * compatible
 167              */
 168             for(j = 4; j < 8; j++) {
 169                 ret[(DWORD_PTR) strided->u.input[i].lpData + j ] += 4;
 170             }
 171         } else {
 172             shift = 0;
 173         }
 174         This->conv_stride += shift;
 175
 176         if(shift) {
 177             orig_type_size = WINED3D_ATR_TYPESIZE(type) * WINED3D_ATR_SIZE(type);
 178             for(j = (DWORD_PTR) strided->u.input[i].lpData + orig_type_size; j < stride; j++) {
 179                 ret[j] += shift;
 180             }
 181         }
 182     }
 183
 184     if(TRACE_ON(d3d)) {
 185         TRACE("Dumping conversion shift:\n");
 186         for(i = 0; i < stride; i++) {
 187             TRACE("[%d]", ret[i]);
 188         }
 189         TRACE("\n");
 190     }
 191     return ret;
 192 }
 193
 194 static inline BOOL process_converted_attribute(IWineD3DVertexBufferImpl *This,
 195                                                const enum vbo_conversion_type conv_type,
 196                                                const WineDirect3DStridedData *attrib,
 197                                                DWORD *stride_this_run, const DWORD type) {
 198     DWORD attrib_size;
 199     BOOL ret = FALSE;
 200     int i;
 201     DWORD offset = This->resource.wineD3DDevice->stateBlock->streamOffset[attrib->streamNo];
 202     DWORD_PTR data;
 203
 204     /* Check for some valid situations which cause us pain. One is if the buffer is used for
 205      * constant attributes(stride = 0), the other one is if the buffer is used on two streams
 206      * with different strides. In the 2nd case we might have to drop conversion entirely,
 207      * it is possible that the same bytes are once read as FLOAT2 and once as UBYTE4N.
 208      */
 209     if(attrib->dwStride == 0) {
 210         FIXME("%s used with stride 0, let's hope we get the vertex stride from somewhere else\n",
 211                 debug_d3ddecltype(type));
 212     } else if(attrib->dwStride != *stride_this_run &&
 213                 *stride_this_run) {
 214         FIXME("Got two concurrent strides, %d and %d\n", attrib->dwStride, *stride_this_run);
 215     } else {
 216         *stride_this_run = attrib->dwStride;
 217         if(This->stride != *stride_this_run) {
 218             /* We rely that this happens only on the first converted attribute that is found,
 219              * if at all. See above check
 220              */
 221             TRACE("Reconverting because converted attributes occur, and the stride changed\n");
 222             This->stride = *stride_this_run;
 223             HeapFree(GetProcessHeap(), HEAP_ZERO_MEMORY, This->conv_map);
 224             This->conv_map = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*This->conv_map) * This->stride);
 225             ret = TRUE;
 226         }
 227     }
 228
 229     data = (((DWORD_PTR) attrib->lpData) + offset) % This->stride;
 230     attrib_size = WINED3D_ATR_SIZE(type) * WINED3D_ATR_TYPESIZE(type);
 231     for(i = 0; i < attrib_size; i++) {
 232         if(This->conv_map[data + i] != conv_type) {
 233             TRACE("Byte %ld in vertex changed\n", i + data);
 234             TRACE("It was type %d, is %d now\n", This->conv_map[data + i], conv_type);
 235             ret = TRUE;
 236             This->conv_map[data + i] = conv_type;
 237         }
 238     }
 239     return ret;
 240 }
 241
 242 static inline BOOL check_attribute(IWineD3DVertexBufferImpl *This, const WineDirect3DStridedData *attrib,
 243                                    const BOOL check_d3dcolor, const BOOL is_ffp_position, const BOOL is_ffp_color,
 244                                    DWORD *stride_this_run, BOOL *float16_used) {
 245     BOOL ret = FALSE;
 246     DWORD type;
 247
 248     /* Ignore attributes that do not have our vbo. After that check we can be sure that the attribute is
 249      * there, on nonexistent attribs the vbo is 0.
 250      */
 251     if(attrib->VBO != This->vbo) return FALSE;
 252
 253     type = attrib->dwType;
 254     /* Look for newly appeared conversion */
 255     if(!GL_SUPPORT(NV_HALF_FLOAT) && (
 256        type == WINED3DDECLTYPE_FLOAT16_2 ||
 257        type == WINED3DDECLTYPE_FLOAT16_4)) {
 258
 259         ret = process_converted_attribute(This, CONV_FLOAT16_2, attrib, stride_this_run, type);
 260
 261         if(is_ffp_position) {
 262             FIXME("Test FLOAT16 fixed function processing positions\n");
 263         } else if(is_ffp_color) {
 264             FIXME("test FLOAT16 fixed function processing colors\n");
 265         }
 266         *float16_used = TRUE;
 267     } else if(check_d3dcolor && type == WINED3DDECLTYPE_D3DCOLOR) {
 268
 269         ret = process_converted_attribute(This, CONV_D3DCOLOR, attrib, stride_this_run, WINED3DDECLTYPE_D3DCOLOR);
 270
 271         if(!is_ffp_color) {
 272             FIXME("Test for non-color fixed function D3DCOLOR type\n");
 273         }
 274     } else if(is_ffp_position && type == WINED3DDECLTYPE_FLOAT4) {
 275         ret = process_converted_attribute(This, CONV_POSITIONT, attrib, stride_this_run, WINED3DDECLTYPE_FLOAT4);
 276     } else if(This->conv_map) {
 277         ret = process_converted_attribute(This, CONV_NONE, attrib, stride_this_run, type);
 278     }
 279     return ret;
 280 }
 281
 282 inline BOOL WINAPI IWineD3DVertexBufferImpl_FindDecl(IWineD3DVertexBufferImpl *This)
 283 {
 284     IWineD3DDeviceImpl *device = This->resource.wineD3DDevice;
 285     BOOL ret = FALSE;
 286     int i;
 287     DWORD stride_this_run = 0;
 288     BOOL float16_used = FALSE;
 289
 290     /* In d3d7 the vertex buffer declaration NEVER changes because it is stored in the d3d7 vertex buffer.
 291      * Once we have our declaration there is no need to look it up again.
 292      */
 293     if(((IWineD3DImpl *)device->wineD3D)->dxVersion == 7 && This->Flags & VBFLAG_HASDESC) {
 294         return FALSE;
 295     }
 296
 297     TRACE("Finding vertex buffer conversion information\n");
 298     /* Certain declaration types need some fixups before we can pass them to opengl. This means D3DCOLOR attributes with fixed
 299      * function vertex processing, FLOAT4 POSITIONT with fixed function, and FLOAT16 if GL_NV_half_float is not supported.
 300      *
 301      * The vertex buffer FVF doesn't help with finding them, we have to use the decoded vertex declaration and pick the things
 302      * that concern the current buffer. A problem with this is that this can change between draws, so we have to validate
 303      * the information and reprocess the buffer if it changes, and avoid false positives for performance reasons.
 304      *
 305      * We have to distinguish between vertex shaders and fixed function to pick the way we access the
 306      * strided vertex information.
 307      *
 308      * This code sets up a per-byte array with the size of the detected stride of the arrays in the
 309      * buffer. For each byte we have a field that marks the conversion needed on this byte. For example,
 310      * the following declaration with fixed function vertex processing:
 311      *
 312      *      POSITIONT, FLOAT4
 313      *      NORMAL, FLOAT3
 314      *      DIFFUSE, FLOAT16_4
 315      *      SPECULAR, D3DCOLOR
 316      *
 317      * Will result in
 318      * {                 POSITIONT                    }{             NORMAL                }{    DIFFUSE          }{SPECULAR }
 319      * [P][P][P][P][P][P][P][P][P][P][P][P][P][P][P][P][0][0][0][0][0][0][0][0][0][0][0][0][F][F][F][F][F][F][F][F][C][C][C][C]
 320      *
 321      * Where in this example map P means 4 component position conversion, 0 means no conversion, F means FLOAT16_2 conversion
 322      * and C means D3DCOLOR conversion(red / blue swizzle).
 323      *
 324      * If we're doing conversion and the stride changes we have to reconvert the whole buffer. Note that we do not mind if the
 325      * semantic changes, we only care for the conversion type. So if the NORMAL is replaced with a TEXCOORD, nothing has to be
 326      * done, or if the DIFFUSE is replaced with a D3DCOLOR BLENDWEIGHT we can happily dismiss the change. Some conversion types
 327      * depend on the semantic as well, for example a FLOAT4 texcoord needs no conversion while a FLOAT4 positiont needs one
 328      */
 329     if(use_vs(device)) {
 330         TRACE("vhsader\n");
 331         /* If the current vertex declaration is marked for no half float conversion don't bother to
 332          * analyse the strided streams in depth, just set them up for no conversion. Return decl changed
 333          * if we used conversion before
 334          */
 335         if(!((IWineD3DVertexDeclarationImpl *) device->stateBlock->vertexDecl)->half_float_conv_needed) {
 336             if(This->conv_map) {
 337                 TRACE("Now using shaders without conversion, but conversion used before\n");
 338                 HeapFree(GetProcessHeap(), 0, This->conv_map);
 339                 HeapFree(GetProcessHeap(), 0, This->conv_shift);
 340                 This->conv_map = NULL;
 341                 This->stride = 0;
 342                 This->conv_shift = NULL;
 343                 This->conv_stride = 0;
 344                 return TRUE;
 345             } else {
 346                 return FALSE;
 347             }
 348         }
 349         for(i = 0; i < MAX_ATTRIBS; i++) {
 350             ret = check_attribute(This, &device->strided_streams.u.input[i], FALSE, FALSE, FALSE, &stride_this_run, &float16_used) || ret;
 351         }
 352
 353         /* Recalculate the conversion shift map if the declaration has changed,
 354          * and we're using float16 conversion or used it on the last run
 355          */
 356         if(ret && (float16_used || This->conv_map)) {
 357             HeapFree(GetProcessHeap(), 0, This->conv_shift);
 358             This->conv_shift = find_conversion_shift(This, &device->strided_streams, This->stride);
 359         }
 360     } else {
 361         /* Fixed function is a bit trickier. We have to take care for D3DCOLOR types, FLOAT4 positions and of course
 362          * FLOAT16s if not supported. Also, we can't iterate over the array, so use macros to generate code for all
 363          * the attributes that our current fixed function pipeline implementation cares for.
 364          */
 365         ret = check_attribute(This, &device->strided_streams.u.s.position,     TRUE, TRUE,  FALSE, &stride_this_run, &float16_used) || ret;
 366         ret = check_attribute(This, &device->strided_streams.u.s.normal,       TRUE, FALSE, FALSE, &stride_this_run, &float16_used) || ret;
 367         ret = check_attribute(This, &device->strided_streams.u.s.diffuse,      TRUE, FALSE, TRUE,  &stride_this_run, &float16_used) || ret;
 368         ret = check_attribute(This, &device->strided_streams.u.s.specular,     TRUE, FALSE, TRUE,  &stride_this_run, &float16_used) || ret;
 369         ret = check_attribute(This, &device->strided_streams.u.s.texCoords[0], TRUE, FALSE, FALSE, &stride_this_run, &float16_used) || ret;
 370         ret = check_attribute(This, &device->strided_streams.u.s.texCoords[1], TRUE, FALSE, FALSE, &stride_this_run, &float16_used) || ret;
 371         ret = check_attribute(This, &device->strided_streams.u.s.texCoords[2], TRUE, FALSE, FALSE, &stride_this_run, &float16_used) || ret;
 372         ret = check_attribute(This, &device->strided_streams.u.s.texCoords[3], TRUE, FALSE, FALSE, &stride_this_run, &float16_used) || ret;
 373         ret = check_attribute(This, &device->strided_streams.u.s.texCoords[4], TRUE, FALSE, FALSE, &stride_this_run, &float16_used) || ret;
 374         ret = check_attribute(This, &device->strided_streams.u.s.texCoords[5], TRUE, FALSE, FALSE, &stride_this_run, &float16_used) || ret;
 375         ret = check_attribute(This, &device->strided_streams.u.s.texCoords[6], TRUE, FALSE, FALSE, &stride_this_run, &float16_used) || ret;
 376         ret = check_attribute(This, &device->strided_streams.u.s.texCoords[7], TRUE, FALSE, FALSE, &stride_this_run, &float16_used) || ret;
 377
 378         if(float16_used) FIXME("Float16 conversion used with fixed function vertex processing\n");
 379     }
 380
 381     if(stride_this_run == 0 && This->conv_map) {
 382         /* Sanity test */
 383         if(ret == FALSE) {
 384             ERR("no converted attributes found, old conversion map exists, and no declaration change???\n");
 385         }
 386         HeapFree(GetProcessHeap(), 0, This->conv_map);
 387         This->conv_map = NULL;
 388         This->stride = 0;
 389     }
 390     This->Flags |= VBFLAG_HASDESC;
 391
 392     if(ret) TRACE("Conversion information changed\n");
 393     return ret;
 394 }
 395
 396 static void check_vbo_size(IWineD3DVertexBufferImpl *This) {
 397     DWORD size = This->conv_stride ? This->conv_stride * (This->resource.size / This->stride) : This->resource.size;
 398     if(This->vbo_size != size) {
 399         TRACE("Old size %d, creating new size %d\n", This->vbo_size, size);
 400         ENTER_GL();
 401         GL_EXTCALL(glBindBufferARB(GL_ARRAY_BUFFER_ARB, This->vbo));
 402         checkGLcall("glBindBufferARB");
 403         GL_EXTCALL(glBufferDataARB(GL_ARRAY_BUFFER_ARB, size, NULL, This->vbo_usage));
 404         This->vbo_size = size;
 405         checkGLcall("glBufferDataARB");
 406         LEAVE_GL();
 407     }
 408 }
 409
 410 static void CreateVBO(IWineD3DVertexBufferImpl *This) {
 411     GLenum error, glUsage;
 412     DWORD vboUsage = This->resource.usage;
 413     IWineD3DDeviceImpl *device = This->resource.wineD3DDevice;
 414
 415     TRACE("Creating an OpenGL vertex buffer object for IWineD3DVertexBuffer %p  Usage(%s)\n", This, debug_d3dusage(vboUsage));
 416
 417     /* Make sure that a context is there. Needed in a multithreaded environment. Otherwise this call is a nop */
 418     ActivateContext(device, device->lastActiveRenderTarget, CTXUSAGE_RESOURCELOAD);
 419     ENTER_GL();
 420
 421     /* Make sure that the gl error is cleared. Do not use checkGLcall
 422     * here because checkGLcall just prints a fixme and continues. However,
 423     * if an error during VBO creation occurs we can fall back to non-vbo operation
 424     * with full functionality(but performance loss)
 425     */
 426     while(glGetError() != GL_NO_ERROR);
 427
 428     /* Basically the FVF parameter passed to CreateVertexBuffer is no good
 429     * It is the FVF set with IWineD3DDevice::SetFVF or the Vertex Declaration set with
 430     * IWineD3DDevice::SetVertexDeclaration that decides how the vertices in the buffer
 431     * look like. This means that on each DrawPrimitive call the vertex buffer has to be verified
 432     * to check if the rhw and color values are in the correct format.
 433     */
 434
 435     GL_EXTCALL(glGenBuffersARB(1, &This->vbo));
 436     error = glGetError();
 437     if(This->vbo == 0 || error != GL_NO_ERROR) {
 438         WARN("Failed to create a VBO with error %s (%#x)\n", debug_glerror(error), error);
 439         goto error;
 440     }
 441
 442     GL_EXTCALL(glBindBufferARB(GL_ARRAY_BUFFER_ARB, This->vbo));
 443     error = glGetError();
 444     if(error != GL_NO_ERROR) {
 445         WARN("Failed to bind the VBO with error %s (%#x)\n", debug_glerror(error), error);
 446         goto error;
 447     }
 448
 449     /* Don't use static, because dx apps tend to update the buffer
 450     * quite often even if they specify 0 usage. Because we always keep the local copy
 451     * we never read from the vbo and can create a write only opengl buffer.
 452     */
 453     switch(vboUsage & (WINED3DUSAGE_WRITEONLY | WINED3DUSAGE_DYNAMIC) ) {
 454         case WINED3DUSAGE_WRITEONLY | WINED3DUSAGE_DYNAMIC:
 455         case WINED3DUSAGE_DYNAMIC:
 456             TRACE("Gl usage = GL_STREAM_DRAW\n");
 457             glUsage = GL_STREAM_DRAW_ARB;
 458             break;
 459         case WINED3DUSAGE_WRITEONLY:
 460         default:
 461             TRACE("Gl usage = GL_DYNAMIC_DRAW\n");
 462             glUsage = GL_DYNAMIC_DRAW_ARB;
 463             break;
 464     }
 465
 466     /* Reserve memory for the buffer. The amount of data won't change
 467     * so we are safe with calling glBufferData once with a NULL ptr and
 468     * calling glBufferSubData on updates
 469     */
 470     GL_EXTCALL(glBufferDataARB(GL_ARRAY_BUFFER_ARB, This->resource.size, NULL, glUsage));
 471     error = glGetError();
 472     if(error != GL_NO_ERROR) {
 473         WARN("glBufferDataARB failed with error %s (%#x)\n", debug_glerror(error), error);
 474         goto error;
 475     }
 476     This->vbo_size = This->resource.size;
 477     This->vbo_usage = glUsage;
 478     This->dirtystart = 0;
 479     This->dirtyend = This->resource.size;
 480     This->Flags |= VBFLAG_DIRTY;
 481
 482     LEAVE_GL();
 483
 484     return;
 485     error:
 486             /* Clean up all vbo init, but continue because we can work without a vbo :-) */
 487             FIXME("Failed to create a vertex buffer object. Continuing, but performance issues can occur\n");
 488     if(This->vbo) GL_EXTCALL(glDeleteBuffersARB(1, &This->vbo));
 489     This->vbo = 0;
 490     LEAVE_GL();
 491     return;
 492 }
 493
 494 static void     WINAPI IWineD3DVertexBufferImpl_PreLoad(IWineD3DVertexBuffer *iface) {
 495     IWineD3DVertexBufferImpl *This = (IWineD3DVertexBufferImpl *) iface;
 496     IWineD3DDeviceImpl *device = This->resource.wineD3DDevice;
 497     BYTE *data;
 498     UINT start = 0, end = 0, vertices;
 499     BOOL declChanged = FALSE;
 500     int i, j;
 501     TRACE("(%p)->()\n", This);
 502
 503     if(!This->vbo) {
 504         /* TODO: Make converting independent from VBOs */
 505         if(This->Flags & VBFLAG_CREATEVBO) {
 506             CreateVBO(This);
 507             This->Flags &= ~VBFLAG_CREATEVBO;
 508         } else {
 509             return; /* Not doing any conversion */
 510         }
 511     }
 512
 513     /* Reading the declaration makes only sense if the stateblock is finalized and the buffer bound to a stream */
 514     if(device->isInDraw && This->bindCount > 0) {
 515         declChanged = IWineD3DVertexBufferImpl_FindDecl(This);
 516     } else if(This->Flags & VBFLAG_HASDESC) {
 517         /* Reuse the declaration stored in the buffer. It will most likely not change, and if it does
 518          * the stream source state handler will call PreLoad again and the change will be caught
 519          */
 520     } else {
 521         /* Cannot get a declaration, and no declaration is stored in the buffer. It is pointless to preload
 522          * now. When the buffer is used, PreLoad will be called by the stream source state handler and a valid
 523          * declaration for the buffer can be found
 524          */
 525         return;
 526     }
 527
 528     /* If applications change the declaration over and over, reconverting all the time is a huge
 529      * performance hit. So count the declaration changes and release the VBO if there are too many
 530      * of them (and thus stop converting)
 531      */
 532     if(declChanged) {
 533         This->declChanges++;
 534         This->draws = 0;
 535
 536         if(This->declChanges > VB_MAXDECLCHANGES) {
 537             FIXME("Too many declaration changes, stopping converting\n");
 538             ActivateContext(device, device->lastActiveRenderTarget, CTXUSAGE_RESOURCELOAD);
 539             ENTER_GL();
 540             GL_EXTCALL(glDeleteBuffersARB(1, &This->vbo));
 541             checkGLcall("glDeleteBuffersARB");
 542             LEAVE_GL();
 543             This->vbo = 0;
 544             HeapFree(GetProcessHeap(), 0, This->conv_shift);
 545
 546             /* The stream source state handler might have read the memory of the vertex buffer already
 547              * and got the memory in the vbo which is not valid any longer. Dirtify the stream source
 548              * to force a reload. This happens only once per changed vertexbuffer and should occur rather
 549              * rarely
 550              */
 551             IWineD3DDeviceImpl_MarkStateDirty(device, STATE_STREAMSRC);
 552
 553             return;
 554         }
 555         check_vbo_size(This);
 556     } else {
 557         /* However, it is perfectly fine to change the declaration every now and then. We don't want a game that
 558          * changes it every minute drop the VBO after VB_MAX_DECL_CHANGES minutes. So count draws without
 559          * decl changes and reset the decl change count after a specific number of them
 560          */
 561         This->draws++;
 562         if(This->draws > VB_RESETDECLCHANGE) This->declChanges = 0;
 563     }
 564
 565     if(declChanged) {
 566         /* The declaration changed, reload the whole buffer */
 567         WARN("Reloading buffer because of decl change\n");
 568         start = 0;
 569         end = This->resource.size;
 570     } else if(This->Flags & VBFLAG_DIRTY) {
 571         /* No decl change, but dirty data, reload the changed stuff */
 572         if(This->conv_shift) {
 573             if(This->dirtystart != 0 || This->dirtyend != 0) {
 574                 FIXME("Implement partial buffer loading with shifted conversion\n");
 575             }
 576         }
 577         start = This->dirtystart;
 578         end = This->dirtyend;
 579     } else {
 580         /* Desc not changed, buffer not dirty, nothing to do :-) */
 581         return;
 582     }
 583
 584     /* Mark the buffer clean */
 585     This->Flags &= ~VBFLAG_DIRTY;
 586     This->dirtystart = 0;
 587     This->dirtyend = 0;
 588
 589     if(!This->conv_map) {
 590         /* That means that there is nothing to fixup. Just upload from This->resource.allocatedMemory
 591          * directly into the vbo. Do not free the system memory copy because drawPrimitive may need it if
 592          * the stride is 0, for instancing emulation, vertex blending emulation or shader emulation.
 593          */
 594         TRACE("No conversion needed\n");
 595
 596         if(!device->isInDraw) {
 597             ActivateContext(device, device->lastActiveRenderTarget, CTXUSAGE_RESOURCELOAD);
 598         }
 599         ENTER_GL();
 600         GL_EXTCALL(glBindBufferARB(GL_ARRAY_BUFFER_ARB, This->vbo));
 601         checkGLcall("glBindBufferARB");
 602         GL_EXTCALL(glBufferSubDataARB(GL_ARRAY_BUFFER_ARB, start, end-start, This->resource.allocatedMemory + start));
 603         checkGLcall("glBufferSubDataARB");
 604         LEAVE_GL();
 605         return;
 606     }
 607
 608     /* Now for each vertex in the buffer that needs conversion */
 609     vertices = This->resource.size / This->stride;
 610
 611     if(This->conv_shift) {
 612         TRACE("Shifted conversion\n");
 613         data = HeapAlloc(GetProcessHeap(), 0, vertices * This->conv_stride);
 614
 615         for(i = start / This->stride; i < min((end / This->stride) + 1, vertices); i++) {
 616             for(j = 0; j < This->stride; j++) {
 617                 switch(This->conv_map[j]) {
 618                     case CONV_NONE:
 619                         data[This->conv_stride * i + j + This->conv_shift[j]] = This->resource.allocatedMemory[This->stride * i + j];
 620                         break;
 621
 622                     case CONV_FLOAT16_2:
 623                     {
 624                         float *out = (float *) (&data[This->conv_stride * i + j + This->conv_shift[j]]);
 625                         WORD *in = (WORD *) (&This->resource.allocatedMemory[i * This->stride + j]);
 626
 627                         out[1] = float_16_to_32(in + 1);
 628                         out[0] = float_16_to_32(in + 0);
 629                         j += 3;    /* Skip 3 additional bytes,as a FLOAT16_2 has 4 bytes */
 630                         break;
 631                     }
 632
 633                     default:
 634                         FIXME("Unimplemented conversion %d in shifted conversion\n", This->conv_map[j]);
 635                 }
 636             }
 637         }
 638
 639         ENTER_GL();
 640         GL_EXTCALL(glBindBufferARB(GL_ARRAY_BUFFER_ARB, This->vbo));
 641         checkGLcall("glBindBufferARB");
 642         GL_EXTCALL(glBufferSubDataARB(GL_ARRAY_BUFFER_ARB, 0, vertices * This->conv_stride, data));
 643         checkGLcall("glBufferSubDataARB");
 644         LEAVE_GL();
 645
 646     } else {
 647         data = HeapAlloc(GetProcessHeap(), 0, This->resource.size);
 648         memcpy(data + start, This->resource.allocatedMemory + start, end - start);
 649         for(i = start / This->stride; i < min((end / This->stride) + 1, vertices); i++) {
 650             for(j = 0; j < This->stride; j++) {
 651                 switch(This->conv_map[j]) {
 652                     case CONV_NONE:
 653                         /* Done already */
 654                         j += 3;
 655                         break;
 656                     case CONV_D3DCOLOR:
 657                         fixup_d3dcolor((DWORD *) (data + i * This->stride + j));
 658                         j += 3;
 659                         break;
 660
 661                     case CONV_POSITIONT:
 662                         fixup_transformed_pos((float *) (data + i * This->stride + j));
 663                         j += 15;
 664                         break;
 665
 666                     case CONV_FLOAT16_2:
 667                         ERR("Did not expect FLOAT16 conversion in unshifted conversion\n");
 668                     default:
 669                         FIXME("Unimplemented conversion %d in shifted conversion\n", This->conv_map[j]);
 670                 }
 671             }
 672         }
 673
 674         ENTER_GL();
 675         GL_EXTCALL(glBindBufferARB(GL_ARRAY_BUFFER_ARB, This->vbo));
 676         checkGLcall("glBindBufferARB");
 677         GL_EXTCALL(glBufferSubDataARB(GL_ARRAY_BUFFER_ARB, start, end - start, data + start));
 678         checkGLcall("glBufferSubDataARB");
 679         LEAVE_GL();
 680     }
 681
 682     HeapFree(GetProcessHeap(), 0, data);
 683 }
 684
 685 static void WINAPI IWineD3DVertexBufferImpl_UnLoad(IWineD3DVertexBuffer *iface) {
 686     IWineD3DVertexBufferImpl *This = (IWineD3DVertexBufferImpl *) iface;
 687     IWineD3DDeviceImpl *device = This->resource.wineD3DDevice;
 688     TRACE("(%p)\n", This);
 689
 690     /* This is easy: The whole content is shadowed in This->resource.allocatedMemory,
 691      * so we only have to destroy the vbo. Only do it if we have a vbo, which implies
 692      * that vbos are supported
 693      */
 694     if(This->vbo) {
 695         ActivateContext(device, device->lastActiveRenderTarget, CTXUSAGE_RESOURCELOAD);
 696         ENTER_GL();
 697         GL_EXTCALL(glDeleteBuffersARB(1, &This->vbo));
 698         checkGLcall("glDeleteBuffersARB");
 699         LEAVE_GL();
 700         This->vbo = 0;
 701         This->Flags |= VBFLAG_CREATEVBO; /* Recreate the VBO next load */
 702     }
 703 }
 704
 705 static WINED3DRESOURCETYPE WINAPI IWineD3DVertexBufferImpl_GetType(IWineD3DVertexBuffer *iface) {
 706     return IWineD3DResourceImpl_GetType((IWineD3DResource *)iface);
 707 }
 708
 709 static HRESULT WINAPI IWineD3DVertexBufferImpl_GetParent(IWineD3DVertexBuffer *iface, IUnknown **pParent) {
 710     return IWineD3DResourceImpl_GetParent((IWineD3DResource *)iface, pParent);
 711 }
 712
 713 /* ******************************************************
 714    IWineD3DVertexBuffer IWineD3DVertexBuffer parts follow
 715    ****************************************************** */
 716 static HRESULT  WINAPI IWineD3DVertexBufferImpl_Lock(IWineD3DVertexBuffer *iface, UINT OffsetToLock, UINT SizeToLock, BYTE** ppbData, DWORD Flags) {
 717     IWineD3DVertexBufferImpl *This = (IWineD3DVertexBufferImpl *)iface;
 718     BYTE *data;
 719     TRACE("(%p)->%d, %d, %p, %08x\n", This, OffsetToLock, SizeToLock, ppbData, Flags);
 720
 721     InterlockedIncrement(&This->lockcount);
 722
 723     if(This->Flags & VBFLAG_DIRTY) {
 724         if(This->dirtystart > OffsetToLock) This->dirtystart = OffsetToLock;
 725         if(SizeToLock) {
 726             if(This->dirtyend < OffsetToLock + SizeToLock) This->dirtyend = OffsetToLock + SizeToLock;
 727         } else {
 728             This->dirtyend = This->resource.size;
 729         }
 730     } else {
 731         This->dirtystart = OffsetToLock;
 732         if(SizeToLock)
 733             This->dirtyend = OffsetToLock + SizeToLock;
 734         else
 735             This->dirtyend = This->resource.size;
 736     }
 737
 738     data = This->resource.allocatedMemory;
 739     This->Flags |= VBFLAG_DIRTY;
 740     *ppbData = data + OffsetToLock;
 741
 742     TRACE("(%p) : returning memory of %p (base:%p,offset:%u)\n", This, data + OffsetToLock, data, OffsetToLock);
 743     /* TODO: check Flags compatibility with This->currentDesc.Usage (see MSDN) */
 744     return WINED3D_OK;
 745 }
 746 HRESULT  WINAPI IWineD3DVertexBufferImpl_Unlock(IWineD3DVertexBuffer *iface) {
 747     IWineD3DVertexBufferImpl *This = (IWineD3DVertexBufferImpl *) iface;
 748     LONG lockcount;
 749     TRACE("(%p)\n", This);
 750
 751     lockcount = InterlockedDecrement(&This->lockcount);
 752     if(lockcount > 0) {
 753         /* Delay loading the buffer until everything is unlocked */
 754         TRACE("Ignoring the unlock\n");
 755         return WINED3D_OK;
 756     }
 757
 758     if(This->Flags & VBFLAG_HASDESC) {
 759         IWineD3DVertexBufferImpl_PreLoad(iface);
 760     }
 761     return WINED3D_OK;
 762 }
 763 static HRESULT  WINAPI IWineD3DVertexBufferImpl_GetDesc(IWineD3DVertexBuffer *iface, WINED3DVERTEXBUFFER_DESC *pDesc) {
 764     IWineD3DVertexBufferImpl *This = (IWineD3DVertexBufferImpl *)iface;
 765
 766     TRACE("(%p)\n", This);
 767     pDesc->Format = This->resource.format;
 768     pDesc->Type   = This->resource.resourceType;
 769     pDesc->Usage  = This->resource.usage;
 770     pDesc->Pool   = This->resource.pool;
 771     pDesc->Size   = This->resource.size;
 772     pDesc->FVF    = This->fvf;
 773     return WINED3D_OK;
 774 }
 775
 776 const IWineD3DVertexBufferVtbl IWineD3DVertexBuffer_Vtbl =
 777 {
 778     /* IUnknown */
 779     IWineD3DVertexBufferImpl_QueryInterface,
 780     IWineD3DVertexBufferImpl_AddRef,
 781     IWineD3DVertexBufferImpl_Release,
 782     /* IWineD3DResource */
 783     IWineD3DVertexBufferImpl_GetParent,
 784     IWineD3DVertexBufferImpl_GetDevice,
 785     IWineD3DVertexBufferImpl_SetPrivateData,
 786     IWineD3DVertexBufferImpl_GetPrivateData,
 787     IWineD3DVertexBufferImpl_FreePrivateData,
 788     IWineD3DVertexBufferImpl_SetPriority,
 789     IWineD3DVertexBufferImpl_GetPriority,
 790     IWineD3DVertexBufferImpl_PreLoad,
 791     IWineD3DVertexBufferImpl_UnLoad,
 792     IWineD3DVertexBufferImpl_GetType,
 793     /* IWineD3DVertexBuffer */
 794     IWineD3DVertexBufferImpl_Lock,
 795     IWineD3DVertexBufferImpl_Unlock,
 796     IWineD3DVertexBufferImpl_GetDesc
 797 };
 798
 799 BYTE* WINAPI IWineD3DVertexBufferImpl_GetMemory(IWineD3DVertexBuffer* iface, DWORD iOffset, GLint *vbo) {
 800     IWineD3DVertexBufferImpl *This = (IWineD3DVertexBufferImpl *)iface;
 801
 802     *vbo = This->vbo;
 803     if(This->vbo == 0) {
 804         if(This->Flags & VBFLAG_CREATEVBO) {
 805             CreateVBO(This);
 806             This->Flags &= ~VBFLAG_CREATEVBO;
 807             if(This->vbo) {
 808                 *vbo = This->vbo;
 809                 return (BYTE *) iOffset;
 810             }
 811         }
 812         return This->resource.allocatedMemory + iOffset;
 813     } else {
 814         return (BYTE *) iOffset;
 815     }
 816 }
 817
 818 HRESULT WINAPI IWineD3DVertexBufferImpl_ReleaseMemory(IWineD3DVertexBuffer* iface) {
 819   return WINED3D_OK;
 820 }