git.oblomov.eu Git - wine/blob - dlls/wined3d/surface.c

   1 /*
   2  * IWineD3DSurface Implementation
   3  *
   4  * Copyright 1998 Lionel Ulmer
   5  * Copyright 2000-2001 TransGaming Technologies Inc.
   6  * Copyright 2002-2005 Jason Edmeades
   7  * Copyright 2002-2003 Raphael Junqueira
   8  * Copyright 2004 Christian Costa
   9  * Copyright 2005 Oliver Stieber
  10  * Copyright 2006-2008 Stefan Dösinger for CodeWeavers
  11  * Copyright 2007-2008 Henri Verbeet
  12  * Copyright 2006-2008 Roderick Colenbrander
  13  * Copyright 2009 Henri Verbeet for CodeWeavers
  14  *
  15  * This library is free software; you can redistribute it and/or
  16  * modify it under the terms of the GNU Lesser General Public
  17  * License as published by the Free Software Foundation; either
  18  * version 2.1 of the License, or (at your option) any later version.
  19  *
  20  * This library is distributed in the hope that it will be useful,
  21  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  22  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  23  * Lesser General Public License for more details.
  24  *
  25  * You should have received a copy of the GNU Lesser General Public
  26  * License along with this library; if not, write to the Free Software
  27  * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA
  28  */
  29
  30 #include "config.h"
  31 #include "wine/port.h"
  32 #include "wined3d_private.h"
  33
  34 WINE_DEFAULT_DEBUG_CHANNEL(d3d_surface);
  35 WINE_DECLARE_DEBUG_CHANNEL(d3d);
  36
  37 static void surface_cleanup(IWineD3DSurfaceImpl *This)
  38 {
  39     IWineD3DDeviceImpl *device = This->resource.device;
  40     const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
  41     struct wined3d_context *context = NULL;
  42     renderbuffer_entry_t *entry, *entry2;
  43
  44     TRACE("(%p) : Cleaning up.\n", This);
  45
  46     /* Need a context to destroy the texture. Use the currently active render
  47      * target, but only if the primary render target exists. Otherwise
  48      * lastActiveRenderTarget is garbage. When destroying the primary render
  49      * target, Uninit3D() will activate a context before doing anything. */
  50     if (device->render_targets && device->render_targets[0])
  51     {
  52         context = context_acquire(device, NULL);
  53     }
  54
  55     ENTER_GL();
  56
  57     if (This->texture_name)
  58     {
  59         /* Release the OpenGL texture. */
  60         TRACE("Deleting texture %u.\n", This->texture_name);
  61         glDeleteTextures(1, &This->texture_name);
  62     }
  63
  64     if (This->Flags & SFLAG_PBO)
  65     {
  66         /* Delete the PBO. */
  67         GL_EXTCALL(glDeleteBuffersARB(1, &This->pbo));
  68     }
  69
  70     LIST_FOR_EACH_ENTRY_SAFE(entry, entry2, &This->renderbuffers, renderbuffer_entry_t, entry)
  71     {
  72         gl_info->fbo_ops.glDeleteRenderbuffers(1, &entry->id);
  73         HeapFree(GetProcessHeap(), 0, entry);
  74     }
  75
  76     LEAVE_GL();
  77
  78     if (This->Flags & SFLAG_DIBSECTION)
  79     {
  80         /* Release the DC. */
  81         SelectObject(This->hDC, This->dib.holdbitmap);
  82         DeleteDC(This->hDC);
  83         /* Release the DIB section. */
  84         DeleteObject(This->dib.DIBsection);
  85         This->dib.bitmap_data = NULL;
  86         This->resource.allocatedMemory = NULL;
  87     }
  88
  89     if (This->Flags & SFLAG_USERPTR) IWineD3DSurface_SetMem((IWineD3DSurface *)This, NULL);
  90     if (This->overlay_dest) list_remove(&This->overlay_entry);
  91
  92     HeapFree(GetProcessHeap(), 0, This->palette9);
  93
  94     resource_cleanup((IWineD3DResource *)This);
  95
  96     if (context) context_release(context);
  97 }
  98
  99 UINT surface_calculate_size(const struct wined3d_format_desc *format_desc, UINT alignment, UINT width, UINT height)
 100 {
 101     UINT size;
 102
 103     if (format_desc->format == WINED3DFMT_UNKNOWN)
 104     {
 105         size = 0;
 106     }
 107     else if (format_desc->Flags & WINED3DFMT_FLAG_COMPRESSED)
 108     {
 109         UINT row_block_count = (width + format_desc->block_width - 1) / format_desc->block_width;
 110         UINT row_count = (height + format_desc->block_height - 1) / format_desc->block_height;
 111         size = row_count * row_block_count * format_desc->block_byte_count;
 112     }
 113     else
 114     {
 115         /* The pitch is a multiple of 4 bytes. */
 116         size = height * (((width * format_desc->byte_count) + alignment - 1) & ~(alignment - 1));
 117     }
 118
 119     if (format_desc->heightscale != 0.0f) size *= format_desc->heightscale;
 120
 121     return size;
 122 }
 123
 124 struct blt_info
 125 {
 126     GLenum binding;
 127     GLenum bind_target;
 128     enum tex_types tex_type;
 129     GLfloat coords[4][3];
 130 };
 131
 132 struct float_rect
 133 {
 134     float l;
 135     float t;
 136     float r;
 137     float b;
 138 };
 139
 140 static inline void cube_coords_float(const RECT *r, UINT w, UINT h, struct float_rect *f)
 141 {
 142     f->l = ((r->left * 2.0f) / w) - 1.0f;
 143     f->t = ((r->top * 2.0f) / h) - 1.0f;
 144     f->r = ((r->right * 2.0f) / w) - 1.0f;
 145     f->b = ((r->bottom * 2.0f) / h) - 1.0f;
 146 }
 147
 148 static void surface_get_blt_info(GLenum target, const RECT *rect_in, GLsizei w, GLsizei h, struct blt_info *info)
 149 {
 150     GLfloat (*coords)[3] = info->coords;
 151     RECT rect;
 152     struct float_rect f;
 153
 154     if (rect_in)
 155         rect = *rect_in;
 156     else
 157     {
 158         rect.left = 0;
 159         rect.top = h;
 160         rect.right = w;
 161         rect.bottom = 0;
 162     }
 163
 164     switch (target)
 165     {
 166         default:
 167             FIXME("Unsupported texture target %#x\n", target);
 168             /* Fall back to GL_TEXTURE_2D */
 169         case GL_TEXTURE_2D:
 170             info->binding = GL_TEXTURE_BINDING_2D;
 171             info->bind_target = GL_TEXTURE_2D;
 172             info->tex_type = tex_2d;
 173             coords[0][0] = (float)rect.left / w;
 174             coords[0][1] = (float)rect.top / h;
 175             coords[0][2] = 0.0f;
 176
 177             coords[1][0] = (float)rect.right / w;
 178             coords[1][1] = (float)rect.top / h;
 179             coords[1][2] = 0.0f;
 180
 181             coords[2][0] = (float)rect.left / w;
 182             coords[2][1] = (float)rect.bottom / h;
 183             coords[2][2] = 0.0f;
 184
 185             coords[3][0] = (float)rect.right / w;
 186             coords[3][1] = (float)rect.bottom / h;
 187             coords[3][2] = 0.0f;
 188             break;
 189
 190         case GL_TEXTURE_RECTANGLE_ARB:
 191             info->binding = GL_TEXTURE_BINDING_RECTANGLE_ARB;
 192             info->bind_target = GL_TEXTURE_RECTANGLE_ARB;
 193             info->tex_type = tex_rect;
 194             coords[0][0] = rect.left;   coords[0][1] = rect.top;     coords[0][2] = 0.0f;
 195             coords[1][0] = rect.right;  coords[1][1] = rect.top;     coords[1][2] = 0.0f;
 196             coords[2][0] = rect.left;   coords[2][1] = rect.bottom;  coords[2][2] = 0.0f;
 197             coords[3][0] = rect.right;  coords[3][1] = rect.bottom;  coords[3][2] = 0.0f;
 198             break;
 199
 200         case GL_TEXTURE_CUBE_MAP_POSITIVE_X:
 201             info->binding = GL_TEXTURE_BINDING_CUBE_MAP_ARB;
 202             info->bind_target = GL_TEXTURE_CUBE_MAP_ARB;
 203             info->tex_type = tex_cube;
 204             cube_coords_float(&rect, w, h, &f);
 205
 206             coords[0][0] =  1.0f;   coords[0][1] = -f.t;   coords[0][2] = -f.l;
 207             coords[1][0] =  1.0f;   coords[1][1] = -f.t;   coords[1][2] = -f.r;
 208             coords[2][0] =  1.0f;   coords[2][1] = -f.b;   coords[2][2] = -f.l;
 209             coords[3][0] =  1.0f;   coords[3][1] = -f.b;   coords[3][2] = -f.r;
 210             break;
 211
 212         case GL_TEXTURE_CUBE_MAP_NEGATIVE_X:
 213             info->binding = GL_TEXTURE_BINDING_CUBE_MAP_ARB;
 214             info->bind_target = GL_TEXTURE_CUBE_MAP_ARB;
 215             info->tex_type = tex_cube;
 216             cube_coords_float(&rect, w, h, &f);
 217
 218             coords[0][0] = -1.0f;   coords[0][1] = -f.t;   coords[0][2] = f.l;
 219             coords[1][0] = -1.0f;   coords[1][1] = -f.t;   coords[1][2] = f.r;
 220             coords[2][0] = -1.0f;   coords[2][1] = -f.b;   coords[2][2] = f.l;
 221             coords[3][0] = -1.0f;   coords[3][1] = -f.b;   coords[3][2] = f.r;
 222             break;
 223
 224         case GL_TEXTURE_CUBE_MAP_POSITIVE_Y:
 225             info->binding = GL_TEXTURE_BINDING_CUBE_MAP_ARB;
 226             info->bind_target = GL_TEXTURE_CUBE_MAP_ARB;
 227             info->tex_type = tex_cube;
 228             cube_coords_float(&rect, w, h, &f);
 229
 230             coords[0][0] = f.l;   coords[0][1] =  1.0f;   coords[0][2] = f.t;
 231             coords[1][0] = f.r;   coords[1][1] =  1.0f;   coords[1][2] = f.t;
 232             coords[2][0] = f.l;   coords[2][1] =  1.0f;   coords[2][2] = f.b;
 233             coords[3][0] = f.r;   coords[3][1] =  1.0f;   coords[3][2] = f.b;
 234             break;
 235
 236         case GL_TEXTURE_CUBE_MAP_NEGATIVE_Y:
 237             info->binding = GL_TEXTURE_BINDING_CUBE_MAP_ARB;
 238             info->bind_target = GL_TEXTURE_CUBE_MAP_ARB;
 239             info->tex_type = tex_cube;
 240             cube_coords_float(&rect, w, h, &f);
 241
 242             coords[0][0] = f.l;   coords[0][1] = -1.0f;   coords[0][2] = -f.t;
 243             coords[1][0] = f.r;   coords[1][1] = -1.0f;   coords[1][2] = -f.t;
 244             coords[2][0] = f.l;   coords[2][1] = -1.0f;   coords[2][2] = -f.b;
 245             coords[3][0] = f.r;   coords[3][1] = -1.0f;   coords[3][2] = -f.b;
 246             break;
 247
 248         case GL_TEXTURE_CUBE_MAP_POSITIVE_Z:
 249             info->binding = GL_TEXTURE_BINDING_CUBE_MAP_ARB;
 250             info->bind_target = GL_TEXTURE_CUBE_MAP_ARB;
 251             info->tex_type = tex_cube;
 252             cube_coords_float(&rect, w, h, &f);
 253
 254             coords[0][0] = f.l;   coords[0][1] = -f.t;   coords[0][2] =  1.0f;
 255             coords[1][0] = f.r;   coords[1][1] = -f.t;   coords[1][2] =  1.0f;
 256             coords[2][0] = f.l;   coords[2][1] = -f.b;   coords[2][2] =  1.0f;
 257             coords[3][0] = f.r;   coords[3][1] = -f.b;   coords[3][2] =  1.0f;
 258             break;
 259
 260         case GL_TEXTURE_CUBE_MAP_NEGATIVE_Z:
 261             info->binding = GL_TEXTURE_BINDING_CUBE_MAP_ARB;
 262             info->bind_target = GL_TEXTURE_CUBE_MAP_ARB;
 263             info->tex_type = tex_cube;
 264             cube_coords_float(&rect, w, h, &f);
 265
 266             coords[0][0] = -f.l;   coords[0][1] = -f.t;   coords[0][2] = -1.0f;
 267             coords[1][0] = -f.r;   coords[1][1] = -f.t;   coords[1][2] = -1.0f;
 268             coords[2][0] = -f.l;   coords[2][1] = -f.b;   coords[2][2] = -1.0f;
 269             coords[3][0] = -f.r;   coords[3][1] = -f.b;   coords[3][2] = -1.0f;
 270             break;
 271     }
 272 }
 273
 274 static inline void surface_get_rect(IWineD3DSurfaceImpl *This, const RECT *rect_in, RECT *rect_out)
 275 {
 276     if (rect_in)
 277         *rect_out = *rect_in;
 278     else
 279     {
 280         rect_out->left = 0;
 281         rect_out->top = 0;
 282         rect_out->right = This->currentDesc.Width;
 283         rect_out->bottom = This->currentDesc.Height;
 284     }
 285 }
 286
 287 /* GL locking and context activation is done by the caller */
 288 void draw_textured_quad(IWineD3DSurfaceImpl *src_surface, const RECT *src_rect, const RECT *dst_rect, WINED3DTEXTUREFILTERTYPE Filter)
 289 {
 290     IWineD3DBaseTextureImpl *texture;
 291     struct blt_info info;
 292
 293     surface_get_blt_info(src_surface->texture_target, src_rect, src_surface->pow2Width, src_surface->pow2Height, &info);
 294
 295     glEnable(info.bind_target);
 296     checkGLcall("glEnable(bind_target)");
 297
 298     /* Bind the texture */
 299     glBindTexture(info.bind_target, src_surface->texture_name);
 300     checkGLcall("glBindTexture");
 301
 302     /* Filtering for StretchRect */
 303     glTexParameteri(info.bind_target, GL_TEXTURE_MAG_FILTER,
 304             wined3d_gl_mag_filter(magLookup, Filter));
 305     checkGLcall("glTexParameteri");
 306     glTexParameteri(info.bind_target, GL_TEXTURE_MIN_FILTER,
 307             wined3d_gl_min_mip_filter(minMipLookup, Filter, WINED3DTEXF_NONE));
 308     checkGLcall("glTexParameteri");
 309     glTexParameteri(info.bind_target, GL_TEXTURE_WRAP_S, GL_CLAMP);
 310     glTexParameteri(info.bind_target, GL_TEXTURE_WRAP_T, GL_CLAMP);
 311     glTexEnvi(GL_TEXTURE_ENV, GL_TEXTURE_ENV_MODE, GL_REPLACE);
 312     checkGLcall("glTexEnvi");
 313
 314     /* Draw a quad */
 315     glBegin(GL_TRIANGLE_STRIP);
 316     glTexCoord3fv(info.coords[0]);
 317     glVertex2i(dst_rect->left, dst_rect->top);
 318
 319     glTexCoord3fv(info.coords[1]);
 320     glVertex2i(dst_rect->right, dst_rect->top);
 321
 322     glTexCoord3fv(info.coords[2]);
 323     glVertex2i(dst_rect->left, dst_rect->bottom);
 324
 325     glTexCoord3fv(info.coords[3]);
 326     glVertex2i(dst_rect->right, dst_rect->bottom);
 327     glEnd();
 328
 329     /* Unbind the texture */
 330     glBindTexture(info.bind_target, 0);
 331     checkGLcall("glBindTexture(info->bind_target, 0)");
 332
 333     /* We changed the filtering settings on the texture. Inform the
 334      * container about this to get the filters reset properly next draw. */
 335     if (SUCCEEDED(IWineD3DSurface_GetContainer((IWineD3DSurface *)src_surface, &IID_IWineD3DBaseTexture, (void **)&texture)))
 336     {
 337         texture->baseTexture.texture_rgb.states[WINED3DTEXSTA_MAGFILTER] = WINED3DTEXF_POINT;
 338         texture->baseTexture.texture_rgb.states[WINED3DTEXSTA_MINFILTER] = WINED3DTEXF_POINT;
 339         texture->baseTexture.texture_rgb.states[WINED3DTEXSTA_MIPFILTER] = WINED3DTEXF_NONE;
 340         IWineD3DBaseTexture_Release((IWineD3DBaseTexture *)texture);
 341     }
 342 }
 343
 344 HRESULT surface_init(IWineD3DSurfaceImpl *surface, WINED3DSURFTYPE surface_type, UINT alignment,
 345         UINT width, UINT height, UINT level, BOOL lockable, BOOL discard, WINED3DMULTISAMPLE_TYPE multisample_type,
 346         UINT multisample_quality, IWineD3DDeviceImpl *device, DWORD usage, WINED3DFORMAT format,
 347         WINED3DPOOL pool, IUnknown *parent, const struct wined3d_parent_ops *parent_ops)
 348 {
 349     const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
 350     const struct wined3d_format_desc *format_desc = getFormatDescEntry(format, gl_info);
 351     void (*cleanup)(IWineD3DSurfaceImpl *This);
 352     unsigned int resource_size;
 353     HRESULT hr;
 354
 355     if (multisample_quality > 0)
 356     {
 357         FIXME("multisample_quality set to %u, substituting 0\n", multisample_quality);
 358         multisample_quality = 0;
 359     }
 360
 361     /* FIXME: Check that the format is supported by the device. */
 362
 363     resource_size = surface_calculate_size(format_desc, alignment, width, height);
 364
 365     /* Look at the implementation and set the correct Vtable. */
 366     switch (surface_type)
 367     {
 368         case SURFACE_OPENGL:
 369             surface->lpVtbl = &IWineD3DSurface_Vtbl;
 370             cleanup = surface_cleanup;
 371             break;
 372
 373         case SURFACE_GDI:
 374             surface->lpVtbl = &IWineGDISurface_Vtbl;
 375             cleanup = surface_gdi_cleanup;
 376             break;
 377
 378         default:
 379             ERR("Requested unknown surface implementation %#x.\n", surface_type);
 380             return WINED3DERR_INVALIDCALL;
 381     }
 382
 383     hr = resource_init((IWineD3DResource *)surface, WINED3DRTYPE_SURFACE,
 384             device, resource_size, usage, format_desc, pool, parent, parent_ops);
 385     if (FAILED(hr))
 386     {
 387         WARN("Failed to initialize resource, returning %#x.\n", hr);
 388         return hr;
 389     }
 390
 391     /* "Standalone" surface. */
 392     IWineD3DSurface_SetContainer((IWineD3DSurface *)surface, NULL);
 393
 394     surface->currentDesc.Width = width;
 395     surface->currentDesc.Height = height;
 396     surface->currentDesc.MultiSampleType = multisample_type;
 397     surface->currentDesc.MultiSampleQuality = multisample_quality;
 398     surface->texture_level = level;
 399     list_init(&surface->overlays);
 400
 401     /* Flags */
 402     surface->Flags = SFLAG_NORMCOORD; /* Default to normalized coords. */
 403     if (discard) surface->Flags |= SFLAG_DISCARD;
 404     if (lockable || format == WINED3DFMT_D16_LOCKABLE) surface->Flags |= SFLAG_LOCKABLE;
 405
 406     /* Quick lockable sanity check.
 407      * TODO: remove this after surfaces, usage and lockability have been debugged properly
 408      * this function is too deep to need to care about things like this.
 409      * Levels need to be checked too, since they all affect what can be done. */
 410     switch (pool)
 411     {
 412         case WINED3DPOOL_SCRATCH:
 413             if(!lockable)
 414             {
 415                 FIXME("Called with a pool of SCRATCH and a lockable of FALSE "
 416                         "which are mutually exclusive, setting lockable to TRUE.\n");
 417                 lockable = TRUE;
 418             }
 419             break;
 420
 421         case WINED3DPOOL_SYSTEMMEM:
 422             if (!lockable)
 423                 FIXME("Called with a pool of SYSTEMMEM and a lockable of FALSE, this is acceptable but unexpected.\n");
 424             break;
 425
 426         case WINED3DPOOL_MANAGED:
 427             if (usage & WINED3DUSAGE_DYNAMIC)
 428                 FIXME("Called with a pool of MANAGED and a usage of DYNAMIC which are mutually exclusive.\n");
 429             break;
 430
 431         case WINED3DPOOL_DEFAULT:
 432             if (lockable && !(usage & (WINED3DUSAGE_DYNAMIC | WINED3DUSAGE_RENDERTARGET | WINED3DUSAGE_DEPTHSTENCIL)))
 433                 WARN("Creating a lockable surface with a POOL of DEFAULT, that doesn't specify DYNAMIC usage.\n");
 434             break;
 435
 436         default:
 437             FIXME("Unknown pool %#x.\n", pool);
 438             break;
 439     };
 440
 441     if (usage & WINED3DUSAGE_RENDERTARGET && pool != WINED3DPOOL_DEFAULT)
 442     {
 443         FIXME("Trying to create a render target that isn't in the default pool.\n");
 444     }
 445
 446     /* Mark the texture as dirty so that it gets loaded first time around. */
 447     surface_add_dirty_rect(surface, NULL);
 448     list_init(&surface->renderbuffers);
 449
 450     TRACE("surface %p, memory %p, size %u\n", surface, surface->resource.allocatedMemory, surface->resource.size);
 451
 452     /* Call the private setup routine */
 453     hr = IWineD3DSurface_PrivateSetup((IWineD3DSurface *)surface);
 454     if (FAILED(hr))
 455     {
 456         ERR("Private setup failed, returning %#x\n", hr);
 457         cleanup(surface);
 458         return hr;
 459     }
 460
 461     return hr;
 462 }
 463
 464 static void surface_force_reload(IWineD3DSurfaceImpl *surface)
 465 {
 466     surface->Flags &= ~(SFLAG_ALLOCATED | SFLAG_SRGBALLOCATED);
 467 }
 468
 469 void surface_set_texture_name(IWineD3DSurfaceImpl *surface, GLuint new_name, BOOL srgb)
 470 {
 471     GLuint *name;
 472     DWORD flag;
 473
 474     TRACE("surface %p, new_name %u, srgb %#x.\n", surface, new_name, srgb);
 475
 476     if(srgb)
 477     {
 478         name = &surface->texture_name_srgb;
 479         flag = SFLAG_INSRGBTEX;
 480     }
 481     else
 482     {
 483         name = &surface->texture_name;
 484         flag = SFLAG_INTEXTURE;
 485     }
 486
 487     if (!*name && new_name)
 488     {
 489         /* FIXME: We shouldn't need to remove SFLAG_INTEXTURE if the
 490          * surface has no texture name yet. See if we can get rid of this. */
 491         if (surface->Flags & flag)
 492             ERR("Surface has SFLAG_INTEXTURE set, but no texture name\n");
 493         surface_modify_location(surface, flag, FALSE);
 494     }
 495
 496     *name = new_name;
 497     surface_force_reload(surface);
 498 }
 499
 500 void surface_set_texture_target(IWineD3DSurfaceImpl *surface, GLenum target)
 501 {
 502     TRACE("surface %p, target %#x.\n", surface, target);
 503
 504     if (surface->texture_target != target)
 505     {
 506         if (target == GL_TEXTURE_RECTANGLE_ARB)
 507         {
 508             surface->Flags &= ~SFLAG_NORMCOORD;
 509         }
 510         else if (surface->texture_target == GL_TEXTURE_RECTANGLE_ARB)
 511         {
 512             surface->Flags |= SFLAG_NORMCOORD;
 513         }
 514     }
 515     surface->texture_target = target;
 516     surface_force_reload(surface);
 517 }
 518
 519 /* Context activation is done by the caller. */
 520 static void surface_bind_and_dirtify(IWineD3DSurfaceImpl *This, BOOL srgb) {
 521     DWORD active_sampler;
 522
 523     /* We don't need a specific texture unit, but after binding the texture the current unit is dirty.
 524      * Read the unit back instead of switching to 0, this avoids messing around with the state manager's
 525      * gl states. The current texture unit should always be a valid one.
 526      *
 527      * To be more specific, this is tricky because we can implicitly be called
 528      * from sampler() in state.c. This means we can't touch anything other than
 529      * whatever happens to be the currently active texture, or we would risk
 530      * marking already applied sampler states dirty again.
 531      *
 532      * TODO: Track the current active texture per GL context instead of using glGet
 533      */
 534     GLint active_texture;
 535     ENTER_GL();
 536     glGetIntegerv(GL_ACTIVE_TEXTURE, &active_texture);
 537     LEAVE_GL();
 538     active_sampler = This->resource.device->rev_tex_unit_map[active_texture - GL_TEXTURE0_ARB];
 539
 540     if (active_sampler != WINED3D_UNMAPPED_STAGE)
 541     {
 542         IWineD3DDeviceImpl_MarkStateDirty(This->resource.device, STATE_SAMPLER(active_sampler));
 543     }
 544     IWineD3DSurface_BindTexture((IWineD3DSurface *)This, srgb);
 545 }
 546
 547 /* This function checks if the primary render target uses the 8bit paletted format. */
 548 static BOOL primary_render_target_is_p8(IWineD3DDeviceImpl *device)
 549 {
 550     if (device->render_targets && device->render_targets[0])
 551     {
 552         IWineD3DSurfaceImpl *render_target = device->render_targets[0];
 553         if ((render_target->resource.usage & WINED3DUSAGE_RENDERTARGET)
 554                 && (render_target->resource.format_desc->format == WINED3DFMT_P8_UINT))
 555             return TRUE;
 556     }
 557     return FALSE;
 558 }
 559
 560 /* This call just downloads data, the caller is responsible for binding the
 561  * correct texture. */
 562 /* Context activation is done by the caller. */
 563 static void surface_download_data(IWineD3DSurfaceImpl *This, const struct wined3d_gl_info *gl_info)
 564 {
 565     const struct wined3d_format_desc *format_desc = This->resource.format_desc;
 566
 567     /* Only support read back of converted P8 surfaces */
 568     if (This->Flags & SFLAG_CONVERTED && format_desc->format != WINED3DFMT_P8_UINT)
 569     {
 570         FIXME("Read back converted textures unsupported, format=%s\n", debug_d3dformat(format_desc->format));
 571         return;
 572     }
 573
 574     ENTER_GL();
 575
 576     if (format_desc->Flags & WINED3DFMT_FLAG_COMPRESSED)
 577     {
 578         TRACE("(%p) : Calling glGetCompressedTexImageARB level %d, format %#x, type %#x, data %p.\n",
 579                 This, This->texture_level, format_desc->glFormat, format_desc->glType,
 580                 This->resource.allocatedMemory);
 581
 582         if (This->Flags & SFLAG_PBO)
 583         {
 584             GL_EXTCALL(glBindBufferARB(GL_PIXEL_PACK_BUFFER_ARB, This->pbo));
 585             checkGLcall("glBindBufferARB");
 586             GL_EXTCALL(glGetCompressedTexImageARB(This->texture_target, This->texture_level, NULL));
 587             checkGLcall("glGetCompressedTexImageARB");
 588             GL_EXTCALL(glBindBufferARB(GL_PIXEL_PACK_BUFFER_ARB, 0));
 589             checkGLcall("glBindBufferARB");
 590         }
 591         else
 592         {
 593             GL_EXTCALL(glGetCompressedTexImageARB(This->texture_target,
 594                     This->texture_level, This->resource.allocatedMemory));
 595             checkGLcall("glGetCompressedTexImageARB");
 596         }
 597
 598         LEAVE_GL();
 599     } else {
 600         void *mem;
 601         GLenum format = format_desc->glFormat;
 602         GLenum type = format_desc->glType;
 603         int src_pitch = 0;
 604         int dst_pitch = 0;
 605
 606         /* In case of P8 the index is stored in the alpha component if the primary render target uses P8 */
 607         if (format_desc->format == WINED3DFMT_P8_UINT && primary_render_target_is_p8(This->resource.device))
 608         {
 609             format = GL_ALPHA;
 610             type = GL_UNSIGNED_BYTE;
 611         }
 612
 613         if (This->Flags & SFLAG_NONPOW2) {
 614             unsigned char alignment = This->resource.device->surface_alignment;
 615             src_pitch = format_desc->byte_count * This->pow2Width;
 616             dst_pitch = IWineD3DSurface_GetPitch((IWineD3DSurface *) This);
 617             src_pitch = (src_pitch + alignment - 1) & ~(alignment - 1);
 618             mem = HeapAlloc(GetProcessHeap(), 0, src_pitch * This->pow2Height);
 619         } else {
 620             mem = This->resource.allocatedMemory;
 621         }
 622
 623         TRACE("(%p) : Calling glGetTexImage level %d, format %#x, type %#x, data %p\n",
 624                 This, This->texture_level, format, type, mem);
 625
 626         if(This->Flags & SFLAG_PBO) {
 627             GL_EXTCALL(glBindBufferARB(GL_PIXEL_PACK_BUFFER_ARB, This->pbo));
 628             checkGLcall("glBindBufferARB");
 629
 630             glGetTexImage(This->texture_target, This->texture_level, format, type, NULL);
 631             checkGLcall("glGetTexImage");
 632
 633             GL_EXTCALL(glBindBufferARB(GL_PIXEL_PACK_BUFFER_ARB, 0));
 634             checkGLcall("glBindBufferARB");
 635         } else {
 636             glGetTexImage(This->texture_target, This->texture_level, format, type, mem);
 637             checkGLcall("glGetTexImage");
 638         }
 639         LEAVE_GL();
 640
 641         if (This->Flags & SFLAG_NONPOW2) {
 642             const BYTE *src_data;
 643             BYTE *dst_data;
 644             UINT y;
 645             /*
 646              * Some games (e.g. warhammer 40k) don't work properly with the odd pitches, preventing
 647              * the surface pitch from being used to box non-power2 textures. Instead we have to use a hack to
 648              * repack the texture so that the bpp * width pitch can be used instead of bpp * pow2width.
 649              *
 650              * We're doing this...
 651              *
 652              * instead of boxing the texture :
 653              * |<-texture width ->|  -->pow2width|   /\
 654              * |111111111111111111|              |   |
 655              * |222 Texture 222222| boxed empty  | texture height
 656              * |3333 Data 33333333|              |   |
 657              * |444444444444444444|              |   \/
 658              * -----------------------------------   |
 659              * |     boxed  empty | boxed empty  | pow2height
 660              * |                  |              |   \/
 661              * -----------------------------------
 662              *
 663              *
 664              * we're repacking the data to the expected texture width
 665              *
 666              * |<-texture width ->|  -->pow2width|   /\
 667              * |111111111111111111222222222222222|   |
 668              * |222333333333333333333444444444444| texture height
 669              * |444444                           |   |
 670              * |                                 |   \/
 671              * |                                 |   |
 672              * |            empty                | pow2height
 673              * |                                 |   \/
 674              * -----------------------------------
 675              *
 676              * == is the same as
 677              *
 678              * |<-texture width ->|    /\
 679              * |111111111111111111|
 680              * |222222222222222222|texture height
 681              * |333333333333333333|
 682              * |444444444444444444|    \/
 683              * --------------------
 684              *
 685              * this also means that any references to allocatedMemory should work with the data as if were a
 686              * standard texture with a non-power2 width instead of texture boxed up to be a power2 texture.
 687              *
 688              * internally the texture is still stored in a boxed format so any references to textureName will
 689              * get a boxed texture with width pow2width and not a texture of width currentDesc.Width.
 690              *
 691              * Performance should not be an issue, because applications normally do not lock the surfaces when
 692              * rendering. If an app does, the SFLAG_DYNLOCK flag will kick in and the memory copy won't be released,
 693              * and doesn't have to be re-read.
 694              */
 695             src_data = mem;
 696             dst_data = This->resource.allocatedMemory;
 697             TRACE("(%p) : Repacking the surface data from pitch %d to pitch %d\n", This, src_pitch, dst_pitch);
 698             for (y = 1 ; y < This->currentDesc.Height; y++) {
 699                 /* skip the first row */
 700                 src_data += src_pitch;
 701                 dst_data += dst_pitch;
 702                 memcpy(dst_data, src_data, dst_pitch);
 703             }
 704
 705             HeapFree(GetProcessHeap(), 0, mem);
 706         }
 707     }
 708
 709     /* Surface has now been downloaded */
 710     This->Flags |= SFLAG_INSYSMEM;
 711 }
 712
 713 /* This call just uploads data, the caller is responsible for binding the
 714  * correct texture. */
 715 /* Context activation is done by the caller. */
 716 static void surface_upload_data(IWineD3DSurfaceImpl *This, const struct wined3d_gl_info *gl_info,
 717         const struct wined3d_format_desc *format_desc, BOOL srgb, const GLvoid *data)
 718 {
 719     GLsizei width = This->currentDesc.Width;
 720     GLsizei height = This->currentDesc.Height;
 721     GLenum internal;
 722
 723     if (srgb)
 724     {
 725         internal = format_desc->glGammaInternal;
 726     }
 727     else if (This->resource.usage & WINED3DUSAGE_RENDERTARGET && surface_is_offscreen(This))
 728     {
 729         internal = format_desc->rtInternal;
 730     }
 731     else
 732     {
 733         internal = format_desc->glInternal;
 734     }
 735
 736     TRACE("This %p, internal %#x, width %d, height %d, format %#x, type %#x, data %p.\n",
 737             This, internal, width, height, format_desc->glFormat, format_desc->glType, data);
 738     TRACE("target %#x, level %u, resource size %u.\n",
 739             This->texture_target, This->texture_level, This->resource.size);
 740
 741     if (format_desc->heightscale != 1.0f && format_desc->heightscale != 0.0f) height *= format_desc->heightscale;
 742
 743     ENTER_GL();
 744
 745     if (This->Flags & SFLAG_PBO)
 746     {
 747         GL_EXTCALL(glBindBufferARB(GL_PIXEL_UNPACK_BUFFER_ARB, This->pbo));
 748         checkGLcall("glBindBufferARB");
 749
 750         TRACE("(%p) pbo: %#x, data: %p.\n", This, This->pbo, data);
 751         data = NULL;
 752     }
 753
 754     if (format_desc->Flags & WINED3DFMT_FLAG_COMPRESSED)
 755     {
 756         TRACE("Calling glCompressedTexSubImage2DARB.\n");
 757
 758         GL_EXTCALL(glCompressedTexSubImage2DARB(This->texture_target, This->texture_level,
 759                 0, 0, width, height, internal, This->resource.size, data));
 760         checkGLcall("glCompressedTexSubImage2DARB");
 761     }
 762     else
 763     {
 764         TRACE("Calling glTexSubImage2D.\n");
 765
 766         glTexSubImage2D(This->texture_target, This->texture_level,
 767                 0, 0, width, height, format_desc->glFormat, format_desc->glType, data);
 768         checkGLcall("glTexSubImage2D");
 769     }
 770
 771     if (This->Flags & SFLAG_PBO)
 772     {
 773         GL_EXTCALL(glBindBufferARB(GL_PIXEL_UNPACK_BUFFER_ARB, 0));
 774         checkGLcall("glBindBufferARB");
 775     }
 776
 777     LEAVE_GL();
 778
 779     if (gl_info->quirks & WINED3D_QUIRK_FBO_TEX_UPDATE)
 780     {
 781         IWineD3DDeviceImpl *device = This->resource.device;
 782         unsigned int i;
 783
 784         for (i = 0; i < device->numContexts; ++i)
 785         {
 786             context_surface_update(device->contexts[i], This);
 787         }
 788     }
 789 }
 790
 791 /* This call just allocates the texture, the caller is responsible for binding
 792  * the correct texture. */
 793 /* Context activation is done by the caller. */
 794 static void surface_allocate_surface(IWineD3DSurfaceImpl *This, const struct wined3d_gl_info *gl_info,
 795         const struct wined3d_format_desc *format_desc, BOOL srgb)
 796 {
 797     BOOL enable_client_storage = FALSE;
 798     GLsizei width = This->pow2Width;
 799     GLsizei height = This->pow2Height;
 800     const BYTE *mem = NULL;
 801     GLenum internal;
 802
 803     if (srgb)
 804     {
 805         internal = format_desc->glGammaInternal;
 806     }
 807     else if (This->resource.usage & WINED3DUSAGE_RENDERTARGET && surface_is_offscreen(This))
 808     {
 809         internal = format_desc->rtInternal;
 810     }
 811     else
 812     {
 813         internal = format_desc->glInternal;
 814     }
 815
 816     if (format_desc->heightscale != 1.0f && format_desc->heightscale != 0.0f) height *= format_desc->heightscale;
 817
 818     TRACE("(%p) : Creating surface (target %#x)  level %d, d3d format %s, internal format %#x, width %d, height %d, gl format %#x, gl type=%#x\n",
 819             This, This->texture_target, This->texture_level, debug_d3dformat(format_desc->format),
 820             internal, width, height, format_desc->glFormat, format_desc->glType);
 821
 822     ENTER_GL();
 823
 824     if (gl_info->supported[APPLE_CLIENT_STORAGE])
 825     {
 826         if(This->Flags & (SFLAG_NONPOW2 | SFLAG_DIBSECTION | SFLAG_CONVERTED) || This->resource.allocatedMemory == NULL) {
 827             /* In some cases we want to disable client storage.
 828              * SFLAG_NONPOW2 has a bigger opengl texture than the client memory, and different pitches
 829              * SFLAG_DIBSECTION: Dibsections may have read / write protections on the memory. Avoid issues...
 830              * SFLAG_CONVERTED: The conversion destination memory is freed after loading the surface
 831              * allocatedMemory == NULL: Not defined in the extension. Seems to disable client storage effectively
 832              */
 833             glPixelStorei(GL_UNPACK_CLIENT_STORAGE_APPLE, GL_FALSE);
 834             checkGLcall("glPixelStorei(GL_UNPACK_CLIENT_STORAGE_APPLE, GL_FALSE)");
 835             This->Flags &= ~SFLAG_CLIENT;
 836             enable_client_storage = TRUE;
 837         } else {
 838             This->Flags |= SFLAG_CLIENT;
 839
 840             /* Point opengl to our allocated texture memory. Do not use resource.allocatedMemory here because
 841              * it might point into a pbo. Instead use heapMemory, but get the alignment right.
 842              */
 843             mem = (BYTE *)(((ULONG_PTR) This->resource.heapMemory + (RESOURCE_ALIGNMENT - 1)) & ~(RESOURCE_ALIGNMENT - 1));
 844         }
 845     }
 846
 847     if (format_desc->Flags & WINED3DFMT_FLAG_COMPRESSED && mem)
 848     {
 849         GL_EXTCALL(glCompressedTexImage2DARB(This->texture_target, This->texture_level,
 850                 internal, width, height, 0, This->resource.size, mem));
 851         checkGLcall("glCompressedTexImage2DARB");
 852     }
 853     else
 854     {
 855         glTexImage2D(This->texture_target, This->texture_level,
 856                 internal, width, height, 0, format_desc->glFormat, format_desc->glType, mem);
 857         checkGLcall("glTexImage2D");
 858     }
 859
 860     if(enable_client_storage) {
 861         glPixelStorei(GL_UNPACK_CLIENT_STORAGE_APPLE, GL_TRUE);
 862         checkGLcall("glPixelStorei(GL_UNPACK_CLIENT_STORAGE_APPLE, GL_TRUE)");
 863     }
 864     LEAVE_GL();
 865 }
 866
 867 /* In D3D the depth stencil dimensions have to be greater than or equal to the
 868  * render target dimensions. With FBOs, the dimensions have to be an exact match. */
 869 /* TODO: We should synchronize the renderbuffer's content with the texture's content. */
 870 /* GL locking is done by the caller */
 871 void surface_set_compatible_renderbuffer(IWineD3DSurfaceImpl *surface, unsigned int width, unsigned int height)
 872 {
 873     const struct wined3d_gl_info *gl_info = &surface->resource.device->adapter->gl_info;
 874     renderbuffer_entry_t *entry;
 875     GLuint renderbuffer = 0;
 876     unsigned int src_width, src_height;
 877
 878     src_width = surface->pow2Width;
 879     src_height = surface->pow2Height;
 880
 881     /* A depth stencil smaller than the render target is not valid */
 882     if (width > src_width || height > src_height) return;
 883
 884     /* Remove any renderbuffer set if the sizes match */
 885     if (gl_info->supported[ARB_FRAMEBUFFER_OBJECT]
 886             || (width == src_width && height == src_height))
 887     {
 888         surface->current_renderbuffer = NULL;
 889         return;
 890     }
 891
 892     /* Look if we've already got a renderbuffer of the correct dimensions */
 893     LIST_FOR_EACH_ENTRY(entry, &surface->renderbuffers, renderbuffer_entry_t, entry)
 894     {
 895         if (entry->width == width && entry->height == height)
 896         {
 897             renderbuffer = entry->id;
 898             surface->current_renderbuffer = entry;
 899             break;
 900         }
 901     }
 902
 903     if (!renderbuffer)
 904     {
 905         gl_info->fbo_ops.glGenRenderbuffers(1, &renderbuffer);
 906         gl_info->fbo_ops.glBindRenderbuffer(GL_RENDERBUFFER, renderbuffer);
 907         gl_info->fbo_ops.glRenderbufferStorage(GL_RENDERBUFFER,
 908                 surface->resource.format_desc->glInternal, width, height);
 909
 910         entry = HeapAlloc(GetProcessHeap(), 0, sizeof(renderbuffer_entry_t));
 911         entry->width = width;
 912         entry->height = height;
 913         entry->id = renderbuffer;
 914         list_add_head(&surface->renderbuffers, &entry->entry);
 915
 916         surface->current_renderbuffer = entry;
 917     }
 918
 919     checkGLcall("set_compatible_renderbuffer");
 920 }
 921
 922 GLenum surface_get_gl_buffer(IWineD3DSurfaceImpl *surface)
 923 {
 924     IWineD3DSwapChainImpl *swapchain = (IWineD3DSwapChainImpl *)surface->container;
 925
 926     TRACE("surface %p.\n", surface);
 927
 928     if (!(surface->Flags & SFLAG_SWAPCHAIN))
 929     {
 930         ERR("Surface %p is not on a swapchain.\n", surface);
 931         return GL_NONE;
 932     }
 933
 934     if (swapchain->back_buffers && swapchain->back_buffers[0] == surface)
 935     {
 936         if (swapchain->render_to_fbo)
 937         {
 938             TRACE("Returning GL_COLOR_ATTACHMENT0\n");
 939             return GL_COLOR_ATTACHMENT0;
 940         }
 941         TRACE("Returning GL_BACK\n");
 942         return GL_BACK;
 943     }
 944     else if (surface == swapchain->front_buffer)
 945     {
 946         TRACE("Returning GL_FRONT\n");
 947         return GL_FRONT;
 948     }
 949
 950     FIXME("Higher back buffer, returning GL_BACK\n");
 951     return GL_BACK;
 952 }
 953
 954 /* Slightly inefficient way to handle multiple dirty rects but it works :) */
 955 void surface_add_dirty_rect(IWineD3DSurfaceImpl *surface, const RECT *dirty_rect)
 956 {
 957     IWineD3DBaseTexture *baseTexture = NULL;
 958
 959     TRACE("surface %p, dirty_rect %s.\n", surface, wine_dbgstr_rect(dirty_rect));
 960
 961     if (!(surface->Flags & SFLAG_INSYSMEM) && (surface->Flags & SFLAG_INTEXTURE))
 962         /* No partial locking for textures yet. */
 963         surface_load_location(surface, SFLAG_INSYSMEM, NULL);
 964
 965     surface_modify_location(surface, SFLAG_INSYSMEM, TRUE);
 966     if (dirty_rect)
 967     {
 968         surface->dirtyRect.left = min(surface->dirtyRect.left, dirty_rect->left);
 969         surface->dirtyRect.top = min(surface->dirtyRect.top, dirty_rect->top);
 970         surface->dirtyRect.right = max(surface->dirtyRect.right, dirty_rect->right);
 971         surface->dirtyRect.bottom = max(surface->dirtyRect.bottom, dirty_rect->bottom);
 972     }
 973     else
 974     {
 975         surface->dirtyRect.left = 0;
 976         surface->dirtyRect.top = 0;
 977         surface->dirtyRect.right = surface->currentDesc.Width;
 978         surface->dirtyRect.bottom = surface->currentDesc.Height;
 979     }
 980
 981     /* if the container is a basetexture then mark it dirty. */
 982     if (SUCCEEDED(IWineD3DSurface_GetContainer((IWineD3DSurface *)surface,
 983             &IID_IWineD3DBaseTexture, (void **)&baseTexture)))
 984     {
 985         TRACE("Passing to container\n");
 986         IWineD3DBaseTexture_SetDirty(baseTexture, TRUE);
 987         IWineD3DBaseTexture_Release(baseTexture);
 988     }
 989 }
 990
 991 static BOOL surface_convert_color_to_argb(IWineD3DSurfaceImpl *This, DWORD color, DWORD *argb_color)
 992 {
 993     IWineD3DDeviceImpl *device = This->resource.device;
 994
 995     switch(This->resource.format_desc->format)
 996     {
 997         case WINED3DFMT_P8_UINT:
 998             {
 999                 DWORD alpha;
1000
1001                 if (primary_render_target_is_p8(device))
1002                     alpha = color << 24;
1003                 else
1004                     alpha = 0xFF000000;
1005
1006                 if (This->palette) {
1007                     *argb_color = (alpha |
1008                             (This->palette->palents[color].peRed << 16) |
1009                             (This->palette->palents[color].peGreen << 8) |
1010                             (This->palette->palents[color].peBlue));
1011                 } else {
1012                     *argb_color = alpha;
1013                 }
1014             }
1015             break;
1016
1017         case WINED3DFMT_B5G6R5_UNORM:
1018             {
1019                 if (color == 0xFFFF) {
1020                     *argb_color = 0xFFFFFFFF;
1021                 } else {
1022                     *argb_color = ((0xFF000000) |
1023                             ((color & 0xF800) << 8) |
1024                             ((color & 0x07E0) << 5) |
1025                             ((color & 0x001F) << 3));
1026                 }
1027             }
1028             break;
1029
1030         case WINED3DFMT_B8G8R8_UNORM:
1031         case WINED3DFMT_B8G8R8X8_UNORM:
1032             *argb_color = 0xFF000000 | color;
1033             break;
1034
1035         case WINED3DFMT_B8G8R8A8_UNORM:
1036             *argb_color = color;
1037             break;
1038
1039         default:
1040             ERR("Unhandled conversion from %s to ARGB!\n", debug_d3dformat(This->resource.format_desc->format));
1041             return FALSE;
1042     }
1043     return TRUE;
1044 }
1045
1046 static ULONG WINAPI IWineD3DSurfaceImpl_Release(IWineD3DSurface *iface)
1047 {
1048     IWineD3DSurfaceImpl *This = (IWineD3DSurfaceImpl *)iface;
1049     ULONG ref = InterlockedDecrement(&This->resource.ref);
1050     TRACE("(%p) : Releasing from %d\n", This, ref + 1);
1051
1052     if (!ref)
1053     {
1054         surface_cleanup(This);
1055         This->resource.parent_ops->wined3d_object_destroyed(This->resource.parent);
1056
1057         TRACE("(%p) Released.\n", This);
1058         HeapFree(GetProcessHeap(), 0, This);
1059     }
1060
1061     return ref;
1062 }
1063
1064 /* ****************************************************
1065    IWineD3DSurface IWineD3DResource parts follow
1066    **************************************************** */
1067
1068 void surface_internal_preload(IWineD3DSurfaceImpl *surface, enum WINED3DSRGB srgb)
1069 {
1070     /* TODO: check for locks */
1071     IWineD3DDeviceImpl *device = surface->resource.device;
1072     IWineD3DBaseTexture *baseTexture = NULL;
1073
1074     TRACE("(%p)Checking to see if the container is a base texture\n", surface);
1075     if (SUCCEEDED(IWineD3DSurface_GetContainer((IWineD3DSurface *)surface,
1076             &IID_IWineD3DBaseTexture, (void **)&baseTexture)))
1077     {
1078         IWineD3DBaseTextureImpl *tex_impl = (IWineD3DBaseTextureImpl *)baseTexture;
1079         TRACE("Passing to container\n");
1080         tex_impl->baseTexture.internal_preload(baseTexture, srgb);
1081         IWineD3DBaseTexture_Release(baseTexture);
1082     } else {
1083         struct wined3d_context *context = NULL;
1084
1085         TRACE("(%p) : About to load surface\n", surface);
1086
1087         if (!device->isInDraw) context = context_acquire(device, NULL);
1088
1089         if (surface->resource.format_desc->format == WINED3DFMT_P8_UINT
1090                 || surface->resource.format_desc->format == WINED3DFMT_P8_UINT_A8_UNORM)
1091         {
1092             if (palette9_changed(surface))
1093             {
1094                 TRACE("Reloading surface because the d3d8/9 palette was changed\n");
1095                 /* TODO: This is not necessarily needed with hw palettized texture support */
1096                 surface_load_location(surface, SFLAG_INSYSMEM, NULL);
1097                 /* Make sure the texture is reloaded because of the palette change, this kills performance though :( */
1098                 surface_modify_location(surface, SFLAG_INTEXTURE, FALSE);
1099             }
1100         }
1101
1102         IWineD3DSurface_LoadTexture((IWineD3DSurface *)surface, srgb == SRGB_SRGB ? TRUE : FALSE);
1103
1104         if (surface->resource.pool == WINED3DPOOL_DEFAULT)
1105         {
1106             /* Tell opengl to try and keep this texture in video ram (well mostly) */
1107             GLclampf tmp;
1108             tmp = 0.9f;
1109             ENTER_GL();
1110             glPrioritizeTextures(1, &surface->texture_name, &tmp);
1111             LEAVE_GL();
1112         }
1113
1114         if (context) context_release(context);
1115     }
1116 }
1117
1118 static void WINAPI IWineD3DSurfaceImpl_PreLoad(IWineD3DSurface *iface)
1119 {
1120     surface_internal_preload((IWineD3DSurfaceImpl *)iface, SRGB_ANY);
1121 }
1122
1123 /* Context activation is done by the caller. */
1124 static void surface_remove_pbo(IWineD3DSurfaceImpl *This, const struct wined3d_gl_info *gl_info)
1125 {
1126     This->resource.heapMemory = HeapAlloc(GetProcessHeap() ,0 , This->resource.size + RESOURCE_ALIGNMENT);
1127     This->resource.allocatedMemory =
1128             (BYTE *)(((ULONG_PTR) This->resource.heapMemory + (RESOURCE_ALIGNMENT - 1)) & ~(RESOURCE_ALIGNMENT - 1));
1129
1130     ENTER_GL();
1131     GL_EXTCALL(glBindBufferARB(GL_PIXEL_UNPACK_BUFFER_ARB, This->pbo));
1132     checkGLcall("glBindBufferARB(GL_PIXEL_UNPACK_BUFFER, This->pbo)");
1133     GL_EXTCALL(glGetBufferSubDataARB(GL_PIXEL_UNPACK_BUFFER_ARB, 0, This->resource.size, This->resource.allocatedMemory));
1134     checkGLcall("glGetBufferSubDataARB");
1135     GL_EXTCALL(glDeleteBuffersARB(1, &This->pbo));
1136     checkGLcall("glDeleteBuffersARB");
1137     LEAVE_GL();
1138
1139     This->pbo = 0;
1140     This->Flags &= ~SFLAG_PBO;
1141 }
1142
1143 BOOL surface_init_sysmem(IWineD3DSurfaceImpl *surface)
1144 {
1145     if (!surface->resource.allocatedMemory)
1146     {
1147         surface->resource.heapMemory = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY,
1148                 surface->resource.size + RESOURCE_ALIGNMENT);
1149         if (!surface->resource.heapMemory)
1150         {
1151             ERR("Out of memory\n");
1152             return FALSE;
1153         }
1154         surface->resource.allocatedMemory =
1155             (BYTE *)(((ULONG_PTR)surface->resource.heapMemory + (RESOURCE_ALIGNMENT - 1)) & ~(RESOURCE_ALIGNMENT - 1));
1156     }
1157     else
1158     {
1159         memset(surface->resource.allocatedMemory, 0, surface->resource.size);
1160     }
1161
1162     surface_modify_location(surface, SFLAG_INSYSMEM, TRUE);
1163
1164     return TRUE;
1165 }
1166
1167 static void WINAPI IWineD3DSurfaceImpl_UnLoad(IWineD3DSurface *iface) {
1168     IWineD3DBaseTexture *texture = NULL;
1169     IWineD3DSurfaceImpl *This = (IWineD3DSurfaceImpl *) iface;
1170     IWineD3DDeviceImpl *device = This->resource.device;
1171     const struct wined3d_gl_info *gl_info;
1172     renderbuffer_entry_t *entry, *entry2;
1173     struct wined3d_context *context;
1174
1175     TRACE("(%p)\n", iface);
1176
1177     if(This->resource.pool == WINED3DPOOL_DEFAULT) {
1178         /* Default pool resources are supposed to be destroyed before Reset is called.
1179          * Implicit resources stay however. So this means we have an implicit render target
1180          * or depth stencil. The content may be destroyed, but we still have to tear down
1181          * opengl resources, so we cannot leave early.
1182          *
1183          * Put the surfaces into sysmem, and reset the content. The D3D content is undefined,
1184          * but we can't set the sysmem INDRAWABLE because when we're rendering the swapchain
1185          * or the depth stencil into an FBO the texture or render buffer will be removed
1186          * and all flags get lost
1187          */
1188         surface_init_sysmem(This);
1189     }
1190     else
1191     {
1192         /* Load the surface into system memory */
1193         surface_load_location(This, SFLAG_INSYSMEM, NULL);
1194         surface_modify_location(This, SFLAG_INDRAWABLE, FALSE);
1195     }
1196     surface_modify_location(This, SFLAG_INTEXTURE, FALSE);
1197     surface_modify_location(This, SFLAG_INSRGBTEX, FALSE);
1198     This->Flags &= ~(SFLAG_ALLOCATED | SFLAG_SRGBALLOCATED);
1199
1200     context = context_acquire(device, NULL);
1201     gl_info = context->gl_info;
1202
1203     /* Destroy PBOs, but load them into real sysmem before */
1204     if (This->Flags & SFLAG_PBO)
1205         surface_remove_pbo(This, gl_info);
1206
1207     /* Destroy fbo render buffers. This is needed for implicit render targets, for
1208      * all application-created targets the application has to release the surface
1209      * before calling _Reset
1210      */
1211     LIST_FOR_EACH_ENTRY_SAFE(entry, entry2, &This->renderbuffers, renderbuffer_entry_t, entry) {
1212         ENTER_GL();
1213         gl_info->fbo_ops.glDeleteRenderbuffers(1, &entry->id);
1214         LEAVE_GL();
1215         list_remove(&entry->entry);
1216         HeapFree(GetProcessHeap(), 0, entry);
1217     }
1218     list_init(&This->renderbuffers);
1219     This->current_renderbuffer = NULL;
1220
1221     /* If we're in a texture, the texture name belongs to the texture. Otherwise,
1222      * destroy it
1223      */
1224     IWineD3DSurface_GetContainer(iface, &IID_IWineD3DBaseTexture, (void **) &texture);
1225     if(!texture) {
1226         ENTER_GL();
1227         glDeleteTextures(1, &This->texture_name);
1228         This->texture_name = 0;
1229         glDeleteTextures(1, &This->texture_name_srgb);
1230         This->texture_name_srgb = 0;
1231         LEAVE_GL();
1232     } else {
1233         IWineD3DBaseTexture_Release(texture);
1234     }
1235
1236     context_release(context);
1237 }
1238
1239 /* ******************************************************
1240    IWineD3DSurface IWineD3DSurface parts follow
1241    ****************************************************** */
1242
1243 /* Read the framebuffer back into the surface */
1244 static void read_from_framebuffer(IWineD3DSurfaceImpl *This, const RECT *rect, void *dest, UINT pitch)
1245 {
1246     IWineD3DDeviceImpl *device = This->resource.device;
1247     const struct wined3d_gl_info *gl_info;
1248     struct wined3d_context *context;
1249     BYTE *mem;
1250     GLint fmt;
1251     GLint type;
1252     BYTE *row, *top, *bottom;
1253     int i;
1254     BOOL bpp;
1255     RECT local_rect;
1256     BOOL srcIsUpsideDown;
1257     GLint rowLen = 0;
1258     GLint skipPix = 0;
1259     GLint skipRow = 0;
1260
1261     if(wined3d_settings.rendertargetlock_mode == RTL_DISABLE) {
1262         static BOOL warned = FALSE;
1263         if(!warned) {
1264             ERR("The application tries to lock the render target, but render target locking is disabled\n");
1265             warned = TRUE;
1266         }
1267         return;
1268     }
1269
1270     /* Activate the surface. Set it up for blitting now, although not necessarily needed for LockRect.
1271      * Certain graphics drivers seem to dislike some enabled states when reading from opengl, the blitting usage
1272      * should help here. Furthermore unlockrect will need the context set up for blitting. The context manager will find
1273      * context->last_was_blit set on the unlock.
1274      */
1275     context = context_acquire(device, This);
1276     context_apply_blit_state(context, device);
1277     gl_info = context->gl_info;
1278
1279     ENTER_GL();
1280
1281     /* Select the correct read buffer, and give some debug output.
1282      * There is no need to keep track of the current read buffer or reset it, every part of the code
1283      * that reads sets the read buffer as desired.
1284      */
1285     if (surface_is_offscreen(This))
1286     {
1287         /* Locking the primary render target which is not on a swapchain(=offscreen render target).
1288          * Read from the back buffer
1289          */
1290         TRACE("Locking offscreen render target\n");
1291         glReadBuffer(device->offscreenBuffer);
1292         srcIsUpsideDown = TRUE;
1293     }
1294     else
1295     {
1296         /* Onscreen surfaces are always part of a swapchain */
1297         GLenum buffer = surface_get_gl_buffer(This);
1298         TRACE("Locking %#x buffer\n", buffer);
1299         glReadBuffer(buffer);
1300         checkGLcall("glReadBuffer");
1301         srcIsUpsideDown = FALSE;
1302     }
1303
1304     /* TODO: Get rid of the extra rectangle comparison and construction of a full surface rectangle */
1305     if(!rect) {
1306         local_rect.left = 0;
1307         local_rect.top = 0;
1308         local_rect.right = This->currentDesc.Width;
1309         local_rect.bottom = This->currentDesc.Height;
1310     } else {
1311         local_rect = *rect;
1312     }
1313     /* TODO: Get rid of the extra GetPitch call, LockRect does that too. Cache the pitch */
1314
1315     switch(This->resource.format_desc->format)
1316     {
1317         case WINED3DFMT_P8_UINT:
1318         {
1319             if (primary_render_target_is_p8(device))
1320             {
1321                 /* In case of P8 render targets the index is stored in the alpha component */
1322                 fmt = GL_ALPHA;
1323                 type = GL_UNSIGNED_BYTE;
1324                 mem = dest;
1325                 bpp = This->resource.format_desc->byte_count;
1326             } else {
1327                 /* GL can't return palettized data, so read ARGB pixels into a
1328                  * separate block of memory and convert them into palettized format
1329                  * in software. Slow, but if the app means to use palettized render
1330                  * targets and locks it...
1331                  *
1332                  * Use GL_RGB, GL_UNSIGNED_BYTE to read the surface for performance reasons
1333                  * Don't use GL_BGR as in the WINED3DFMT_R8G8B8 case, instead watch out
1334                  * for the color channels when palettizing the colors.
1335                  */
1336                 fmt = GL_RGB;
1337                 type = GL_UNSIGNED_BYTE;
1338                 pitch *= 3;
1339                 mem = HeapAlloc(GetProcessHeap(), 0, This->resource.size * 3);
1340                 if(!mem) {
1341                     ERR("Out of memory\n");
1342                     LEAVE_GL();
1343                     return;
1344                 }
1345                 bpp = This->resource.format_desc->byte_count * 3;
1346             }
1347         }
1348         break;
1349
1350         default:
1351             mem = dest;
1352             fmt = This->resource.format_desc->glFormat;
1353             type = This->resource.format_desc->glType;
1354             bpp = This->resource.format_desc->byte_count;
1355     }
1356
1357     if(This->Flags & SFLAG_PBO) {
1358         GL_EXTCALL(glBindBufferARB(GL_PIXEL_PACK_BUFFER_ARB, This->pbo));
1359         checkGLcall("glBindBufferARB");
1360         if(mem != NULL) {
1361             ERR("mem not null for pbo -- unexpected\n");
1362             mem = NULL;
1363         }
1364     }
1365
1366     /* Save old pixel store pack state */
1367     glGetIntegerv(GL_PACK_ROW_LENGTH, &rowLen);
1368     checkGLcall("glGetIntegerv");
1369     glGetIntegerv(GL_PACK_SKIP_PIXELS, &skipPix);
1370     checkGLcall("glGetIntegerv");
1371     glGetIntegerv(GL_PACK_SKIP_ROWS, &skipRow);
1372     checkGLcall("glGetIntegerv");
1373
1374     /* Setup pixel store pack state -- to glReadPixels into the correct place */
1375     glPixelStorei(GL_PACK_ROW_LENGTH, This->currentDesc.Width);
1376     checkGLcall("glPixelStorei");
1377     glPixelStorei(GL_PACK_SKIP_PIXELS, local_rect.left);
1378     checkGLcall("glPixelStorei");
1379     glPixelStorei(GL_PACK_SKIP_ROWS, local_rect.top);
1380     checkGLcall("glPixelStorei");
1381
1382     glReadPixels(local_rect.left, (!srcIsUpsideDown) ? (This->currentDesc.Height - local_rect.bottom) : local_rect.top ,
1383                  local_rect.right - local_rect.left,
1384                  local_rect.bottom - local_rect.top,
1385                  fmt, type, mem);
1386     checkGLcall("glReadPixels");
1387
1388     /* Reset previous pixel store pack state */
1389     glPixelStorei(GL_PACK_ROW_LENGTH, rowLen);
1390     checkGLcall("glPixelStorei");
1391     glPixelStorei(GL_PACK_SKIP_PIXELS, skipPix);
1392     checkGLcall("glPixelStorei");
1393     glPixelStorei(GL_PACK_SKIP_ROWS, skipRow);
1394     checkGLcall("glPixelStorei");
1395
1396     if(This->Flags & SFLAG_PBO) {
1397         GL_EXTCALL(glBindBufferARB(GL_PIXEL_PACK_BUFFER_ARB, 0));
1398         checkGLcall("glBindBufferARB");
1399
1400         /* Check if we need to flip the image. If we need to flip use glMapBufferARB
1401          * to get a pointer to it and perform the flipping in software. This is a lot
1402          * faster than calling glReadPixels for each line. In case we want more speed
1403          * we should rerender it flipped in a FBO and read the data back from the FBO. */
1404         if(!srcIsUpsideDown) {
1405             GL_EXTCALL(glBindBufferARB(GL_PIXEL_UNPACK_BUFFER_ARB, This->pbo));
1406             checkGLcall("glBindBufferARB");
1407
1408             mem = GL_EXTCALL(glMapBufferARB(GL_PIXEL_UNPACK_BUFFER_ARB, GL_READ_WRITE_ARB));
1409             checkGLcall("glMapBufferARB");
1410         }
1411     }
1412
1413     /* TODO: Merge this with the palettization loop below for P8 targets */
1414     if(!srcIsUpsideDown) {
1415         UINT len, off;
1416         /* glReadPixels returns the image upside down, and there is no way to prevent this.
1417             Flip the lines in software */
1418         len = (local_rect.right - local_rect.left) * bpp;
1419         off = local_rect.left * bpp;
1420
1421         row = HeapAlloc(GetProcessHeap(), 0, len);
1422         if(!row) {
1423             ERR("Out of memory\n");
1424             if (This->resource.format_desc->format == WINED3DFMT_P8_UINT) HeapFree(GetProcessHeap(), 0, mem);
1425             LEAVE_GL();
1426             return;
1427         }
1428
1429         top = mem + pitch * local_rect.top;
1430         bottom = mem + pitch * (local_rect.bottom - 1);
1431         for(i = 0; i < (local_rect.bottom - local_rect.top) / 2; i++) {
1432             memcpy(row, top + off, len);
1433             memcpy(top + off, bottom + off, len);
1434             memcpy(bottom + off, row, len);
1435             top += pitch;
1436             bottom -= pitch;
1437         }
1438         HeapFree(GetProcessHeap(), 0, row);
1439
1440         /* Unmap the temp PBO buffer */
1441         if(This->Flags & SFLAG_PBO) {
1442             GL_EXTCALL(glUnmapBufferARB(GL_PIXEL_UNPACK_BUFFER_ARB));
1443             GL_EXTCALL(glBindBufferARB(GL_PIXEL_UNPACK_BUFFER_ARB, 0));
1444         }
1445     }
1446
1447     LEAVE_GL();
1448     context_release(context);
1449
1450     /* For P8 textures we need to perform an inverse palette lookup. This is done by searching for a palette
1451      * index which matches the RGB value. Note this isn't guaranteed to work when there are multiple entries for
1452      * the same color but we have no choice.
1453      * In case of P8 render targets, the index is stored in the alpha component so no conversion is needed.
1454      */
1455     if (This->resource.format_desc->format == WINED3DFMT_P8_UINT && !primary_render_target_is_p8(device))
1456     {
1457         const PALETTEENTRY *pal = NULL;
1458         DWORD width = pitch / 3;
1459         int x, y, c;
1460
1461         if(This->palette) {
1462             pal = This->palette->palents;
1463         } else {
1464             ERR("Palette is missing, cannot perform inverse palette lookup\n");
1465             HeapFree(GetProcessHeap(), 0, mem);
1466             return ;
1467         }
1468
1469         for(y = local_rect.top; y < local_rect.bottom; y++) {
1470             for(x = local_rect.left; x < local_rect.right; x++) {
1471                 /*                      start              lines            pixels      */
1472                 const BYTE *blue = mem + y * pitch + x * (sizeof(BYTE) * 3);
1473                 const BYTE *green = blue  + 1;
1474                 const BYTE *red = green + 1;
1475
1476                 for(c = 0; c < 256; c++) {
1477                     if(*red   == pal[c].peRed   &&
1478                        *green == pal[c].peGreen &&
1479                        *blue  == pal[c].peBlue)
1480                     {
1481                         *((BYTE *) dest + y * width + x) = c;
1482                         break;
1483                     }
1484                 }
1485             }
1486         }
1487         HeapFree(GetProcessHeap(), 0, mem);
1488     }
1489 }
1490
1491 /* Read the framebuffer contents into a texture */
1492 static void read_from_framebuffer_texture(IWineD3DSurfaceImpl *This, BOOL srgb)
1493 {
1494     IWineD3DDeviceImpl *device = This->resource.device;
1495     const struct wined3d_gl_info *gl_info;
1496     struct wined3d_context *context;
1497
1498     if (!surface_is_offscreen(This))
1499     {
1500         /* We would need to flip onscreen surfaces, but there's no efficient
1501          * way to do that here. It makes more sense for the caller to
1502          * explicitly go through sysmem. */
1503         ERR("Not supported for onscreen targets.\n");
1504         return;
1505     }
1506
1507     /* Activate the surface to read from. In some situations it isn't the currently active target(e.g. backbuffer
1508      * locking during offscreen rendering). RESOURCELOAD is ok because glCopyTexSubImage2D isn't affected by any
1509      * states in the stateblock, and no driver was found yet that had bugs in that regard.
1510      */
1511     context = context_acquire(device, This);
1512     gl_info = context->gl_info;
1513
1514     surface_prepare_texture(This, gl_info, srgb);
1515     surface_bind_and_dirtify(This, srgb);
1516
1517     TRACE("Reading back offscreen render target %p.\n", This);
1518
1519     ENTER_GL();
1520
1521     glReadBuffer(device->offscreenBuffer);
1522     checkGLcall("glReadBuffer");
1523
1524     glCopyTexSubImage2D(This->texture_target, This->texture_level,
1525             0, 0, 0, 0, This->currentDesc.Width, This->currentDesc.Height);
1526     checkGLcall("glCopyTexSubImage2D");
1527
1528     LEAVE_GL();
1529
1530     context_release(context);
1531 }
1532
1533 /* Context activation is done by the caller. */
1534 static void surface_prepare_texture_internal(IWineD3DSurfaceImpl *surface,
1535         const struct wined3d_gl_info *gl_info, BOOL srgb)
1536 {
1537     DWORD alloc_flag = srgb ? SFLAG_SRGBALLOCATED : SFLAG_ALLOCATED;
1538     CONVERT_TYPES convert;
1539     struct wined3d_format_desc desc;
1540
1541     if (surface->Flags & alloc_flag) return;
1542
1543     d3dfmt_get_conv(surface, TRUE, TRUE, &desc, &convert);
1544     if(convert != NO_CONVERSION || desc.convert) surface->Flags |= SFLAG_CONVERTED;
1545     else surface->Flags &= ~SFLAG_CONVERTED;
1546
1547     surface_bind_and_dirtify(surface, srgb);
1548     surface_allocate_surface(surface, gl_info, &desc, srgb);
1549     surface->Flags |= alloc_flag;
1550 }
1551
1552 /* Context activation is done by the caller. */
1553 void surface_prepare_texture(IWineD3DSurfaceImpl *surface, const struct wined3d_gl_info *gl_info, BOOL srgb)
1554 {
1555     IWineD3DBaseTextureImpl *texture;
1556
1557     if (SUCCEEDED(IWineD3DSurface_GetContainer((IWineD3DSurface *)surface,
1558             &IID_IWineD3DBaseTexture, (void **)&texture)))
1559     {
1560         UINT sub_count = texture->baseTexture.level_count * texture->baseTexture.layer_count;
1561         UINT i;
1562
1563         TRACE("surface %p is a subresource of texture %p.\n", surface, texture);
1564
1565         for (i = 0; i < sub_count; ++i)
1566         {
1567             IWineD3DSurfaceImpl *s = (IWineD3DSurfaceImpl *)texture->baseTexture.sub_resources[i];
1568             surface_prepare_texture_internal(s, gl_info, srgb);
1569         }
1570
1571         IWineD3DBaseTexture_Release((IWineD3DBaseTexture *)texture);
1572
1573         return;
1574     }
1575
1576     surface_prepare_texture_internal(surface, gl_info, srgb);
1577 }
1578
1579 static void surface_prepare_system_memory(IWineD3DSurfaceImpl *This)
1580 {
1581     IWineD3DDeviceImpl *device = This->resource.device;
1582     const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
1583
1584     /* Performance optimization: Count how often a surface is locked, if it is locked regularly do not throw away the system memory copy.
1585      * This avoids the need to download the surface from opengl all the time. The surface is still downloaded if the opengl texture is
1586      * changed
1587      */
1588     if(!(This->Flags & SFLAG_DYNLOCK)) {
1589         This->lockCount++;
1590         /* MAXLOCKCOUNT is defined in wined3d_private.h */
1591         if(This->lockCount > MAXLOCKCOUNT) {
1592             TRACE("Surface is locked regularly, not freeing the system memory copy any more\n");
1593             This->Flags |= SFLAG_DYNLOCK;
1594         }
1595     }
1596
1597     /* Create a PBO for dynamically locked surfaces but don't do it for converted or non-pow2 surfaces.
1598      * Also don't create a PBO for systemmem surfaces.
1599      */
1600     if (gl_info->supported[ARB_PIXEL_BUFFER_OBJECT] && (This->Flags & SFLAG_DYNLOCK)
1601             && !(This->Flags & (SFLAG_PBO | SFLAG_CONVERTED | SFLAG_NONPOW2))
1602             && (This->resource.pool != WINED3DPOOL_SYSTEMMEM))
1603     {
1604         GLenum error;
1605         struct wined3d_context *context;
1606
1607         context = context_acquire(device, NULL);
1608         ENTER_GL();
1609
1610         GL_EXTCALL(glGenBuffersARB(1, &This->pbo));
1611         error = glGetError();
1612         if(This->pbo == 0 || error != GL_NO_ERROR) {
1613             ERR("Failed to bind the PBO with error %s (%#x)\n", debug_glerror(error), error);
1614         }
1615
1616         TRACE("Attaching pbo=%#x to (%p)\n", This->pbo, This);
1617
1618         GL_EXTCALL(glBindBufferARB(GL_PIXEL_UNPACK_BUFFER_ARB, This->pbo));
1619         checkGLcall("glBindBufferARB");
1620
1621         GL_EXTCALL(glBufferDataARB(GL_PIXEL_UNPACK_BUFFER_ARB, This->resource.size + 4, This->resource.allocatedMemory, GL_STREAM_DRAW_ARB));
1622         checkGLcall("glBufferDataARB");
1623
1624         GL_EXTCALL(glBindBufferARB(GL_PIXEL_UNPACK_BUFFER_ARB, 0));
1625         checkGLcall("glBindBufferARB");
1626
1627         /* We don't need the system memory anymore and we can't even use it for PBOs */
1628         if(!(This->Flags & SFLAG_CLIENT)) {
1629             HeapFree(GetProcessHeap(), 0, This->resource.heapMemory);
1630             This->resource.heapMemory = NULL;
1631         }
1632         This->resource.allocatedMemory = NULL;
1633         This->Flags |= SFLAG_PBO;
1634         LEAVE_GL();
1635         context_release(context);
1636     }
1637     else if (!(This->resource.allocatedMemory || This->Flags & SFLAG_PBO))
1638     {
1639         /* Whatever surface we have, make sure that there is memory allocated for the downloaded copy,
1640          * or a pbo to map
1641          */
1642         if(!This->resource.heapMemory) {
1643             This->resource.heapMemory = HeapAlloc(GetProcessHeap() ,0 , This->resource.size + RESOURCE_ALIGNMENT);
1644         }
1645         This->resource.allocatedMemory =
1646                 (BYTE *)(((ULONG_PTR) This->resource.heapMemory + (RESOURCE_ALIGNMENT - 1)) & ~(RESOURCE_ALIGNMENT - 1));
1647         if(This->Flags & SFLAG_INSYSMEM) {
1648             ERR("Surface without memory or pbo has SFLAG_INSYSMEM set!\n");
1649         }
1650     }
1651 }
1652
1653 static HRESULT WINAPI IWineD3DSurfaceImpl_LockRect(IWineD3DSurface *iface, WINED3DLOCKED_RECT* pLockedRect, CONST RECT* pRect, DWORD Flags) {
1654     IWineD3DSurfaceImpl *This = (IWineD3DSurfaceImpl *)iface;
1655     IWineD3DDeviceImpl *device = This->resource.device;
1656     const RECT *pass_rect = pRect;
1657
1658     TRACE("iface %p, locked_rect %p, rect %s, flags %#x.\n",
1659             iface, pLockedRect, wine_dbgstr_rect(pRect), Flags);
1660
1661     /* This is also done in the base class, but we have to verify this before loading any data from
1662      * gl into the sysmem copy. The PBO may be mapped, a different rectangle locked, the discard flag
1663      * may interfere, and all other bad things may happen
1664      */
1665     if (This->Flags & SFLAG_LOCKED) {
1666         WARN("Surface is already locked, returning D3DERR_INVALIDCALL\n");
1667         return WINED3DERR_INVALIDCALL;
1668     }
1669     This->Flags |= SFLAG_LOCKED;
1670
1671     if (!(This->Flags & SFLAG_LOCKABLE))
1672     {
1673         TRACE("Warning: trying to lock unlockable surf@%p\n", This);
1674     }
1675
1676     if (Flags & WINED3DLOCK_DISCARD) {
1677         /* Set SFLAG_INSYSMEM, so we'll never try to download the data from the texture. */
1678         TRACE("WINED3DLOCK_DISCARD flag passed, marking local copy as up to date\n");
1679         surface_prepare_system_memory(This); /* Makes sure memory is allocated */
1680         This->Flags |= SFLAG_INSYSMEM;
1681         goto lock_end;
1682     }
1683
1684     if (This->Flags & SFLAG_INSYSMEM) {
1685         TRACE("Local copy is up to date, not downloading data\n");
1686         surface_prepare_system_memory(This); /* Makes sure memory is allocated */
1687         goto lock_end;
1688     }
1689
1690     /* surface_load_location() does not check if the rectangle specifies
1691      * the full surface. Most callers don't need that, so do it here. */
1692     if (pRect && pRect->top == 0 && pRect->left == 0
1693             && pRect->right == This->currentDesc.Width
1694             && pRect->bottom == This->currentDesc.Height)
1695     {
1696         pass_rect = NULL;
1697     }
1698
1699     if (!(wined3d_settings.rendertargetlock_mode == RTL_DISABLE
1700             && ((This->Flags & SFLAG_SWAPCHAIN) || This == device->render_targets[0])))
1701     {
1702         surface_load_location(This, SFLAG_INSYSMEM, pass_rect);
1703     }
1704
1705 lock_end:
1706     if (This->Flags & SFLAG_PBO)
1707     {
1708         const struct wined3d_gl_info *gl_info;
1709         struct wined3d_context *context;
1710
1711         context = context_acquire(device, NULL);
1712         gl_info = context->gl_info;
1713
1714         ENTER_GL();
1715         GL_EXTCALL(glBindBufferARB(GL_PIXEL_UNPACK_BUFFER_ARB, This->pbo));
1716         checkGLcall("glBindBufferARB");
1717
1718         /* This shouldn't happen but could occur if some other function didn't handle the PBO properly */
1719         if(This->resource.allocatedMemory) {
1720             ERR("The surface already has PBO memory allocated!\n");
1721         }
1722
1723         This->resource.allocatedMemory = GL_EXTCALL(glMapBufferARB(GL_PIXEL_UNPACK_BUFFER_ARB, GL_READ_WRITE_ARB));
1724         checkGLcall("glMapBufferARB");
1725
1726         /* Make sure the pbo isn't set anymore in order not to break non-pbo calls */
1727         GL_EXTCALL(glBindBufferARB(GL_PIXEL_UNPACK_BUFFER_ARB, 0));
1728         checkGLcall("glBindBufferARB");
1729
1730         LEAVE_GL();
1731         context_release(context);
1732     }
1733
1734     if (Flags & (WINED3DLOCK_NO_DIRTY_UPDATE | WINED3DLOCK_READONLY)) {
1735         /* Don't dirtify */
1736     } else {
1737         IWineD3DBaseTexture *pBaseTexture;
1738         /**
1739          * Dirtify on lock
1740          * as seen in msdn docs
1741          */
1742         surface_add_dirty_rect(This, pRect);
1743
1744         /** Dirtify Container if needed */
1745         if (SUCCEEDED(IWineD3DSurface_GetContainer(iface, &IID_IWineD3DBaseTexture, (void **)&pBaseTexture))) {
1746             TRACE("Making container dirty\n");
1747             IWineD3DBaseTexture_SetDirty(pBaseTexture, TRUE);
1748             IWineD3DBaseTexture_Release(pBaseTexture);
1749         } else {
1750             TRACE("Surface is standalone, no need to dirty the container\n");
1751         }
1752     }
1753
1754     return IWineD3DBaseSurfaceImpl_LockRect(iface, pLockedRect, pRect, Flags);
1755 }
1756
1757 static void flush_to_framebuffer_drawpixels(IWineD3DSurfaceImpl *This, GLenum fmt, GLenum type, UINT bpp, const BYTE *mem) {
1758     GLint  prev_store;
1759     GLint  prev_rasterpos[4];
1760     GLint skipBytes = 0;
1761     UINT pitch = IWineD3DSurface_GetPitch((IWineD3DSurface *) This);    /* target is argb, 4 byte */
1762     IWineD3DDeviceImpl *device = This->resource.device;
1763     const struct wined3d_gl_info *gl_info;
1764     struct wined3d_context *context;
1765
1766     /* Activate the correct context for the render target */
1767     context = context_acquire(device, This);
1768     context_apply_blit_state(context, device);
1769     gl_info = context->gl_info;
1770
1771     ENTER_GL();
1772
1773     if (!surface_is_offscreen(This))
1774     {
1775         GLenum buffer = surface_get_gl_buffer(This);
1776         TRACE("Unlocking %#x buffer.\n", buffer);
1777         context_set_draw_buffer(context, buffer);
1778     }
1779     else
1780     {
1781         /* Primary offscreen render target */
1782         TRACE("Offscreen render target.\n");
1783         context_set_draw_buffer(context, device->offscreenBuffer);
1784     }
1785
1786     glGetIntegerv(GL_PACK_SWAP_BYTES, &prev_store);
1787     checkGLcall("glGetIntegerv");
1788     glGetIntegerv(GL_CURRENT_RASTER_POSITION, &prev_rasterpos[0]);
1789     checkGLcall("glGetIntegerv");
1790     glPixelZoom(1.0f, -1.0f);
1791     checkGLcall("glPixelZoom");
1792
1793     /* If not fullscreen, we need to skip a number of bytes to find the next row of data */
1794     glGetIntegerv(GL_UNPACK_ROW_LENGTH, &skipBytes);
1795     glPixelStorei(GL_UNPACK_ROW_LENGTH, This->currentDesc.Width);
1796
1797     glRasterPos3i(This->lockedRect.left, This->lockedRect.top, 1);
1798     checkGLcall("glRasterPos3i");
1799
1800     /* Some drivers(radeon dri, others?) don't like exceptions during
1801      * glDrawPixels. If the surface is a DIB section, it might be in GDIMode
1802      * after ReleaseDC. Reading it will cause an exception, which x11drv will
1803      * catch to put the dib section in InSync mode, which leads to a crash
1804      * and a blocked x server on my radeon card.
1805      *
1806      * The following lines read the dib section so it is put in InSync mode
1807      * before glDrawPixels is called and the crash is prevented. There won't
1808      * be any interfering gdi accesses, because UnlockRect is called from
1809      * ReleaseDC, and the app won't use the dc any more afterwards.
1810      */
1811     if((This->Flags & SFLAG_DIBSECTION) && !(This->Flags & SFLAG_PBO)) {
1812         volatile BYTE read;
1813         read = This->resource.allocatedMemory[0];
1814     }
1815
1816     if(This->Flags & SFLAG_PBO) {
1817         GL_EXTCALL(glBindBufferARB(GL_PIXEL_UNPACK_BUFFER_ARB, This->pbo));
1818         checkGLcall("glBindBufferARB");
1819     }
1820
1821     /* When the surface is locked we only have to refresh the locked part else we need to update the whole image */
1822     if(This->Flags & SFLAG_LOCKED) {
1823         glDrawPixels(This->lockedRect.right - This->lockedRect.left,
1824                      (This->lockedRect.bottom - This->lockedRect.top)-1,
1825                      fmt, type,
1826                      mem + bpp * This->lockedRect.left + pitch * This->lockedRect.top);
1827         checkGLcall("glDrawPixels");
1828     } else {
1829         glDrawPixels(This->currentDesc.Width,
1830                      This->currentDesc.Height,
1831                      fmt, type, mem);
1832         checkGLcall("glDrawPixels");
1833     }
1834
1835     if(This->Flags & SFLAG_PBO) {
1836         GL_EXTCALL(glBindBufferARB(GL_PIXEL_UNPACK_BUFFER_ARB, 0));
1837         checkGLcall("glBindBufferARB");
1838     }
1839
1840     glPixelZoom(1.0f, 1.0f);
1841     checkGLcall("glPixelZoom");
1842
1843     glRasterPos3iv(&prev_rasterpos[0]);
1844     checkGLcall("glRasterPos3iv");
1845
1846     /* Reset to previous pack row length */
1847     glPixelStorei(GL_UNPACK_ROW_LENGTH, skipBytes);
1848     checkGLcall("glPixelStorei(GL_UNPACK_ROW_LENGTH)");
1849
1850     LEAVE_GL();
1851     context_release(context);
1852 }
1853
1854 static HRESULT WINAPI IWineD3DSurfaceImpl_UnlockRect(IWineD3DSurface *iface) {
1855     IWineD3DSurfaceImpl *This = (IWineD3DSurfaceImpl *)iface;
1856     IWineD3DDeviceImpl *device = This->resource.device;
1857     BOOL fullsurface;
1858
1859     if (!(This->Flags & SFLAG_LOCKED)) {
1860         WARN("trying to Unlock an unlocked surf@%p\n", This);
1861         return WINEDDERR_NOTLOCKED;
1862     }
1863
1864     if (This->Flags & SFLAG_PBO)
1865     {
1866         const struct wined3d_gl_info *gl_info;
1867         struct wined3d_context *context;
1868
1869         TRACE("Freeing PBO memory\n");
1870
1871         context = context_acquire(device, NULL);
1872         gl_info = context->gl_info;
1873
1874         ENTER_GL();
1875         GL_EXTCALL(glBindBufferARB(GL_PIXEL_UNPACK_BUFFER_ARB, This->pbo));
1876         GL_EXTCALL(glUnmapBufferARB(GL_PIXEL_UNPACK_BUFFER_ARB));
1877         GL_EXTCALL(glBindBufferARB(GL_PIXEL_UNPACK_BUFFER_ARB, 0));
1878         checkGLcall("glUnmapBufferARB");
1879         LEAVE_GL();
1880         context_release(context);
1881
1882         This->resource.allocatedMemory = NULL;
1883     }
1884
1885     TRACE("(%p) : dirtyfied(%d)\n", This, This->Flags & (SFLAG_INDRAWABLE | SFLAG_INTEXTURE) ? 0 : 1);
1886
1887     if (This->Flags & (SFLAG_INDRAWABLE | SFLAG_INTEXTURE)) {
1888         TRACE("(%p) : Not Dirtified so nothing to do, return now\n", This);
1889         goto unlock_end;
1890     }
1891
1892     if ((This->Flags & SFLAG_SWAPCHAIN) || (device->render_targets && This == device->render_targets[0]))
1893     {
1894         if(wined3d_settings.rendertargetlock_mode == RTL_DISABLE) {
1895             static BOOL warned = FALSE;
1896             if(!warned) {
1897                 ERR("The application tries to write to the render target, but render target locking is disabled\n");
1898                 warned = TRUE;
1899             }
1900             goto unlock_end;
1901         }
1902
1903         if(This->dirtyRect.left   == 0 &&
1904            This->dirtyRect.top    == 0 &&
1905            This->dirtyRect.right  == This->currentDesc.Width &&
1906            This->dirtyRect.bottom == This->currentDesc.Height) {
1907             fullsurface = TRUE;
1908         } else {
1909             /* TODO: Proper partial rectangle tracking */
1910             fullsurface = FALSE;
1911             This->Flags |= SFLAG_INSYSMEM;
1912         }
1913
1914         switch(wined3d_settings.rendertargetlock_mode) {
1915             case RTL_READTEX:
1916                 surface_load_location(This, SFLAG_INTEXTURE, NULL /* partial texture loading not supported yet */);
1917                 /* drop through */
1918
1919             case RTL_READDRAW:
1920                 surface_load_location(This, SFLAG_INDRAWABLE, fullsurface ? NULL : &This->dirtyRect);
1921                 break;
1922         }
1923
1924         if(!fullsurface) {
1925             /* Partial rectangle tracking is not commonly implemented, it is only done for render targets. Overwrite
1926              * the flags to bring them back into a sane state. INSYSMEM was set before to tell LoadLocation where
1927              * to read the rectangle from. Indrawable is set because all modifications from the partial sysmem copy
1928              * are written back to the drawable, thus the surface is merged again in the drawable. The sysmem copy is
1929              * not fully up to date because only a subrectangle was read in LockRect.
1930              */
1931             This->Flags &= ~SFLAG_INSYSMEM;
1932             This->Flags |= SFLAG_INDRAWABLE;
1933         }
1934
1935         This->dirtyRect.left   = This->currentDesc.Width;
1936         This->dirtyRect.top    = This->currentDesc.Height;
1937         This->dirtyRect.right  = 0;
1938         This->dirtyRect.bottom = 0;
1939     }
1940     else if (This == device->depth_stencil)
1941     {
1942         FIXME("Depth Stencil buffer locking is not implemented\n");
1943     } else {
1944         /* The rest should be a normal texture */
1945         IWineD3DBaseTextureImpl *impl;
1946         /* Check if the texture is bound, if yes dirtify the sampler to force a re-upload of the texture
1947          * Can't load the texture here because PreLoad may destroy and recreate the gl texture, so sampler
1948          * states need resetting
1949          */
1950         if(IWineD3DSurface_GetContainer(iface, &IID_IWineD3DBaseTexture, (void **)&impl) == WINED3D_OK) {
1951             if (impl->baseTexture.bindCount)
1952                 IWineD3DDeviceImpl_MarkStateDirty(device, STATE_SAMPLER(impl->baseTexture.sampler));
1953             IWineD3DBaseTexture_Release((IWineD3DBaseTexture *) impl);
1954         }
1955     }
1956
1957     unlock_end:
1958     This->Flags &= ~SFLAG_LOCKED;
1959     memset(&This->lockedRect, 0, sizeof(RECT));
1960
1961     /* Overlays have to be redrawn manually after changes with the GL implementation */
1962     if(This->overlay_dest) {
1963         IWineD3DSurface_DrawOverlay(iface);
1964     }
1965     return WINED3D_OK;
1966 }
1967
1968 static void surface_release_client_storage(IWineD3DSurfaceImpl *surface)
1969 {
1970     struct wined3d_context *context;
1971
1972     context = context_acquire(surface->resource.device, NULL);
1973
1974     ENTER_GL();
1975     glPixelStorei(GL_UNPACK_CLIENT_STORAGE_APPLE, GL_FALSE);
1976     if (surface->texture_name)
1977     {
1978         surface_bind_and_dirtify(surface, FALSE);
1979         glTexImage2D(surface->texture_target, surface->texture_level,
1980                 GL_RGB, 1, 1, 0, GL_RGB, GL_UNSIGNED_BYTE, NULL);
1981     }
1982     if (surface->texture_name_srgb)
1983     {
1984         surface_bind_and_dirtify(surface, TRUE);
1985         glTexImage2D(surface->texture_target, surface->texture_level,
1986                 GL_RGB, 1, 1, 0, GL_RGB, GL_UNSIGNED_BYTE, NULL);
1987     }
1988     glPixelStorei(GL_UNPACK_CLIENT_STORAGE_APPLE, GL_TRUE);
1989
1990     LEAVE_GL();
1991     context_release(context);
1992
1993     surface_modify_location(surface, SFLAG_INSRGBTEX, FALSE);
1994     surface_modify_location(surface, SFLAG_INTEXTURE, FALSE);
1995     surface_force_reload(surface);
1996 }
1997
1998 static HRESULT WINAPI IWineD3DSurfaceImpl_GetDC(IWineD3DSurface *iface, HDC *pHDC)
1999 {
2000     IWineD3DSurfaceImpl *This = (IWineD3DSurfaceImpl *)iface;
2001     WINED3DLOCKED_RECT lock;
2002     HRESULT hr;
2003     RGBQUAD col[256];
2004
2005     TRACE("(%p)->(%p)\n",This,pHDC);
2006
2007     if(This->Flags & SFLAG_USERPTR) {
2008         ERR("Not supported on surfaces with an application-provided surfaces\n");
2009         return WINEDDERR_NODC;
2010     }
2011
2012     /* Give more detailed info for ddraw */
2013     if (This->Flags & SFLAG_DCINUSE)
2014         return WINEDDERR_DCALREADYCREATED;
2015
2016     /* Can't GetDC if the surface is locked */
2017     if (This->Flags & SFLAG_LOCKED)
2018         return WINED3DERR_INVALIDCALL;
2019
2020     memset(&lock, 0, sizeof(lock)); /* To be sure */
2021
2022     /* Create a DIB section if there isn't a hdc yet */
2023     if (!This->hDC)
2024     {
2025         if (This->Flags & SFLAG_CLIENT)
2026         {
2027             surface_load_location(This, SFLAG_INSYSMEM, NULL);
2028             surface_release_client_storage(This);
2029         }
2030         hr = IWineD3DBaseSurfaceImpl_CreateDIBSection(iface);
2031         if(FAILED(hr)) return WINED3DERR_INVALIDCALL;
2032
2033         /* Use the dib section from now on if we are not using a PBO */
2034         if(!(This->Flags & SFLAG_PBO))
2035             This->resource.allocatedMemory = This->dib.bitmap_data;
2036     }
2037
2038     /* Lock the surface */
2039     hr = IWineD3DSurface_LockRect(iface,
2040                                   &lock,
2041                                   NULL,
2042                                   0);
2043
2044     if(This->Flags & SFLAG_PBO) {
2045         /* Sync the DIB with the PBO. This can't be done earlier because LockRect activates the allocatedMemory */
2046         memcpy(This->dib.bitmap_data, This->resource.allocatedMemory, This->dib.bitmap_size);
2047     }
2048
2049     if(FAILED(hr)) {
2050         ERR("IWineD3DSurface_LockRect failed with hr = %08x\n", hr);
2051         /* keep the dib section */
2052         return hr;
2053     }
2054
2055     if (This->resource.format_desc->format == WINED3DFMT_P8_UINT
2056             || This->resource.format_desc->format == WINED3DFMT_P8_UINT_A8_UNORM)
2057     {
2058         /* GetDC on palettized formats is unsupported in D3D9, and the method is missing in
2059             D3D8, so this should only be used for DX <=7 surfaces (with non-device palettes) */
2060         unsigned int n;
2061         const PALETTEENTRY *pal = NULL;
2062
2063         if(This->palette) {
2064             pal = This->palette->palents;
2065         } else {
2066             IWineD3DSurfaceImpl *dds_primary;
2067             IWineD3DSwapChainImpl *swapchain;
2068             swapchain = (IWineD3DSwapChainImpl *)This->resource.device->swapchains[0];
2069             dds_primary = swapchain->front_buffer;
2070             if (dds_primary && dds_primary->palette)
2071                 pal = dds_primary->palette->palents;
2072         }
2073
2074         if (pal) {
2075             for (n=0; n<256; n++) {
2076                 col[n].rgbRed   = pal[n].peRed;
2077                 col[n].rgbGreen = pal[n].peGreen;
2078                 col[n].rgbBlue  = pal[n].peBlue;
2079                 col[n].rgbReserved = 0;
2080             }
2081             SetDIBColorTable(This->hDC, 0, 256, col);
2082         }
2083     }
2084
2085     *pHDC = This->hDC;
2086     TRACE("returning %p\n",*pHDC);
2087     This->Flags |= SFLAG_DCINUSE;
2088
2089     return WINED3D_OK;
2090 }
2091
2092 static HRESULT WINAPI IWineD3DSurfaceImpl_ReleaseDC(IWineD3DSurface *iface, HDC hDC)
2093 {
2094     IWineD3DSurfaceImpl *This = (IWineD3DSurfaceImpl *)iface;
2095
2096     TRACE("(%p)->(%p)\n",This,hDC);
2097
2098     if (!(This->Flags & SFLAG_DCINUSE))
2099         return WINEDDERR_NODC;
2100
2101     if (This->hDC !=hDC) {
2102         WARN("Application tries to release an invalid DC(%p), surface dc is %p\n", hDC, This->hDC);
2103         return WINEDDERR_NODC;
2104     }
2105
2106     if((This->Flags & SFLAG_PBO) && This->resource.allocatedMemory) {
2107         /* Copy the contents of the DIB over to the PBO */
2108         memcpy(This->resource.allocatedMemory, This->dib.bitmap_data, This->dib.bitmap_size);
2109     }
2110
2111     /* we locked first, so unlock now */
2112     IWineD3DSurface_UnlockRect(iface);
2113
2114     This->Flags &= ~SFLAG_DCINUSE;
2115
2116     return WINED3D_OK;
2117 }
2118
2119 /* ******************************************************
2120    IWineD3DSurface Internal (No mapping to directx api) parts follow
2121    ****************************************************** */
2122
2123 HRESULT d3dfmt_get_conv(IWineD3DSurfaceImpl *This, BOOL need_alpha_ck, BOOL use_texturing, struct wined3d_format_desc *desc, CONVERT_TYPES *convert)
2124 {
2125     BOOL colorkey_active = need_alpha_ck && (This->CKeyFlags & WINEDDSD_CKSRCBLT);
2126     IWineD3DDeviceImpl *device = This->resource.device;
2127     const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
2128     BOOL blit_supported = FALSE;
2129
2130     /* Copy the default values from the surface. Below we might perform fixups */
2131     /* TODO: get rid of color keying desc fixups by using e.g. a table. */
2132     *desc = *This->resource.format_desc;
2133     *convert = NO_CONVERSION;
2134
2135     /* Ok, now look if we have to do any conversion */
2136     switch(This->resource.format_desc->format)
2137     {
2138         case WINED3DFMT_P8_UINT:
2139             /* ****************
2140                 Paletted Texture
2141                 **************** */
2142
2143             /* Below the call to blit_supported is disabled for Wine 1.2 because the function isn't operating correctly yet.
2144              * At the moment 8-bit blits are handled in software and if certain GL extensions are around, surface conversion
2145              * is performed at upload time. The blit_supported call recognizes it as a destination fixup. This type of upload 'fixup'
2146              * and 8-bit to 8-bit blits need to be handled by the blit_shader.
2147              * TODO: get rid of this #if 0
2148              */
2149 #if 0
2150             blit_supported = device->blitter->blit_supported(&device->adapter->gl_info, BLIT_OP_BLIT,
2151                                                              &rect, This->resource.usage, This->resource.pool,
2152                                                              This->resource.format_desc, &rect, This->resource.usage,
2153                                                              This->resource.pool, This->resource.format_desc);
2154 #endif
2155             blit_supported = gl_info->supported[EXT_PALETTED_TEXTURE] || gl_info->supported[ARB_FRAGMENT_PROGRAM];
2156
2157             /* Use conversion when the blit_shader backend supports it. It only supports this in case of
2158              * texturing. Further also use conversion in case of color keying.
2159              * Paletted textures can be emulated using shaders but only do that for 2D purposes e.g. situations
2160              * in which the main render target uses p8. Some games like GTA Vice City use P8 for texturing which
2161              * conflicts with this.
2162              */
2163             if (!((blit_supported && device->render_targets && This == device->render_targets[0]))
2164                     || colorkey_active || !use_texturing)
2165             {
2166                 desc->glFormat = GL_RGBA;
2167                 desc->glInternal = GL_RGBA;
2168                 desc->glType = GL_UNSIGNED_BYTE;
2169                 desc->conv_byte_count = 4;
2170                 if(colorkey_active) {
2171                     *convert = CONVERT_PALETTED_CK;
2172                 } else {
2173                     *convert = CONVERT_PALETTED;
2174                 }
2175             }
2176             break;
2177
2178         case WINED3DFMT_B2G3R3_UNORM:
2179             /* **********************
2180                 GL_UNSIGNED_BYTE_3_3_2
2181                 ********************** */
2182             if (colorkey_active) {
2183                 /* This texture format will never be used.. So do not care about color keying
2184                     up until the point in time it will be needed :-) */
2185                 FIXME(" ColorKeying not supported in the RGB 332 format !\n");
2186             }
2187             break;
2188
2189         case WINED3DFMT_B5G6R5_UNORM:
2190             if (colorkey_active) {
2191                 *convert = CONVERT_CK_565;
2192                 desc->glFormat = GL_RGBA;
2193                 desc->glInternal = GL_RGB5_A1;
2194                 desc->glType = GL_UNSIGNED_SHORT_5_5_5_1;
2195                 desc->conv_byte_count = 2;
2196             }
2197             break;
2198
2199         case WINED3DFMT_B5G5R5X1_UNORM:
2200             if (colorkey_active) {
2201                 *convert = CONVERT_CK_5551;
2202                 desc->glFormat = GL_BGRA;
2203                 desc->glInternal = GL_RGB5_A1;
2204                 desc->glType = GL_UNSIGNED_SHORT_1_5_5_5_REV;
2205                 desc->conv_byte_count = 2;
2206             }
2207             break;
2208
2209         case WINED3DFMT_B8G8R8_UNORM:
2210             if (colorkey_active) {
2211                 *convert = CONVERT_CK_RGB24;
2212                 desc->glFormat = GL_RGBA;
2213                 desc->glInternal = GL_RGBA8;
2214                 desc->glType = GL_UNSIGNED_INT_8_8_8_8;
2215                 desc->conv_byte_count = 4;
2216             }
2217             break;
2218
2219         case WINED3DFMT_B8G8R8X8_UNORM:
2220             if (colorkey_active) {
2221                 *convert = CONVERT_RGB32_888;
2222                 desc->glFormat = GL_RGBA;
2223                 desc->glInternal = GL_RGBA8;
2224                 desc->glType = GL_UNSIGNED_INT_8_8_8_8;
2225                 desc->conv_byte_count = 4;
2226             }
2227             break;
2228
2229         default:
2230             break;
2231     }
2232
2233     return WINED3D_OK;
2234 }
2235
2236 void d3dfmt_p8_init_palette(IWineD3DSurfaceImpl *This, BYTE table[256][4], BOOL colorkey)
2237 {
2238     IWineD3DDeviceImpl *device = This->resource.device;
2239     IWineD3DPaletteImpl *pal = This->palette;
2240     BOOL index_in_alpha = FALSE;
2241     unsigned int i;
2242
2243     /* Old games like StarCraft, C&C, Red Alert and others use P8 render targets.
2244      * Reading back the RGB output each lockrect (each frame as they lock the whole screen)
2245      * is slow. Further RGB->P8 conversion is not possible because palettes can have
2246      * duplicate entries. Store the color key in the unused alpha component to speed the
2247      * download up and to make conversion unneeded. */
2248     index_in_alpha = primary_render_target_is_p8(device);
2249
2250     if (!pal)
2251     {
2252         UINT dxVersion = ((IWineD3DImpl *)device->wined3d)->dxVersion;
2253
2254         /* In DirectDraw the palette is a property of the surface, there are no such things as device palettes. */
2255         if (dxVersion <= 7)
2256         {
2257             ERR("This code should never get entered for DirectDraw!, expect problems\n");
2258             if (index_in_alpha)
2259             {
2260                 /* Guarantees that memory representation remains correct after sysmem<->texture transfers even if
2261                  * there's no palette at this time. */
2262                 for (i = 0; i < 256; i++) table[i][3] = i;
2263             }
2264         }
2265         else
2266         {
2267             /* Direct3D >= 8 palette usage style: P8 textures use device palettes, palette entry format is A8R8G8B8,
2268              * alpha is stored in peFlags and may be used by the app if D3DPTEXTURECAPS_ALPHAPALETTE device
2269              * capability flag is present (wine does advertise this capability) */
2270             for (i = 0; i < 256; ++i)
2271             {
2272                 table[i][0] = device->palettes[device->currentPalette][i].peRed;
2273                 table[i][1] = device->palettes[device->currentPalette][i].peGreen;
2274                 table[i][2] = device->palettes[device->currentPalette][i].peBlue;
2275                 table[i][3] = device->palettes[device->currentPalette][i].peFlags;
2276             }
2277         }
2278     }
2279     else
2280     {
2281         TRACE("Using surface palette %p\n", pal);
2282         /* Get the surface's palette */
2283         for (i = 0; i < 256; ++i)
2284         {
2285             table[i][0] = pal->palents[i].peRed;
2286             table[i][1] = pal->palents[i].peGreen;
2287             table[i][2] = pal->palents[i].peBlue;
2288
2289             /* When index_in_alpha is set the palette index is stored in the
2290              * alpha component. In case of a readback we can then read
2291              * GL_ALPHA. Color keying is handled in BltOverride using a
2292              * GL_ALPHA_TEST using GL_NOT_EQUAL. In case of index_in_alpha the
2293              * color key itself is passed to glAlphaFunc in other cases the
2294              * alpha component of pixels that should be masked away is set to 0. */
2295             if (index_in_alpha)
2296             {
2297                 table[i][3] = i;
2298             }
2299             else if (colorkey && (i >= This->SrcBltCKey.dwColorSpaceLowValue)
2300                     && (i <= This->SrcBltCKey.dwColorSpaceHighValue))
2301             {
2302                 table[i][3] = 0x00;
2303             }
2304             else if(pal->Flags & WINEDDPCAPS_ALPHA)
2305             {
2306                 table[i][3] = pal->palents[i].peFlags;
2307             }
2308             else
2309             {
2310                 table[i][3] = 0xFF;
2311             }
2312         }
2313     }
2314 }
2315
2316 static HRESULT d3dfmt_convert_surface(const BYTE *src, BYTE *dst, UINT pitch, UINT width,
2317         UINT height, UINT outpitch, CONVERT_TYPES convert, IWineD3DSurfaceImpl *This)
2318 {
2319     const BYTE *source;
2320     BYTE *dest;
2321     TRACE("(%p)->(%p),(%d,%d,%d,%d,%p)\n", src, dst, pitch, height, outpitch, convert,This);
2322
2323     switch (convert) {
2324         case NO_CONVERSION:
2325         {
2326             memcpy(dst, src, pitch * height);
2327             break;
2328         }
2329         case CONVERT_PALETTED:
2330         case CONVERT_PALETTED_CK:
2331         {
2332             IWineD3DPaletteImpl* pal = This->palette;
2333             BYTE table[256][4];
2334             unsigned int x, y;
2335
2336             if( pal == NULL) {
2337                 /* TODO: If we are a sublevel, try to get the palette from level 0 */
2338             }
2339
2340             d3dfmt_p8_init_palette(This, table, (convert == CONVERT_PALETTED_CK));
2341
2342             for (y = 0; y < height; y++)
2343             {
2344                 source = src + pitch * y;
2345                 dest = dst + outpitch * y;
2346                 /* This is an 1 bpp format, using the width here is fine */
2347                 for (x = 0; x < width; x++) {
2348                     BYTE color = *source++;
2349                     *dest++ = table[color][0];
2350                     *dest++ = table[color][1];
2351                     *dest++ = table[color][2];
2352                     *dest++ = table[color][3];
2353                 }
2354             }
2355         }
2356         break;
2357
2358         case CONVERT_CK_565:
2359         {
2360             /* Converting the 565 format in 5551 packed to emulate color-keying.
2361
2362               Note : in all these conversion, it would be best to average the averaging
2363                       pixels to get the color of the pixel that will be color-keyed to
2364                       prevent 'color bleeding'. This will be done later on if ever it is
2365                       too visible.
2366
2367               Note2: Nvidia documents say that their driver does not support alpha + color keying
2368                      on the same surface and disables color keying in such a case
2369             */
2370             unsigned int x, y;
2371             const WORD *Source;
2372             WORD *Dest;
2373
2374             TRACE("Color keyed 565\n");
2375
2376             for (y = 0; y < height; y++) {
2377                 Source = (const WORD *)(src + y * pitch);
2378                 Dest = (WORD *) (dst + y * outpitch);
2379                 for (x = 0; x < width; x++ ) {
2380                     WORD color = *Source++;
2381                     *Dest = ((color & 0xFFC0) | ((color & 0x1F) << 1));
2382                     if ((color < This->SrcBltCKey.dwColorSpaceLowValue) ||
2383                         (color > This->SrcBltCKey.dwColorSpaceHighValue)) {
2384                         *Dest |= 0x0001;
2385                     }
2386                     Dest++;
2387                 }
2388             }
2389         }
2390         break;
2391
2392         case CONVERT_CK_5551:
2393         {
2394             /* Converting X1R5G5B5 format to R5G5B5A1 to emulate color-keying. */
2395             unsigned int x, y;
2396             const WORD *Source;
2397             WORD *Dest;
2398             TRACE("Color keyed 5551\n");
2399             for (y = 0; y < height; y++) {
2400                 Source = (const WORD *)(src + y * pitch);
2401                 Dest = (WORD *) (dst + y * outpitch);
2402                 for (x = 0; x < width; x++ ) {
2403                     WORD color = *Source++;
2404                     *Dest = color;
2405                     if ((color < This->SrcBltCKey.dwColorSpaceLowValue) ||
2406                         (color > This->SrcBltCKey.dwColorSpaceHighValue)) {
2407                         *Dest |= (1 << 15);
2408                     }
2409                     else {
2410                         *Dest &= ~(1 << 15);
2411                     }
2412                     Dest++;
2413                 }
2414             }
2415         }
2416         break;
2417
2418         case CONVERT_CK_RGB24:
2419         {
2420             /* Converting R8G8B8 format to R8G8B8A8 with color-keying. */
2421             unsigned int x, y;
2422             for (y = 0; y < height; y++)
2423             {
2424                 source = src + pitch * y;
2425                 dest = dst + outpitch * y;
2426                 for (x = 0; x < width; x++) {
2427                     DWORD color = ((DWORD)source[0] << 16) + ((DWORD)source[1] << 8) + (DWORD)source[2] ;
2428                     DWORD dstcolor = color << 8;
2429                     if ((color < This->SrcBltCKey.dwColorSpaceLowValue) ||
2430                         (color > This->SrcBltCKey.dwColorSpaceHighValue)) {
2431                         dstcolor |= 0xff;
2432                     }
2433                     *(DWORD*)dest = dstcolor;
2434                     source += 3;
2435                     dest += 4;
2436                 }
2437             }
2438         }
2439         break;
2440
2441         case CONVERT_RGB32_888:
2442         {
2443             /* Converting X8R8G8B8 format to R8G8B8A8 with color-keying. */
2444             unsigned int x, y;
2445             for (y = 0; y < height; y++)
2446             {
2447                 source = src + pitch * y;
2448                 dest = dst + outpitch * y;
2449                 for (x = 0; x < width; x++) {
2450                     DWORD color = 0xffffff & *(const DWORD*)source;
2451                     DWORD dstcolor = color << 8;
2452                     if ((color < This->SrcBltCKey.dwColorSpaceLowValue) ||
2453                         (color > This->SrcBltCKey.dwColorSpaceHighValue)) {
2454                         dstcolor |= 0xff;
2455                     }
2456                     *(DWORD*)dest = dstcolor;
2457                     source += 4;
2458                     dest += 4;
2459                 }
2460             }
2461         }
2462         break;
2463
2464         default:
2465             ERR("Unsupported conversion type %#x.\n", convert);
2466     }
2467     return WINED3D_OK;
2468 }
2469
2470 BOOL palette9_changed(IWineD3DSurfaceImpl *This)
2471 {
2472     IWineD3DDeviceImpl *device = This->resource.device;
2473
2474     if (This->palette || (This->resource.format_desc->format != WINED3DFMT_P8_UINT
2475             && This->resource.format_desc->format != WINED3DFMT_P8_UINT_A8_UNORM))
2476     {
2477         /* If a ddraw-style palette is attached assume no d3d9 palette change.
2478          * Also the palette isn't interesting if the surface format isn't P8 or A8P8
2479          */
2480         return FALSE;
2481     }
2482
2483     if (This->palette9)
2484     {
2485         if (!memcmp(This->palette9, device->palettes[device->currentPalette], sizeof(PALETTEENTRY) * 256))
2486         {
2487             return FALSE;
2488         }
2489     } else {
2490         This->palette9 = HeapAlloc(GetProcessHeap(), 0, sizeof(PALETTEENTRY) * 256);
2491     }
2492     memcpy(This->palette9, device->palettes[device->currentPalette], sizeof(PALETTEENTRY) * 256);
2493     return TRUE;
2494 }
2495
2496 static HRESULT WINAPI IWineD3DSurfaceImpl_LoadTexture(IWineD3DSurface *iface, BOOL srgb_mode) {
2497     IWineD3DSurfaceImpl *This = (IWineD3DSurfaceImpl *)iface;
2498     DWORD flag = srgb_mode ? SFLAG_INSRGBTEX : SFLAG_INTEXTURE;
2499
2500     if (!(This->Flags & flag)) {
2501         TRACE("Reloading because surface is dirty\n");
2502     } else if(/* Reload: gl texture has ck, now no ckey is set OR */
2503               ((This->Flags & SFLAG_GLCKEY) && (!(This->CKeyFlags & WINEDDSD_CKSRCBLT))) ||
2504               /* Reload: vice versa  OR */
2505               ((!(This->Flags & SFLAG_GLCKEY)) && (This->CKeyFlags & WINEDDSD_CKSRCBLT)) ||
2506               /* Also reload: Color key is active AND the color key has changed */
2507               ((This->CKeyFlags & WINEDDSD_CKSRCBLT) && (
2508                 (This->glCKey.dwColorSpaceLowValue != This->SrcBltCKey.dwColorSpaceLowValue) ||
2509                 (This->glCKey.dwColorSpaceHighValue != This->SrcBltCKey.dwColorSpaceHighValue)))) {
2510         TRACE("Reloading because of color keying\n");
2511         /* To perform the color key conversion we need a sysmem copy of
2512          * the surface. Make sure we have it
2513          */
2514
2515         surface_load_location(This, SFLAG_INSYSMEM, NULL);
2516         /* Make sure the texture is reloaded because of the color key change, this kills performance though :( */
2517         /* TODO: This is not necessarily needed with hw palettized texture support */
2518         surface_modify_location(This, SFLAG_INSYSMEM, TRUE);
2519     } else {
2520         TRACE("surface is already in texture\n");
2521         return WINED3D_OK;
2522     }
2523
2524     /* Resources are placed in system RAM and do not need to be recreated when a device is lost.
2525      *  These resources are not bound by device size or format restrictions. Because of this,
2526      *  these resources cannot be accessed by the Direct3D device nor set as textures or render targets.
2527      *  However, these resources can always be created, locked, and copied.
2528      */
2529     if (This->resource.pool == WINED3DPOOL_SCRATCH )
2530     {
2531         FIXME("(%p) Operation not supported for scratch textures\n",This);
2532         return WINED3DERR_INVALIDCALL;
2533     }
2534
2535     surface_load_location(This, flag, NULL /* no partial locking for textures yet */);
2536
2537     if (!(This->Flags & SFLAG_DONOTFREE)) {
2538         HeapFree(GetProcessHeap(), 0, This->resource.heapMemory);
2539         This->resource.allocatedMemory = NULL;
2540         This->resource.heapMemory = NULL;
2541         surface_modify_location(This, SFLAG_INSYSMEM, FALSE);
2542     }
2543
2544     return WINED3D_OK;
2545 }
2546
2547 /* Context activation is done by the caller. */
2548 static void WINAPI IWineD3DSurfaceImpl_BindTexture(IWineD3DSurface *iface, BOOL srgb) {
2549     /* TODO: check for locks */
2550     IWineD3DSurfaceImpl *This = (IWineD3DSurfaceImpl *)iface;
2551     IWineD3DBaseTexture *baseTexture = NULL;
2552
2553     TRACE("(%p)Checking to see if the container is a base texture\n", This);
2554     if (IWineD3DSurface_GetContainer(iface, &IID_IWineD3DBaseTexture, (void **)&baseTexture) == WINED3D_OK) {
2555         TRACE("Passing to container\n");
2556         IWineD3DBaseTexture_BindTexture(baseTexture, srgb);
2557         IWineD3DBaseTexture_Release(baseTexture);
2558     }
2559     else
2560     {
2561         GLuint *name;
2562
2563         TRACE("(%p) : Binding surface\n", This);
2564
2565         name = srgb ? &This->texture_name_srgb : &This->texture_name;
2566
2567         ENTER_GL();
2568
2569         if (!This->texture_level)
2570         {
2571             if (!*name) {
2572                 glGenTextures(1, name);
2573                 checkGLcall("glGenTextures");
2574                 TRACE("Surface %p given name %d\n", This, *name);
2575
2576                 glBindTexture(This->texture_target, *name);
2577                 checkGLcall("glBindTexture");
2578                 glTexParameteri(This->texture_target, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
2579                 checkGLcall("glTexParameteri(dimension, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE)");
2580                 glTexParameteri(This->texture_target, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
2581                 checkGLcall("glTexParameteri(dimension, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE)");
2582                 glTexParameteri(This->texture_target, GL_TEXTURE_WRAP_R, GL_CLAMP_TO_EDGE);
2583                 checkGLcall("glTexParameteri(dimension, GL_TEXTURE_WRAP_R, GL_CLAMP_TO_EDGE)");
2584                 glTexParameteri(This->texture_target, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
2585                 checkGLcall("glTexParameteri(dimension, GL_TEXTURE_MIN_FILTER, GL_NEAREST)");
2586                 glTexParameteri(This->texture_target, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
2587                 checkGLcall("glTexParameteri(dimension, GL_TEXTURE_MAG_FILTER, GL_NEAREST)");
2588             }
2589             /* This is where we should be reducing the amount of GLMemoryUsed */
2590         } else if (*name) {
2591             /* Mipmap surfaces should have a base texture container */
2592             ERR("Mipmap surface has a glTexture bound to it!\n");
2593         }
2594
2595         glBindTexture(This->texture_target, *name);
2596         checkGLcall("glBindTexture");
2597
2598         LEAVE_GL();
2599     }
2600 }
2601
2602 static HRESULT WINAPI IWineD3DSurfaceImpl_SetFormat(IWineD3DSurface *iface, WINED3DFORMAT format) {
2603     IWineD3DSurfaceImpl *This = (IWineD3DSurfaceImpl *)iface;
2604     HRESULT hr;
2605
2606     TRACE("(%p) : Calling base function first\n", This);
2607     hr = IWineD3DBaseSurfaceImpl_SetFormat(iface, format);
2608     if(SUCCEEDED(hr)) {
2609         This->Flags &= ~(SFLAG_ALLOCATED | SFLAG_SRGBALLOCATED);
2610         TRACE("(%p) : glFormat %d, glFormatInternal %d, glType %d\n", This, This->resource.format_desc->glFormat,
2611                 This->resource.format_desc->glInternal, This->resource.format_desc->glType);
2612     }
2613     return hr;
2614 }
2615
2616 static HRESULT WINAPI IWineD3DSurfaceImpl_SetMem(IWineD3DSurface *iface, void *Mem) {
2617     IWineD3DSurfaceImpl *This = (IWineD3DSurfaceImpl *) iface;
2618
2619     if(This->Flags & (SFLAG_LOCKED | SFLAG_DCINUSE)) {
2620         WARN("Surface is locked or the HDC is in use\n");
2621         return WINED3DERR_INVALIDCALL;
2622     }
2623
2624     if(Mem && Mem != This->resource.allocatedMemory) {
2625         void *release = NULL;
2626
2627         /* Do I have to copy the old surface content? */
2628         if(This->Flags & SFLAG_DIBSECTION) {
2629                 /* Release the DC. No need to hold the critical section for the update
2630                  * Thread because this thread runs only on front buffers, but this method
2631                  * fails for render targets in the check above.
2632                  */
2633                 SelectObject(This->hDC, This->dib.holdbitmap);
2634                 DeleteDC(This->hDC);
2635                 /* Release the DIB section */
2636                 DeleteObject(This->dib.DIBsection);
2637                 This->dib.bitmap_data = NULL;
2638                 This->resource.allocatedMemory = NULL;
2639                 This->hDC = NULL;
2640                 This->Flags &= ~SFLAG_DIBSECTION;
2641         } else if(!(This->Flags & SFLAG_USERPTR)) {
2642             release = This->resource.heapMemory;
2643             This->resource.heapMemory = NULL;
2644         }
2645         This->resource.allocatedMemory = Mem;
2646         This->Flags |= SFLAG_USERPTR | SFLAG_INSYSMEM;
2647
2648         /* Now the surface memory is most up do date. Invalidate drawable and texture */
2649         surface_modify_location(This, SFLAG_INSYSMEM, TRUE);
2650
2651         /* For client textures opengl has to be notified */
2652         if (This->Flags & SFLAG_CLIENT)
2653             surface_release_client_storage(This);
2654
2655         /* Now free the old memory if any */
2656         HeapFree(GetProcessHeap(), 0, release);
2657     } else if(This->Flags & SFLAG_USERPTR) {
2658         /* LockRect and GetDC will re-create the dib section and allocated memory */
2659         This->resource.allocatedMemory = NULL;
2660         /* HeapMemory should be NULL already */
2661         if(This->resource.heapMemory != NULL) ERR("User pointer surface has heap memory allocated\n");
2662         This->Flags &= ~SFLAG_USERPTR;
2663
2664         if (This->Flags & SFLAG_CLIENT)
2665             surface_release_client_storage(This);
2666     }
2667     return WINED3D_OK;
2668 }
2669
2670 void flip_surface(IWineD3DSurfaceImpl *front, IWineD3DSurfaceImpl *back) {
2671
2672     /* Flip the surface contents */
2673     /* Flip the DC */
2674     {
2675         HDC tmp;
2676         tmp = front->hDC;
2677         front->hDC = back->hDC;
2678         back->hDC = tmp;
2679     }
2680
2681     /* Flip the DIBsection */
2682     {
2683         HBITMAP tmp;
2684         BOOL hasDib = front->Flags & SFLAG_DIBSECTION;
2685         tmp = front->dib.DIBsection;
2686         front->dib.DIBsection = back->dib.DIBsection;
2687         back->dib.DIBsection = tmp;
2688
2689         if(back->Flags & SFLAG_DIBSECTION) front->Flags |= SFLAG_DIBSECTION;
2690         else front->Flags &= ~SFLAG_DIBSECTION;
2691         if(hasDib) back->Flags |= SFLAG_DIBSECTION;
2692         else back->Flags &= ~SFLAG_DIBSECTION;
2693     }
2694
2695     /* Flip the surface data */
2696     {
2697         void* tmp;
2698
2699         tmp = front->dib.bitmap_data;
2700         front->dib.bitmap_data = back->dib.bitmap_data;
2701         back->dib.bitmap_data = tmp;
2702
2703         tmp = front->resource.allocatedMemory;
2704         front->resource.allocatedMemory = back->resource.allocatedMemory;
2705         back->resource.allocatedMemory = tmp;
2706
2707         tmp = front->resource.heapMemory;
2708         front->resource.heapMemory = back->resource.heapMemory;
2709         back->resource.heapMemory = tmp;
2710     }
2711
2712     /* Flip the PBO */
2713     {
2714         GLuint tmp_pbo = front->pbo;
2715         front->pbo = back->pbo;
2716         back->pbo = tmp_pbo;
2717     }
2718
2719     /* client_memory should not be different, but just in case */
2720     {
2721         BOOL tmp;
2722         tmp = front->dib.client_memory;
2723         front->dib.client_memory = back->dib.client_memory;
2724         back->dib.client_memory = tmp;
2725     }
2726
2727     /* Flip the opengl texture */
2728     {
2729         GLuint tmp;
2730
2731         tmp = back->texture_name;
2732         back->texture_name = front->texture_name;
2733         front->texture_name = tmp;
2734
2735         tmp = back->texture_name_srgb;
2736         back->texture_name_srgb = front->texture_name_srgb;
2737         front->texture_name_srgb = tmp;
2738     }
2739
2740     {
2741         DWORD tmp_flags = back->Flags;
2742         back->Flags = front->Flags;
2743         front->Flags = tmp_flags;
2744     }
2745 }
2746
2747 static HRESULT WINAPI IWineD3DSurfaceImpl_Flip(IWineD3DSurface *iface, IWineD3DSurface *override, DWORD Flags) {
2748     IWineD3DSurfaceImpl *This = (IWineD3DSurfaceImpl *)iface;
2749     IWineD3DSwapChainImpl *swapchain = NULL;
2750     HRESULT hr;
2751     TRACE("(%p)->(%p,%x)\n", This, override, Flags);
2752
2753     /* Flipping is only supported on RenderTargets and overlays*/
2754     if( !(This->resource.usage & (WINED3DUSAGE_RENDERTARGET | WINED3DUSAGE_OVERLAY)) ) {
2755         WARN("Tried to flip a non-render target, non-overlay surface\n");
2756         return WINEDDERR_NOTFLIPPABLE;
2757     }
2758
2759     if(This->resource.usage & WINED3DUSAGE_OVERLAY) {
2760         flip_surface(This, (IWineD3DSurfaceImpl *) override);
2761
2762         /* Update the overlay if it is visible */
2763         if(This->overlay_dest) {
2764             return IWineD3DSurface_DrawOverlay((IWineD3DSurface *) This);
2765         } else {
2766             return WINED3D_OK;
2767         }
2768     }
2769
2770     if(override) {
2771         /* DDraw sets this for the X11 surfaces, so don't confuse the user
2772          * FIXME("(%p) Target override is not supported by now\n", This);
2773          * Additionally, it isn't really possible to support triple-buffering
2774          * properly on opengl at all
2775          */
2776     }
2777
2778     IWineD3DSurface_GetContainer(iface, &IID_IWineD3DSwapChain, (void **) &swapchain);
2779     if(!swapchain) {
2780         ERR("Flipped surface is not on a swapchain\n");
2781         return WINEDDERR_NOTFLIPPABLE;
2782     }
2783
2784     /* Just overwrite the swapchain presentation interval. This is ok because only ddraw apps can call Flip,
2785      * and only d3d8 and d3d9 apps specify the presentation interval
2786      */
2787     if((Flags & (WINEDDFLIP_NOVSYNC | WINEDDFLIP_INTERVAL2 | WINEDDFLIP_INTERVAL3 | WINEDDFLIP_INTERVAL4)) == 0) {
2788         /* Most common case first to avoid wasting time on all the other cases */
2789         swapchain->presentParms.PresentationInterval = WINED3DPRESENT_INTERVAL_ONE;
2790     } else if(Flags & WINEDDFLIP_NOVSYNC) {
2791         swapchain->presentParms.PresentationInterval = WINED3DPRESENT_INTERVAL_IMMEDIATE;
2792     } else if(Flags & WINEDDFLIP_INTERVAL2) {
2793         swapchain->presentParms.PresentationInterval = WINED3DPRESENT_INTERVAL_TWO;
2794     } else if(Flags & WINEDDFLIP_INTERVAL3) {
2795         swapchain->presentParms.PresentationInterval = WINED3DPRESENT_INTERVAL_THREE;
2796     } else {
2797         swapchain->presentParms.PresentationInterval = WINED3DPRESENT_INTERVAL_FOUR;
2798     }
2799
2800     /* Flipping a OpenGL surface -> Use WineD3DDevice::Present */
2801     hr = IWineD3DSwapChain_Present((IWineD3DSwapChain *)swapchain,
2802             NULL, NULL, swapchain->win_handle, NULL, 0);
2803     IWineD3DSwapChain_Release((IWineD3DSwapChain *) swapchain);
2804     return hr;
2805 }
2806
2807 /* Does a direct frame buffer -> texture copy. Stretching is done
2808  * with single pixel copy calls
2809  */
2810 static void fb_copy_to_texture_direct(IWineD3DSurfaceImpl *dst_surface, IWineD3DSurfaceImpl *src_surface,
2811         const RECT *src_rect, const RECT *dst_rect_in, WINED3DTEXTUREFILTERTYPE Filter)
2812 {
2813     IWineD3DDeviceImpl *device = dst_surface->resource.device;
2814     float xrel, yrel;
2815     UINT row;
2816     struct wined3d_context *context;
2817     BOOL upsidedown = FALSE;
2818     RECT dst_rect = *dst_rect_in;
2819
2820     /* Make sure that the top pixel is always above the bottom pixel, and keep a separate upside down flag
2821      * glCopyTexSubImage is a bit picky about the parameters we pass to it
2822      */
2823     if(dst_rect.top > dst_rect.bottom) {
2824         UINT tmp = dst_rect.bottom;
2825         dst_rect.bottom = dst_rect.top;
2826         dst_rect.top = tmp;
2827         upsidedown = TRUE;
2828     }
2829
2830     context = context_acquire(device, src_surface);
2831     context_apply_blit_state(context, device);
2832     surface_internal_preload(dst_surface, SRGB_RGB);
2833     ENTER_GL();
2834
2835     /* Bind the target texture */
2836     glBindTexture(dst_surface->texture_target, dst_surface->texture_name);
2837     checkGLcall("glBindTexture");
2838     if (surface_is_offscreen(src_surface))
2839     {
2840         TRACE("Reading from an offscreen target\n");
2841         upsidedown = !upsidedown;
2842         glReadBuffer(device->offscreenBuffer);
2843     }
2844     else
2845     {
2846         glReadBuffer(surface_get_gl_buffer(src_surface));
2847     }
2848     checkGLcall("glReadBuffer");
2849
2850     xrel = (float) (src_rect->right - src_rect->left) / (float) (dst_rect.right - dst_rect.left);
2851     yrel = (float) (src_rect->bottom - src_rect->top) / (float) (dst_rect.bottom - dst_rect.top);
2852
2853     if ((xrel - 1.0f < -eps) || (xrel - 1.0f > eps))
2854     {
2855         FIXME("Doing a pixel by pixel copy from the framebuffer to a texture, expect major performance issues\n");
2856
2857         if(Filter != WINED3DTEXF_NONE && Filter != WINED3DTEXF_POINT) {
2858             ERR("Texture filtering not supported in direct blit\n");
2859         }
2860     }
2861     else if ((Filter != WINED3DTEXF_NONE && Filter != WINED3DTEXF_POINT)
2862             && ((yrel - 1.0f < -eps) || (yrel - 1.0f > eps)))
2863     {
2864         ERR("Texture filtering not supported in direct blit\n");
2865     }
2866
2867     if (upsidedown
2868             && !((xrel - 1.0f < -eps) || (xrel - 1.0f > eps))
2869             && !((yrel - 1.0f < -eps) || (yrel - 1.0f > eps)))
2870     {
2871         /* Upside down copy without stretching is nice, one glCopyTexSubImage call will do */
2872
2873         glCopyTexSubImage2D(dst_surface->texture_target, dst_surface->texture_level,
2874                 dst_rect.left /*xoffset */, dst_rect.top /* y offset */,
2875                 src_rect->left, src_surface->currentDesc.Height - src_rect->bottom,
2876                 dst_rect.right - dst_rect.left, dst_rect.bottom - dst_rect.top);
2877     } else {
2878         UINT yoffset = src_surface->currentDesc.Height - src_rect->top + dst_rect.top - 1;
2879         /* I have to process this row by row to swap the image,
2880          * otherwise it would be upside down, so stretching in y direction
2881          * doesn't cost extra time
2882          *
2883          * However, stretching in x direction can be avoided if not necessary
2884          */
2885         for(row = dst_rect.top; row < dst_rect.bottom; row++) {
2886             if ((xrel - 1.0f < -eps) || (xrel - 1.0f > eps))
2887             {
2888                 /* Well, that stuff works, but it's very slow.
2889                  * find a better way instead
2890                  */
2891                 UINT col;
2892
2893                 for (col = dst_rect.left; col < dst_rect.right; ++col)
2894                 {
2895                     glCopyTexSubImage2D(dst_surface->texture_target, dst_surface->texture_level,
2896                             dst_rect.left + col /* x offset */, row /* y offset */,
2897                             src_rect->left + col * xrel, yoffset - (int) (row * yrel), 1, 1);
2898                 }
2899             }
2900             else
2901             {
2902                 glCopyTexSubImage2D(dst_surface->texture_target, dst_surface->texture_level,
2903                         dst_rect.left /* x offset */, row /* y offset */,
2904                         src_rect->left, yoffset - (int) (row * yrel), dst_rect.right - dst_rect.left, 1);
2905             }
2906         }
2907     }
2908     checkGLcall("glCopyTexSubImage2D");
2909
2910     LEAVE_GL();
2911     context_release(context);
2912
2913     /* The texture is now most up to date - If the surface is a render target and has a drawable, this
2914      * path is never entered
2915      */
2916     surface_modify_location(dst_surface, SFLAG_INTEXTURE, TRUE);
2917 }
2918
2919 /* Uses the hardware to stretch and flip the image */
2920 static void fb_copy_to_texture_hwstretch(IWineD3DSurfaceImpl *dst_surface, IWineD3DSurfaceImpl *src_surface,
2921         const RECT *src_rect, const RECT *dst_rect_in, WINED3DTEXTUREFILTERTYPE Filter)
2922 {
2923     IWineD3DDeviceImpl *device = dst_surface->resource.device;
2924     GLuint src, backup = 0;
2925     IWineD3DSwapChainImpl *src_swapchain = NULL;
2926     float left, right, top, bottom; /* Texture coordinates */
2927     UINT fbwidth = src_surface->currentDesc.Width;
2928     UINT fbheight = src_surface->currentDesc.Height;
2929     struct wined3d_context *context;
2930     GLenum drawBuffer = GL_BACK;
2931     GLenum texture_target;
2932     BOOL noBackBufferBackup;
2933     BOOL src_offscreen;
2934     BOOL upsidedown = FALSE;
2935     RECT dst_rect = *dst_rect_in;
2936
2937     TRACE("Using hwstretch blit\n");
2938     /* Activate the Proper context for reading from the source surface, set it up for blitting */
2939     context = context_acquire(device, src_surface);
2940     context_apply_blit_state(context, device);
2941     surface_internal_preload(dst_surface, SRGB_RGB);
2942
2943     src_offscreen = surface_is_offscreen(src_surface);
2944     noBackBufferBackup = src_offscreen && wined3d_settings.offscreen_rendering_mode == ORM_FBO;
2945     if (!noBackBufferBackup && !src_surface->texture_name)
2946     {
2947         /* Get it a description */
2948         surface_internal_preload(src_surface, SRGB_RGB);
2949     }
2950     ENTER_GL();
2951
2952     /* Try to use an aux buffer for drawing the rectangle. This way it doesn't need restoring.
2953      * This way we don't have to wait for the 2nd readback to finish to leave this function.
2954      */
2955     if (context->aux_buffers >= 2)
2956     {
2957         /* Got more than one aux buffer? Use the 2nd aux buffer */
2958         drawBuffer = GL_AUX1;
2959     }
2960     else if ((!src_offscreen || device->offscreenBuffer == GL_BACK) && context->aux_buffers >= 1)
2961     {
2962         /* Only one aux buffer, but it isn't used (Onscreen rendering, or non-aux orm)? Use it! */
2963         drawBuffer = GL_AUX0;
2964     }
2965
2966     if(noBackBufferBackup) {
2967         glGenTextures(1, &backup);
2968         checkGLcall("glGenTextures");
2969         glBindTexture(GL_TEXTURE_2D, backup);
2970         checkGLcall("glBindTexture(GL_TEXTURE_2D, backup)");
2971         texture_target = GL_TEXTURE_2D;
2972     } else {
2973         /* Backup the back buffer and copy the source buffer into a texture to draw an upside down stretched quad. If
2974          * we are reading from the back buffer, the backup can be used as source texture
2975          */
2976         texture_target = src_surface->texture_target;
2977         glBindTexture(texture_target, src_surface->texture_name);
2978         checkGLcall("glBindTexture(texture_target, src_surface->texture_name)");
2979         glEnable(texture_target);
2980         checkGLcall("glEnable(texture_target)");
2981
2982         /* For now invalidate the texture copy of the back buffer. Drawable and sysmem copy are untouched */
2983         src_surface->Flags &= ~SFLAG_INTEXTURE;
2984     }
2985
2986     /* Make sure that the top pixel is always above the bottom pixel, and keep a separate upside down flag
2987      * glCopyTexSubImage is a bit picky about the parameters we pass to it
2988      */
2989     if(dst_rect.top > dst_rect.bottom) {
2990         UINT tmp = dst_rect.bottom;
2991         dst_rect.bottom = dst_rect.top;
2992         dst_rect.top = tmp;
2993         upsidedown = TRUE;
2994     }
2995
2996     if (src_offscreen)
2997     {
2998         TRACE("Reading from an offscreen target\n");
2999         upsidedown = !upsidedown;
3000         glReadBuffer(device->offscreenBuffer);
3001     }
3002     else
3003     {
3004         glReadBuffer(surface_get_gl_buffer(src_surface));
3005     }
3006
3007     /* TODO: Only back up the part that will be overwritten */
3008     glCopyTexSubImage2D(texture_target, 0,
3009                         0, 0 /* read offsets */,
3010                         0, 0,
3011                         fbwidth,
3012                         fbheight);
3013
3014     checkGLcall("glCopyTexSubImage2D");
3015
3016     /* No issue with overriding these - the sampler is dirty due to blit usage */
3017     glTexParameteri(texture_target, GL_TEXTURE_MAG_FILTER,
3018             wined3d_gl_mag_filter(magLookup, Filter));
3019     checkGLcall("glTexParameteri");
3020     glTexParameteri(texture_target, GL_TEXTURE_MIN_FILTER,
3021             wined3d_gl_min_mip_filter(minMipLookup, Filter, WINED3DTEXF_NONE));
3022     checkGLcall("glTexParameteri");
3023
3024     IWineD3DSurface_GetContainer((IWineD3DSurface *)src_surface, &IID_IWineD3DSwapChain, (void **)&src_swapchain);
3025     if (src_swapchain) IWineD3DSwapChain_Release((IWineD3DSwapChain *)src_swapchain);
3026     if (!src_swapchain || src_surface == src_swapchain->back_buffers[0])
3027     {
3028         src = backup ? backup : src_surface->texture_name;
3029     }
3030     else
3031     {
3032         glReadBuffer(GL_FRONT);
3033         checkGLcall("glReadBuffer(GL_FRONT)");
3034
3035         glGenTextures(1, &src);
3036         checkGLcall("glGenTextures(1, &src)");
3037         glBindTexture(GL_TEXTURE_2D, src);
3038         checkGLcall("glBindTexture(GL_TEXTURE_2D, src)");
3039
3040         /* TODO: Only copy the part that will be read. Use src_rect->left, src_rect->bottom as origin, but with the width watch
3041          * out for power of 2 sizes
3042          */
3043         glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA, src_surface->pow2Width,
3044                 src_surface->pow2Height, 0, GL_RGBA, GL_UNSIGNED_BYTE, NULL);
3045         checkGLcall("glTexImage2D");
3046         glCopyTexSubImage2D(GL_TEXTURE_2D, 0,
3047                             0, 0 /* read offsets */,
3048                             0, 0,
3049                             fbwidth,
3050                             fbheight);
3051
3052         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
3053         checkGLcall("glTexParameteri");
3054         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
3055         checkGLcall("glTexParameteri");
3056
3057         glReadBuffer(GL_BACK);
3058         checkGLcall("glReadBuffer(GL_BACK)");
3059
3060         if(texture_target != GL_TEXTURE_2D) {
3061             glDisable(texture_target);
3062             glEnable(GL_TEXTURE_2D);
3063             texture_target = GL_TEXTURE_2D;
3064         }
3065     }
3066     checkGLcall("glEnd and previous");
3067
3068     left = src_rect->left;
3069     right = src_rect->right;
3070
3071     if (upsidedown)
3072     {
3073         top = src_surface->currentDesc.Height - src_rect->top;
3074         bottom = src_surface->currentDesc.Height - src_rect->bottom;
3075     }
3076     else
3077     {
3078         top = src_surface->currentDesc.Height - src_rect->bottom;
3079         bottom = src_surface->currentDesc.Height - src_rect->top;
3080     }
3081
3082     if (src_surface->Flags & SFLAG_NORMCOORD)
3083     {
3084         left /= src_surface->pow2Width;
3085         right /= src_surface->pow2Width;
3086         top /= src_surface->pow2Height;
3087         bottom /= src_surface->pow2Height;
3088     }
3089
3090     /* draw the source texture stretched and upside down. The correct surface is bound already */
3091     glTexParameteri(texture_target, GL_TEXTURE_WRAP_S, GL_CLAMP);
3092     glTexParameteri(texture_target, GL_TEXTURE_WRAP_T, GL_CLAMP);
3093
3094     context_set_draw_buffer(context, drawBuffer);
3095     glReadBuffer(drawBuffer);
3096
3097     glBegin(GL_QUADS);
3098         /* bottom left */
3099         glTexCoord2f(left, bottom);
3100         glVertex2i(0, fbheight);
3101
3102         /* top left */
3103         glTexCoord2f(left, top);
3104         glVertex2i(0, fbheight - dst_rect.bottom - dst_rect.top);
3105
3106         /* top right */
3107         glTexCoord2f(right, top);
3108         glVertex2i(dst_rect.right - dst_rect.left, fbheight - dst_rect.bottom - dst_rect.top);
3109
3110         /* bottom right */
3111         glTexCoord2f(right, bottom);
3112         glVertex2i(dst_rect.right - dst_rect.left, fbheight);
3113     glEnd();
3114     checkGLcall("glEnd and previous");
3115
3116     if (texture_target != dst_surface->texture_target)
3117     {
3118         glDisable(texture_target);
3119         glEnable(dst_surface->texture_target);
3120         texture_target = dst_surface->texture_target;
3121     }
3122
3123     /* Now read the stretched and upside down image into the destination texture */
3124     glBindTexture(texture_target, dst_surface->texture_name);
3125     checkGLcall("glBindTexture");
3126     glCopyTexSubImage2D(texture_target,
3127                         0,
3128                         dst_rect.left, dst_rect.top, /* xoffset, yoffset */
3129                         0, 0, /* We blitted the image to the origin */
3130                         dst_rect.right - dst_rect.left, dst_rect.bottom - dst_rect.top);
3131     checkGLcall("glCopyTexSubImage2D");
3132
3133     if(drawBuffer == GL_BACK) {
3134         /* Write the back buffer backup back */
3135         if(backup) {
3136             if(texture_target != GL_TEXTURE_2D) {
3137                 glDisable(texture_target);
3138                 glEnable(GL_TEXTURE_2D);
3139                 texture_target = GL_TEXTURE_2D;
3140             }
3141             glBindTexture(GL_TEXTURE_2D, backup);
3142             checkGLcall("glBindTexture(GL_TEXTURE_2D, backup)");
3143         }
3144         else
3145         {
3146             if (texture_target != src_surface->texture_target)
3147             {
3148                 glDisable(texture_target);
3149                 glEnable(src_surface->texture_target);
3150                 texture_target = src_surface->texture_target;
3151             }
3152             glBindTexture(src_surface->texture_target, src_surface->texture_name);
3153             checkGLcall("glBindTexture(src_surface->texture_target, src_surface->texture_name)");
3154         }
3155
3156         glBegin(GL_QUADS);
3157             /* top left */
3158             glTexCoord2f(0.0f, (float)fbheight / (float)src_surface->pow2Height);
3159             glVertex2i(0, 0);
3160
3161             /* bottom left */
3162             glTexCoord2f(0.0f, 0.0f);
3163             glVertex2i(0, fbheight);
3164
3165             /* bottom right */
3166             glTexCoord2f((float)fbwidth / (float)src_surface->pow2Width, 0.0f);
3167             glVertex2i(fbwidth, src_surface->currentDesc.Height);
3168
3169             /* top right */
3170             glTexCoord2f((float)fbwidth / (float)src_surface->pow2Width,
3171                     (float)fbheight / (float)src_surface->pow2Height);
3172             glVertex2i(fbwidth, 0);
3173         glEnd();
3174     }
3175     glDisable(texture_target);
3176     checkGLcall("glDisable(texture_target)");
3177
3178     /* Cleanup */
3179     if (src != src_surface->texture_name && src != backup)
3180     {
3181         glDeleteTextures(1, &src);
3182         checkGLcall("glDeleteTextures(1, &src)");
3183     }
3184     if(backup) {
3185         glDeleteTextures(1, &backup);
3186         checkGLcall("glDeleteTextures(1, &backup)");
3187     }
3188
3189     LEAVE_GL();
3190
3191     if (wined3d_settings.strict_draw_ordering) wglFlush(); /* Flush to ensure ordering across contexts. */
3192
3193     context_release(context);
3194
3195     /* The texture is now most up to date - If the surface is a render target and has a drawable, this
3196      * path is never entered
3197      */
3198     surface_modify_location(dst_surface, SFLAG_INTEXTURE, TRUE);
3199 }
3200
3201 /* Until the blit_shader is ready, define some prototypes here. */
3202 static BOOL fbo_blit_supported(const struct wined3d_gl_info *gl_info, enum blit_operation blit_op,
3203                                const RECT *src_rect, DWORD src_usage, WINED3DPOOL src_pool,
3204                                const struct wined3d_format_desc *src_format_desc,
3205                                const RECT *dst_rect, DWORD dst_usage, WINED3DPOOL dst_pool,
3206                                const struct wined3d_format_desc *dst_format_desc);
3207
3208 /* Front buffer coordinates are always full screen coordinates, but our GL
3209  * drawable is limited to the window's client area. The sysmem and texture
3210  * copies do have the full screen size. Note that GL has a bottom-left
3211  * origin, while D3D has a top-left origin. */
3212 void surface_translate_frontbuffer_coords(IWineD3DSurfaceImpl *surface, HWND window, RECT *rect)
3213 {
3214     POINT offset = {0, surface->currentDesc.Height};
3215     RECT windowsize;
3216
3217     GetClientRect(window, &windowsize);
3218     offset.y -= windowsize.bottom - windowsize.top;
3219     ScreenToClient(window, &offset);
3220     OffsetRect(rect, offset.x, offset.y);
3221 }
3222
3223 static BOOL surface_is_full_rect(IWineD3DSurfaceImpl *surface, const RECT *r)
3224 {
3225     if ((r->left && r->right) || abs(r->right - r->left) != surface->currentDesc.Width)
3226         return FALSE;
3227     if ((r->top && r->bottom) || abs(r->bottom - r->top) != surface->currentDesc.Height)
3228         return FALSE;
3229     return TRUE;
3230 }
3231
3232 /* blit between surface locations. onscreen on different swapchains is not supported.
3233  * depth / stencil is not supported. */
3234 static void surface_blt_fbo(IWineD3DDeviceImpl *device, const WINED3DTEXTUREFILTERTYPE filter,
3235         IWineD3DSurfaceImpl *src_surface, DWORD src_location, const RECT *src_rect_in,
3236         IWineD3DSurfaceImpl *dst_surface, DWORD dst_location, const RECT *dst_rect_in)
3237 {
3238     const struct wined3d_gl_info *gl_info;
3239     struct wined3d_context *context;
3240     RECT src_rect, dst_rect;
3241     GLenum gl_filter;
3242
3243     TRACE("device %p, filter %s,\n", device, debug_d3dtexturefiltertype(filter));
3244     TRACE("src_surface %p, src_location %s, src_rect %s,\n",
3245             src_surface, debug_surflocation(src_location), wine_dbgstr_rect(src_rect_in));
3246     TRACE("dst_surface %p, dst_location %s, dst_rect %s.\n",
3247             dst_surface, debug_surflocation(dst_location), wine_dbgstr_rect(dst_rect_in));
3248
3249     src_rect = *src_rect_in;
3250     dst_rect = *dst_rect_in;
3251
3252     switch (filter)
3253     {
3254         case WINED3DTEXF_LINEAR:
3255             gl_filter = GL_LINEAR;
3256             break;
3257
3258         default:
3259             FIXME("Unsupported filter mode %s (%#x).\n", debug_d3dtexturefiltertype(filter), filter);
3260         case WINED3DTEXF_NONE:
3261         case WINED3DTEXF_POINT:
3262             gl_filter = GL_NEAREST;
3263             break;
3264     }
3265
3266     if (src_location == SFLAG_INDRAWABLE && surface_is_offscreen(src_surface))
3267         src_location = SFLAG_INTEXTURE;
3268     if (dst_location == SFLAG_INDRAWABLE && surface_is_offscreen(dst_surface))
3269         dst_location = SFLAG_INTEXTURE;
3270
3271     /* Make sure the locations are up-to-date. Loading the destination
3272      * surface isn't required if the entire surface is overwritten. (And is
3273      * in fact harmful if we're being called by surface_load_location() with
3274      * the purpose of loading the destination surface.) */
3275     surface_load_location(src_surface, src_location, NULL);
3276     if (!surface_is_full_rect(dst_surface, &dst_rect))
3277         surface_load_location(dst_surface, dst_location, NULL);
3278
3279     if (src_location == SFLAG_INDRAWABLE) context = context_acquire(device, src_surface);
3280     else if (dst_location == SFLAG_INDRAWABLE) context = context_acquire(device, dst_surface);
3281     else context = context_acquire(device, NULL);
3282
3283     if (!context->valid)
3284     {
3285         context_release(context);
3286         WARN("Invalid context, skipping blit.\n");
3287         return;
3288     }
3289
3290     gl_info = context->gl_info;
3291
3292     if (src_location == SFLAG_INDRAWABLE)
3293     {
3294         GLenum buffer = surface_get_gl_buffer(src_surface);
3295
3296         TRACE("Source surface %p is onscreen.\n", src_surface);
3297
3298         if (buffer == GL_FRONT)
3299             surface_translate_frontbuffer_coords(src_surface, context->win_handle, &src_rect);
3300
3301         src_rect.top = src_surface->currentDesc.Height - src_rect.top;
3302         src_rect.bottom = src_surface->currentDesc.Height - src_rect.bottom;
3303
3304         ENTER_GL();
3305         context_bind_fbo(context, GL_READ_FRAMEBUFFER, NULL);
3306         glReadBuffer(buffer);
3307         checkGLcall("glReadBuffer()");
3308     }
3309     else
3310     {
3311         TRACE("Source surface %p is offscreen.\n", src_surface);
3312         ENTER_GL();
3313         context_apply_fbo_state_blit(context, GL_READ_FRAMEBUFFER, src_surface, NULL, src_location);
3314         glReadBuffer(GL_COLOR_ATTACHMENT0);
3315         checkGLcall("glReadBuffer()");
3316     }
3317     LEAVE_GL();
3318
3319     if (dst_location == SFLAG_INDRAWABLE)
3320     {
3321         GLenum buffer = surface_get_gl_buffer(dst_surface);
3322
3323         TRACE("Destination surface %p is onscreen.\n", dst_surface);
3324
3325         if (buffer == GL_FRONT)
3326             surface_translate_frontbuffer_coords(dst_surface, context->win_handle, &dst_rect);
3327
3328         dst_rect.top = dst_surface->currentDesc.Height - dst_rect.top;
3329         dst_rect.bottom = dst_surface->currentDesc.Height - dst_rect.bottom;
3330
3331         ENTER_GL();
3332         context_bind_fbo(context, GL_DRAW_FRAMEBUFFER, NULL);
3333         context_set_draw_buffer(context, buffer);
3334     }
3335     else
3336     {
3337         TRACE("Destination surface %p is offscreen.\n", dst_surface);
3338
3339         ENTER_GL();
3340         context_apply_fbo_state_blit(context, GL_DRAW_FRAMEBUFFER, dst_surface, NULL, dst_location);
3341         context_set_draw_buffer(context, GL_COLOR_ATTACHMENT0);
3342     }
3343
3344     glColorMask(GL_TRUE, GL_TRUE, GL_TRUE, GL_TRUE);
3345     IWineD3DDeviceImpl_MarkStateDirty(device, STATE_RENDER(WINED3DRS_COLORWRITEENABLE));
3346     IWineD3DDeviceImpl_MarkStateDirty(device, STATE_RENDER(WINED3DRS_COLORWRITEENABLE1));
3347     IWineD3DDeviceImpl_MarkStateDirty(device, STATE_RENDER(WINED3DRS_COLORWRITEENABLE2));
3348     IWineD3DDeviceImpl_MarkStateDirty(device, STATE_RENDER(WINED3DRS_COLORWRITEENABLE3));
3349
3350     glDisable(GL_SCISSOR_TEST);
3351     IWineD3DDeviceImpl_MarkStateDirty(device, STATE_RENDER(WINED3DRS_SCISSORTESTENABLE));
3352
3353     gl_info->fbo_ops.glBlitFramebuffer(src_rect.left, src_rect.top, src_rect.right, src_rect.bottom,
3354             dst_rect.left, dst_rect.top, dst_rect.right, dst_rect.bottom, GL_COLOR_BUFFER_BIT, gl_filter);
3355     checkGLcall("glBlitFramebuffer()");
3356
3357     LEAVE_GL();
3358
3359     if (wined3d_settings.strict_draw_ordering) wglFlush(); /* Flush to ensure ordering across contexts. */
3360
3361     context_release(context);
3362 }
3363
3364 /* Not called from the VTable */
3365 static HRESULT IWineD3DSurfaceImpl_BltOverride(IWineD3DSurfaceImpl *dst_surface, const RECT *DestRect,
3366         IWineD3DSurfaceImpl *src_surface, const RECT *SrcRect, DWORD Flags, const WINEDDBLTFX *DDBltFx,
3367         WINED3DTEXTUREFILTERTYPE Filter)
3368 {
3369     IWineD3DDeviceImpl *device = dst_surface->resource.device;
3370     IWineD3DSwapChainImpl *srcSwapchain = NULL, *dstSwapchain = NULL;
3371     RECT dst_rect, src_rect;
3372
3373     TRACE("dst_surface %p, dst_rect %s, src_surface %p, src_rect %s, flags %#x, blt_fx %p, filter %s.\n",
3374             dst_surface, wine_dbgstr_rect(DestRect), src_surface, wine_dbgstr_rect(SrcRect),
3375             Flags, DDBltFx, debug_d3dtexturefiltertype(Filter));
3376
3377     /* Get the swapchain. One of the surfaces has to be a primary surface */
3378     if (dst_surface->resource.pool == WINED3DPOOL_SYSTEMMEM)
3379     {
3380         WARN("Destination is in sysmem, rejecting gl blt\n");
3381         return WINED3DERR_INVALIDCALL;
3382     }
3383     IWineD3DSurface_GetContainer((IWineD3DSurface *)dst_surface, &IID_IWineD3DSwapChain, (void **)&dstSwapchain);
3384     if (dstSwapchain) IWineD3DSwapChain_Release((IWineD3DSwapChain *)dstSwapchain);
3385     if (src_surface)
3386     {
3387         if (src_surface->resource.pool == WINED3DPOOL_SYSTEMMEM)
3388         {
3389             WARN("Src is in sysmem, rejecting gl blt\n");
3390             return WINED3DERR_INVALIDCALL;
3391         }
3392         IWineD3DSurface_GetContainer((IWineD3DSurface *)src_surface, &IID_IWineD3DSwapChain, (void **)&srcSwapchain);
3393         if (srcSwapchain) IWineD3DSwapChain_Release((IWineD3DSwapChain *)srcSwapchain);
3394     }
3395
3396     /* Early sort out of cases where no render target is used */
3397     if (!dstSwapchain && !srcSwapchain
3398             && src_surface != device->render_targets[0]
3399             && dst_surface != device->render_targets[0])
3400     {
3401         TRACE("No surface is render target, not using hardware blit.\n");
3402         return WINED3DERR_INVALIDCALL;
3403     }
3404
3405     /* No destination color keying supported */
3406     if(Flags & (WINEDDBLT_KEYDEST | WINEDDBLT_KEYDESTOVERRIDE)) {
3407         /* Can we support that with glBlendFunc if blitting to the frame buffer? */
3408         TRACE("Destination color key not supported in accelerated Blit, falling back to software\n");
3409         return WINED3DERR_INVALIDCALL;
3410     }
3411
3412     surface_get_rect(dst_surface, DestRect, &dst_rect);
3413     if (src_surface) surface_get_rect(src_surface, SrcRect, &src_rect);
3414
3415     /* The only case where both surfaces on a swapchain are supported is a back buffer -> front buffer blit on the same swapchain */
3416     if (dstSwapchain && dstSwapchain == srcSwapchain && dstSwapchain->back_buffers
3417             && dst_surface == dstSwapchain->front_buffer
3418             && src_surface == dstSwapchain->back_buffers[0])
3419     {
3420         /* Half-Life does a Blt from the back buffer to the front buffer,
3421          * Full surface size, no flags... Use present instead
3422          *
3423          * This path will only be entered for d3d7 and ddraw apps, because d3d8/9 offer no way to blit TO the front buffer
3424          */
3425
3426         /* Check rects - IWineD3DDevice_Present doesn't handle them */
3427         while(1)
3428         {
3429             TRACE("Looking if a Present can be done...\n");
3430             /* Source Rectangle must be full surface */
3431             if (src_rect.left || src_rect.top
3432                     || src_rect.right != src_surface->currentDesc.Width
3433                     || src_rect.bottom != src_surface->currentDesc.Height)
3434             {
3435                 TRACE("No, Source rectangle doesn't match\n");
3436                 break;
3437             }
3438
3439             /* No stretching may occur */
3440             if(src_rect.right != dst_rect.right - dst_rect.left ||
3441                src_rect.bottom != dst_rect.bottom - dst_rect.top) {
3442                 TRACE("No, stretching is done\n");
3443                 break;
3444             }
3445
3446             /* Destination must be full surface or match the clipping rectangle */
3447             if (dst_surface->clipper && ((IWineD3DClipperImpl *)dst_surface->clipper)->hWnd)
3448             {
3449                 RECT cliprect;
3450                 POINT pos[2];
3451                 GetClientRect(((IWineD3DClipperImpl *)dst_surface->clipper)->hWnd, &cliprect);
3452                 pos[0].x = dst_rect.left;
3453                 pos[0].y = dst_rect.top;
3454                 pos[1].x = dst_rect.right;
3455                 pos[1].y = dst_rect.bottom;
3456                 MapWindowPoints(GetDesktopWindow(), ((IWineD3DClipperImpl *)dst_surface->clipper)->hWnd, pos, 2);
3457
3458                 if(pos[0].x != cliprect.left  || pos[0].y != cliprect.top   ||
3459                    pos[1].x != cliprect.right || pos[1].y != cliprect.bottom)
3460                 {
3461                     TRACE("No, dest rectangle doesn't match(clipper)\n");
3462                     TRACE("Clip rect at %s\n", wine_dbgstr_rect(&cliprect));
3463                     TRACE("Blt dest: %s\n", wine_dbgstr_rect(&dst_rect));
3464                     break;
3465                 }
3466             }
3467             else if (dst_rect.left || dst_rect.top
3468                     || dst_rect.right != dst_surface->currentDesc.Width
3469                     || dst_rect.bottom != dst_surface->currentDesc.Height)
3470             {
3471                 TRACE("No, dest rectangle doesn't match(surface size)\n");
3472                 break;
3473             }
3474
3475             TRACE("Yes\n");
3476
3477             /* These flags are unimportant for the flag check, remove them */
3478             if((Flags & ~(WINEDDBLT_DONOTWAIT | WINEDDBLT_WAIT)) == 0) {
3479                 WINED3DSWAPEFFECT orig_swap = dstSwapchain->presentParms.SwapEffect;
3480
3481                 /* The idea behind this is that a glReadPixels and a glDrawPixels call
3482                     * take very long, while a flip is fast.
3483                     * This applies to Half-Life, which does such Blts every time it finished
3484                     * a frame, and to Prince of Persia 3D, which uses this to draw at least the main
3485                     * menu. This is also used by all apps when they do windowed rendering
3486                     *
3487                     * The problem is that flipping is not really the same as copying. After a
3488                     * Blt the front buffer is a copy of the back buffer, and the back buffer is
3489                     * untouched. Therefore it's necessary to override the swap effect
3490                     * and to set it back after the flip.
3491                     *
3492                     * Windowed Direct3D < 7 apps do the same. The D3D7 sdk demos are nice
3493                     * testcases.
3494                     */
3495
3496                 dstSwapchain->presentParms.SwapEffect = WINED3DSWAPEFFECT_COPY;
3497                 dstSwapchain->presentParms.PresentationInterval = WINED3DPRESENT_INTERVAL_IMMEDIATE;
3498
3499                 TRACE("Full screen back buffer -> front buffer blt, performing a flip instead\n");
3500                 IWineD3DSwapChain_Present((IWineD3DSwapChain *)dstSwapchain,
3501                         NULL, NULL, dstSwapchain->win_handle, NULL, 0);
3502
3503                 dstSwapchain->presentParms.SwapEffect = orig_swap;
3504
3505                 return WINED3D_OK;
3506             }
3507             break;
3508         }
3509
3510         TRACE("Unsupported blit between buffers on the same swapchain\n");
3511         return WINED3DERR_INVALIDCALL;
3512     } else if(dstSwapchain && dstSwapchain == srcSwapchain) {
3513         FIXME("Implement hardware blit between two surfaces on the same swapchain\n");
3514         return WINED3DERR_INVALIDCALL;
3515     } else if(dstSwapchain && srcSwapchain) {
3516         FIXME("Implement hardware blit between two different swapchains\n");
3517         return WINED3DERR_INVALIDCALL;
3518     }
3519     else if (dstSwapchain)
3520     {
3521         /* Handled with regular texture -> swapchain blit */
3522         if (src_surface == device->render_targets[0])
3523             TRACE("Blit from active render target to a swapchain\n");
3524     }
3525     else if (srcSwapchain && dst_surface == device->render_targets[0])
3526     {
3527         FIXME("Implement blit from a swapchain to the active render target\n");
3528         return WINED3DERR_INVALIDCALL;
3529     }
3530
3531     if ((srcSwapchain || src_surface == device->render_targets[0]) && !dstSwapchain)
3532     {
3533         /* Blit from render target to texture */
3534         BOOL stretchx;
3535
3536         /* P8 read back is not implemented */
3537         if (src_surface->resource.format_desc->format == WINED3DFMT_P8_UINT
3538                 || dst_surface->resource.format_desc->format == WINED3DFMT_P8_UINT)
3539         {
3540             TRACE("P8 read back not supported by frame buffer to texture blit\n");
3541             return WINED3DERR_INVALIDCALL;
3542         }
3543
3544         if(Flags & (WINEDDBLT_KEYSRC | WINEDDBLT_KEYSRCOVERRIDE)) {
3545             TRACE("Color keying not supported by frame buffer to texture blit\n");
3546             return WINED3DERR_INVALIDCALL;
3547             /* Destination color key is checked above */
3548         }
3549
3550         if(dst_rect.right - dst_rect.left != src_rect.right - src_rect.left) {
3551             stretchx = TRUE;
3552         } else {
3553             stretchx = FALSE;
3554         }
3555
3556         /* Blt is a pretty powerful call, while glCopyTexSubImage2D is not. glCopyTexSubImage cannot
3557          * flip the image nor scale it.
3558          *
3559          * -> If the app asks for a unscaled, upside down copy, just perform one glCopyTexSubImage2D call
3560          * -> If the app wants a image width an unscaled width, copy it line per line
3561          * -> If the app wants a image that is scaled on the x axis, and the destination rectangle is smaller
3562          *    than the frame buffer, draw an upside down scaled image onto the fb, read it back and restore the
3563          *    back buffer. This is slower than reading line per line, thus not used for flipping
3564          * -> If the app wants a scaled image with a dest rect that is bigger than the fb, it has to be copied
3565          *    pixel by pixel
3566          *
3567          * If EXT_framebuffer_blit is supported that can be used instead. Note that EXT_framebuffer_blit implies
3568          * FBO support, so it doesn't really make sense to try and make it work with different offscreen rendering
3569          * backends.
3570          */
3571         if (fbo_blit_supported(&device->adapter->gl_info, BLIT_OP_BLIT,
3572                 &src_rect, src_surface->resource.usage, src_surface->resource.pool, src_surface->resource.format_desc,
3573                 &dst_rect, dst_surface->resource.usage, dst_surface->resource.pool, dst_surface->resource.format_desc))
3574         {
3575             surface_blt_fbo(device, Filter,
3576                     src_surface, SFLAG_INDRAWABLE, &src_rect,
3577                     dst_surface, SFLAG_INDRAWABLE, &dst_rect);
3578             surface_modify_location(dst_surface, SFLAG_INDRAWABLE, TRUE);
3579         }
3580         else if (!stretchx || dst_rect.right - dst_rect.left > src_surface->currentDesc.Width
3581                 || dst_rect.bottom - dst_rect.top > src_surface->currentDesc.Height)
3582         {
3583             TRACE("No stretching in x direction, using direct framebuffer -> texture copy\n");
3584             fb_copy_to_texture_direct(dst_surface, src_surface, &src_rect, &dst_rect, Filter);
3585         } else {
3586             TRACE("Using hardware stretching to flip / stretch the texture\n");
3587             fb_copy_to_texture_hwstretch(dst_surface, src_surface, &src_rect, &dst_rect, Filter);
3588         }
3589
3590         if (!(dst_surface->Flags & SFLAG_DONOTFREE))
3591         {
3592             HeapFree(GetProcessHeap(), 0, dst_surface->resource.heapMemory);
3593             dst_surface->resource.allocatedMemory = NULL;
3594             dst_surface->resource.heapMemory = NULL;
3595         }
3596         else
3597         {
3598             dst_surface->Flags &= ~SFLAG_INSYSMEM;
3599         }
3600
3601         return WINED3D_OK;
3602     }
3603     else if (src_surface)
3604     {
3605         /* Blit from offscreen surface to render target */
3606         DWORD oldCKeyFlags = src_surface->CKeyFlags;
3607         WINEDDCOLORKEY oldBltCKey = src_surface->SrcBltCKey;
3608         struct wined3d_context *context;
3609
3610         TRACE("Blt from surface %p to rendertarget %p\n", src_surface, dst_surface);
3611
3612         if (!(Flags & (WINEDDBLT_KEYSRC | WINEDDBLT_KEYSRCOVERRIDE))
3613                 && fbo_blit_supported(&device->adapter->gl_info, BLIT_OP_BLIT,
3614                         &src_rect, src_surface->resource.usage, src_surface->resource.pool,
3615                         src_surface->resource.format_desc,
3616                         &dst_rect, dst_surface->resource.usage, dst_surface->resource.pool,
3617                         dst_surface->resource.format_desc))
3618         {
3619             TRACE("Using surface_blt_fbo.\n");
3620             /* The source is always a texture, but never the currently active render target, and the texture
3621              * contents are never upside down. */
3622             surface_blt_fbo(device, Filter,
3623                     src_surface, SFLAG_INDRAWABLE, &src_rect,
3624                     dst_surface, SFLAG_INDRAWABLE, &dst_rect);
3625             surface_modify_location(dst_surface, SFLAG_INDRAWABLE, TRUE);
3626             return WINED3D_OK;
3627         }
3628
3629         if (!(Flags & (WINEDDBLT_KEYSRC | WINEDDBLT_KEYSRCOVERRIDE))
3630                 && arbfp_blit.blit_supported(&device->adapter->gl_info, BLIT_OP_BLIT,
3631                         &src_rect, src_surface->resource.usage, src_surface->resource.pool,
3632                         src_surface->resource.format_desc,
3633                         &dst_rect, dst_surface->resource.usage, dst_surface->resource.pool,
3634                         dst_surface->resource.format_desc))
3635         {
3636             return arbfp_blit_surface(device, src_surface, &src_rect, dst_surface, &dst_rect, BLIT_OP_BLIT, Filter);
3637         }
3638
3639         /* Color keying: Check if we have to do a color keyed blt,
3640          * and if not check if a color key is activated.
3641          *
3642          * Just modify the color keying parameters in the surface and restore them afterwards
3643          * The surface keeps track of the color key last used to load the opengl surface.
3644          * PreLoad will catch the change to the flags and color key and reload if necessary.
3645          */
3646         if(Flags & WINEDDBLT_KEYSRC) {
3647             /* Use color key from surface */
3648         } else if(Flags & WINEDDBLT_KEYSRCOVERRIDE) {
3649             /* Use color key from DDBltFx */
3650             src_surface->CKeyFlags |= WINEDDSD_CKSRCBLT;
3651             src_surface->SrcBltCKey = DDBltFx->ddckSrcColorkey;
3652         } else {
3653             /* Do not use color key */
3654             src_surface->CKeyFlags &= ~WINEDDSD_CKSRCBLT;
3655         }
3656
3657         /* Now load the surface */
3658         surface_internal_preload(src_surface, SRGB_RGB);
3659
3660         /* Activate the destination context, set it up for blitting */
3661         context = context_acquire(device, dst_surface);
3662         context_apply_blit_state(context, device);
3663
3664         if (dstSwapchain && dst_surface == dstSwapchain->front_buffer)
3665             surface_translate_frontbuffer_coords(dst_surface, context->win_handle, &dst_rect);
3666
3667         if (!device->blitter->blit_supported(&device->adapter->gl_info, BLIT_OP_BLIT,
3668                 &src_rect, src_surface->resource.usage, src_surface->resource.pool, src_surface->resource.format_desc,
3669                 &dst_rect, dst_surface->resource.usage, dst_surface->resource.pool, dst_surface->resource.format_desc))
3670         {
3671             FIXME("Unsupported blit operation falling back to software\n");
3672             return WINED3DERR_INVALIDCALL;
3673         }
3674
3675         device->blitter->set_shader((IWineD3DDevice *)device, src_surface);
3676
3677         ENTER_GL();
3678
3679         /* This is for color keying */
3680         if(Flags & (WINEDDBLT_KEYSRC | WINEDDBLT_KEYSRCOVERRIDE)) {
3681             glEnable(GL_ALPHA_TEST);
3682             checkGLcall("glEnable(GL_ALPHA_TEST)");
3683
3684             /* When the primary render target uses P8, the alpha component contains the palette index.
3685              * Which means that the colorkey is one of the palette entries. In other cases pixels that
3686              * should be masked away have alpha set to 0. */
3687             if (primary_render_target_is_p8(device))
3688                 glAlphaFunc(GL_NOTEQUAL, (float)src_surface->SrcBltCKey.dwColorSpaceLowValue / 256.0f);
3689             else
3690                 glAlphaFunc(GL_NOTEQUAL, 0.0f);
3691             checkGLcall("glAlphaFunc");
3692         } else {
3693             glDisable(GL_ALPHA_TEST);
3694             checkGLcall("glDisable(GL_ALPHA_TEST)");
3695         }
3696
3697         /* Draw a textured quad
3698          */
3699         draw_textured_quad(src_surface, &src_rect, &dst_rect, Filter);
3700
3701         if(Flags & (WINEDDBLT_KEYSRC | WINEDDBLT_KEYSRCOVERRIDE)) {
3702             glDisable(GL_ALPHA_TEST);
3703             checkGLcall("glDisable(GL_ALPHA_TEST)");
3704         }
3705
3706         /* Restore the color key parameters */
3707         src_surface->CKeyFlags = oldCKeyFlags;
3708         src_surface->SrcBltCKey = oldBltCKey;
3709
3710         LEAVE_GL();
3711
3712         /* Leave the opengl state valid for blitting */
3713         device->blitter->unset_shader((IWineD3DDevice *)device);
3714
3715         if (wined3d_settings.strict_draw_ordering || (dstSwapchain
3716                 && (dst_surface == dstSwapchain->front_buffer
3717                 || dstSwapchain->num_contexts > 1)))
3718             wglFlush(); /* Flush to ensure ordering across contexts. */
3719
3720         context_release(context);
3721
3722         /* TODO: If the surface is locked often, perform the Blt in software on the memory instead */
3723         /* The surface is now in the drawable. On onscreen surfaces or without fbos the texture
3724          * is outdated now
3725          */
3726         surface_modify_location(dst_surface, SFLAG_INDRAWABLE, TRUE);
3727
3728         return WINED3D_OK;
3729     } else {
3730         /* Source-Less Blit to render target */
3731         if (Flags & WINEDDBLT_COLORFILL) {
3732             DWORD color;
3733
3734             TRACE("Colorfill\n");
3735
3736             /* The color as given in the Blt function is in the format of the frame-buffer...
3737              * 'clear' expect it in ARGB format => we need to do some conversion :-)
3738              */
3739             if (!surface_convert_color_to_argb(dst_surface, DDBltFx->u5.dwFillColor, &color))
3740             {
3741                 /* The color conversion function already prints an error, so need to do it here */
3742                 return WINED3DERR_INVALIDCALL;
3743             }
3744
3745             if (ffp_blit.blit_supported(&device->adapter->gl_info, BLIT_OP_COLOR_FILL,
3746                     NULL, 0, 0, NULL,
3747                     &dst_rect, dst_surface->resource.usage, dst_surface->resource.pool,
3748                     dst_surface->resource.format_desc))
3749             {
3750                 return ffp_blit.color_fill(device, dst_surface, &dst_rect, color);
3751             }
3752             else if (cpu_blit.blit_supported(&device->adapter->gl_info, BLIT_OP_COLOR_FILL,
3753                     NULL, 0, 0, NULL,
3754                     &dst_rect, dst_surface->resource.usage, dst_surface->resource.pool,
3755                     dst_surface->resource.format_desc))
3756             {
3757                 return cpu_blit.color_fill(device, dst_surface, &dst_rect, color);
3758             }
3759             return WINED3DERR_INVALIDCALL;
3760         }
3761     }
3762
3763     /* Default: Fall back to the generic blt. Not an error, a TRACE is enough */
3764     TRACE("Didn't find any usable render target setup for hw blit, falling back to software\n");
3765     return WINED3DERR_INVALIDCALL;
3766 }
3767
3768 static HRESULT IWineD3DSurfaceImpl_BltZ(IWineD3DSurfaceImpl *This, const RECT *DestRect,
3769         IWineD3DSurface *SrcSurface, const RECT *SrcRect, DWORD Flags, const WINEDDBLTFX *DDBltFx)
3770 {
3771     IWineD3DDeviceImpl *device = This->resource.device;
3772     float depth;
3773
3774     if (Flags & WINEDDBLT_DEPTHFILL) {
3775         switch(This->resource.format_desc->format)
3776         {
3777             case WINED3DFMT_D16_UNORM:
3778                 depth = (float) DDBltFx->u5.dwFillDepth / (float) 0x0000ffff;
3779                 break;
3780             case WINED3DFMT_S1_UINT_D15_UNORM:
3781                 depth = (float) DDBltFx->u5.dwFillDepth / (float) 0x00007fff;
3782                 break;
3783             case WINED3DFMT_D24_UNORM_S8_UINT:
3784             case WINED3DFMT_X8D24_UNORM:
3785                 depth = (float) DDBltFx->u5.dwFillDepth / (float) 0x00ffffff;
3786                 break;
3787             case WINED3DFMT_D32_UNORM:
3788                 depth = (float) DDBltFx->u5.dwFillDepth / (float) 0xffffffff;
3789                 break;
3790             default:
3791                 depth = 0.0f;
3792                 ERR("Unexpected format for depth fill: %s\n", debug_d3dformat(This->resource.format_desc->format));
3793         }
3794
3795         return IWineD3DDevice_Clear((IWineD3DDevice *)device, DestRect ? 1 : 0, (const WINED3DRECT *)DestRect,
3796                 WINED3DCLEAR_ZBUFFER, 0x00000000, depth, 0x00000000);
3797     }
3798
3799     FIXME("(%p): Unsupp depthstencil blit\n", This);
3800     return WINED3DERR_INVALIDCALL;
3801 }
3802
3803 static HRESULT WINAPI IWineD3DSurfaceImpl_Blt(IWineD3DSurface *iface, const RECT *DestRect, IWineD3DSurface *SrcSurface,
3804         const RECT *SrcRect, DWORD Flags, const WINEDDBLTFX *DDBltFx, WINED3DTEXTUREFILTERTYPE Filter) {
3805     IWineD3DSurfaceImpl *This = (IWineD3DSurfaceImpl *)iface;
3806     IWineD3DSurfaceImpl *Src = (IWineD3DSurfaceImpl *) SrcSurface;
3807     IWineD3DDeviceImpl *device = This->resource.device;
3808
3809     TRACE("(%p)->(%p,%p,%p,%x,%p)\n", This, DestRect, SrcSurface, SrcRect, Flags, DDBltFx);
3810     TRACE("(%p): Usage is %s\n", This, debug_d3dusage(This->resource.usage));
3811
3812     if ( (This->Flags & SFLAG_LOCKED) || ((Src != NULL) && (Src->Flags & SFLAG_LOCKED)))
3813     {
3814         WARN(" Surface is busy, returning DDERR_SURFACEBUSY\n");
3815         return WINEDDERR_SURFACEBUSY;
3816     }
3817
3818     /* Accessing the depth stencil is supposed to fail between a BeginScene and EndScene pair,
3819      * except depth blits, which seem to work
3820      */
3821     if (This == device->depth_stencil || (Src && Src == device->depth_stencil))
3822     {
3823         if (device->inScene && !(Flags & WINEDDBLT_DEPTHFILL))
3824         {
3825             TRACE("Attempt to access the depth stencil surface in a BeginScene / EndScene pair, returning WINED3DERR_INVALIDCALL\n");
3826             return WINED3DERR_INVALIDCALL;
3827         } else if(IWineD3DSurfaceImpl_BltZ(This, DestRect, SrcSurface, SrcRect, Flags, DDBltFx) == WINED3D_OK) {
3828             TRACE("Z Blit override handled the blit\n");
3829             return WINED3D_OK;
3830         }
3831     }
3832
3833     /* Special cases for RenderTargets */
3834     if ((This->resource.usage & WINED3DUSAGE_RENDERTARGET)
3835             || (Src && (Src->resource.usage & WINED3DUSAGE_RENDERTARGET)))
3836     {
3837         if (SUCCEEDED(IWineD3DSurfaceImpl_BltOverride(This, DestRect, Src, SrcRect, Flags, DDBltFx, Filter)))
3838             return WINED3D_OK;
3839     }
3840
3841     /* For the rest call the X11 surface implementation.
3842      * For RenderTargets this should be implemented OpenGL accelerated in BltOverride,
3843      * other Blts are rather rare
3844      */
3845     return IWineD3DBaseSurfaceImpl_Blt(iface, DestRect, SrcSurface, SrcRect, Flags, DDBltFx, Filter);
3846 }
3847
3848 static HRESULT WINAPI IWineD3DSurfaceImpl_BltFast(IWineD3DSurface *iface, DWORD dstx, DWORD dsty,
3849         IWineD3DSurface *Source, const RECT *rsrc, DWORD trans)
3850 {
3851     IWineD3DSurfaceImpl *This = (IWineD3DSurfaceImpl *) iface;
3852     IWineD3DSurfaceImpl *srcImpl = (IWineD3DSurfaceImpl *) Source;
3853     IWineD3DDeviceImpl *device = This->resource.device;
3854
3855     TRACE("(%p)->(%d, %d, %p, %p, %08x\n", iface, dstx, dsty, Source, rsrc, trans);
3856
3857     if ( (This->Flags & SFLAG_LOCKED) || (srcImpl->Flags & SFLAG_LOCKED))
3858     {
3859         WARN(" Surface is busy, returning DDERR_SURFACEBUSY\n");
3860         return WINEDDERR_SURFACEBUSY;
3861     }
3862
3863     if (device->inScene && (This == device->depth_stencil || srcImpl == device->depth_stencil))
3864     {
3865         TRACE("Attempt to access the depth stencil surface in a BeginScene / EndScene pair, returning WINED3DERR_INVALIDCALL\n");
3866         return WINED3DERR_INVALIDCALL;
3867     }
3868
3869     /* Special cases for RenderTargets */
3870     if( (This->resource.usage & WINED3DUSAGE_RENDERTARGET) ||
3871         (srcImpl->resource.usage & WINED3DUSAGE_RENDERTARGET) ) {
3872
3873         RECT SrcRect, DstRect;
3874         DWORD Flags=0;
3875
3876         surface_get_rect(srcImpl, rsrc, &SrcRect);
3877
3878         DstRect.left = dstx;
3879         DstRect.top=dsty;
3880         DstRect.right = dstx + SrcRect.right - SrcRect.left;
3881         DstRect.bottom = dsty + SrcRect.bottom - SrcRect.top;
3882
3883         /* Convert BltFast flags into Btl ones because it is called from SurfaceImpl_Blt as well */
3884         if(trans & WINEDDBLTFAST_SRCCOLORKEY)
3885             Flags |= WINEDDBLT_KEYSRC;
3886         if(trans & WINEDDBLTFAST_DESTCOLORKEY)
3887             Flags |= WINEDDBLT_KEYDEST;
3888         if(trans & WINEDDBLTFAST_WAIT)
3889             Flags |= WINEDDBLT_WAIT;
3890         if(trans & WINEDDBLTFAST_DONOTWAIT)
3891             Flags |= WINEDDBLT_DONOTWAIT;
3892
3893         if (SUCCEEDED(IWineD3DSurfaceImpl_BltOverride(This,
3894                 &DstRect, srcImpl, &SrcRect, Flags, NULL, WINED3DTEXF_POINT)))
3895             return WINED3D_OK;
3896     }
3897
3898
3899     return IWineD3DBaseSurfaceImpl_BltFast(iface, dstx, dsty, Source, rsrc, trans);
3900 }
3901
3902 static HRESULT WINAPI IWineD3DSurfaceImpl_RealizePalette(IWineD3DSurface *iface)
3903 {
3904     IWineD3DSurfaceImpl *This = (IWineD3DSurfaceImpl *) iface;
3905     RGBQUAD col[256];
3906     IWineD3DPaletteImpl *pal = This->palette;
3907     unsigned int n;
3908     TRACE("(%p)\n", This);
3909
3910     if (!pal) return WINED3D_OK;
3911
3912     if (This->resource.format_desc->format == WINED3DFMT_P8_UINT
3913             || This->resource.format_desc->format == WINED3DFMT_P8_UINT_A8_UNORM)
3914     {
3915         if (This->resource.usage & WINED3DUSAGE_RENDERTARGET)
3916         {
3917             /* Make sure the texture is up to date. This call doesn't do
3918              * anything if the texture is already up to date. */
3919             surface_load_location(This, SFLAG_INTEXTURE, NULL);
3920
3921             /* We want to force a palette refresh, so mark the drawable as not being up to date */
3922             surface_modify_location(This, SFLAG_INDRAWABLE, FALSE);
3923         }
3924         else
3925         {
3926             if (!(This->Flags & SFLAG_INSYSMEM))
3927             {
3928                 TRACE("Palette changed with surface that does not have an up to date system memory copy.\n");
3929                 surface_load_location(This, SFLAG_INSYSMEM, NULL);
3930             }
3931             TRACE("Dirtifying surface\n");
3932             surface_modify_location(This, SFLAG_INSYSMEM, TRUE);
3933         }
3934     }
3935
3936     if(This->Flags & SFLAG_DIBSECTION) {
3937         TRACE("(%p): Updating the hdc's palette\n", This);
3938         for (n=0; n<256; n++) {
3939             col[n].rgbRed   = pal->palents[n].peRed;
3940             col[n].rgbGreen = pal->palents[n].peGreen;
3941             col[n].rgbBlue  = pal->palents[n].peBlue;
3942             col[n].rgbReserved = 0;
3943         }
3944         SetDIBColorTable(This->hDC, 0, 256, col);
3945     }
3946
3947     /* Propagate the changes to the drawable when we have a palette. */
3948     if (This->resource.usage & WINED3DUSAGE_RENDERTARGET)
3949         surface_load_location(This, SFLAG_INDRAWABLE, NULL);
3950
3951     return WINED3D_OK;
3952 }
3953
3954 static HRESULT WINAPI IWineD3DSurfaceImpl_PrivateSetup(IWineD3DSurface *iface) {
3955     /** Check against the maximum texture sizes supported by the video card **/
3956     IWineD3DSurfaceImpl *This = (IWineD3DSurfaceImpl *) iface;
3957     const struct wined3d_gl_info *gl_info = &This->resource.device->adapter->gl_info;
3958     unsigned int pow2Width, pow2Height;
3959
3960     This->texture_name = 0;
3961     This->texture_target = GL_TEXTURE_2D;
3962
3963     /* Non-power2 support */
3964     if (gl_info->supported[ARB_TEXTURE_NON_POWER_OF_TWO] || gl_info->supported[WINED3D_GL_NORMALIZED_TEXRECT])
3965     {
3966         pow2Width = This->currentDesc.Width;
3967         pow2Height = This->currentDesc.Height;
3968     }
3969     else
3970     {
3971         /* Find the nearest pow2 match */
3972         pow2Width = pow2Height = 1;
3973         while (pow2Width < This->currentDesc.Width) pow2Width <<= 1;
3974         while (pow2Height < This->currentDesc.Height) pow2Height <<= 1;
3975     }
3976     This->pow2Width  = pow2Width;
3977     This->pow2Height = pow2Height;
3978
3979     if (pow2Width > This->currentDesc.Width || pow2Height > This->currentDesc.Height) {
3980         /** TODO: add support for non power two compressed textures **/
3981         if (This->resource.format_desc->Flags & WINED3DFMT_FLAG_COMPRESSED)
3982         {
3983             FIXME("(%p) Compressed non-power-two textures are not supported w(%d) h(%d)\n",
3984                   This, This->currentDesc.Width, This->currentDesc.Height);
3985             return WINED3DERR_NOTAVAILABLE;
3986         }
3987     }
3988
3989     if(pow2Width != This->currentDesc.Width ||
3990        pow2Height != This->currentDesc.Height) {
3991         This->Flags |= SFLAG_NONPOW2;
3992     }
3993
3994     TRACE("%p\n", This);
3995     if ((This->pow2Width > gl_info->limits.texture_size || This->pow2Height > gl_info->limits.texture_size)
3996             && !(This->resource.usage & (WINED3DUSAGE_RENDERTARGET | WINED3DUSAGE_DEPTHSTENCIL)))
3997     {
3998         /* one of three options
3999         1: Do the same as we do with nonpow 2 and scale the texture, (any texture ops would require the texture to be scaled which is potentially slow)
4000         2: Set the texture to the maximum size (bad idea)
4001         3:    WARN and return WINED3DERR_NOTAVAILABLE;
4002         4: Create the surface, but allow it to be used only for DirectDraw Blts. Some apps(e.g. Swat 3) create textures with a Height of 16 and a Width > 3000 and blt 16x16 letter areas from them to the render target.
4003         */
4004         if(This->resource.pool == WINED3DPOOL_DEFAULT || This->resource.pool == WINED3DPOOL_MANAGED)
4005         {
4006             WARN("(%p) Unable to allocate a surface which exceeds the maximum OpenGL texture size\n", This);
4007             return WINED3DERR_NOTAVAILABLE;
4008         }
4009
4010         /* We should never use this surface in combination with OpenGL! */
4011         TRACE("(%p) Creating an oversized surface: %ux%u\n", This, This->pow2Width, This->pow2Height);
4012     }
4013     else
4014     {
4015         /* Don't use ARB_TEXTURE_RECTANGLE in case the surface format is P8 and EXT_PALETTED_TEXTURE
4016            is used in combination with texture uploads (RTL_READTEX/RTL_TEXTEX). The reason is that EXT_PALETTED_TEXTURE
4017            doesn't work in combination with ARB_TEXTURE_RECTANGLE.
4018         */
4019         if (This->Flags & SFLAG_NONPOW2 && gl_info->supported[ARB_TEXTURE_RECTANGLE]
4020                 && !(This->resource.format_desc->format == WINED3DFMT_P8_UINT
4021                 && gl_info->supported[EXT_PALETTED_TEXTURE]
4022                 && wined3d_settings.rendertargetlock_mode == RTL_READTEX))
4023         {
4024             This->texture_target = GL_TEXTURE_RECTANGLE_ARB;
4025             This->pow2Width  = This->currentDesc.Width;
4026             This->pow2Height = This->currentDesc.Height;
4027             This->Flags &= ~(SFLAG_NONPOW2 | SFLAG_NORMCOORD);
4028         }
4029     }
4030
4031     if(This->resource.usage & WINED3DUSAGE_RENDERTARGET) {
4032         switch(wined3d_settings.offscreen_rendering_mode) {
4033             case ORM_FBO:        This->get_drawable_size = get_drawable_size_fbo;        break;
4034             case ORM_BACKBUFFER: This->get_drawable_size = get_drawable_size_backbuffer; break;
4035         }
4036     }
4037
4038     This->Flags |= SFLAG_INSYSMEM;
4039
4040     return WINED3D_OK;
4041 }
4042
4043 /* GL locking is done by the caller */
4044 static void surface_depth_blt(IWineD3DSurfaceImpl *This, const struct wined3d_gl_info *gl_info,
4045         GLuint texture, GLsizei w, GLsizei h, GLenum target)
4046 {
4047     IWineD3DDeviceImpl *device = This->resource.device;
4048     GLint compare_mode = GL_NONE;
4049     struct blt_info info;
4050     GLint old_binding = 0;
4051
4052     glPushAttrib(GL_ENABLE_BIT | GL_DEPTH_BUFFER_BIT | GL_COLOR_BUFFER_BIT | GL_VIEWPORT_BIT);
4053
4054     glDisable(GL_CULL_FACE);
4055     glDisable(GL_BLEND);
4056     glDisable(GL_ALPHA_TEST);
4057     glDisable(GL_SCISSOR_TEST);
4058     glDisable(GL_STENCIL_TEST);
4059     glEnable(GL_DEPTH_TEST);
4060     glDepthFunc(GL_ALWAYS);
4061     glDepthMask(GL_TRUE);
4062     glColorMask(GL_FALSE, GL_FALSE, GL_FALSE, GL_FALSE);
4063     glViewport(0, 0, w, h);
4064
4065     surface_get_blt_info(target, NULL, w, h, &info);
4066     GL_EXTCALL(glActiveTextureARB(GL_TEXTURE0_ARB));
4067     glGetIntegerv(info.binding, &old_binding);
4068     glBindTexture(info.bind_target, texture);
4069     if (gl_info->supported[ARB_SHADOW])
4070     {
4071         glGetTexParameteriv(info.bind_target, GL_TEXTURE_COMPARE_MODE_ARB, &compare_mode);
4072         if (compare_mode != GL_NONE) glTexParameteri(info.bind_target, GL_TEXTURE_COMPARE_MODE_ARB, GL_NONE);
4073     }
4074
4075     device->shader_backend->shader_select_depth_blt((IWineD3DDevice *)device,
4076             info.tex_type, &This->ds_current_size);
4077
4078     glBegin(GL_TRIANGLE_STRIP);
4079     glTexCoord3fv(info.coords[0]);
4080     glVertex2f(-1.0f, -1.0f);
4081     glTexCoord3fv(info.coords[1]);
4082     glVertex2f(1.0f, -1.0f);
4083     glTexCoord3fv(info.coords[2]);
4084     glVertex2f(-1.0f, 1.0f);
4085     glTexCoord3fv(info.coords[3]);
4086     glVertex2f(1.0f, 1.0f);
4087     glEnd();
4088
4089     if (compare_mode != GL_NONE) glTexParameteri(info.bind_target, GL_TEXTURE_COMPARE_MODE_ARB, compare_mode);
4090     glBindTexture(info.bind_target, old_binding);
4091
4092     glPopAttrib();
4093
4094     device->shader_backend->shader_deselect_depth_blt((IWineD3DDevice *)device);
4095 }
4096
4097 void surface_modify_ds_location(IWineD3DSurfaceImpl *surface,
4098         DWORD location, UINT w, UINT h)
4099 {
4100     TRACE("surface %p, new location %#x, w %u, h %u.\n", surface, location, w, h);
4101
4102     if (location & ~SFLAG_DS_LOCATIONS)
4103         FIXME("Invalid location (%#x) specified.\n", location);
4104
4105     surface->ds_current_size.cx = w;
4106     surface->ds_current_size.cy = h;
4107     surface->Flags &= ~SFLAG_DS_LOCATIONS;
4108     surface->Flags |= location;
4109 }
4110
4111 /* Context activation is done by the caller. */
4112 void surface_load_ds_location(IWineD3DSurfaceImpl *surface, struct wined3d_context *context, DWORD location)
4113 {
4114     IWineD3DDeviceImpl *device = surface->resource.device;
4115     const struct wined3d_gl_info *gl_info = context->gl_info;
4116
4117     TRACE("surface %p, new location %#x.\n", surface, location);
4118
4119     /* TODO: Make this work for modes other than FBO */
4120     if (wined3d_settings.offscreen_rendering_mode != ORM_FBO) return;
4121
4122     if (!(surface->Flags & location))
4123     {
4124         surface->ds_current_size.cx = 0;
4125         surface->ds_current_size.cy = 0;
4126     }
4127
4128     if (surface->ds_current_size.cx == surface->currentDesc.Width
4129             && surface->ds_current_size.cy == surface->currentDesc.Height)
4130     {
4131         TRACE("Location (%#x) is already up to date.\n", location);
4132         return;
4133     }
4134
4135     if (surface->current_renderbuffer)
4136     {
4137         FIXME("Not supported with fixed up depth stencil.\n");
4138         return;
4139     }
4140
4141     if (!(surface->Flags & SFLAG_LOCATIONS))
4142     {
4143         FIXME("No up to date depth stencil location.\n");
4144         surface->Flags |= location;
4145         return;
4146     }
4147
4148     if (location == SFLAG_DS_OFFSCREEN)
4149     {
4150         GLint old_binding = 0;
4151         GLenum bind_target;
4152
4153         TRACE("Copying onscreen depth buffer to depth texture.\n");
4154
4155         ENTER_GL();
4156
4157         if (!device->depth_blt_texture)
4158         {
4159             glGenTextures(1, &device->depth_blt_texture);
4160         }
4161
4162         /* Note that we use depth_blt here as well, rather than glCopyTexImage2D
4163          * directly on the FBO texture. That's because we need to flip. */
4164         context_bind_fbo(context, GL_FRAMEBUFFER, NULL);
4165         if (surface->texture_target == GL_TEXTURE_RECTANGLE_ARB)
4166         {
4167             glGetIntegerv(GL_TEXTURE_BINDING_RECTANGLE_ARB, &old_binding);
4168             bind_target = GL_TEXTURE_RECTANGLE_ARB;
4169         }
4170         else
4171         {
4172             glGetIntegerv(GL_TEXTURE_BINDING_2D, &old_binding);
4173             bind_target = GL_TEXTURE_2D;
4174         }
4175         glBindTexture(bind_target, device->depth_blt_texture);
4176         glCopyTexImage2D(bind_target, surface->texture_level, surface->resource.format_desc->glInternal,
4177                 0, 0, surface->currentDesc.Width, surface->currentDesc.Height, 0);
4178         glTexParameteri(bind_target, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
4179         glTexParameteri(bind_target, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
4180         glTexParameteri(bind_target, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
4181         glTexParameteri(bind_target, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
4182         glTexParameteri(bind_target, GL_TEXTURE_WRAP_R, GL_CLAMP_TO_EDGE);
4183         glTexParameteri(bind_target, GL_DEPTH_TEXTURE_MODE_ARB, GL_LUMINANCE);
4184         glBindTexture(bind_target, old_binding);
4185
4186         /* Setup the destination */
4187         if (!device->depth_blt_rb)
4188         {
4189             gl_info->fbo_ops.glGenRenderbuffers(1, &device->depth_blt_rb);
4190             checkGLcall("glGenRenderbuffersEXT");
4191         }
4192         if (device->depth_blt_rb_w != surface->currentDesc.Width
4193                 || device->depth_blt_rb_h != surface->currentDesc.Height)
4194         {
4195             gl_info->fbo_ops.glBindRenderbuffer(GL_RENDERBUFFER, device->depth_blt_rb);
4196             checkGLcall("glBindRenderbufferEXT");
4197             gl_info->fbo_ops.glRenderbufferStorage(GL_RENDERBUFFER, GL_RGBA8,
4198                     surface->currentDesc.Width, surface->currentDesc.Height);
4199             checkGLcall("glRenderbufferStorageEXT");
4200             device->depth_blt_rb_w = surface->currentDesc.Width;
4201             device->depth_blt_rb_h = surface->currentDesc.Height;
4202         }
4203
4204         context_bind_fbo(context, GL_FRAMEBUFFER, &context->dst_fbo);
4205         gl_info->fbo_ops.glFramebufferRenderbuffer(GL_FRAMEBUFFER,
4206                 GL_COLOR_ATTACHMENT0, GL_RENDERBUFFER, device->depth_blt_rb);
4207         checkGLcall("glFramebufferRenderbufferEXT");
4208         context_attach_depth_stencil_fbo(context, GL_FRAMEBUFFER, surface, FALSE);
4209
4210         /* Do the actual blit */
4211         surface_depth_blt(surface, gl_info, device->depth_blt_texture,
4212                 surface->currentDesc.Width, surface->currentDesc.Height, bind_target);
4213         checkGLcall("depth_blt");
4214
4215         if (context->current_fbo) context_bind_fbo(context, GL_FRAMEBUFFER, &context->current_fbo->id);
4216         else context_bind_fbo(context, GL_FRAMEBUFFER, NULL);
4217
4218         LEAVE_GL();
4219
4220         if (wined3d_settings.strict_draw_ordering) wglFlush(); /* Flush to ensure ordering across contexts. */
4221     }
4222     else if (location == SFLAG_DS_ONSCREEN)
4223     {
4224         TRACE("Copying depth texture to onscreen depth buffer.\n");
4225
4226         ENTER_GL();
4227
4228         context_bind_fbo(context, GL_FRAMEBUFFER, NULL);
4229         surface_depth_blt(surface, gl_info, surface->texture_name,
4230                 surface->currentDesc.Width, surface->currentDesc.Height, surface->texture_target);
4231         checkGLcall("depth_blt");
4232
4233         if (context->current_fbo) context_bind_fbo(context, GL_FRAMEBUFFER, &context->current_fbo->id);
4234
4235         LEAVE_GL();
4236
4237         if (wined3d_settings.strict_draw_ordering) wglFlush(); /* Flush to ensure ordering across contexts. */
4238     }
4239     else
4240     {
4241         ERR("Invalid location (%#x) specified.\n", location);
4242     }
4243
4244     surface->Flags |= location;
4245     surface->ds_current_size.cx = surface->currentDesc.Width;
4246     surface->ds_current_size.cy = surface->currentDesc.Height;
4247 }
4248
4249 void surface_modify_location(IWineD3DSurfaceImpl *surface, DWORD flag, BOOL persistent)
4250 {
4251     IWineD3DBaseTexture *texture;
4252     IWineD3DSurfaceImpl *overlay;
4253
4254     TRACE("surface %p, location %s, persitent %#x.\n",
4255             surface, debug_surflocation(flag), persistent);
4256
4257     if (wined3d_settings.offscreen_rendering_mode == ORM_FBO)
4258     {
4259         if (surface_is_offscreen(surface))
4260         {
4261             /* With ORM_FBO, SFLAG_INTEXTURE and SFLAG_INDRAWABLE are the same for offscreen targets. */
4262             if (flag & (SFLAG_INTEXTURE | SFLAG_INDRAWABLE)) flag |= (SFLAG_INTEXTURE | SFLAG_INDRAWABLE);
4263         }
4264         else
4265         {
4266             TRACE("Surface %p is an onscreen surface.\n", surface);
4267         }
4268     }
4269
4270     if (persistent)
4271     {
4272         if (((surface->Flags & SFLAG_INTEXTURE) && !(flag & SFLAG_INTEXTURE))
4273                 || ((surface->Flags & SFLAG_INSRGBTEX) && !(flag & SFLAG_INSRGBTEX)))
4274         {
4275             if (SUCCEEDED(IWineD3DSurface_GetContainer((IWineD3DSurface *)surface,
4276                     &IID_IWineD3DBaseTexture, (void **)&texture)))
4277             {
4278                 TRACE("Passing to container.\n");
4279                 IWineD3DBaseTexture_SetDirty(texture, TRUE);
4280                 IWineD3DBaseTexture_Release(texture);
4281             }
4282         }
4283         surface->Flags &= ~SFLAG_LOCATIONS;
4284         surface->Flags |= flag;
4285
4286         /* Redraw emulated overlays, if any */
4287         if (flag & SFLAG_INDRAWABLE && !list_empty(&surface->overlays))
4288         {
4289             LIST_FOR_EACH_ENTRY(overlay, &surface->overlays, IWineD3DSurfaceImpl, overlay_entry)
4290             {
4291                 IWineD3DSurface_DrawOverlay((IWineD3DSurface *)overlay);
4292             }
4293         }
4294     }
4295     else
4296     {
4297         if ((surface->Flags & (SFLAG_INTEXTURE | SFLAG_INSRGBTEX)) && (flag & (SFLAG_INTEXTURE | SFLAG_INSRGBTEX)))
4298         {
4299             if (SUCCEEDED(IWineD3DSurface_GetContainer((IWineD3DSurface *)surface,
4300                     &IID_IWineD3DBaseTexture, (void **)&texture)))
4301             {
4302                 TRACE("Passing to container\n");
4303                 IWineD3DBaseTexture_SetDirty(texture, TRUE);
4304                 IWineD3DBaseTexture_Release(texture);
4305             }
4306         }
4307         surface->Flags &= ~flag;
4308     }
4309
4310     if (!(surface->Flags & SFLAG_LOCATIONS))
4311     {
4312         ERR("Surface %p does not have any up to date location.\n", surface);
4313     }
4314 }
4315
4316 static inline void surface_blt_to_drawable(IWineD3DSurfaceImpl *This, const RECT *rect_in)
4317 {
4318     IWineD3DDeviceImpl *device = This->resource.device;
4319     IWineD3DSwapChainImpl *swapchain;
4320     struct wined3d_context *context;
4321     RECT src_rect, dst_rect;
4322
4323     surface_get_rect(This, rect_in, &src_rect);
4324
4325     context = context_acquire(device, This);
4326     context_apply_blit_state(context, device);
4327     if (context->render_offscreen)
4328     {
4329         dst_rect.left = src_rect.left;
4330         dst_rect.right = src_rect.right;
4331         dst_rect.top = src_rect.bottom;
4332         dst_rect.bottom = src_rect.top;
4333     }
4334     else
4335     {
4336         dst_rect = src_rect;
4337     }
4338
4339     if ((This->Flags & SFLAG_SWAPCHAIN) && This == ((IWineD3DSwapChainImpl *)This->container)->front_buffer)
4340         surface_translate_frontbuffer_coords(This, context->win_handle, &dst_rect);
4341
4342     device->blitter->set_shader((IWineD3DDevice *) device, This);
4343
4344     ENTER_GL();
4345     draw_textured_quad(This, &src_rect, &dst_rect, WINED3DTEXF_POINT);
4346     LEAVE_GL();
4347
4348     device->blitter->unset_shader((IWineD3DDevice *) device);
4349
4350     swapchain = (This->Flags & SFLAG_SWAPCHAIN) ? (IWineD3DSwapChainImpl *)This->container : NULL;
4351     if (wined3d_settings.strict_draw_ordering || (swapchain
4352             && (This == swapchain->front_buffer || swapchain->num_contexts > 1)))
4353         wglFlush(); /* Flush to ensure ordering across contexts. */
4354
4355     context_release(context);
4356 }
4357
4358 HRESULT surface_load_location(IWineD3DSurfaceImpl *surface, DWORD flag, const RECT *rect)
4359 {
4360     IWineD3DDeviceImpl *device = surface->resource.device;
4361     const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
4362     BOOL drawable_read_ok = surface_is_offscreen(surface);
4363     struct wined3d_format_desc desc;
4364     CONVERT_TYPES convert;
4365     int width, pitch, outpitch;
4366     BYTE *mem;
4367     BOOL in_fbo = FALSE;
4368
4369     TRACE("surface %p, location %s, rect %s.\n", surface, debug_surflocation(flag), wine_dbgstr_rect(rect));
4370
4371     if (surface->resource.usage & WINED3DUSAGE_DEPTHSTENCIL)
4372     {
4373         if (flag == SFLAG_INTEXTURE)
4374         {
4375             struct wined3d_context *context = context_acquire(device, NULL);
4376             surface_load_ds_location(surface, context, SFLAG_DS_OFFSCREEN);
4377             context_release(context);
4378             return WINED3D_OK;
4379         }
4380         else
4381         {
4382             FIXME("Unimplemented location %#x for depth/stencil buffers.\n", flag);
4383             return WINED3DERR_INVALIDCALL;
4384         }
4385     }
4386
4387     if (wined3d_settings.offscreen_rendering_mode == ORM_FBO)
4388     {
4389         if (surface_is_offscreen(surface))
4390         {
4391             /* With ORM_FBO, SFLAG_INTEXTURE and SFLAG_INDRAWABLE are the same for offscreen targets.
4392              * Prefer SFLAG_INTEXTURE. */
4393             if (flag == SFLAG_INDRAWABLE) flag = SFLAG_INTEXTURE;
4394             drawable_read_ok = FALSE;
4395             in_fbo = TRUE;
4396         }
4397         else
4398         {
4399             TRACE("Surface %p is an onscreen surface.\n", surface);
4400         }
4401     }
4402
4403     if (surface->Flags & flag)
4404     {
4405         TRACE("Location already up to date\n");
4406         return WINED3D_OK;
4407     }
4408
4409     if (!(surface->Flags & SFLAG_LOCATIONS))
4410     {
4411         ERR("Surface %p does not have any up to date location.\n", surface);
4412         surface->Flags |= SFLAG_LOST;
4413         return WINED3DERR_DEVICELOST;
4414     }
4415
4416     if (flag == SFLAG_INSYSMEM)
4417     {
4418         surface_prepare_system_memory(surface);
4419
4420         /* Download the surface to system memory */
4421         if (surface->Flags & (SFLAG_INTEXTURE | SFLAG_INSRGBTEX))
4422         {
4423             struct wined3d_context *context = NULL;
4424
4425             if (!device->isInDraw) context = context_acquire(device, NULL);
4426
4427             surface_bind_and_dirtify(surface, !(surface->Flags & SFLAG_INTEXTURE));
4428             surface_download_data(surface, gl_info);
4429
4430             if (context) context_release(context);
4431         }
4432         else
4433         {
4434             /* Note: It might be faster to download into a texture first. */
4435             read_from_framebuffer(surface, rect, surface->resource.allocatedMemory,
4436                     IWineD3DSurface_GetPitch((IWineD3DSurface *)surface));
4437         }
4438     }
4439     else if (flag == SFLAG_INDRAWABLE)
4440     {
4441         if (surface->Flags & SFLAG_INTEXTURE)
4442         {
4443             surface_blt_to_drawable(surface, rect);
4444         }
4445         else
4446         {
4447             int byte_count;
4448             if ((surface->Flags & SFLAG_LOCATIONS) == SFLAG_INSRGBTEX)
4449             {
4450                 /* This needs a shader to convert the srgb data sampled from the GL texture into RGB
4451                  * values, otherwise we get incorrect values in the target. For now go the slow way
4452                  * via a system memory copy
4453                  */
4454                 surface_load_location(surface, SFLAG_INSYSMEM, rect);
4455             }
4456
4457             d3dfmt_get_conv(surface, FALSE /* We need color keying */,
4458                     FALSE /* We won't use textures */, &desc, &convert);
4459
4460             /* The width is in 'length' not in bytes */
4461             width = surface->currentDesc.Width;
4462             pitch = IWineD3DSurface_GetPitch((IWineD3DSurface *)surface);
4463
4464             /* Don't use PBOs for converted surfaces. During PBO conversion we look at SFLAG_CONVERTED
4465              * but it isn't set (yet) in all cases it is getting called. */
4466             if ((convert != NO_CONVERSION) && (surface->Flags & SFLAG_PBO))
4467             {
4468                 struct wined3d_context *context = NULL;
4469
4470                 TRACE("Removing the pbo attached to surface %p.\n", surface);
4471
4472                 if (!device->isInDraw) context = context_acquire(device, NULL);
4473                 surface_remove_pbo(surface, gl_info);
4474                 if (context) context_release(context);
4475             }
4476
4477             if ((convert != NO_CONVERSION) && surface->resource.allocatedMemory)
4478             {
4479                 int height = surface->currentDesc.Height;
4480                 byte_count = desc.conv_byte_count;
4481
4482                 /* Stick to the alignment for the converted surface too, makes it easier to load the surface */
4483                 outpitch = width * byte_count;
4484                 outpitch = (outpitch + device->surface_alignment - 1) & ~(device->surface_alignment - 1);
4485
4486                 mem = HeapAlloc(GetProcessHeap(), 0, outpitch * height);
4487                 if(!mem) {
4488                     ERR("Out of memory %d, %d!\n", outpitch, height);
4489                     return WINED3DERR_OUTOFVIDEOMEMORY;
4490                 }
4491                 d3dfmt_convert_surface(surface->resource.allocatedMemory, mem, pitch,
4492                         width, height, outpitch, convert, surface);
4493
4494                 surface->Flags |= SFLAG_CONVERTED;
4495             }
4496             else
4497             {
4498                 surface->Flags &= ~SFLAG_CONVERTED;
4499                 mem = surface->resource.allocatedMemory;
4500                 byte_count = desc.byte_count;
4501             }
4502
4503             flush_to_framebuffer_drawpixels(surface, desc.glFormat, desc.glType, byte_count, mem);
4504
4505             /* Don't delete PBO memory */
4506             if ((mem != surface->resource.allocatedMemory) && !(surface->Flags & SFLAG_PBO))
4507                 HeapFree(GetProcessHeap(), 0, mem);
4508         }
4509     }
4510     else /* if(flag & (SFLAG_INTEXTURE | SFLAG_INSRGBTEX)) */
4511     {
4512         const DWORD attach_flags = WINED3DFMT_FLAG_FBO_ATTACHABLE | WINED3DFMT_FLAG_FBO_ATTACHABLE_SRGB;
4513
4514         if (drawable_read_ok && (surface->Flags & SFLAG_INDRAWABLE))
4515         {
4516             read_from_framebuffer_texture(surface, flag == SFLAG_INSRGBTEX);
4517         }
4518         else if (surface->Flags & (SFLAG_INSRGBTEX | SFLAG_INTEXTURE)
4519                 && (surface->resource.format_desc->Flags & attach_flags) == attach_flags
4520                 && fbo_blit_supported(gl_info, BLIT_OP_BLIT,
4521                         NULL, surface->resource.usage, surface->resource.pool, surface->resource.format_desc,
4522                         NULL, surface->resource.usage, surface->resource.pool, surface->resource.format_desc))
4523         {
4524             DWORD src_location = flag == SFLAG_INSRGBTEX ? SFLAG_INTEXTURE : SFLAG_INSRGBTEX;
4525             RECT rect = {0, 0, surface->currentDesc.Width, surface->currentDesc.Height};
4526
4527             surface_blt_fbo(surface->resource.device, WINED3DTEXF_POINT,
4528                     surface, src_location, &rect, surface, flag, &rect);
4529         }
4530         else
4531         {
4532             /* Upload from system memory */
4533             BOOL srgb = flag == SFLAG_INSRGBTEX;
4534             struct wined3d_context *context = NULL;
4535
4536             d3dfmt_get_conv(surface, TRUE /* We need color keying */,
4537                     TRUE /* We will use textures */, &desc, &convert);
4538
4539             if (srgb)
4540             {
4541                 if ((surface->Flags & (SFLAG_INTEXTURE | SFLAG_INSYSMEM)) == SFLAG_INTEXTURE)
4542                 {
4543                     /* Performance warning... */
4544                     FIXME("Downloading RGB surface %p to reload it as sRGB.\n", surface);
4545                     surface_load_location(surface, SFLAG_INSYSMEM, rect);
4546                 }
4547             }
4548             else
4549             {
4550                 if ((surface->Flags & (SFLAG_INSRGBTEX | SFLAG_INSYSMEM)) == SFLAG_INSRGBTEX)
4551                 {
4552                     /* Performance warning... */
4553                     FIXME("Downloading sRGB surface %p to reload it as RGB.\n", surface);
4554                     surface_load_location(surface, SFLAG_INSYSMEM, rect);
4555                 }
4556             }
4557             if (!(surface->Flags & SFLAG_INSYSMEM))
4558             {
4559                 WARN("Trying to load a texture from sysmem, but SFLAG_INSYSMEM is not set.\n");
4560                 /* Lets hope we get it from somewhere... */
4561                 surface_load_location(surface, SFLAG_INSYSMEM, rect);
4562             }
4563
4564             if (!device->isInDraw) context = context_acquire(device, NULL);
4565
4566             surface_prepare_texture(surface, gl_info, srgb);
4567             surface_bind_and_dirtify(surface, srgb);
4568
4569             if (surface->CKeyFlags & WINEDDSD_CKSRCBLT)
4570             {
4571                 surface->Flags |= SFLAG_GLCKEY;
4572                 surface->glCKey = surface->SrcBltCKey;
4573             }
4574             else surface->Flags &= ~SFLAG_GLCKEY;
4575
4576             /* The width is in 'length' not in bytes */
4577             width = surface->currentDesc.Width;
4578             pitch = IWineD3DSurface_GetPitch((IWineD3DSurface *)surface);
4579
4580             /* Don't use PBOs for converted surfaces. During PBO conversion we look at SFLAG_CONVERTED
4581              * but it isn't set (yet) in all cases it is getting called. */
4582             if ((convert != NO_CONVERSION || desc.convert) && (surface->Flags & SFLAG_PBO))
4583             {
4584                 TRACE("Removing the pbo attached to surface %p.\n", surface);
4585                 surface_remove_pbo(surface, gl_info);
4586             }
4587
4588             if (desc.convert)
4589             {
4590                 /* This code is entered for texture formats which need a fixup. */
4591                 int height = surface->currentDesc.Height;
4592
4593                 /* Stick to the alignment for the converted surface too, makes it easier to load the surface */
4594                 outpitch = width * desc.conv_byte_count;
4595                 outpitch = (outpitch + device->surface_alignment - 1) & ~(device->surface_alignment - 1);
4596
4597                 mem = HeapAlloc(GetProcessHeap(), 0, outpitch * height);
4598                 if(!mem) {
4599                     ERR("Out of memory %d, %d!\n", outpitch, height);
4600                     if (context) context_release(context);
4601                     return WINED3DERR_OUTOFVIDEOMEMORY;
4602                 }
4603                 desc.convert(surface->resource.allocatedMemory, mem, pitch, width, height);
4604             }
4605             else if (convert != NO_CONVERSION && surface->resource.allocatedMemory)
4606             {
4607                 /* This code is only entered for color keying fixups */
4608                 int height = surface->currentDesc.Height;
4609
4610                 /* Stick to the alignment for the converted surface too, makes it easier to load the surface */
4611                 outpitch = width * desc.conv_byte_count;
4612                 outpitch = (outpitch + device->surface_alignment - 1) & ~(device->surface_alignment - 1);
4613
4614                 mem = HeapAlloc(GetProcessHeap(), 0, outpitch * height);
4615                 if(!mem) {
4616                     ERR("Out of memory %d, %d!\n", outpitch, height);
4617                     if (context) context_release(context);
4618                     return WINED3DERR_OUTOFVIDEOMEMORY;
4619                 }
4620                 d3dfmt_convert_surface(surface->resource.allocatedMemory, mem, pitch,
4621                         width, height, outpitch, convert, surface);
4622             }
4623             else
4624             {
4625                 mem = surface->resource.allocatedMemory;
4626             }
4627
4628             /* Make sure the correct pitch is used */
4629             ENTER_GL();
4630             glPixelStorei(GL_UNPACK_ROW_LENGTH, width);
4631             LEAVE_GL();
4632
4633             if (mem || (surface->Flags & SFLAG_PBO))
4634                 surface_upload_data(surface, gl_info, &desc, srgb, mem);
4635
4636             /* Restore the default pitch */
4637             ENTER_GL();
4638             glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
4639             LEAVE_GL();
4640
4641             if (context) context_release(context);
4642
4643             /* Don't delete PBO memory */
4644             if ((mem != surface->resource.allocatedMemory) && !(surface->Flags & SFLAG_PBO))
4645                 HeapFree(GetProcessHeap(), 0, mem);
4646         }
4647     }
4648
4649     if (!rect) surface->Flags |= flag;
4650
4651     if (in_fbo && (surface->Flags & (SFLAG_INTEXTURE | SFLAG_INDRAWABLE)))
4652     {
4653         /* With ORM_FBO, SFLAG_INTEXTURE and SFLAG_INDRAWABLE are the same for offscreen targets. */
4654         surface->Flags |= (SFLAG_INTEXTURE | SFLAG_INDRAWABLE);
4655     }
4656
4657     return WINED3D_OK;
4658 }
4659
4660 static HRESULT WINAPI IWineD3DSurfaceImpl_SetContainer(IWineD3DSurface *iface, IWineD3DBase *container)
4661 {
4662     IWineD3DSurfaceImpl *This = (IWineD3DSurfaceImpl *) iface;
4663     IWineD3DSwapChain *swapchain = NULL;
4664
4665     /* Update the drawable size method */
4666     if(container) {
4667         IWineD3DBase_QueryInterface(container, &IID_IWineD3DSwapChain, (void **) &swapchain);
4668     }
4669     if(swapchain) {
4670         This->get_drawable_size = get_drawable_size_swapchain;
4671         IWineD3DSwapChain_Release(swapchain);
4672     } else if(This->resource.usage & WINED3DUSAGE_RENDERTARGET) {
4673         switch(wined3d_settings.offscreen_rendering_mode) {
4674             case ORM_FBO:        This->get_drawable_size = get_drawable_size_fbo;        break;
4675             case ORM_BACKBUFFER: This->get_drawable_size = get_drawable_size_backbuffer; break;
4676         }
4677     }
4678
4679     return IWineD3DBaseSurfaceImpl_SetContainer(iface, container);
4680 }
4681
4682 static WINED3DSURFTYPE WINAPI IWineD3DSurfaceImpl_GetImplType(IWineD3DSurface *iface) {
4683     return SURFACE_OPENGL;
4684 }
4685
4686 static HRESULT WINAPI IWineD3DSurfaceImpl_DrawOverlay(IWineD3DSurface *iface) {
4687     IWineD3DSurfaceImpl *This = (IWineD3DSurfaceImpl *) iface;
4688     HRESULT hr;
4689
4690     /* If there's no destination surface there is nothing to do */
4691     if(!This->overlay_dest) return WINED3D_OK;
4692
4693     /* Blt calls ModifyLocation on the dest surface, which in turn calls DrawOverlay to
4694      * update the overlay. Prevent an endless recursion
4695      */
4696     if(This->overlay_dest->Flags & SFLAG_INOVERLAYDRAW) {
4697         return WINED3D_OK;
4698     }
4699     This->overlay_dest->Flags |= SFLAG_INOVERLAYDRAW;
4700     hr = IWineD3DSurfaceImpl_Blt((IWineD3DSurface *) This->overlay_dest, &This->overlay_destrect,
4701                                  iface, &This->overlay_srcrect, WINEDDBLT_WAIT,
4702                                  NULL, WINED3DTEXF_LINEAR);
4703     This->overlay_dest->Flags &= ~SFLAG_INOVERLAYDRAW;
4704
4705     return hr;
4706 }
4707
4708 BOOL surface_is_offscreen(IWineD3DSurfaceImpl *surface)
4709 {
4710     IWineD3DSwapChainImpl *swapchain = (IWineD3DSwapChainImpl *)surface->container;
4711
4712     /* Not on a swapchain - must be offscreen */
4713     if (!(surface->Flags & SFLAG_SWAPCHAIN)) return TRUE;
4714
4715     /* The front buffer is always onscreen */
4716     if (surface == swapchain->front_buffer) return FALSE;
4717
4718     /* If the swapchain is rendered to an FBO, the backbuffer is
4719      * offscreen, otherwise onscreen */
4720     return swapchain->render_to_fbo;
4721 }
4722
4723 const IWineD3DSurfaceVtbl IWineD3DSurface_Vtbl =
4724 {
4725     /* IUnknown */
4726     IWineD3DBaseSurfaceImpl_QueryInterface,
4727     IWineD3DBaseSurfaceImpl_AddRef,
4728     IWineD3DSurfaceImpl_Release,
4729     /* IWineD3DResource */
4730     IWineD3DBaseSurfaceImpl_GetParent,
4731     IWineD3DBaseSurfaceImpl_SetPrivateData,
4732     IWineD3DBaseSurfaceImpl_GetPrivateData,
4733     IWineD3DBaseSurfaceImpl_FreePrivateData,
4734     IWineD3DBaseSurfaceImpl_SetPriority,
4735     IWineD3DBaseSurfaceImpl_GetPriority,
4736     IWineD3DSurfaceImpl_PreLoad,
4737     IWineD3DSurfaceImpl_UnLoad,
4738     IWineD3DBaseSurfaceImpl_GetType,
4739     /* IWineD3DSurface */
4740     IWineD3DBaseSurfaceImpl_GetContainer,
4741     IWineD3DBaseSurfaceImpl_GetDesc,
4742     IWineD3DSurfaceImpl_LockRect,
4743     IWineD3DSurfaceImpl_UnlockRect,
4744     IWineD3DSurfaceImpl_GetDC,
4745     IWineD3DSurfaceImpl_ReleaseDC,
4746     IWineD3DSurfaceImpl_Flip,
4747     IWineD3DSurfaceImpl_Blt,
4748     IWineD3DBaseSurfaceImpl_GetBltStatus,
4749     IWineD3DBaseSurfaceImpl_GetFlipStatus,
4750     IWineD3DBaseSurfaceImpl_IsLost,
4751     IWineD3DBaseSurfaceImpl_Restore,
4752     IWineD3DSurfaceImpl_BltFast,
4753     IWineD3DBaseSurfaceImpl_GetPalette,
4754     IWineD3DBaseSurfaceImpl_SetPalette,
4755     IWineD3DSurfaceImpl_RealizePalette,
4756     IWineD3DBaseSurfaceImpl_SetColorKey,
4757     IWineD3DBaseSurfaceImpl_GetPitch,
4758     IWineD3DSurfaceImpl_SetMem,
4759     IWineD3DBaseSurfaceImpl_SetOverlayPosition,
4760     IWineD3DBaseSurfaceImpl_GetOverlayPosition,
4761     IWineD3DBaseSurfaceImpl_UpdateOverlayZOrder,
4762     IWineD3DBaseSurfaceImpl_UpdateOverlay,
4763     IWineD3DBaseSurfaceImpl_SetClipper,
4764     IWineD3DBaseSurfaceImpl_GetClipper,
4765     /* Internal use: */
4766     IWineD3DSurfaceImpl_LoadTexture,
4767     IWineD3DSurfaceImpl_BindTexture,
4768     IWineD3DSurfaceImpl_SetContainer,
4769     IWineD3DBaseSurfaceImpl_GetData,
4770     IWineD3DSurfaceImpl_SetFormat,
4771     IWineD3DSurfaceImpl_PrivateSetup,
4772     IWineD3DSurfaceImpl_GetImplType,
4773     IWineD3DSurfaceImpl_DrawOverlay
4774 };
4775
4776 static HRESULT ffp_blit_alloc(IWineD3DDevice *iface) { return WINED3D_OK; }
4777 /* Context activation is done by the caller. */
4778 static void ffp_blit_free(IWineD3DDevice *iface) { }
4779
4780 /* This function is used in case of 8bit paletted textures using GL_EXT_paletted_texture */
4781 /* Context activation is done by the caller. */
4782 static void ffp_blit_p8_upload_palette(IWineD3DSurfaceImpl *surface, const struct wined3d_gl_info *gl_info)
4783 {
4784     BYTE table[256][4];
4785     BOOL colorkey_active = (surface->CKeyFlags & WINEDDSD_CKSRCBLT) ? TRUE : FALSE;
4786
4787     d3dfmt_p8_init_palette(surface, table, colorkey_active);
4788
4789     TRACE("Using GL_EXT_PALETTED_TEXTURE for 8-bit paletted texture support\n");
4790     ENTER_GL();
4791     GL_EXTCALL(glColorTableEXT(surface->texture_target, GL_RGBA, 256, GL_RGBA, GL_UNSIGNED_BYTE, table));
4792     LEAVE_GL();
4793 }
4794
4795 /* Context activation is done by the caller. */
4796 static HRESULT ffp_blit_set(IWineD3DDevice *iface, IWineD3DSurfaceImpl *surface)
4797 {
4798     IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *) iface;
4799     const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
4800     enum complex_fixup fixup = get_complex_fixup(surface->resource.format_desc->color_fixup);
4801
4802     /* When EXT_PALETTED_TEXTURE is around, palette conversion is done by the GPU
4803      * else the surface is converted in software at upload time in LoadLocation.
4804      */
4805     if(fixup == COMPLEX_FIXUP_P8 && gl_info->supported[EXT_PALETTED_TEXTURE])
4806         ffp_blit_p8_upload_palette(surface, gl_info);
4807
4808     ENTER_GL();
4809     glEnable(surface->texture_target);
4810     checkGLcall("glEnable(surface->texture_target)");
4811     LEAVE_GL();
4812     return WINED3D_OK;
4813 }
4814
4815 /* Context activation is done by the caller. */
4816 static void ffp_blit_unset(IWineD3DDevice *iface)
4817 {
4818     IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *) iface;
4819     const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
4820
4821     ENTER_GL();
4822     glDisable(GL_TEXTURE_2D);
4823     checkGLcall("glDisable(GL_TEXTURE_2D)");
4824     if (gl_info->supported[ARB_TEXTURE_CUBE_MAP])
4825     {
4826         glDisable(GL_TEXTURE_CUBE_MAP_ARB);
4827         checkGLcall("glDisable(GL_TEXTURE_CUBE_MAP_ARB)");
4828     }
4829     if (gl_info->supported[ARB_TEXTURE_RECTANGLE])
4830     {
4831         glDisable(GL_TEXTURE_RECTANGLE_ARB);
4832         checkGLcall("glDisable(GL_TEXTURE_RECTANGLE_ARB)");
4833     }
4834     LEAVE_GL();
4835 }
4836
4837 static BOOL ffp_blit_supported(const struct wined3d_gl_info *gl_info, enum blit_operation blit_op,
4838                                const RECT *src_rect, DWORD src_usage, WINED3DPOOL src_pool,
4839                                const struct wined3d_format_desc *src_format_desc,
4840                                const RECT *dst_rect, DWORD dst_usage, WINED3DPOOL dst_pool,
4841                                const struct wined3d_format_desc *dst_format_desc)
4842 {
4843     enum complex_fixup src_fixup;
4844
4845     if (blit_op == BLIT_OP_COLOR_FILL)
4846     {
4847         if (!(dst_usage & WINED3DUSAGE_RENDERTARGET))
4848         {
4849             TRACE("Color fill not supported\n");
4850             return FALSE;
4851         }
4852
4853         return TRUE;
4854     }
4855
4856     src_fixup = get_complex_fixup(src_format_desc->color_fixup);
4857     if (TRACE_ON(d3d_surface) && TRACE_ON(d3d))
4858     {
4859         TRACE("Checking support for fixup:\n");
4860         dump_color_fixup_desc(src_format_desc->color_fixup);
4861     }
4862
4863     if (blit_op != BLIT_OP_BLIT)
4864     {
4865         TRACE("Unsupported blit_op=%d\n", blit_op);
4866         return FALSE;
4867      }
4868
4869     if (!is_identity_fixup(dst_format_desc->color_fixup))
4870     {
4871         TRACE("Destination fixups are not supported\n");
4872         return FALSE;
4873     }
4874
4875     if (src_fixup == COMPLEX_FIXUP_P8 && gl_info->supported[EXT_PALETTED_TEXTURE])
4876     {
4877         TRACE("P8 fixup supported\n");
4878         return TRUE;
4879     }
4880
4881     /* We only support identity conversions. */
4882     if (is_identity_fixup(src_format_desc->color_fixup))
4883     {
4884         TRACE("[OK]\n");
4885         return TRUE;
4886     }
4887
4888     TRACE("[FAILED]\n");
4889     return FALSE;
4890 }
4891
4892 static HRESULT ffp_blit_color_fill(IWineD3DDeviceImpl *device,
4893         IWineD3DSurfaceImpl *dst_surface, const RECT *dst_rect, DWORD fill_color)
4894 {
4895     return device_clear_render_targets(device, 1 /* rt_count */, &dst_surface, 1 /* rect_count */,
4896             (const WINED3DRECT *)dst_rect, WINED3DCLEAR_TARGET, fill_color, 0.0f /* depth */, 0 /* stencil */);
4897 }
4898
4899 const struct blit_shader ffp_blit =  {
4900     ffp_blit_alloc,
4901     ffp_blit_free,
4902     ffp_blit_set,
4903     ffp_blit_unset,
4904     ffp_blit_supported,
4905     ffp_blit_color_fill
4906 };
4907
4908 static HRESULT cpu_blit_alloc(IWineD3DDevice *iface)
4909 {
4910     return WINED3D_OK;
4911 }
4912
4913 /* Context activation is done by the caller. */
4914 static void cpu_blit_free(IWineD3DDevice *iface)
4915 {
4916 }
4917
4918 /* Context activation is done by the caller. */
4919 static HRESULT cpu_blit_set(IWineD3DDevice *iface, IWineD3DSurfaceImpl *surface)
4920 {
4921     return WINED3D_OK;
4922 }
4923
4924 /* Context activation is done by the caller. */
4925 static void cpu_blit_unset(IWineD3DDevice *iface)
4926 {
4927 }
4928
4929 static BOOL cpu_blit_supported(const struct wined3d_gl_info *gl_info, enum blit_operation blit_op,
4930                                const RECT *src_rect, DWORD src_usage, WINED3DPOOL src_pool,
4931                                const struct wined3d_format_desc *src_format_desc,
4932                                const RECT *dst_rect, DWORD dst_usage, WINED3DPOOL dst_pool,
4933                                const struct wined3d_format_desc *dst_format_desc)
4934 {
4935     if (blit_op == BLIT_OP_COLOR_FILL)
4936     {
4937         return TRUE;
4938     }
4939
4940     return FALSE;
4941 }
4942
4943 static HRESULT cpu_blit_color_fill(IWineD3DDeviceImpl *device, IWineD3DSurfaceImpl *dst_surface, const RECT *dst_rect, DWORD fill_color)
4944 {
4945     WINEDDBLTFX BltFx;
4946     memset(&BltFx, 0, sizeof(BltFx));
4947     BltFx.dwSize = sizeof(BltFx);
4948     BltFx.u5.dwFillColor = color_convert_argb_to_fmt(fill_color, dst_surface->resource.format_desc->format);
4949     return IWineD3DBaseSurfaceImpl_Blt((IWineD3DSurface*)dst_surface, dst_rect, NULL, NULL, WINEDDBLT_COLORFILL, &BltFx, WINED3DTEXF_POINT);
4950 }
4951
4952 const struct blit_shader cpu_blit =  {
4953     cpu_blit_alloc,
4954     cpu_blit_free,
4955     cpu_blit_set,
4956     cpu_blit_unset,
4957     cpu_blit_supported,
4958     cpu_blit_color_fill
4959 };
4960
4961 static BOOL fbo_blit_supported(const struct wined3d_gl_info *gl_info, enum blit_operation blit_op,
4962                                const RECT *src_rect, DWORD src_usage, WINED3DPOOL src_pool,
4963                                const struct wined3d_format_desc *src_format_desc,
4964                                const RECT *dst_rect, DWORD dst_usage, WINED3DPOOL dst_pool,
4965                                const struct wined3d_format_desc *dst_format_desc)
4966 {
4967     if ((wined3d_settings.offscreen_rendering_mode != ORM_FBO) || !gl_info->fbo_ops.glBlitFramebuffer)
4968         return FALSE;
4969
4970     /* We only support blitting. Things like color keying / color fill should
4971      * be handled by other blitters.
4972      */
4973     if (blit_op != BLIT_OP_BLIT)
4974         return FALSE;
4975
4976     /* Source and/or destination need to be on the GL side */
4977     if (src_pool == WINED3DPOOL_SYSTEMMEM || dst_pool == WINED3DPOOL_SYSTEMMEM)
4978         return FALSE;
4979
4980     if(!((src_format_desc->Flags & WINED3DFMT_FLAG_FBO_ATTACHABLE) || (src_usage & WINED3DUSAGE_RENDERTARGET))
4981         && ((dst_format_desc->Flags & WINED3DFMT_FLAG_FBO_ATTACHABLE) || (dst_usage & WINED3DUSAGE_RENDERTARGET)))
4982         return FALSE;
4983
4984     if (!is_identity_fixup(src_format_desc->color_fixup) ||
4985         !is_identity_fixup(dst_format_desc->color_fixup))
4986         return FALSE;
4987
4988     if (!(src_format_desc->format == dst_format_desc->format
4989         || (is_identity_fixup(src_format_desc->color_fixup)
4990         && is_identity_fixup(dst_format_desc->color_fixup))))
4991         return FALSE;
4992
4993     return TRUE;
4994 }