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     TRACE("(%p) : Cleaning up.\n", This);
  40
  41     if (This->texture_name || (This->Flags & SFLAG_PBO) || !list_empty(&This->renderbuffers))
  42     {
  43         const struct wined3d_gl_info *gl_info;
  44         renderbuffer_entry_t *entry, *entry2;
  45         struct wined3d_context *context;
  46
  47         context = context_acquire(This->resource.device, NULL);
  48         gl_info = context->gl_info;
  49
  50         ENTER_GL();
  51
  52         if (This->texture_name)
  53         {
  54             TRACE("Deleting texture %u.\n", This->texture_name);
  55             glDeleteTextures(1, &This->texture_name);
  56         }
  57
  58         if (This->Flags & SFLAG_PBO)
  59         {
  60             TRACE("Deleting PBO %u.\n", This->pbo);
  61             GL_EXTCALL(glDeleteBuffersARB(1, &This->pbo));
  62         }
  63
  64         LIST_FOR_EACH_ENTRY_SAFE(entry, entry2, &This->renderbuffers, renderbuffer_entry_t, entry)
  65         {
  66             TRACE("Deleting renderbuffer %u.\n", entry->id);
  67             gl_info->fbo_ops.glDeleteRenderbuffers(1, &entry->id);
  68             HeapFree(GetProcessHeap(), 0, entry);
  69         }
  70
  71         LEAVE_GL();
  72
  73         context_release(context);
  74     }
  75
  76     if (This->Flags & SFLAG_DIBSECTION)
  77     {
  78         /* Release the DC. */
  79         SelectObject(This->hDC, This->dib.holdbitmap);
  80         DeleteDC(This->hDC);
  81         /* Release the DIB section. */
  82         DeleteObject(This->dib.DIBsection);
  83         This->dib.bitmap_data = NULL;
  84         This->resource.allocatedMemory = NULL;
  85     }
  86
  87     if (This->Flags & SFLAG_USERPTR) IWineD3DSurface_SetMem((IWineD3DSurface *)This, NULL);
  88     if (This->overlay_dest) list_remove(&This->overlay_entry);
  89
  90     HeapFree(GetProcessHeap(), 0, This->palette9);
  91
  92     resource_cleanup((IWineD3DResource *)This);
  93 }
  94
  95 void surface_set_container(IWineD3DSurfaceImpl *surface, enum wined3d_container_type type, IWineD3DBase *container)
  96 {
  97     TRACE("surface %p, container %p.\n", surface, container);
  98
  99     if (type == WINED3D_CONTAINER_SWAPCHAIN)
 100     {
 101         surface->get_drawable_size = get_drawable_size_swapchain;
 102     }
 103     else
 104     {
 105         switch (wined3d_settings.offscreen_rendering_mode)
 106         {
 107             case ORM_FBO:
 108                 surface->get_drawable_size = get_drawable_size_fbo;
 109                 break;
 110
 111             case ORM_BACKBUFFER:
 112                 surface->get_drawable_size = get_drawable_size_backbuffer;
 113                 break;
 114
 115             default:
 116                 ERR("Unhandled offscreen rendering mode %#x.\n", wined3d_settings.offscreen_rendering_mode);
 117                 return;
 118         }
 119     }
 120
 121     surface->container.type = type;
 122     surface->container.u.base = container;
 123 }
 124
 125 struct blt_info
 126 {
 127     GLenum binding;
 128     GLenum bind_target;
 129     enum tex_types tex_type;
 130     GLfloat coords[4][3];
 131 };
 132
 133 struct float_rect
 134 {
 135     float l;
 136     float t;
 137     float r;
 138     float b;
 139 };
 140
 141 static inline void cube_coords_float(const RECT *r, UINT w, UINT h, struct float_rect *f)
 142 {
 143     f->l = ((r->left * 2.0f) / w) - 1.0f;
 144     f->t = ((r->top * 2.0f) / h) - 1.0f;
 145     f->r = ((r->right * 2.0f) / w) - 1.0f;
 146     f->b = ((r->bottom * 2.0f) / h) - 1.0f;
 147 }
 148
 149 static void surface_get_blt_info(GLenum target, const RECT *rect_in, GLsizei w, GLsizei h, struct blt_info *info)
 150 {
 151     GLfloat (*coords)[3] = info->coords;
 152     RECT rect;
 153     struct float_rect f;
 154
 155     if (rect_in)
 156         rect = *rect_in;
 157     else
 158     {
 159         rect.left = 0;
 160         rect.top = h;
 161         rect.right = w;
 162         rect.bottom = 0;
 163     }
 164
 165     switch (target)
 166     {
 167         default:
 168             FIXME("Unsupported texture target %#x\n", target);
 169             /* Fall back to GL_TEXTURE_2D */
 170         case GL_TEXTURE_2D:
 171             info->binding = GL_TEXTURE_BINDING_2D;
 172             info->bind_target = GL_TEXTURE_2D;
 173             info->tex_type = tex_2d;
 174             coords[0][0] = (float)rect.left / w;
 175             coords[0][1] = (float)rect.top / h;
 176             coords[0][2] = 0.0f;
 177
 178             coords[1][0] = (float)rect.right / w;
 179             coords[1][1] = (float)rect.top / h;
 180             coords[1][2] = 0.0f;
 181
 182             coords[2][0] = (float)rect.left / w;
 183             coords[2][1] = (float)rect.bottom / h;
 184             coords[2][2] = 0.0f;
 185
 186             coords[3][0] = (float)rect.right / w;
 187             coords[3][1] = (float)rect.bottom / h;
 188             coords[3][2] = 0.0f;
 189             break;
 190
 191         case GL_TEXTURE_RECTANGLE_ARB:
 192             info->binding = GL_TEXTURE_BINDING_RECTANGLE_ARB;
 193             info->bind_target = GL_TEXTURE_RECTANGLE_ARB;
 194             info->tex_type = tex_rect;
 195             coords[0][0] = rect.left;   coords[0][1] = rect.top;     coords[0][2] = 0.0f;
 196             coords[1][0] = rect.right;  coords[1][1] = rect.top;     coords[1][2] = 0.0f;
 197             coords[2][0] = rect.left;   coords[2][1] = rect.bottom;  coords[2][2] = 0.0f;
 198             coords[3][0] = rect.right;  coords[3][1] = rect.bottom;  coords[3][2] = 0.0f;
 199             break;
 200
 201         case GL_TEXTURE_CUBE_MAP_POSITIVE_X:
 202             info->binding = GL_TEXTURE_BINDING_CUBE_MAP_ARB;
 203             info->bind_target = GL_TEXTURE_CUBE_MAP_ARB;
 204             info->tex_type = tex_cube;
 205             cube_coords_float(&rect, w, h, &f);
 206
 207             coords[0][0] =  1.0f;   coords[0][1] = -f.t;   coords[0][2] = -f.l;
 208             coords[1][0] =  1.0f;   coords[1][1] = -f.t;   coords[1][2] = -f.r;
 209             coords[2][0] =  1.0f;   coords[2][1] = -f.b;   coords[2][2] = -f.l;
 210             coords[3][0] =  1.0f;   coords[3][1] = -f.b;   coords[3][2] = -f.r;
 211             break;
 212
 213         case GL_TEXTURE_CUBE_MAP_NEGATIVE_X:
 214             info->binding = GL_TEXTURE_BINDING_CUBE_MAP_ARB;
 215             info->bind_target = GL_TEXTURE_CUBE_MAP_ARB;
 216             info->tex_type = tex_cube;
 217             cube_coords_float(&rect, w, h, &f);
 218
 219             coords[0][0] = -1.0f;   coords[0][1] = -f.t;   coords[0][2] = f.l;
 220             coords[1][0] = -1.0f;   coords[1][1] = -f.t;   coords[1][2] = f.r;
 221             coords[2][0] = -1.0f;   coords[2][1] = -f.b;   coords[2][2] = f.l;
 222             coords[3][0] = -1.0f;   coords[3][1] = -f.b;   coords[3][2] = f.r;
 223             break;
 224
 225         case GL_TEXTURE_CUBE_MAP_POSITIVE_Y:
 226             info->binding = GL_TEXTURE_BINDING_CUBE_MAP_ARB;
 227             info->bind_target = GL_TEXTURE_CUBE_MAP_ARB;
 228             info->tex_type = tex_cube;
 229             cube_coords_float(&rect, w, h, &f);
 230
 231             coords[0][0] = f.l;   coords[0][1] =  1.0f;   coords[0][2] = f.t;
 232             coords[1][0] = f.r;   coords[1][1] =  1.0f;   coords[1][2] = f.t;
 233             coords[2][0] = f.l;   coords[2][1] =  1.0f;   coords[2][2] = f.b;
 234             coords[3][0] = f.r;   coords[3][1] =  1.0f;   coords[3][2] = f.b;
 235             break;
 236
 237         case GL_TEXTURE_CUBE_MAP_NEGATIVE_Y:
 238             info->binding = GL_TEXTURE_BINDING_CUBE_MAP_ARB;
 239             info->bind_target = GL_TEXTURE_CUBE_MAP_ARB;
 240             info->tex_type = tex_cube;
 241             cube_coords_float(&rect, w, h, &f);
 242
 243             coords[0][0] = f.l;   coords[0][1] = -1.0f;   coords[0][2] = -f.t;
 244             coords[1][0] = f.r;   coords[1][1] = -1.0f;   coords[1][2] = -f.t;
 245             coords[2][0] = f.l;   coords[2][1] = -1.0f;   coords[2][2] = -f.b;
 246             coords[3][0] = f.r;   coords[3][1] = -1.0f;   coords[3][2] = -f.b;
 247             break;
 248
 249         case GL_TEXTURE_CUBE_MAP_POSITIVE_Z:
 250             info->binding = GL_TEXTURE_BINDING_CUBE_MAP_ARB;
 251             info->bind_target = GL_TEXTURE_CUBE_MAP_ARB;
 252             info->tex_type = tex_cube;
 253             cube_coords_float(&rect, w, h, &f);
 254
 255             coords[0][0] = f.l;   coords[0][1] = -f.t;   coords[0][2] =  1.0f;
 256             coords[1][0] = f.r;   coords[1][1] = -f.t;   coords[1][2] =  1.0f;
 257             coords[2][0] = f.l;   coords[2][1] = -f.b;   coords[2][2] =  1.0f;
 258             coords[3][0] = f.r;   coords[3][1] = -f.b;   coords[3][2] =  1.0f;
 259             break;
 260
 261         case GL_TEXTURE_CUBE_MAP_NEGATIVE_Z:
 262             info->binding = GL_TEXTURE_BINDING_CUBE_MAP_ARB;
 263             info->bind_target = GL_TEXTURE_CUBE_MAP_ARB;
 264             info->tex_type = tex_cube;
 265             cube_coords_float(&rect, w, h, &f);
 266
 267             coords[0][0] = -f.l;   coords[0][1] = -f.t;   coords[0][2] = -1.0f;
 268             coords[1][0] = -f.r;   coords[1][1] = -f.t;   coords[1][2] = -1.0f;
 269             coords[2][0] = -f.l;   coords[2][1] = -f.b;   coords[2][2] = -1.0f;
 270             coords[3][0] = -f.r;   coords[3][1] = -f.b;   coords[3][2] = -1.0f;
 271             break;
 272     }
 273 }
 274
 275 static inline void surface_get_rect(IWineD3DSurfaceImpl *This, const RECT *rect_in, RECT *rect_out)
 276 {
 277     if (rect_in)
 278         *rect_out = *rect_in;
 279     else
 280     {
 281         rect_out->left = 0;
 282         rect_out->top = 0;
 283         rect_out->right = This->currentDesc.Width;
 284         rect_out->bottom = This->currentDesc.Height;
 285     }
 286 }
 287
 288 /* GL locking and context activation is done by the caller */
 289 void draw_textured_quad(IWineD3DSurfaceImpl *src_surface, const RECT *src_rect, const RECT *dst_rect, WINED3DTEXTUREFILTERTYPE Filter)
 290 {
 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 (src_surface->container.type == WINED3D_CONTAINER_TEXTURE)
 336     {
 337         IWineD3DBaseTextureImpl *texture = src_surface->container.u.texture;
 338         texture->baseTexture.texture_rgb.states[WINED3DTEXSTA_MAGFILTER] = WINED3DTEXF_POINT;
 339         texture->baseTexture.texture_rgb.states[WINED3DTEXSTA_MINFILTER] = WINED3DTEXF_POINT;
 340         texture->baseTexture.texture_rgb.states[WINED3DTEXSTA_MIPFILTER] = WINED3DTEXF_NONE;
 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, enum wined3d_format_id format_id,
 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_id, 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 = wined3d_format_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     surface_set_container(surface, WINED3D_CONTAINER_NONE, 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_id == 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 %s set, but no texture name.\n", debug_surflocation(flag));
 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 = surface->container.u.swapchain;
 925
 926     TRACE("surface %p.\n", surface);
 927
 928     if (surface->container.type != WINED3D_CONTAINER_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     TRACE("surface %p, dirty_rect %s.\n", surface, wine_dbgstr_rect(dirty_rect));
 958
 959     if (!(surface->Flags & SFLAG_INSYSMEM) && (surface->Flags & SFLAG_INTEXTURE))
 960         /* No partial locking for textures yet. */
 961         surface_load_location(surface, SFLAG_INSYSMEM, NULL);
 962
 963     surface_modify_location(surface, SFLAG_INSYSMEM, TRUE);
 964     if (dirty_rect)
 965     {
 966         surface->dirtyRect.left = min(surface->dirtyRect.left, dirty_rect->left);
 967         surface->dirtyRect.top = min(surface->dirtyRect.top, dirty_rect->top);
 968         surface->dirtyRect.right = max(surface->dirtyRect.right, dirty_rect->right);
 969         surface->dirtyRect.bottom = max(surface->dirtyRect.bottom, dirty_rect->bottom);
 970     }
 971     else
 972     {
 973         surface->dirtyRect.left = 0;
 974         surface->dirtyRect.top = 0;
 975         surface->dirtyRect.right = surface->currentDesc.Width;
 976         surface->dirtyRect.bottom = surface->currentDesc.Height;
 977     }
 978
 979     /* if the container is a basetexture then mark it dirty. */
 980     if (surface->container.type == WINED3D_CONTAINER_TEXTURE)
 981     {
 982         TRACE("Passing to container.\n");
 983         IWineD3DBaseTexture_SetDirty((IWineD3DBaseTexture *)surface->container.u.texture, TRUE);
 984     }
 985 }
 986
 987 static BOOL surface_convert_color_to_float(IWineD3DSurfaceImpl *surface, DWORD color, WINED3DCOLORVALUE *float_color)
 988 {
 989     const struct wined3d_format_desc *format = surface->resource.format_desc;
 990     IWineD3DDeviceImpl *device = surface->resource.device;
 991
 992     switch (format->format)
 993     {
 994         case WINED3DFMT_P8_UINT:
 995             if (surface->palette)
 996             {
 997                 float_color->r = surface->palette->palents[color].peRed / 255.0f;
 998                 float_color->g = surface->palette->palents[color].peGreen / 255.0f;
 999                 float_color->b = surface->palette->palents[color].peBlue / 255.0f;
1000             }
1001             else
1002             {
1003                 float_color->r = 0.0f;
1004                 float_color->g = 0.0f;
1005                 float_color->b = 0.0f;
1006             }
1007             float_color->a = primary_render_target_is_p8(device) ? color / 255.0f : 1.0f;
1008             break;
1009
1010         case WINED3DFMT_B5G6R5_UNORM:
1011             float_color->r = ((color >> 11) & 0x1f) / 31.0f;
1012             float_color->g = ((color >> 5) & 0x3f) / 63.0f;
1013             float_color->b = (color & 0x1f) / 31.0f;
1014             float_color->a = 1.0f;
1015             break;
1016
1017         case WINED3DFMT_B8G8R8_UNORM:
1018         case WINED3DFMT_B8G8R8X8_UNORM:
1019             float_color->r = D3DCOLOR_R(color);
1020             float_color->g = D3DCOLOR_G(color);
1021             float_color->b = D3DCOLOR_B(color);
1022             float_color->a = 1.0f;
1023             break;
1024
1025         case WINED3DFMT_B8G8R8A8_UNORM:
1026             float_color->r = D3DCOLOR_R(color);
1027             float_color->g = D3DCOLOR_G(color);
1028             float_color->b = D3DCOLOR_B(color);
1029             float_color->a = D3DCOLOR_A(color);
1030             break;
1031
1032         default:
1033             ERR("Unhandled conversion from %s to floating point.\n", debug_d3dformat(format->format));
1034             return FALSE;
1035     }
1036
1037     return TRUE;
1038 }
1039
1040 static ULONG WINAPI IWineD3DSurfaceImpl_Release(IWineD3DSurface *iface)
1041 {
1042     IWineD3DSurfaceImpl *This = (IWineD3DSurfaceImpl *)iface;
1043     ULONG ref = InterlockedDecrement(&This->resource.ref);
1044     TRACE("(%p) : Releasing from %d\n", This, ref + 1);
1045
1046     if (!ref)
1047     {
1048         surface_cleanup(This);
1049         This->resource.parent_ops->wined3d_object_destroyed(This->resource.parent);
1050
1051         TRACE("(%p) Released.\n", This);
1052         HeapFree(GetProcessHeap(), 0, This);
1053     }
1054
1055     return ref;
1056 }
1057
1058 /* ****************************************************
1059    IWineD3DSurface IWineD3DResource parts follow
1060    **************************************************** */
1061
1062 void surface_internal_preload(IWineD3DSurfaceImpl *surface, enum WINED3DSRGB srgb)
1063 {
1064     IWineD3DDeviceImpl *device = surface->resource.device;
1065
1066     TRACE("iface %p, srgb %#x.\n", surface, srgb);
1067
1068     if (surface->container.type == WINED3D_CONTAINER_TEXTURE)
1069     {
1070         IWineD3DBaseTextureImpl *texture = surface->container.u.texture;
1071
1072         TRACE("Passing to container.\n");
1073         texture->baseTexture.internal_preload((IWineD3DBaseTexture *)texture, srgb);
1074     }
1075     else
1076     {
1077         struct wined3d_context *context = NULL;
1078
1079         TRACE("(%p) : About to load surface\n", surface);
1080
1081         if (!device->isInDraw) context = context_acquire(device, NULL);
1082
1083         if (surface->resource.format_desc->format == WINED3DFMT_P8_UINT
1084                 || surface->resource.format_desc->format == WINED3DFMT_P8_UINT_A8_UNORM)
1085         {
1086             if (palette9_changed(surface))
1087             {
1088                 TRACE("Reloading surface because the d3d8/9 palette was changed\n");
1089                 /* TODO: This is not necessarily needed with hw palettized texture support */
1090                 surface_load_location(surface, SFLAG_INSYSMEM, NULL);
1091                 /* Make sure the texture is reloaded because of the palette change, this kills performance though :( */
1092                 surface_modify_location(surface, SFLAG_INTEXTURE, FALSE);
1093             }
1094         }
1095
1096         IWineD3DSurface_LoadTexture((IWineD3DSurface *)surface, srgb == SRGB_SRGB ? TRUE : FALSE);
1097
1098         if (surface->resource.pool == WINED3DPOOL_DEFAULT)
1099         {
1100             /* Tell opengl to try and keep this texture in video ram (well mostly) */
1101             GLclampf tmp;
1102             tmp = 0.9f;
1103             ENTER_GL();
1104             glPrioritizeTextures(1, &surface->texture_name, &tmp);
1105             LEAVE_GL();
1106         }
1107
1108         if (context) context_release(context);
1109     }
1110 }
1111
1112 static void WINAPI IWineD3DSurfaceImpl_PreLoad(IWineD3DSurface *iface)
1113 {
1114     surface_internal_preload((IWineD3DSurfaceImpl *)iface, SRGB_ANY);
1115 }
1116
1117 /* Context activation is done by the caller. */
1118 static void surface_remove_pbo(IWineD3DSurfaceImpl *This, const struct wined3d_gl_info *gl_info)
1119 {
1120     This->resource.heapMemory = HeapAlloc(GetProcessHeap() ,0 , This->resource.size + RESOURCE_ALIGNMENT);
1121     This->resource.allocatedMemory =
1122             (BYTE *)(((ULONG_PTR) This->resource.heapMemory + (RESOURCE_ALIGNMENT - 1)) & ~(RESOURCE_ALIGNMENT - 1));
1123
1124     ENTER_GL();
1125     GL_EXTCALL(glBindBufferARB(GL_PIXEL_UNPACK_BUFFER_ARB, This->pbo));
1126     checkGLcall("glBindBufferARB(GL_PIXEL_UNPACK_BUFFER, This->pbo)");
1127     GL_EXTCALL(glGetBufferSubDataARB(GL_PIXEL_UNPACK_BUFFER_ARB, 0, This->resource.size, This->resource.allocatedMemory));
1128     checkGLcall("glGetBufferSubDataARB");
1129     GL_EXTCALL(glDeleteBuffersARB(1, &This->pbo));
1130     checkGLcall("glDeleteBuffersARB");
1131     LEAVE_GL();
1132
1133     This->pbo = 0;
1134     This->Flags &= ~SFLAG_PBO;
1135 }
1136
1137 BOOL surface_init_sysmem(IWineD3DSurfaceImpl *surface)
1138 {
1139     if (!surface->resource.allocatedMemory)
1140     {
1141         surface->resource.heapMemory = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY,
1142                 surface->resource.size + RESOURCE_ALIGNMENT);
1143         if (!surface->resource.heapMemory)
1144         {
1145             ERR("Out of memory\n");
1146             return FALSE;
1147         }
1148         surface->resource.allocatedMemory =
1149             (BYTE *)(((ULONG_PTR)surface->resource.heapMemory + (RESOURCE_ALIGNMENT - 1)) & ~(RESOURCE_ALIGNMENT - 1));
1150     }
1151     else
1152     {
1153         memset(surface->resource.allocatedMemory, 0, surface->resource.size);
1154     }
1155
1156     surface_modify_location(surface, SFLAG_INSYSMEM, TRUE);
1157
1158     return TRUE;
1159 }
1160
1161 static void WINAPI IWineD3DSurfaceImpl_UnLoad(IWineD3DSurface *iface)
1162 {
1163     IWineD3DSurfaceImpl *This = (IWineD3DSurfaceImpl *) iface;
1164     IWineD3DDeviceImpl *device = This->resource.device;
1165     const struct wined3d_gl_info *gl_info;
1166     renderbuffer_entry_t *entry, *entry2;
1167     struct wined3d_context *context;
1168
1169     TRACE("(%p)\n", iface);
1170
1171     if(This->resource.pool == WINED3DPOOL_DEFAULT) {
1172         /* Default pool resources are supposed to be destroyed before Reset is called.
1173          * Implicit resources stay however. So this means we have an implicit render target
1174          * or depth stencil. The content may be destroyed, but we still have to tear down
1175          * opengl resources, so we cannot leave early.
1176          *
1177          * Put the surfaces into sysmem, and reset the content. The D3D content is undefined,
1178          * but we can't set the sysmem INDRAWABLE because when we're rendering the swapchain
1179          * or the depth stencil into an FBO the texture or render buffer will be removed
1180          * and all flags get lost
1181          */
1182         surface_init_sysmem(This);
1183     }
1184     else
1185     {
1186         /* Load the surface into system memory */
1187         surface_load_location(This, SFLAG_INSYSMEM, NULL);
1188         surface_modify_location(This, SFLAG_INDRAWABLE, FALSE);
1189     }
1190     surface_modify_location(This, SFLAG_INTEXTURE, FALSE);
1191     surface_modify_location(This, SFLAG_INSRGBTEX, FALSE);
1192     This->Flags &= ~(SFLAG_ALLOCATED | SFLAG_SRGBALLOCATED);
1193
1194     context = context_acquire(device, NULL);
1195     gl_info = context->gl_info;
1196
1197     /* Destroy PBOs, but load them into real sysmem before */
1198     if (This->Flags & SFLAG_PBO)
1199         surface_remove_pbo(This, gl_info);
1200
1201     /* Destroy fbo render buffers. This is needed for implicit render targets, for
1202      * all application-created targets the application has to release the surface
1203      * before calling _Reset
1204      */
1205     LIST_FOR_EACH_ENTRY_SAFE(entry, entry2, &This->renderbuffers, renderbuffer_entry_t, entry) {
1206         ENTER_GL();
1207         gl_info->fbo_ops.glDeleteRenderbuffers(1, &entry->id);
1208         LEAVE_GL();
1209         list_remove(&entry->entry);
1210         HeapFree(GetProcessHeap(), 0, entry);
1211     }
1212     list_init(&This->renderbuffers);
1213     This->current_renderbuffer = NULL;
1214
1215     /* If we're in a texture, the texture name belongs to the texture.
1216      * Otherwise, destroy it. */
1217     if (This->container.type != WINED3D_CONTAINER_TEXTURE)
1218     {
1219         ENTER_GL();
1220         glDeleteTextures(1, &This->texture_name);
1221         This->texture_name = 0;
1222         glDeleteTextures(1, &This->texture_name_srgb);
1223         This->texture_name_srgb = 0;
1224         LEAVE_GL();
1225     }
1226
1227     context_release(context);
1228
1229     resource_unload((IWineD3DResourceImpl *)This);
1230 }
1231
1232 /* ******************************************************
1233    IWineD3DSurface IWineD3DSurface parts follow
1234    ****************************************************** */
1235
1236 /* Read the framebuffer back into the surface */
1237 static void read_from_framebuffer(IWineD3DSurfaceImpl *This, const RECT *rect, void *dest, UINT pitch)
1238 {
1239     IWineD3DDeviceImpl *device = This->resource.device;
1240     const struct wined3d_gl_info *gl_info;
1241     struct wined3d_context *context;
1242     BYTE *mem;
1243     GLint fmt;
1244     GLint type;
1245     BYTE *row, *top, *bottom;
1246     int i;
1247     BOOL bpp;
1248     RECT local_rect;
1249     BOOL srcIsUpsideDown;
1250     GLint rowLen = 0;
1251     GLint skipPix = 0;
1252     GLint skipRow = 0;
1253
1254     if(wined3d_settings.rendertargetlock_mode == RTL_DISABLE) {
1255         static BOOL warned = FALSE;
1256         if(!warned) {
1257             ERR("The application tries to lock the render target, but render target locking is disabled\n");
1258             warned = TRUE;
1259         }
1260         return;
1261     }
1262
1263     /* Activate the surface. Set it up for blitting now, although not necessarily needed for LockRect.
1264      * Certain graphics drivers seem to dislike some enabled states when reading from opengl, the blitting usage
1265      * should help here. Furthermore unlockrect will need the context set up for blitting. The context manager will find
1266      * context->last_was_blit set on the unlock.
1267      */
1268     context = context_acquire(device, This);
1269     context_apply_blit_state(context, device);
1270     gl_info = context->gl_info;
1271
1272     ENTER_GL();
1273
1274     /* Select the correct read buffer, and give some debug output.
1275      * There is no need to keep track of the current read buffer or reset it, every part of the code
1276      * that reads sets the read buffer as desired.
1277      */
1278     if (surface_is_offscreen(This))
1279     {
1280         /* Locking the primary render target which is not on a swapchain(=offscreen render target).
1281          * Read from the back buffer
1282          */
1283         TRACE("Locking offscreen render target\n");
1284         glReadBuffer(device->offscreenBuffer);
1285         srcIsUpsideDown = TRUE;
1286     }
1287     else
1288     {
1289         /* Onscreen surfaces are always part of a swapchain */
1290         GLenum buffer = surface_get_gl_buffer(This);
1291         TRACE("Locking %#x buffer\n", buffer);
1292         glReadBuffer(buffer);
1293         checkGLcall("glReadBuffer");
1294         srcIsUpsideDown = FALSE;
1295     }
1296
1297     /* TODO: Get rid of the extra rectangle comparison and construction of a full surface rectangle */
1298     if(!rect) {
1299         local_rect.left = 0;
1300         local_rect.top = 0;
1301         local_rect.right = This->currentDesc.Width;
1302         local_rect.bottom = This->currentDesc.Height;
1303     } else {
1304         local_rect = *rect;
1305     }
1306     /* TODO: Get rid of the extra GetPitch call, LockRect does that too. Cache the pitch */
1307
1308     switch(This->resource.format_desc->format)
1309     {
1310         case WINED3DFMT_P8_UINT:
1311         {
1312             if (primary_render_target_is_p8(device))
1313             {
1314                 /* In case of P8 render targets the index is stored in the alpha component */
1315                 fmt = GL_ALPHA;
1316                 type = GL_UNSIGNED_BYTE;
1317                 mem = dest;
1318                 bpp = This->resource.format_desc->byte_count;
1319             } else {
1320                 /* GL can't return palettized data, so read ARGB pixels into a
1321                  * separate block of memory and convert them into palettized format
1322                  * in software. Slow, but if the app means to use palettized render
1323                  * targets and locks it...
1324                  *
1325                  * Use GL_RGB, GL_UNSIGNED_BYTE to read the surface for performance reasons
1326                  * Don't use GL_BGR as in the WINED3DFMT_R8G8B8 case, instead watch out
1327                  * for the color channels when palettizing the colors.
1328                  */
1329                 fmt = GL_RGB;
1330                 type = GL_UNSIGNED_BYTE;
1331                 pitch *= 3;
1332                 mem = HeapAlloc(GetProcessHeap(), 0, This->resource.size * 3);
1333                 if(!mem) {
1334                     ERR("Out of memory\n");
1335                     LEAVE_GL();
1336                     return;
1337                 }
1338                 bpp = This->resource.format_desc->byte_count * 3;
1339             }
1340         }
1341         break;
1342
1343         default:
1344             mem = dest;
1345             fmt = This->resource.format_desc->glFormat;
1346             type = This->resource.format_desc->glType;
1347             bpp = This->resource.format_desc->byte_count;
1348     }
1349
1350     if(This->Flags & SFLAG_PBO) {
1351         GL_EXTCALL(glBindBufferARB(GL_PIXEL_PACK_BUFFER_ARB, This->pbo));
1352         checkGLcall("glBindBufferARB");
1353         if(mem != NULL) {
1354             ERR("mem not null for pbo -- unexpected\n");
1355             mem = NULL;
1356         }
1357     }
1358
1359     /* Save old pixel store pack state */
1360     glGetIntegerv(GL_PACK_ROW_LENGTH, &rowLen);
1361     checkGLcall("glGetIntegerv");
1362     glGetIntegerv(GL_PACK_SKIP_PIXELS, &skipPix);
1363     checkGLcall("glGetIntegerv");
1364     glGetIntegerv(GL_PACK_SKIP_ROWS, &skipRow);
1365     checkGLcall("glGetIntegerv");
1366
1367     /* Setup pixel store pack state -- to glReadPixels into the correct place */
1368     glPixelStorei(GL_PACK_ROW_LENGTH, This->currentDesc.Width);
1369     checkGLcall("glPixelStorei");
1370     glPixelStorei(GL_PACK_SKIP_PIXELS, local_rect.left);
1371     checkGLcall("glPixelStorei");
1372     glPixelStorei(GL_PACK_SKIP_ROWS, local_rect.top);
1373     checkGLcall("glPixelStorei");
1374
1375     glReadPixels(local_rect.left, (!srcIsUpsideDown) ? (This->currentDesc.Height - local_rect.bottom) : local_rect.top ,
1376                  local_rect.right - local_rect.left,
1377                  local_rect.bottom - local_rect.top,
1378                  fmt, type, mem);
1379     checkGLcall("glReadPixels");
1380
1381     /* Reset previous pixel store pack state */
1382     glPixelStorei(GL_PACK_ROW_LENGTH, rowLen);
1383     checkGLcall("glPixelStorei");
1384     glPixelStorei(GL_PACK_SKIP_PIXELS, skipPix);
1385     checkGLcall("glPixelStorei");
1386     glPixelStorei(GL_PACK_SKIP_ROWS, skipRow);
1387     checkGLcall("glPixelStorei");
1388
1389     if(This->Flags & SFLAG_PBO) {
1390         GL_EXTCALL(glBindBufferARB(GL_PIXEL_PACK_BUFFER_ARB, 0));
1391         checkGLcall("glBindBufferARB");
1392
1393         /* Check if we need to flip the image. If we need to flip use glMapBufferARB
1394          * to get a pointer to it and perform the flipping in software. This is a lot
1395          * faster than calling glReadPixels for each line. In case we want more speed
1396          * we should rerender it flipped in a FBO and read the data back from the FBO. */
1397         if(!srcIsUpsideDown) {
1398             GL_EXTCALL(glBindBufferARB(GL_PIXEL_UNPACK_BUFFER_ARB, This->pbo));
1399             checkGLcall("glBindBufferARB");
1400
1401             mem = GL_EXTCALL(glMapBufferARB(GL_PIXEL_UNPACK_BUFFER_ARB, GL_READ_WRITE_ARB));
1402             checkGLcall("glMapBufferARB");
1403         }
1404     }
1405
1406     /* TODO: Merge this with the palettization loop below for P8 targets */
1407     if(!srcIsUpsideDown) {
1408         UINT len, off;
1409         /* glReadPixels returns the image upside down, and there is no way to prevent this.
1410             Flip the lines in software */
1411         len = (local_rect.right - local_rect.left) * bpp;
1412         off = local_rect.left * bpp;
1413
1414         row = HeapAlloc(GetProcessHeap(), 0, len);
1415         if(!row) {
1416             ERR("Out of memory\n");
1417             if (This->resource.format_desc->format == WINED3DFMT_P8_UINT) HeapFree(GetProcessHeap(), 0, mem);
1418             LEAVE_GL();
1419             return;
1420         }
1421
1422         top = mem + pitch * local_rect.top;
1423         bottom = mem + pitch * (local_rect.bottom - 1);
1424         for(i = 0; i < (local_rect.bottom - local_rect.top) / 2; i++) {
1425             memcpy(row, top + off, len);
1426             memcpy(top + off, bottom + off, len);
1427             memcpy(bottom + off, row, len);
1428             top += pitch;
1429             bottom -= pitch;
1430         }
1431         HeapFree(GetProcessHeap(), 0, row);
1432
1433         /* Unmap the temp PBO buffer */
1434         if(This->Flags & SFLAG_PBO) {
1435             GL_EXTCALL(glUnmapBufferARB(GL_PIXEL_UNPACK_BUFFER_ARB));
1436             GL_EXTCALL(glBindBufferARB(GL_PIXEL_UNPACK_BUFFER_ARB, 0));
1437         }
1438     }
1439
1440     LEAVE_GL();
1441     context_release(context);
1442
1443     /* For P8 textures we need to perform an inverse palette lookup. This is done by searching for a palette
1444      * index which matches the RGB value. Note this isn't guaranteed to work when there are multiple entries for
1445      * the same color but we have no choice.
1446      * In case of P8 render targets, the index is stored in the alpha component so no conversion is needed.
1447      */
1448     if (This->resource.format_desc->format == WINED3DFMT_P8_UINT && !primary_render_target_is_p8(device))
1449     {
1450         const PALETTEENTRY *pal = NULL;
1451         DWORD width = pitch / 3;
1452         int x, y, c;
1453
1454         if(This->palette) {
1455             pal = This->palette->palents;
1456         } else {
1457             ERR("Palette is missing, cannot perform inverse palette lookup\n");
1458             HeapFree(GetProcessHeap(), 0, mem);
1459             return ;
1460         }
1461
1462         for(y = local_rect.top; y < local_rect.bottom; y++) {
1463             for(x = local_rect.left; x < local_rect.right; x++) {
1464                 /*                      start              lines            pixels      */
1465                 const BYTE *blue = mem + y * pitch + x * (sizeof(BYTE) * 3);
1466                 const BYTE *green = blue  + 1;
1467                 const BYTE *red = green + 1;
1468
1469                 for(c = 0; c < 256; c++) {
1470                     if(*red   == pal[c].peRed   &&
1471                        *green == pal[c].peGreen &&
1472                        *blue  == pal[c].peBlue)
1473                     {
1474                         *((BYTE *) dest + y * width + x) = c;
1475                         break;
1476                     }
1477                 }
1478             }
1479         }
1480         HeapFree(GetProcessHeap(), 0, mem);
1481     }
1482 }
1483
1484 /* Read the framebuffer contents into a texture */
1485 static void read_from_framebuffer_texture(IWineD3DSurfaceImpl *This, BOOL srgb)
1486 {
1487     IWineD3DDeviceImpl *device = This->resource.device;
1488     const struct wined3d_gl_info *gl_info;
1489     struct wined3d_context *context;
1490
1491     if (!surface_is_offscreen(This))
1492     {
1493         /* We would need to flip onscreen surfaces, but there's no efficient
1494          * way to do that here. It makes more sense for the caller to
1495          * explicitly go through sysmem. */
1496         ERR("Not supported for onscreen targets.\n");
1497         return;
1498     }
1499
1500     /* Activate the surface to read from. In some situations it isn't the currently active target(e.g. backbuffer
1501      * locking during offscreen rendering). RESOURCELOAD is ok because glCopyTexSubImage2D isn't affected by any
1502      * states in the stateblock, and no driver was found yet that had bugs in that regard.
1503      */
1504     context = context_acquire(device, This);
1505     gl_info = context->gl_info;
1506
1507     surface_prepare_texture(This, gl_info, srgb);
1508     surface_bind_and_dirtify(This, srgb);
1509
1510     TRACE("Reading back offscreen render target %p.\n", This);
1511
1512     ENTER_GL();
1513
1514     glReadBuffer(device->offscreenBuffer);
1515     checkGLcall("glReadBuffer");
1516
1517     glCopyTexSubImage2D(This->texture_target, This->texture_level,
1518             0, 0, 0, 0, This->currentDesc.Width, This->currentDesc.Height);
1519     checkGLcall("glCopyTexSubImage2D");
1520
1521     LEAVE_GL();
1522
1523     context_release(context);
1524 }
1525
1526 /* Context activation is done by the caller. */
1527 static void surface_prepare_texture_internal(IWineD3DSurfaceImpl *surface,
1528         const struct wined3d_gl_info *gl_info, BOOL srgb)
1529 {
1530     DWORD alloc_flag = srgb ? SFLAG_SRGBALLOCATED : SFLAG_ALLOCATED;
1531     CONVERT_TYPES convert;
1532     struct wined3d_format_desc desc;
1533
1534     if (surface->Flags & alloc_flag) return;
1535
1536     d3dfmt_get_conv(surface, TRUE, TRUE, &desc, &convert);
1537     if(convert != NO_CONVERSION || desc.convert) surface->Flags |= SFLAG_CONVERTED;
1538     else surface->Flags &= ~SFLAG_CONVERTED;
1539
1540     surface_bind_and_dirtify(surface, srgb);
1541     surface_allocate_surface(surface, gl_info, &desc, srgb);
1542     surface->Flags |= alloc_flag;
1543 }
1544
1545 /* Context activation is done by the caller. */
1546 void surface_prepare_texture(IWineD3DSurfaceImpl *surface, const struct wined3d_gl_info *gl_info, BOOL srgb)
1547 {
1548     if (surface->container.type == WINED3D_CONTAINER_TEXTURE)
1549     {
1550         IWineD3DBaseTextureImpl *texture = surface->container.u.texture;
1551         UINT sub_count = texture->baseTexture.level_count * texture->baseTexture.layer_count;
1552         UINT i;
1553
1554         TRACE("surface %p is a subresource of texture %p.\n", surface, texture);
1555
1556         for (i = 0; i < sub_count; ++i)
1557         {
1558             IWineD3DSurfaceImpl *s = (IWineD3DSurfaceImpl *)texture->baseTexture.sub_resources[i];
1559             surface_prepare_texture_internal(s, gl_info, srgb);
1560         }
1561
1562         return;
1563     }
1564
1565     surface_prepare_texture_internal(surface, gl_info, srgb);
1566 }
1567
1568 static void surface_prepare_system_memory(IWineD3DSurfaceImpl *This)
1569 {
1570     IWineD3DDeviceImpl *device = This->resource.device;
1571     const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
1572
1573     /* Performance optimization: Count how often a surface is locked, if it is locked regularly do not throw away the system memory copy.
1574      * This avoids the need to download the surface from opengl all the time. The surface is still downloaded if the opengl texture is
1575      * changed
1576      */
1577     if(!(This->Flags & SFLAG_DYNLOCK)) {
1578         This->lockCount++;
1579         /* MAXLOCKCOUNT is defined in wined3d_private.h */
1580         if(This->lockCount > MAXLOCKCOUNT) {
1581             TRACE("Surface is locked regularly, not freeing the system memory copy any more\n");
1582             This->Flags |= SFLAG_DYNLOCK;
1583         }
1584     }
1585
1586     /* Create a PBO for dynamically locked surfaces but don't do it for converted or non-pow2 surfaces.
1587      * Also don't create a PBO for systemmem surfaces.
1588      */
1589     if (gl_info->supported[ARB_PIXEL_BUFFER_OBJECT] && (This->Flags & SFLAG_DYNLOCK)
1590             && !(This->Flags & (SFLAG_PBO | SFLAG_CONVERTED | SFLAG_NONPOW2))
1591             && (This->resource.pool != WINED3DPOOL_SYSTEMMEM))
1592     {
1593         GLenum error;
1594         struct wined3d_context *context;
1595
1596         context = context_acquire(device, NULL);
1597         ENTER_GL();
1598
1599         GL_EXTCALL(glGenBuffersARB(1, &This->pbo));
1600         error = glGetError();
1601         if(This->pbo == 0 || error != GL_NO_ERROR) {
1602             ERR("Failed to bind the PBO with error %s (%#x)\n", debug_glerror(error), error);
1603         }
1604
1605         TRACE("Attaching pbo=%#x to (%p)\n", This->pbo, This);
1606
1607         GL_EXTCALL(glBindBufferARB(GL_PIXEL_UNPACK_BUFFER_ARB, This->pbo));
1608         checkGLcall("glBindBufferARB");
1609
1610         GL_EXTCALL(glBufferDataARB(GL_PIXEL_UNPACK_BUFFER_ARB, This->resource.size + 4, This->resource.allocatedMemory, GL_STREAM_DRAW_ARB));
1611         checkGLcall("glBufferDataARB");
1612
1613         GL_EXTCALL(glBindBufferARB(GL_PIXEL_UNPACK_BUFFER_ARB, 0));
1614         checkGLcall("glBindBufferARB");
1615
1616         /* We don't need the system memory anymore and we can't even use it for PBOs */
1617         if(!(This->Flags & SFLAG_CLIENT)) {
1618             HeapFree(GetProcessHeap(), 0, This->resource.heapMemory);
1619             This->resource.heapMemory = NULL;
1620         }
1621         This->resource.allocatedMemory = NULL;
1622         This->Flags |= SFLAG_PBO;
1623         LEAVE_GL();
1624         context_release(context);
1625     }
1626     else if (!(This->resource.allocatedMemory || This->Flags & SFLAG_PBO))
1627     {
1628         /* Whatever surface we have, make sure that there is memory allocated for the downloaded copy,
1629          * or a pbo to map
1630          */
1631         if(!This->resource.heapMemory) {
1632             This->resource.heapMemory = HeapAlloc(GetProcessHeap() ,0 , This->resource.size + RESOURCE_ALIGNMENT);
1633         }
1634         This->resource.allocatedMemory =
1635                 (BYTE *)(((ULONG_PTR) This->resource.heapMemory + (RESOURCE_ALIGNMENT - 1)) & ~(RESOURCE_ALIGNMENT - 1));
1636         if(This->Flags & SFLAG_INSYSMEM) {
1637             ERR("Surface without memory or pbo has SFLAG_INSYSMEM set!\n");
1638         }
1639     }
1640 }
1641
1642 static HRESULT WINAPI IWineD3DSurfaceImpl_LockRect(IWineD3DSurface *iface, WINED3DLOCKED_RECT* pLockedRect, CONST RECT* pRect, DWORD Flags) {
1643     IWineD3DSurfaceImpl *This = (IWineD3DSurfaceImpl *)iface;
1644     IWineD3DDeviceImpl *device = This->resource.device;
1645     const RECT *pass_rect = pRect;
1646
1647     TRACE("iface %p, locked_rect %p, rect %s, flags %#x.\n",
1648             iface, pLockedRect, wine_dbgstr_rect(pRect), Flags);
1649
1650     /* This is also done in the base class, but we have to verify this before loading any data from
1651      * gl into the sysmem copy. The PBO may be mapped, a different rectangle locked, the discard flag
1652      * may interfere, and all other bad things may happen
1653      */
1654     if (This->Flags & SFLAG_LOCKED) {
1655         WARN("Surface is already locked, returning D3DERR_INVALIDCALL\n");
1656         return WINED3DERR_INVALIDCALL;
1657     }
1658     This->Flags |= SFLAG_LOCKED;
1659
1660     if (!(This->Flags & SFLAG_LOCKABLE))
1661     {
1662         TRACE("Warning: trying to lock unlockable surf@%p\n", This);
1663     }
1664
1665     if (Flags & WINED3DLOCK_DISCARD) {
1666         /* Set SFLAG_INSYSMEM, so we'll never try to download the data from the texture. */
1667         TRACE("WINED3DLOCK_DISCARD flag passed, marking local copy as up to date\n");
1668         surface_prepare_system_memory(This); /* Makes sure memory is allocated */
1669         This->Flags |= SFLAG_INSYSMEM;
1670         goto lock_end;
1671     }
1672
1673     if (This->Flags & SFLAG_INSYSMEM) {
1674         TRACE("Local copy is up to date, not downloading data\n");
1675         surface_prepare_system_memory(This); /* Makes sure memory is allocated */
1676         goto lock_end;
1677     }
1678
1679     /* surface_load_location() does not check if the rectangle specifies
1680      * the full surface. Most callers don't need that, so do it here. */
1681     if (pRect && pRect->top == 0 && pRect->left == 0
1682             && pRect->right == This->currentDesc.Width
1683             && pRect->bottom == This->currentDesc.Height)
1684     {
1685         pass_rect = NULL;
1686     }
1687
1688     if (!(wined3d_settings.rendertargetlock_mode == RTL_DISABLE
1689             && ((This->container.type == WINED3D_CONTAINER_SWAPCHAIN) || This == device->render_targets[0])))
1690     {
1691         surface_load_location(This, SFLAG_INSYSMEM, pass_rect);
1692     }
1693
1694 lock_end:
1695     if (This->Flags & SFLAG_PBO)
1696     {
1697         const struct wined3d_gl_info *gl_info;
1698         struct wined3d_context *context;
1699
1700         context = context_acquire(device, NULL);
1701         gl_info = context->gl_info;
1702
1703         ENTER_GL();
1704         GL_EXTCALL(glBindBufferARB(GL_PIXEL_UNPACK_BUFFER_ARB, This->pbo));
1705         checkGLcall("glBindBufferARB");
1706
1707         /* This shouldn't happen but could occur if some other function didn't handle the PBO properly */
1708         if(This->resource.allocatedMemory) {
1709             ERR("The surface already has PBO memory allocated!\n");
1710         }
1711
1712         This->resource.allocatedMemory = GL_EXTCALL(glMapBufferARB(GL_PIXEL_UNPACK_BUFFER_ARB, GL_READ_WRITE_ARB));
1713         checkGLcall("glMapBufferARB");
1714
1715         /* Make sure the pbo isn't set anymore in order not to break non-pbo calls */
1716         GL_EXTCALL(glBindBufferARB(GL_PIXEL_UNPACK_BUFFER_ARB, 0));
1717         checkGLcall("glBindBufferARB");
1718
1719         LEAVE_GL();
1720         context_release(context);
1721     }
1722
1723     if (Flags & (WINED3DLOCK_NO_DIRTY_UPDATE | WINED3DLOCK_READONLY)) {
1724         /* Don't dirtify */
1725     }
1726     else
1727     {
1728         surface_add_dirty_rect(This, pRect);
1729
1730         if (This->container.type == WINED3D_CONTAINER_TEXTURE)
1731         {
1732             TRACE("Making container dirty.\n");
1733             IWineD3DBaseTexture_SetDirty((IWineD3DBaseTexture *)This->container.u.texture, TRUE);
1734         }
1735         else
1736         {
1737             TRACE("Surface is standalone, no need to dirty the container\n");
1738         }
1739     }
1740
1741     return IWineD3DBaseSurfaceImpl_LockRect(iface, pLockedRect, pRect, Flags);
1742 }
1743
1744 static void flush_to_framebuffer_drawpixels(IWineD3DSurfaceImpl *This, GLenum fmt, GLenum type, UINT bpp, const BYTE *mem) {
1745     GLint  prev_store;
1746     GLint  prev_rasterpos[4];
1747     GLint skipBytes = 0;
1748     UINT pitch = IWineD3DSurface_GetPitch((IWineD3DSurface *) This);    /* target is argb, 4 byte */
1749     IWineD3DDeviceImpl *device = This->resource.device;
1750     const struct wined3d_gl_info *gl_info;
1751     struct wined3d_context *context;
1752
1753     /* Activate the correct context for the render target */
1754     context = context_acquire(device, This);
1755     context_apply_blit_state(context, device);
1756     gl_info = context->gl_info;
1757
1758     ENTER_GL();
1759
1760     if (!surface_is_offscreen(This))
1761     {
1762         GLenum buffer = surface_get_gl_buffer(This);
1763         TRACE("Unlocking %#x buffer.\n", buffer);
1764         context_set_draw_buffer(context, buffer);
1765     }
1766     else
1767     {
1768         /* Primary offscreen render target */
1769         TRACE("Offscreen render target.\n");
1770         context_set_draw_buffer(context, device->offscreenBuffer);
1771     }
1772
1773     glGetIntegerv(GL_PACK_SWAP_BYTES, &prev_store);
1774     checkGLcall("glGetIntegerv");
1775     glGetIntegerv(GL_CURRENT_RASTER_POSITION, &prev_rasterpos[0]);
1776     checkGLcall("glGetIntegerv");
1777     glPixelZoom(1.0f, -1.0f);
1778     checkGLcall("glPixelZoom");
1779
1780     /* If not fullscreen, we need to skip a number of bytes to find the next row of data */
1781     glGetIntegerv(GL_UNPACK_ROW_LENGTH, &skipBytes);
1782     glPixelStorei(GL_UNPACK_ROW_LENGTH, This->currentDesc.Width);
1783
1784     glRasterPos3i(This->lockedRect.left, This->lockedRect.top, 1);
1785     checkGLcall("glRasterPos3i");
1786
1787     /* Some drivers(radeon dri, others?) don't like exceptions during
1788      * glDrawPixels. If the surface is a DIB section, it might be in GDIMode
1789      * after ReleaseDC. Reading it will cause an exception, which x11drv will
1790      * catch to put the dib section in InSync mode, which leads to a crash
1791      * and a blocked x server on my radeon card.
1792      *
1793      * The following lines read the dib section so it is put in InSync mode
1794      * before glDrawPixels is called and the crash is prevented. There won't
1795      * be any interfering gdi accesses, because UnlockRect is called from
1796      * ReleaseDC, and the app won't use the dc any more afterwards.
1797      */
1798     if((This->Flags & SFLAG_DIBSECTION) && !(This->Flags & SFLAG_PBO)) {
1799         volatile BYTE read;
1800         read = This->resource.allocatedMemory[0];
1801     }
1802
1803     if(This->Flags & SFLAG_PBO) {
1804         GL_EXTCALL(glBindBufferARB(GL_PIXEL_UNPACK_BUFFER_ARB, This->pbo));
1805         checkGLcall("glBindBufferARB");
1806     }
1807
1808     /* When the surface is locked we only have to refresh the locked part else we need to update the whole image */
1809     if(This->Flags & SFLAG_LOCKED) {
1810         glDrawPixels(This->lockedRect.right - This->lockedRect.left,
1811                      (This->lockedRect.bottom - This->lockedRect.top)-1,
1812                      fmt, type,
1813                      mem + bpp * This->lockedRect.left + pitch * This->lockedRect.top);
1814         checkGLcall("glDrawPixels");
1815     } else {
1816         glDrawPixels(This->currentDesc.Width,
1817                      This->currentDesc.Height,
1818                      fmt, type, mem);
1819         checkGLcall("glDrawPixels");
1820     }
1821
1822     if(This->Flags & SFLAG_PBO) {
1823         GL_EXTCALL(glBindBufferARB(GL_PIXEL_UNPACK_BUFFER_ARB, 0));
1824         checkGLcall("glBindBufferARB");
1825     }
1826
1827     glPixelZoom(1.0f, 1.0f);
1828     checkGLcall("glPixelZoom");
1829
1830     glRasterPos3iv(&prev_rasterpos[0]);
1831     checkGLcall("glRasterPos3iv");
1832
1833     /* Reset to previous pack row length */
1834     glPixelStorei(GL_UNPACK_ROW_LENGTH, skipBytes);
1835     checkGLcall("glPixelStorei(GL_UNPACK_ROW_LENGTH)");
1836
1837     LEAVE_GL();
1838     context_release(context);
1839 }
1840
1841 static HRESULT WINAPI IWineD3DSurfaceImpl_UnlockRect(IWineD3DSurface *iface) {
1842     IWineD3DSurfaceImpl *This = (IWineD3DSurfaceImpl *)iface;
1843     IWineD3DDeviceImpl *device = This->resource.device;
1844     BOOL fullsurface;
1845
1846     if (!(This->Flags & SFLAG_LOCKED)) {
1847         WARN("trying to Unlock an unlocked surf@%p\n", This);
1848         return WINEDDERR_NOTLOCKED;
1849     }
1850
1851     if (This->Flags & SFLAG_PBO)
1852     {
1853         const struct wined3d_gl_info *gl_info;
1854         struct wined3d_context *context;
1855
1856         TRACE("Freeing PBO memory\n");
1857
1858         context = context_acquire(device, NULL);
1859         gl_info = context->gl_info;
1860
1861         ENTER_GL();
1862         GL_EXTCALL(glBindBufferARB(GL_PIXEL_UNPACK_BUFFER_ARB, This->pbo));
1863         GL_EXTCALL(glUnmapBufferARB(GL_PIXEL_UNPACK_BUFFER_ARB));
1864         GL_EXTCALL(glBindBufferARB(GL_PIXEL_UNPACK_BUFFER_ARB, 0));
1865         checkGLcall("glUnmapBufferARB");
1866         LEAVE_GL();
1867         context_release(context);
1868
1869         This->resource.allocatedMemory = NULL;
1870     }
1871
1872     TRACE("(%p) : dirtyfied(%d)\n", This, This->Flags & (SFLAG_INDRAWABLE | SFLAG_INTEXTURE) ? 0 : 1);
1873
1874     if (This->Flags & (SFLAG_INDRAWABLE | SFLAG_INTEXTURE)) {
1875         TRACE("(%p) : Not Dirtified so nothing to do, return now\n", This);
1876         goto unlock_end;
1877     }
1878
1879     if (This->container.type == WINED3D_CONTAINER_SWAPCHAIN
1880             || (device->render_targets && This == device->render_targets[0]))
1881     {
1882         if(wined3d_settings.rendertargetlock_mode == RTL_DISABLE) {
1883             static BOOL warned = FALSE;
1884             if(!warned) {
1885                 ERR("The application tries to write to the render target, but render target locking is disabled\n");
1886                 warned = TRUE;
1887             }
1888             goto unlock_end;
1889         }
1890
1891         if(This->dirtyRect.left   == 0 &&
1892            This->dirtyRect.top    == 0 &&
1893            This->dirtyRect.right  == This->currentDesc.Width &&
1894            This->dirtyRect.bottom == This->currentDesc.Height) {
1895             fullsurface = TRUE;
1896         } else {
1897             /* TODO: Proper partial rectangle tracking */
1898             fullsurface = FALSE;
1899             This->Flags |= SFLAG_INSYSMEM;
1900         }
1901
1902         switch(wined3d_settings.rendertargetlock_mode) {
1903             case RTL_READTEX:
1904                 surface_load_location(This, SFLAG_INTEXTURE, NULL /* partial texture loading not supported yet */);
1905                 /* drop through */
1906
1907             case RTL_READDRAW:
1908                 surface_load_location(This, SFLAG_INDRAWABLE, fullsurface ? NULL : &This->dirtyRect);
1909                 break;
1910         }
1911
1912         if(!fullsurface) {
1913             /* Partial rectangle tracking is not commonly implemented, it is only done for render targets. Overwrite
1914              * the flags to bring them back into a sane state. INSYSMEM was set before to tell LoadLocation where
1915              * to read the rectangle from. Indrawable is set because all modifications from the partial sysmem copy
1916              * are written back to the drawable, thus the surface is merged again in the drawable. The sysmem copy is
1917              * not fully up to date because only a subrectangle was read in LockRect.
1918              */
1919             This->Flags &= ~SFLAG_INSYSMEM;
1920             This->Flags |= SFLAG_INDRAWABLE;
1921         }
1922
1923         This->dirtyRect.left   = This->currentDesc.Width;
1924         This->dirtyRect.top    = This->currentDesc.Height;
1925         This->dirtyRect.right  = 0;
1926         This->dirtyRect.bottom = 0;
1927     }
1928     else if (This == device->depth_stencil)
1929     {
1930         FIXME("Depth Stencil buffer locking is not implemented\n");
1931     } else {
1932         /* The rest should be a normal texture */
1933         /* Check if the texture is bound, if yes dirtify the sampler to force a re-upload of the texture
1934          * Can't load the texture here because PreLoad may destroy and recreate the gl texture, so sampler
1935          * states need resetting
1936          */
1937         if (This->container.type == WINED3D_CONTAINER_TEXTURE)
1938         {
1939             IWineD3DBaseTextureImpl *texture = This->container.u.texture;
1940             if (texture->baseTexture.bindCount)
1941                 IWineD3DDeviceImpl_MarkStateDirty(device, STATE_SAMPLER(texture->baseTexture.sampler));
1942         }
1943     }
1944
1945     unlock_end:
1946     This->Flags &= ~SFLAG_LOCKED;
1947     memset(&This->lockedRect, 0, sizeof(RECT));
1948
1949     /* Overlays have to be redrawn manually after changes with the GL implementation */
1950     if(This->overlay_dest) {
1951         IWineD3DSurface_DrawOverlay(iface);
1952     }
1953     return WINED3D_OK;
1954 }
1955
1956 static void surface_release_client_storage(IWineD3DSurfaceImpl *surface)
1957 {
1958     struct wined3d_context *context;
1959
1960     context = context_acquire(surface->resource.device, NULL);
1961
1962     ENTER_GL();
1963     glPixelStorei(GL_UNPACK_CLIENT_STORAGE_APPLE, GL_FALSE);
1964     if (surface->texture_name)
1965     {
1966         surface_bind_and_dirtify(surface, FALSE);
1967         glTexImage2D(surface->texture_target, surface->texture_level,
1968                 GL_RGB, 1, 1, 0, GL_RGB, GL_UNSIGNED_BYTE, NULL);
1969     }
1970     if (surface->texture_name_srgb)
1971     {
1972         surface_bind_and_dirtify(surface, TRUE);
1973         glTexImage2D(surface->texture_target, surface->texture_level,
1974                 GL_RGB, 1, 1, 0, GL_RGB, GL_UNSIGNED_BYTE, NULL);
1975     }
1976     glPixelStorei(GL_UNPACK_CLIENT_STORAGE_APPLE, GL_TRUE);
1977
1978     LEAVE_GL();
1979     context_release(context);
1980
1981     surface_modify_location(surface, SFLAG_INSRGBTEX, FALSE);
1982     surface_modify_location(surface, SFLAG_INTEXTURE, FALSE);
1983     surface_force_reload(surface);
1984 }
1985
1986 static HRESULT WINAPI IWineD3DSurfaceImpl_GetDC(IWineD3DSurface *iface, HDC *pHDC)
1987 {
1988     IWineD3DSurfaceImpl *This = (IWineD3DSurfaceImpl *)iface;
1989     WINED3DLOCKED_RECT lock;
1990     HRESULT hr;
1991     RGBQUAD col[256];
1992
1993     TRACE("(%p)->(%p)\n",This,pHDC);
1994
1995     if(This->Flags & SFLAG_USERPTR) {
1996         ERR("Not supported on surfaces with an application-provided surfaces\n");
1997         return WINEDDERR_NODC;
1998     }
1999
2000     /* Give more detailed info for ddraw */
2001     if (This->Flags & SFLAG_DCINUSE)
2002         return WINEDDERR_DCALREADYCREATED;
2003
2004     /* Can't GetDC if the surface is locked */
2005     if (This->Flags & SFLAG_LOCKED)
2006         return WINED3DERR_INVALIDCALL;
2007
2008     memset(&lock, 0, sizeof(lock)); /* To be sure */
2009
2010     /* Create a DIB section if there isn't a hdc yet */
2011     if (!This->hDC)
2012     {
2013         if (This->Flags & SFLAG_CLIENT)
2014         {
2015             surface_load_location(This, SFLAG_INSYSMEM, NULL);
2016             surface_release_client_storage(This);
2017         }
2018         hr = IWineD3DBaseSurfaceImpl_CreateDIBSection(iface);
2019         if(FAILED(hr)) return WINED3DERR_INVALIDCALL;
2020
2021         /* Use the dib section from now on if we are not using a PBO */
2022         if(!(This->Flags & SFLAG_PBO))
2023             This->resource.allocatedMemory = This->dib.bitmap_data;
2024     }
2025
2026     /* Lock the surface */
2027     hr = IWineD3DSurface_LockRect(iface,
2028                                   &lock,
2029                                   NULL,
2030                                   0);
2031
2032     if(This->Flags & SFLAG_PBO) {
2033         /* Sync the DIB with the PBO. This can't be done earlier because LockRect activates the allocatedMemory */
2034         memcpy(This->dib.bitmap_data, This->resource.allocatedMemory, This->dib.bitmap_size);
2035     }
2036
2037     if(FAILED(hr)) {
2038         ERR("IWineD3DSurface_LockRect failed with hr = %08x\n", hr);
2039         /* keep the dib section */
2040         return hr;
2041     }
2042
2043     if (This->resource.format_desc->format == WINED3DFMT_P8_UINT
2044             || This->resource.format_desc->format == WINED3DFMT_P8_UINT_A8_UNORM)
2045     {
2046         /* GetDC on palettized formats is unsupported in D3D9, and the method is missing in
2047             D3D8, so this should only be used for DX <=7 surfaces (with non-device palettes) */
2048         unsigned int n;
2049         const PALETTEENTRY *pal = NULL;
2050
2051         if(This->palette) {
2052             pal = This->palette->palents;
2053         } else {
2054             IWineD3DSurfaceImpl *dds_primary;
2055             IWineD3DSwapChainImpl *swapchain;
2056             swapchain = (IWineD3DSwapChainImpl *)This->resource.device->swapchains[0];
2057             dds_primary = swapchain->front_buffer;
2058             if (dds_primary && dds_primary->palette)
2059                 pal = dds_primary->palette->palents;
2060         }
2061
2062         if (pal) {
2063             for (n=0; n<256; n++) {
2064                 col[n].rgbRed   = pal[n].peRed;
2065                 col[n].rgbGreen = pal[n].peGreen;
2066                 col[n].rgbBlue  = pal[n].peBlue;
2067                 col[n].rgbReserved = 0;
2068             }
2069             SetDIBColorTable(This->hDC, 0, 256, col);
2070         }
2071     }
2072
2073     *pHDC = This->hDC;
2074     TRACE("returning %p\n",*pHDC);
2075     This->Flags |= SFLAG_DCINUSE;
2076
2077     return WINED3D_OK;
2078 }
2079
2080 static HRESULT WINAPI IWineD3DSurfaceImpl_ReleaseDC(IWineD3DSurface *iface, HDC hDC)
2081 {
2082     IWineD3DSurfaceImpl *This = (IWineD3DSurfaceImpl *)iface;
2083
2084     TRACE("(%p)->(%p)\n",This,hDC);
2085
2086     if (!(This->Flags & SFLAG_DCINUSE))
2087         return WINEDDERR_NODC;
2088
2089     if (This->hDC !=hDC) {
2090         WARN("Application tries to release an invalid DC(%p), surface dc is %p\n", hDC, This->hDC);
2091         return WINEDDERR_NODC;
2092     }
2093
2094     if((This->Flags & SFLAG_PBO) && This->resource.allocatedMemory) {
2095         /* Copy the contents of the DIB over to the PBO */
2096         memcpy(This->resource.allocatedMemory, This->dib.bitmap_data, This->dib.bitmap_size);
2097     }
2098
2099     /* we locked first, so unlock now */
2100     IWineD3DSurface_UnlockRect(iface);
2101
2102     This->Flags &= ~SFLAG_DCINUSE;
2103
2104     return WINED3D_OK;
2105 }
2106
2107 /* ******************************************************
2108    IWineD3DSurface Internal (No mapping to directx api) parts follow
2109    ****************************************************** */
2110
2111 HRESULT d3dfmt_get_conv(IWineD3DSurfaceImpl *This, BOOL need_alpha_ck, BOOL use_texturing, struct wined3d_format_desc *desc, CONVERT_TYPES *convert)
2112 {
2113     BOOL colorkey_active = need_alpha_ck && (This->CKeyFlags & WINEDDSD_CKSRCBLT);
2114     IWineD3DDeviceImpl *device = This->resource.device;
2115     const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
2116     BOOL blit_supported = FALSE;
2117
2118     /* Copy the default values from the surface. Below we might perform fixups */
2119     /* TODO: get rid of color keying desc fixups by using e.g. a table. */
2120     *desc = *This->resource.format_desc;
2121     *convert = NO_CONVERSION;
2122
2123     /* Ok, now look if we have to do any conversion */
2124     switch(This->resource.format_desc->format)
2125     {
2126         case WINED3DFMT_P8_UINT:
2127             /* ****************
2128                 Paletted Texture
2129                 **************** */
2130
2131             /* Below the call to blit_supported is disabled for Wine 1.2 because the function isn't operating correctly yet.
2132              * At the moment 8-bit blits are handled in software and if certain GL extensions are around, surface conversion
2133              * is performed at upload time. The blit_supported call recognizes it as a destination fixup. This type of upload 'fixup'
2134              * and 8-bit to 8-bit blits need to be handled by the blit_shader.
2135              * TODO: get rid of this #if 0
2136              */
2137 #if 0
2138             blit_supported = device->blitter->blit_supported(&device->adapter->gl_info, BLIT_OP_BLIT,
2139                                                              &rect, This->resource.usage, This->resource.pool,
2140                                                              This->resource.format_desc, &rect, This->resource.usage,
2141                                                              This->resource.pool, This->resource.format_desc);
2142 #endif
2143             blit_supported = gl_info->supported[EXT_PALETTED_TEXTURE] || gl_info->supported[ARB_FRAGMENT_PROGRAM];
2144
2145             /* Use conversion when the blit_shader backend supports it. It only supports this in case of
2146              * texturing. Further also use conversion in case of color keying.
2147              * Paletted textures can be emulated using shaders but only do that for 2D purposes e.g. situations
2148              * in which the main render target uses p8. Some games like GTA Vice City use P8 for texturing which
2149              * conflicts with this.
2150              */
2151             if (!((blit_supported && device->render_targets && This == device->render_targets[0]))
2152                     || colorkey_active || !use_texturing)
2153             {
2154                 desc->glFormat = GL_RGBA;
2155                 desc->glInternal = GL_RGBA;
2156                 desc->glType = GL_UNSIGNED_BYTE;
2157                 desc->conv_byte_count = 4;
2158                 if(colorkey_active) {
2159                     *convert = CONVERT_PALETTED_CK;
2160                 } else {
2161                     *convert = CONVERT_PALETTED;
2162                 }
2163             }
2164             break;
2165
2166         case WINED3DFMT_B2G3R3_UNORM:
2167             /* **********************
2168                 GL_UNSIGNED_BYTE_3_3_2
2169                 ********************** */
2170             if (colorkey_active) {
2171                 /* This texture format will never be used.. So do not care about color keying
2172                     up until the point in time it will be needed :-) */
2173                 FIXME(" ColorKeying not supported in the RGB 332 format !\n");
2174             }
2175             break;
2176
2177         case WINED3DFMT_B5G6R5_UNORM:
2178             if (colorkey_active) {
2179                 *convert = CONVERT_CK_565;
2180                 desc->glFormat = GL_RGBA;
2181                 desc->glInternal = GL_RGB5_A1;
2182                 desc->glType = GL_UNSIGNED_SHORT_5_5_5_1;
2183                 desc->conv_byte_count = 2;
2184             }
2185             break;
2186
2187         case WINED3DFMT_B5G5R5X1_UNORM:
2188             if (colorkey_active) {
2189                 *convert = CONVERT_CK_5551;
2190                 desc->glFormat = GL_BGRA;
2191                 desc->glInternal = GL_RGB5_A1;
2192                 desc->glType = GL_UNSIGNED_SHORT_1_5_5_5_REV;
2193                 desc->conv_byte_count = 2;
2194             }
2195             break;
2196
2197         case WINED3DFMT_B8G8R8_UNORM:
2198             if (colorkey_active) {
2199                 *convert = CONVERT_CK_RGB24;
2200                 desc->glFormat = GL_RGBA;
2201                 desc->glInternal = GL_RGBA8;
2202                 desc->glType = GL_UNSIGNED_INT_8_8_8_8;
2203                 desc->conv_byte_count = 4;
2204             }
2205             break;
2206
2207         case WINED3DFMT_B8G8R8X8_UNORM:
2208             if (colorkey_active) {
2209                 *convert = CONVERT_RGB32_888;
2210                 desc->glFormat = GL_RGBA;
2211                 desc->glInternal = GL_RGBA8;
2212                 desc->glType = GL_UNSIGNED_INT_8_8_8_8;
2213                 desc->conv_byte_count = 4;
2214             }
2215             break;
2216
2217         default:
2218             break;
2219     }
2220
2221     return WINED3D_OK;
2222 }
2223
2224 void d3dfmt_p8_init_palette(IWineD3DSurfaceImpl *This, BYTE table[256][4], BOOL colorkey)
2225 {
2226     IWineD3DDeviceImpl *device = This->resource.device;
2227     IWineD3DPaletteImpl *pal = This->palette;
2228     BOOL index_in_alpha = FALSE;
2229     unsigned int i;
2230
2231     /* Old games like StarCraft, C&C, Red Alert and others use P8 render targets.
2232      * Reading back the RGB output each lockrect (each frame as they lock the whole screen)
2233      * is slow. Further RGB->P8 conversion is not possible because palettes can have
2234      * duplicate entries. Store the color key in the unused alpha component to speed the
2235      * download up and to make conversion unneeded. */
2236     index_in_alpha = primary_render_target_is_p8(device);
2237
2238     if (!pal)
2239     {
2240         UINT dxVersion = ((IWineD3DImpl *)device->wined3d)->dxVersion;
2241
2242         /* In DirectDraw the palette is a property of the surface, there are no such things as device palettes. */
2243         if (dxVersion <= 7)
2244         {
2245             ERR("This code should never get entered for DirectDraw!, expect problems\n");
2246             if (index_in_alpha)
2247             {
2248                 /* Guarantees that memory representation remains correct after sysmem<->texture transfers even if
2249                  * there's no palette at this time. */
2250                 for (i = 0; i < 256; i++) table[i][3] = i;
2251             }
2252         }
2253         else
2254         {
2255             /* Direct3D >= 8 palette usage style: P8 textures use device palettes, palette entry format is A8R8G8B8,
2256              * alpha is stored in peFlags and may be used by the app if D3DPTEXTURECAPS_ALPHAPALETTE device
2257              * capability flag is present (wine does advertise this capability) */
2258             for (i = 0; i < 256; ++i)
2259             {
2260                 table[i][0] = device->palettes[device->currentPalette][i].peRed;
2261                 table[i][1] = device->palettes[device->currentPalette][i].peGreen;
2262                 table[i][2] = device->palettes[device->currentPalette][i].peBlue;
2263                 table[i][3] = device->palettes[device->currentPalette][i].peFlags;
2264             }
2265         }
2266     }
2267     else
2268     {
2269         TRACE("Using surface palette %p\n", pal);
2270         /* Get the surface's palette */
2271         for (i = 0; i < 256; ++i)
2272         {
2273             table[i][0] = pal->palents[i].peRed;
2274             table[i][1] = pal->palents[i].peGreen;
2275             table[i][2] = pal->palents[i].peBlue;
2276
2277             /* When index_in_alpha is set the palette index is stored in the
2278              * alpha component. In case of a readback we can then read
2279              * GL_ALPHA. Color keying is handled in BltOverride using a
2280              * GL_ALPHA_TEST using GL_NOT_EQUAL. In case of index_in_alpha the
2281              * color key itself is passed to glAlphaFunc in other cases the
2282              * alpha component of pixels that should be masked away is set to 0. */
2283             if (index_in_alpha)
2284             {
2285                 table[i][3] = i;
2286             }
2287             else if (colorkey && (i >= This->SrcBltCKey.dwColorSpaceLowValue)
2288                     && (i <= This->SrcBltCKey.dwColorSpaceHighValue))
2289             {
2290                 table[i][3] = 0x00;
2291             }
2292             else if(pal->Flags & WINEDDPCAPS_ALPHA)
2293             {
2294                 table[i][3] = pal->palents[i].peFlags;
2295             }
2296             else
2297             {
2298                 table[i][3] = 0xFF;
2299             }
2300         }
2301     }
2302 }
2303
2304 static HRESULT d3dfmt_convert_surface(const BYTE *src, BYTE *dst, UINT pitch, UINT width,
2305         UINT height, UINT outpitch, CONVERT_TYPES convert, IWineD3DSurfaceImpl *This)
2306 {
2307     const BYTE *source;
2308     BYTE *dest;
2309     TRACE("(%p)->(%p),(%d,%d,%d,%d,%p)\n", src, dst, pitch, height, outpitch, convert,This);
2310
2311     switch (convert) {
2312         case NO_CONVERSION:
2313         {
2314             memcpy(dst, src, pitch * height);
2315             break;
2316         }
2317         case CONVERT_PALETTED:
2318         case CONVERT_PALETTED_CK:
2319         {
2320             IWineD3DPaletteImpl* pal = This->palette;
2321             BYTE table[256][4];
2322             unsigned int x, y;
2323
2324             if( pal == NULL) {
2325                 /* TODO: If we are a sublevel, try to get the palette from level 0 */
2326             }
2327
2328             d3dfmt_p8_init_palette(This, table, (convert == CONVERT_PALETTED_CK));
2329
2330             for (y = 0; y < height; y++)
2331             {
2332                 source = src + pitch * y;
2333                 dest = dst + outpitch * y;
2334                 /* This is an 1 bpp format, using the width here is fine */
2335                 for (x = 0; x < width; x++) {
2336                     BYTE color = *source++;
2337                     *dest++ = table[color][0];
2338                     *dest++ = table[color][1];
2339                     *dest++ = table[color][2];
2340                     *dest++ = table[color][3];
2341                 }
2342             }
2343         }
2344         break;
2345
2346         case CONVERT_CK_565:
2347         {
2348             /* Converting the 565 format in 5551 packed to emulate color-keying.
2349
2350               Note : in all these conversion, it would be best to average the averaging
2351                       pixels to get the color of the pixel that will be color-keyed to
2352                       prevent 'color bleeding'. This will be done later on if ever it is
2353                       too visible.
2354
2355               Note2: Nvidia documents say that their driver does not support alpha + color keying
2356                      on the same surface and disables color keying in such a case
2357             */
2358             unsigned int x, y;
2359             const WORD *Source;
2360             WORD *Dest;
2361
2362             TRACE("Color keyed 565\n");
2363
2364             for (y = 0; y < height; y++) {
2365                 Source = (const WORD *)(src + y * pitch);
2366                 Dest = (WORD *) (dst + y * outpitch);
2367                 for (x = 0; x < width; x++ ) {
2368                     WORD color = *Source++;
2369                     *Dest = ((color & 0xFFC0) | ((color & 0x1F) << 1));
2370                     if ((color < This->SrcBltCKey.dwColorSpaceLowValue) ||
2371                         (color > This->SrcBltCKey.dwColorSpaceHighValue)) {
2372                         *Dest |= 0x0001;
2373                     }
2374                     Dest++;
2375                 }
2376             }
2377         }
2378         break;
2379
2380         case CONVERT_CK_5551:
2381         {
2382             /* Converting X1R5G5B5 format to R5G5B5A1 to emulate color-keying. */
2383             unsigned int x, y;
2384             const WORD *Source;
2385             WORD *Dest;
2386             TRACE("Color keyed 5551\n");
2387             for (y = 0; y < height; y++) {
2388                 Source = (const WORD *)(src + y * pitch);
2389                 Dest = (WORD *) (dst + y * outpitch);
2390                 for (x = 0; x < width; x++ ) {
2391                     WORD color = *Source++;
2392                     *Dest = color;
2393                     if ((color < This->SrcBltCKey.dwColorSpaceLowValue) ||
2394                         (color > This->SrcBltCKey.dwColorSpaceHighValue)) {
2395                         *Dest |= (1 << 15);
2396                     }
2397                     else {
2398                         *Dest &= ~(1 << 15);
2399                     }
2400                     Dest++;
2401                 }
2402             }
2403         }
2404         break;
2405
2406         case CONVERT_CK_RGB24:
2407         {
2408             /* Converting R8G8B8 format to R8G8B8A8 with color-keying. */
2409             unsigned int x, y;
2410             for (y = 0; y < height; y++)
2411             {
2412                 source = src + pitch * y;
2413                 dest = dst + outpitch * y;
2414                 for (x = 0; x < width; x++) {
2415                     DWORD color = ((DWORD)source[0] << 16) + ((DWORD)source[1] << 8) + (DWORD)source[2] ;
2416                     DWORD dstcolor = color << 8;
2417                     if ((color < This->SrcBltCKey.dwColorSpaceLowValue) ||
2418                         (color > This->SrcBltCKey.dwColorSpaceHighValue)) {
2419                         dstcolor |= 0xff;
2420                     }
2421                     *(DWORD*)dest = dstcolor;
2422                     source += 3;
2423                     dest += 4;
2424                 }
2425             }
2426         }
2427         break;
2428
2429         case CONVERT_RGB32_888:
2430         {
2431             /* Converting X8R8G8B8 format to R8G8B8A8 with color-keying. */
2432             unsigned int x, y;
2433             for (y = 0; y < height; y++)
2434             {
2435                 source = src + pitch * y;
2436                 dest = dst + outpitch * y;
2437                 for (x = 0; x < width; x++) {
2438                     DWORD color = 0xffffff & *(const DWORD*)source;
2439                     DWORD dstcolor = color << 8;
2440                     if ((color < This->SrcBltCKey.dwColorSpaceLowValue) ||
2441                         (color > This->SrcBltCKey.dwColorSpaceHighValue)) {
2442                         dstcolor |= 0xff;
2443                     }
2444                     *(DWORD*)dest = dstcolor;
2445                     source += 4;
2446                     dest += 4;
2447                 }
2448             }
2449         }
2450         break;
2451
2452         default:
2453             ERR("Unsupported conversion type %#x.\n", convert);
2454     }
2455     return WINED3D_OK;
2456 }
2457
2458 BOOL palette9_changed(IWineD3DSurfaceImpl *This)
2459 {
2460     IWineD3DDeviceImpl *device = This->resource.device;
2461
2462     if (This->palette || (This->resource.format_desc->format != WINED3DFMT_P8_UINT
2463             && This->resource.format_desc->format != WINED3DFMT_P8_UINT_A8_UNORM))
2464     {
2465         /* If a ddraw-style palette is attached assume no d3d9 palette change.
2466          * Also the palette isn't interesting if the surface format isn't P8 or A8P8
2467          */
2468         return FALSE;
2469     }
2470
2471     if (This->palette9)
2472     {
2473         if (!memcmp(This->palette9, device->palettes[device->currentPalette], sizeof(PALETTEENTRY) * 256))
2474         {
2475             return FALSE;
2476         }
2477     } else {
2478         This->palette9 = HeapAlloc(GetProcessHeap(), 0, sizeof(PALETTEENTRY) * 256);
2479     }
2480     memcpy(This->palette9, device->palettes[device->currentPalette], sizeof(PALETTEENTRY) * 256);
2481     return TRUE;
2482 }
2483
2484 static HRESULT WINAPI IWineD3DSurfaceImpl_LoadTexture(IWineD3DSurface *iface, BOOL srgb_mode) {
2485     IWineD3DSurfaceImpl *This = (IWineD3DSurfaceImpl *)iface;
2486     DWORD flag = srgb_mode ? SFLAG_INSRGBTEX : SFLAG_INTEXTURE;
2487
2488     TRACE("iface %p, srgb %#x.\n", iface, srgb_mode);
2489
2490     if (!(This->Flags & flag)) {
2491         TRACE("Reloading because surface is dirty\n");
2492     } else if(/* Reload: gl texture has ck, now no ckey is set OR */
2493               ((This->Flags & SFLAG_GLCKEY) && (!(This->CKeyFlags & WINEDDSD_CKSRCBLT))) ||
2494               /* Reload: vice versa  OR */
2495               ((!(This->Flags & SFLAG_GLCKEY)) && (This->CKeyFlags & WINEDDSD_CKSRCBLT)) ||
2496               /* Also reload: Color key is active AND the color key has changed */
2497               ((This->CKeyFlags & WINEDDSD_CKSRCBLT) && (
2498                 (This->glCKey.dwColorSpaceLowValue != This->SrcBltCKey.dwColorSpaceLowValue) ||
2499                 (This->glCKey.dwColorSpaceHighValue != This->SrcBltCKey.dwColorSpaceHighValue)))) {
2500         TRACE("Reloading because of color keying\n");
2501         /* To perform the color key conversion we need a sysmem copy of
2502          * the surface. Make sure we have it
2503          */
2504
2505         surface_load_location(This, SFLAG_INSYSMEM, NULL);
2506         /* Make sure the texture is reloaded because of the color key change, this kills performance though :( */
2507         /* TODO: This is not necessarily needed with hw palettized texture support */
2508         surface_modify_location(This, SFLAG_INSYSMEM, TRUE);
2509     } else {
2510         TRACE("surface is already in texture\n");
2511         return WINED3D_OK;
2512     }
2513
2514     /* Resources are placed in system RAM and do not need to be recreated when a device is lost.
2515      *  These resources are not bound by device size or format restrictions. Because of this,
2516      *  these resources cannot be accessed by the Direct3D device nor set as textures or render targets.
2517      *  However, these resources can always be created, locked, and copied.
2518      */
2519     if (This->resource.pool == WINED3DPOOL_SCRATCH )
2520     {
2521         FIXME("(%p) Operation not supported for scratch textures\n",This);
2522         return WINED3DERR_INVALIDCALL;
2523     }
2524
2525     surface_load_location(This, flag, NULL /* no partial locking for textures yet */);
2526
2527     if (!(This->Flags & SFLAG_DONOTFREE)) {
2528         HeapFree(GetProcessHeap(), 0, This->resource.heapMemory);
2529         This->resource.allocatedMemory = NULL;
2530         This->resource.heapMemory = NULL;
2531         surface_modify_location(This, SFLAG_INSYSMEM, FALSE);
2532     }
2533
2534     return WINED3D_OK;
2535 }
2536
2537 /* Context activation is done by the caller. */
2538 static void WINAPI IWineD3DSurfaceImpl_BindTexture(IWineD3DSurface *iface, BOOL srgb)
2539 {
2540     IWineD3DSurfaceImpl *This = (IWineD3DSurfaceImpl *)iface;
2541
2542     TRACE("iface %p, srgb %#x.\n", iface, srgb);
2543
2544     if (This->container.type == WINED3D_CONTAINER_TEXTURE)
2545     {
2546         TRACE("Passing to container.\n");
2547         IWineD3DBaseTexture_BindTexture((IWineD3DBaseTexture *)This->container.u.texture, srgb);
2548     }
2549     else
2550     {
2551         GLuint *name;
2552
2553         TRACE("(%p) : Binding surface\n", This);
2554
2555         name = srgb ? &This->texture_name_srgb : &This->texture_name;
2556
2557         ENTER_GL();
2558
2559         if (!This->texture_level)
2560         {
2561             if (!*name) {
2562                 glGenTextures(1, name);
2563                 checkGLcall("glGenTextures");
2564                 TRACE("Surface %p given name %d\n", This, *name);
2565
2566                 glBindTexture(This->texture_target, *name);
2567                 checkGLcall("glBindTexture");
2568                 glTexParameteri(This->texture_target, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
2569                 checkGLcall("glTexParameteri(dimension, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE)");
2570                 glTexParameteri(This->texture_target, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
2571                 checkGLcall("glTexParameteri(dimension, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE)");
2572                 glTexParameteri(This->texture_target, GL_TEXTURE_WRAP_R, GL_CLAMP_TO_EDGE);
2573                 checkGLcall("glTexParameteri(dimension, GL_TEXTURE_WRAP_R, GL_CLAMP_TO_EDGE)");
2574                 glTexParameteri(This->texture_target, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
2575                 checkGLcall("glTexParameteri(dimension, GL_TEXTURE_MIN_FILTER, GL_NEAREST)");
2576                 glTexParameteri(This->texture_target, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
2577                 checkGLcall("glTexParameteri(dimension, GL_TEXTURE_MAG_FILTER, GL_NEAREST)");
2578             }
2579             /* This is where we should be reducing the amount of GLMemoryUsed */
2580         } else if (*name) {
2581             /* Mipmap surfaces should have a base texture container */
2582             ERR("Mipmap surface has a glTexture bound to it!\n");
2583         }
2584
2585         glBindTexture(This->texture_target, *name);
2586         checkGLcall("glBindTexture");
2587
2588         LEAVE_GL();
2589     }
2590 }
2591
2592 static HRESULT WINAPI IWineD3DSurfaceImpl_SetFormat(IWineD3DSurface *iface, enum wined3d_format_id format)
2593 {
2594     IWineD3DSurfaceImpl *This = (IWineD3DSurfaceImpl *)iface;
2595     HRESULT hr;
2596
2597     TRACE("(%p) : Calling base function first\n", This);
2598     hr = IWineD3DBaseSurfaceImpl_SetFormat(iface, format);
2599     if(SUCCEEDED(hr)) {
2600         This->Flags &= ~(SFLAG_ALLOCATED | SFLAG_SRGBALLOCATED);
2601         TRACE("(%p) : glFormat %d, glFormatInternal %d, glType %d\n", This, This->resource.format_desc->glFormat,
2602                 This->resource.format_desc->glInternal, This->resource.format_desc->glType);
2603     }
2604     return hr;
2605 }
2606
2607 static HRESULT WINAPI IWineD3DSurfaceImpl_SetMem(IWineD3DSurface *iface, void *Mem) {
2608     IWineD3DSurfaceImpl *This = (IWineD3DSurfaceImpl *) iface;
2609
2610     if(This->Flags & (SFLAG_LOCKED | SFLAG_DCINUSE)) {
2611         WARN("Surface is locked or the HDC is in use\n");
2612         return WINED3DERR_INVALIDCALL;
2613     }
2614
2615     if(Mem && Mem != This->resource.allocatedMemory) {
2616         void *release = NULL;
2617
2618         /* Do I have to copy the old surface content? */
2619         if(This->Flags & SFLAG_DIBSECTION) {
2620                 /* Release the DC. No need to hold the critical section for the update
2621                  * Thread because this thread runs only on front buffers, but this method
2622                  * fails for render targets in the check above.
2623                  */
2624                 SelectObject(This->hDC, This->dib.holdbitmap);
2625                 DeleteDC(This->hDC);
2626                 /* Release the DIB section */
2627                 DeleteObject(This->dib.DIBsection);
2628                 This->dib.bitmap_data = NULL;
2629                 This->resource.allocatedMemory = NULL;
2630                 This->hDC = NULL;
2631                 This->Flags &= ~SFLAG_DIBSECTION;
2632         } else if(!(This->Flags & SFLAG_USERPTR)) {
2633             release = This->resource.heapMemory;
2634             This->resource.heapMemory = NULL;
2635         }
2636         This->resource.allocatedMemory = Mem;
2637         This->Flags |= SFLAG_USERPTR | SFLAG_INSYSMEM;
2638
2639         /* Now the surface memory is most up do date. Invalidate drawable and texture */
2640         surface_modify_location(This, SFLAG_INSYSMEM, TRUE);
2641
2642         /* For client textures opengl has to be notified */
2643         if (This->Flags & SFLAG_CLIENT)
2644             surface_release_client_storage(This);
2645
2646         /* Now free the old memory if any */
2647         HeapFree(GetProcessHeap(), 0, release);
2648     } else if(This->Flags & SFLAG_USERPTR) {
2649         /* LockRect and GetDC will re-create the dib section and allocated memory */
2650         This->resource.allocatedMemory = NULL;
2651         /* HeapMemory should be NULL already */
2652         if(This->resource.heapMemory != NULL) ERR("User pointer surface has heap memory allocated\n");
2653         This->Flags &= ~SFLAG_USERPTR;
2654
2655         if (This->Flags & SFLAG_CLIENT)
2656             surface_release_client_storage(This);
2657     }
2658     return WINED3D_OK;
2659 }
2660
2661 void flip_surface(IWineD3DSurfaceImpl *front, IWineD3DSurfaceImpl *back) {
2662
2663     /* Flip the surface contents */
2664     /* Flip the DC */
2665     {
2666         HDC tmp;
2667         tmp = front->hDC;
2668         front->hDC = back->hDC;
2669         back->hDC = tmp;
2670     }
2671
2672     /* Flip the DIBsection */
2673     {
2674         HBITMAP tmp;
2675         BOOL hasDib = front->Flags & SFLAG_DIBSECTION;
2676         tmp = front->dib.DIBsection;
2677         front->dib.DIBsection = back->dib.DIBsection;
2678         back->dib.DIBsection = tmp;
2679
2680         if(back->Flags & SFLAG_DIBSECTION) front->Flags |= SFLAG_DIBSECTION;
2681         else front->Flags &= ~SFLAG_DIBSECTION;
2682         if(hasDib) back->Flags |= SFLAG_DIBSECTION;
2683         else back->Flags &= ~SFLAG_DIBSECTION;
2684     }
2685
2686     /* Flip the surface data */
2687     {
2688         void* tmp;
2689
2690         tmp = front->dib.bitmap_data;
2691         front->dib.bitmap_data = back->dib.bitmap_data;
2692         back->dib.bitmap_data = tmp;
2693
2694         tmp = front->resource.allocatedMemory;
2695         front->resource.allocatedMemory = back->resource.allocatedMemory;
2696         back->resource.allocatedMemory = tmp;
2697
2698         tmp = front->resource.heapMemory;
2699         front->resource.heapMemory = back->resource.heapMemory;
2700         back->resource.heapMemory = tmp;
2701     }
2702
2703     /* Flip the PBO */
2704     {
2705         GLuint tmp_pbo = front->pbo;
2706         front->pbo = back->pbo;
2707         back->pbo = tmp_pbo;
2708     }
2709
2710     /* client_memory should not be different, but just in case */
2711     {
2712         BOOL tmp;
2713         tmp = front->dib.client_memory;
2714         front->dib.client_memory = back->dib.client_memory;
2715         back->dib.client_memory = tmp;
2716     }
2717
2718     /* Flip the opengl texture */
2719     {
2720         GLuint tmp;
2721
2722         tmp = back->texture_name;
2723         back->texture_name = front->texture_name;
2724         front->texture_name = tmp;
2725
2726         tmp = back->texture_name_srgb;
2727         back->texture_name_srgb = front->texture_name_srgb;
2728         front->texture_name_srgb = tmp;
2729     }
2730
2731     {
2732         DWORD tmp_flags = back->Flags;
2733         back->Flags = front->Flags;
2734         front->Flags = tmp_flags;
2735     }
2736 }
2737
2738 static HRESULT WINAPI IWineD3DSurfaceImpl_Flip(IWineD3DSurface *iface, IWineD3DSurface *override, DWORD Flags) {
2739     IWineD3DSurfaceImpl *This = (IWineD3DSurfaceImpl *)iface;
2740     IWineD3DSwapChainImpl *swapchain = NULL;
2741
2742     TRACE("(%p)->(%p,%x)\n", This, override, Flags);
2743
2744     /* Flipping is only supported on RenderTargets and overlays*/
2745     if( !(This->resource.usage & (WINED3DUSAGE_RENDERTARGET | WINED3DUSAGE_OVERLAY)) ) {
2746         WARN("Tried to flip a non-render target, non-overlay surface\n");
2747         return WINEDDERR_NOTFLIPPABLE;
2748     }
2749
2750     if(This->resource.usage & WINED3DUSAGE_OVERLAY) {
2751         flip_surface(This, (IWineD3DSurfaceImpl *) override);
2752
2753         /* Update the overlay if it is visible */
2754         if(This->overlay_dest) {
2755             return IWineD3DSurface_DrawOverlay((IWineD3DSurface *) This);
2756         } else {
2757             return WINED3D_OK;
2758         }
2759     }
2760
2761     if(override) {
2762         /* DDraw sets this for the X11 surfaces, so don't confuse the user
2763          * FIXME("(%p) Target override is not supported by now\n", This);
2764          * Additionally, it isn't really possible to support triple-buffering
2765          * properly on opengl at all
2766          */
2767     }
2768
2769     if (This->container.type != WINED3D_CONTAINER_SWAPCHAIN)
2770     {
2771         ERR("Flipped surface is not on a swapchain\n");
2772         return WINEDDERR_NOTFLIPPABLE;
2773     }
2774     swapchain = This->container.u.swapchain;
2775
2776     /* Just overwrite the swapchain presentation interval. This is ok because only ddraw apps can call Flip,
2777      * and only d3d8 and d3d9 apps specify the presentation interval
2778      */
2779     if((Flags & (WINEDDFLIP_NOVSYNC | WINEDDFLIP_INTERVAL2 | WINEDDFLIP_INTERVAL3 | WINEDDFLIP_INTERVAL4)) == 0) {
2780         /* Most common case first to avoid wasting time on all the other cases */
2781         swapchain->presentParms.PresentationInterval = WINED3DPRESENT_INTERVAL_ONE;
2782     } else if(Flags & WINEDDFLIP_NOVSYNC) {
2783         swapchain->presentParms.PresentationInterval = WINED3DPRESENT_INTERVAL_IMMEDIATE;
2784     } else if(Flags & WINEDDFLIP_INTERVAL2) {
2785         swapchain->presentParms.PresentationInterval = WINED3DPRESENT_INTERVAL_TWO;
2786     } else if(Flags & WINEDDFLIP_INTERVAL3) {
2787         swapchain->presentParms.PresentationInterval = WINED3DPRESENT_INTERVAL_THREE;
2788     } else {
2789         swapchain->presentParms.PresentationInterval = WINED3DPRESENT_INTERVAL_FOUR;
2790     }
2791
2792     /* Flipping a OpenGL surface -> Use WineD3DDevice::Present */
2793     return IWineD3DSwapChain_Present((IWineD3DSwapChain *)swapchain,
2794             NULL, NULL, swapchain->win_handle, NULL, 0);
2795 }
2796
2797 /* Does a direct frame buffer -> texture copy. Stretching is done
2798  * with single pixel copy calls
2799  */
2800 static void fb_copy_to_texture_direct(IWineD3DSurfaceImpl *dst_surface, IWineD3DSurfaceImpl *src_surface,
2801         const RECT *src_rect, const RECT *dst_rect_in, WINED3DTEXTUREFILTERTYPE Filter)
2802 {
2803     IWineD3DDeviceImpl *device = dst_surface->resource.device;
2804     float xrel, yrel;
2805     UINT row;
2806     struct wined3d_context *context;
2807     BOOL upsidedown = FALSE;
2808     RECT dst_rect = *dst_rect_in;
2809
2810     /* Make sure that the top pixel is always above the bottom pixel, and keep a separate upside down flag
2811      * glCopyTexSubImage is a bit picky about the parameters we pass to it
2812      */
2813     if(dst_rect.top > dst_rect.bottom) {
2814         UINT tmp = dst_rect.bottom;
2815         dst_rect.bottom = dst_rect.top;
2816         dst_rect.top = tmp;
2817         upsidedown = TRUE;
2818     }
2819
2820     context = context_acquire(device, src_surface);
2821     context_apply_blit_state(context, device);
2822     surface_internal_preload(dst_surface, SRGB_RGB);
2823     ENTER_GL();
2824
2825     /* Bind the target texture */
2826     glBindTexture(dst_surface->texture_target, dst_surface->texture_name);
2827     checkGLcall("glBindTexture");
2828     if (surface_is_offscreen(src_surface))
2829     {
2830         TRACE("Reading from an offscreen target\n");
2831         upsidedown = !upsidedown;
2832         glReadBuffer(device->offscreenBuffer);
2833     }
2834     else
2835     {
2836         glReadBuffer(surface_get_gl_buffer(src_surface));
2837     }
2838     checkGLcall("glReadBuffer");
2839
2840     xrel = (float) (src_rect->right - src_rect->left) / (float) (dst_rect.right - dst_rect.left);
2841     yrel = (float) (src_rect->bottom - src_rect->top) / (float) (dst_rect.bottom - dst_rect.top);
2842
2843     if ((xrel - 1.0f < -eps) || (xrel - 1.0f > eps))
2844     {
2845         FIXME("Doing a pixel by pixel copy from the framebuffer to a texture, expect major performance issues\n");
2846
2847         if(Filter != WINED3DTEXF_NONE && Filter != WINED3DTEXF_POINT) {
2848             ERR("Texture filtering not supported in direct blit\n");
2849         }
2850     }
2851     else if ((Filter != WINED3DTEXF_NONE && Filter != WINED3DTEXF_POINT)
2852             && ((yrel - 1.0f < -eps) || (yrel - 1.0f > eps)))
2853     {
2854         ERR("Texture filtering not supported in direct blit\n");
2855     }
2856
2857     if (upsidedown
2858             && !((xrel - 1.0f < -eps) || (xrel - 1.0f > eps))
2859             && !((yrel - 1.0f < -eps) || (yrel - 1.0f > eps)))
2860     {
2861         /* Upside down copy without stretching is nice, one glCopyTexSubImage call will do */
2862
2863         glCopyTexSubImage2D(dst_surface->texture_target, dst_surface->texture_level,
2864                 dst_rect.left /*xoffset */, dst_rect.top /* y offset */,
2865                 src_rect->left, src_surface->currentDesc.Height - src_rect->bottom,
2866                 dst_rect.right - dst_rect.left, dst_rect.bottom - dst_rect.top);
2867     } else {
2868         UINT yoffset = src_surface->currentDesc.Height - src_rect->top + dst_rect.top - 1;
2869         /* I have to process this row by row to swap the image,
2870          * otherwise it would be upside down, so stretching in y direction
2871          * doesn't cost extra time
2872          *
2873          * However, stretching in x direction can be avoided if not necessary
2874          */
2875         for(row = dst_rect.top; row < dst_rect.bottom; row++) {
2876             if ((xrel - 1.0f < -eps) || (xrel - 1.0f > eps))
2877             {
2878                 /* Well, that stuff works, but it's very slow.
2879                  * find a better way instead
2880                  */
2881                 UINT col;
2882
2883                 for (col = dst_rect.left; col < dst_rect.right; ++col)
2884                 {
2885                     glCopyTexSubImage2D(dst_surface->texture_target, dst_surface->texture_level,
2886                             dst_rect.left + col /* x offset */, row /* y offset */,
2887                             src_rect->left + col * xrel, yoffset - (int) (row * yrel), 1, 1);
2888                 }
2889             }
2890             else
2891             {
2892                 glCopyTexSubImage2D(dst_surface->texture_target, dst_surface->texture_level,
2893                         dst_rect.left /* x offset */, row /* y offset */,
2894                         src_rect->left, yoffset - (int) (row * yrel), dst_rect.right - dst_rect.left, 1);
2895             }
2896         }
2897     }
2898     checkGLcall("glCopyTexSubImage2D");
2899
2900     LEAVE_GL();
2901     context_release(context);
2902
2903     /* The texture is now most up to date - If the surface is a render target and has a drawable, this
2904      * path is never entered
2905      */
2906     surface_modify_location(dst_surface, SFLAG_INTEXTURE, TRUE);
2907 }
2908
2909 /* Uses the hardware to stretch and flip the image */
2910 static void fb_copy_to_texture_hwstretch(IWineD3DSurfaceImpl *dst_surface, IWineD3DSurfaceImpl *src_surface,
2911         const RECT *src_rect, const RECT *dst_rect_in, WINED3DTEXTUREFILTERTYPE Filter)
2912 {
2913     IWineD3DDeviceImpl *device = dst_surface->resource.device;
2914     GLuint src, backup = 0;
2915     IWineD3DSwapChainImpl *src_swapchain = NULL;
2916     float left, right, top, bottom; /* Texture coordinates */
2917     UINT fbwidth = src_surface->currentDesc.Width;
2918     UINT fbheight = src_surface->currentDesc.Height;
2919     struct wined3d_context *context;
2920     GLenum drawBuffer = GL_BACK;
2921     GLenum texture_target;
2922     BOOL noBackBufferBackup;
2923     BOOL src_offscreen;
2924     BOOL upsidedown = FALSE;
2925     RECT dst_rect = *dst_rect_in;
2926
2927     TRACE("Using hwstretch blit\n");
2928     /* Activate the Proper context for reading from the source surface, set it up for blitting */
2929     context = context_acquire(device, src_surface);
2930     context_apply_blit_state(context, device);
2931     surface_internal_preload(dst_surface, SRGB_RGB);
2932
2933     src_offscreen = surface_is_offscreen(src_surface);
2934     noBackBufferBackup = src_offscreen && wined3d_settings.offscreen_rendering_mode == ORM_FBO;
2935     if (!noBackBufferBackup && !src_surface->texture_name)
2936     {
2937         /* Get it a description */
2938         surface_internal_preload(src_surface, SRGB_RGB);
2939     }
2940     ENTER_GL();
2941
2942     /* Try to use an aux buffer for drawing the rectangle. This way it doesn't need restoring.
2943      * This way we don't have to wait for the 2nd readback to finish to leave this function.
2944      */
2945     if (context->aux_buffers >= 2)
2946     {
2947         /* Got more than one aux buffer? Use the 2nd aux buffer */
2948         drawBuffer = GL_AUX1;
2949     }
2950     else if ((!src_offscreen || device->offscreenBuffer == GL_BACK) && context->aux_buffers >= 1)
2951     {
2952         /* Only one aux buffer, but it isn't used (Onscreen rendering, or non-aux orm)? Use it! */
2953         drawBuffer = GL_AUX0;
2954     }
2955
2956     if(noBackBufferBackup) {
2957         glGenTextures(1, &backup);
2958         checkGLcall("glGenTextures");
2959         glBindTexture(GL_TEXTURE_2D, backup);
2960         checkGLcall("glBindTexture(GL_TEXTURE_2D, backup)");
2961         texture_target = GL_TEXTURE_2D;
2962     } else {
2963         /* Backup the back buffer and copy the source buffer into a texture to draw an upside down stretched quad. If
2964          * we are reading from the back buffer, the backup can be used as source texture
2965          */
2966         texture_target = src_surface->texture_target;
2967         glBindTexture(texture_target, src_surface->texture_name);
2968         checkGLcall("glBindTexture(texture_target, src_surface->texture_name)");
2969         glEnable(texture_target);
2970         checkGLcall("glEnable(texture_target)");
2971
2972         /* For now invalidate the texture copy of the back buffer. Drawable and sysmem copy are untouched */
2973         src_surface->Flags &= ~SFLAG_INTEXTURE;
2974     }
2975
2976     /* Make sure that the top pixel is always above the bottom pixel, and keep a separate upside down flag
2977      * glCopyTexSubImage is a bit picky about the parameters we pass to it
2978      */
2979     if(dst_rect.top > dst_rect.bottom) {
2980         UINT tmp = dst_rect.bottom;
2981         dst_rect.bottom = dst_rect.top;
2982         dst_rect.top = tmp;
2983         upsidedown = TRUE;
2984     }
2985
2986     if (src_offscreen)
2987     {
2988         TRACE("Reading from an offscreen target\n");
2989         upsidedown = !upsidedown;
2990         glReadBuffer(device->offscreenBuffer);
2991     }
2992     else
2993     {
2994         glReadBuffer(surface_get_gl_buffer(src_surface));
2995     }
2996
2997     /* TODO: Only back up the part that will be overwritten */
2998     glCopyTexSubImage2D(texture_target, 0,
2999                         0, 0 /* read offsets */,
3000                         0, 0,
3001                         fbwidth,
3002                         fbheight);
3003
3004     checkGLcall("glCopyTexSubImage2D");
3005
3006     /* No issue with overriding these - the sampler is dirty due to blit usage */
3007     glTexParameteri(texture_target, GL_TEXTURE_MAG_FILTER,
3008             wined3d_gl_mag_filter(magLookup, Filter));
3009     checkGLcall("glTexParameteri");
3010     glTexParameteri(texture_target, GL_TEXTURE_MIN_FILTER,
3011             wined3d_gl_min_mip_filter(minMipLookup, Filter, WINED3DTEXF_NONE));
3012     checkGLcall("glTexParameteri");
3013
3014     if (src_surface->container.type == WINED3D_CONTAINER_SWAPCHAIN)
3015         src_swapchain = src_surface->container.u.swapchain;
3016     if (!src_swapchain || src_surface == src_swapchain->back_buffers[0])
3017     {
3018         src = backup ? backup : src_surface->texture_name;
3019     }
3020     else
3021     {
3022         glReadBuffer(GL_FRONT);
3023         checkGLcall("glReadBuffer(GL_FRONT)");
3024
3025         glGenTextures(1, &src);
3026         checkGLcall("glGenTextures(1, &src)");
3027         glBindTexture(GL_TEXTURE_2D, src);
3028         checkGLcall("glBindTexture(GL_TEXTURE_2D, src)");
3029
3030         /* TODO: Only copy the part that will be read. Use src_rect->left, src_rect->bottom as origin, but with the width watch
3031          * out for power of 2 sizes
3032          */
3033         glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA, src_surface->pow2Width,
3034                 src_surface->pow2Height, 0, GL_RGBA, GL_UNSIGNED_BYTE, NULL);
3035         checkGLcall("glTexImage2D");
3036         glCopyTexSubImage2D(GL_TEXTURE_2D, 0,
3037                             0, 0 /* read offsets */,
3038                             0, 0,
3039                             fbwidth,
3040                             fbheight);
3041
3042         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
3043         checkGLcall("glTexParameteri");
3044         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
3045         checkGLcall("glTexParameteri");
3046
3047         glReadBuffer(GL_BACK);
3048         checkGLcall("glReadBuffer(GL_BACK)");
3049
3050         if(texture_target != GL_TEXTURE_2D) {
3051             glDisable(texture_target);
3052             glEnable(GL_TEXTURE_2D);
3053             texture_target = GL_TEXTURE_2D;
3054         }
3055     }
3056     checkGLcall("glEnd and previous");
3057
3058     left = src_rect->left;
3059     right = src_rect->right;
3060
3061     if (upsidedown)
3062     {
3063         top = src_surface->currentDesc.Height - src_rect->top;
3064         bottom = src_surface->currentDesc.Height - src_rect->bottom;
3065     }
3066     else
3067     {
3068         top = src_surface->currentDesc.Height - src_rect->bottom;
3069         bottom = src_surface->currentDesc.Height - src_rect->top;
3070     }
3071
3072     if (src_surface->Flags & SFLAG_NORMCOORD)
3073     {
3074         left /= src_surface->pow2Width;
3075         right /= src_surface->pow2Width;
3076         top /= src_surface->pow2Height;
3077         bottom /= src_surface->pow2Height;
3078     }
3079
3080     /* draw the source texture stretched and upside down. The correct surface is bound already */
3081     glTexParameteri(texture_target, GL_TEXTURE_WRAP_S, GL_CLAMP);
3082     glTexParameteri(texture_target, GL_TEXTURE_WRAP_T, GL_CLAMP);
3083
3084     context_set_draw_buffer(context, drawBuffer);
3085     glReadBuffer(drawBuffer);
3086
3087     glBegin(GL_QUADS);
3088         /* bottom left */
3089         glTexCoord2f(left, bottom);
3090         glVertex2i(0, fbheight);
3091
3092         /* top left */
3093         glTexCoord2f(left, top);
3094         glVertex2i(0, fbheight - dst_rect.bottom - dst_rect.top);
3095
3096         /* top right */
3097         glTexCoord2f(right, top);
3098         glVertex2i(dst_rect.right - dst_rect.left, fbheight - dst_rect.bottom - dst_rect.top);
3099
3100         /* bottom right */
3101         glTexCoord2f(right, bottom);
3102         glVertex2i(dst_rect.right - dst_rect.left, fbheight);
3103     glEnd();
3104     checkGLcall("glEnd and previous");
3105
3106     if (texture_target != dst_surface->texture_target)
3107     {
3108         glDisable(texture_target);
3109         glEnable(dst_surface->texture_target);
3110         texture_target = dst_surface->texture_target;
3111     }
3112
3113     /* Now read the stretched and upside down image into the destination texture */
3114     glBindTexture(texture_target, dst_surface->texture_name);
3115     checkGLcall("glBindTexture");
3116     glCopyTexSubImage2D(texture_target,
3117                         0,
3118                         dst_rect.left, dst_rect.top, /* xoffset, yoffset */
3119                         0, 0, /* We blitted the image to the origin */
3120                         dst_rect.right - dst_rect.left, dst_rect.bottom - dst_rect.top);
3121     checkGLcall("glCopyTexSubImage2D");
3122
3123     if(drawBuffer == GL_BACK) {
3124         /* Write the back buffer backup back */
3125         if(backup) {
3126             if(texture_target != GL_TEXTURE_2D) {
3127                 glDisable(texture_target);
3128                 glEnable(GL_TEXTURE_2D);
3129                 texture_target = GL_TEXTURE_2D;
3130             }
3131             glBindTexture(GL_TEXTURE_2D, backup);
3132             checkGLcall("glBindTexture(GL_TEXTURE_2D, backup)");
3133         }
3134         else
3135         {
3136             if (texture_target != src_surface->texture_target)
3137             {
3138                 glDisable(texture_target);
3139                 glEnable(src_surface->texture_target);
3140                 texture_target = src_surface->texture_target;
3141             }
3142             glBindTexture(src_surface->texture_target, src_surface->texture_name);
3143             checkGLcall("glBindTexture(src_surface->texture_target, src_surface->texture_name)");
3144         }
3145
3146         glBegin(GL_QUADS);
3147             /* top left */
3148             glTexCoord2f(0.0f, (float)fbheight / (float)src_surface->pow2Height);
3149             glVertex2i(0, 0);
3150
3151             /* bottom left */
3152             glTexCoord2f(0.0f, 0.0f);
3153             glVertex2i(0, fbheight);
3154
3155             /* bottom right */
3156             glTexCoord2f((float)fbwidth / (float)src_surface->pow2Width, 0.0f);
3157             glVertex2i(fbwidth, src_surface->currentDesc.Height);
3158
3159             /* top right */
3160             glTexCoord2f((float)fbwidth / (float)src_surface->pow2Width,
3161                     (float)fbheight / (float)src_surface->pow2Height);
3162             glVertex2i(fbwidth, 0);
3163         glEnd();
3164     }
3165     glDisable(texture_target);
3166     checkGLcall("glDisable(texture_target)");
3167
3168     /* Cleanup */
3169     if (src != src_surface->texture_name && src != backup)
3170     {
3171         glDeleteTextures(1, &src);
3172         checkGLcall("glDeleteTextures(1, &src)");
3173     }
3174     if(backup) {
3175         glDeleteTextures(1, &backup);
3176         checkGLcall("glDeleteTextures(1, &backup)");
3177     }
3178
3179     LEAVE_GL();
3180
3181     if (wined3d_settings.strict_draw_ordering) wglFlush(); /* Flush to ensure ordering across contexts. */
3182
3183     context_release(context);
3184
3185     /* The texture is now most up to date - If the surface is a render target and has a drawable, this
3186      * path is never entered
3187      */
3188     surface_modify_location(dst_surface, SFLAG_INTEXTURE, TRUE);
3189 }
3190
3191 /* Until the blit_shader is ready, define some prototypes here. */
3192 static BOOL fbo_blit_supported(const struct wined3d_gl_info *gl_info, enum blit_operation blit_op,
3193                                const RECT *src_rect, DWORD src_usage, WINED3DPOOL src_pool,
3194                                const struct wined3d_format_desc *src_format_desc,
3195                                const RECT *dst_rect, DWORD dst_usage, WINED3DPOOL dst_pool,
3196                                const struct wined3d_format_desc *dst_format_desc);
3197
3198 /* Front buffer coordinates are always full screen coordinates, but our GL
3199  * drawable is limited to the window's client area. The sysmem and texture
3200  * copies do have the full screen size. Note that GL has a bottom-left
3201  * origin, while D3D has a top-left origin. */
3202 void surface_translate_frontbuffer_coords(IWineD3DSurfaceImpl *surface, HWND window, RECT *rect)
3203 {
3204     POINT offset = {0, surface->currentDesc.Height};
3205     RECT windowsize;
3206
3207     GetClientRect(window, &windowsize);
3208     offset.y -= windowsize.bottom - windowsize.top;
3209     ScreenToClient(window, &offset);
3210     OffsetRect(rect, offset.x, offset.y);
3211 }
3212
3213 static BOOL surface_is_full_rect(IWineD3DSurfaceImpl *surface, const RECT *r)
3214 {
3215     if ((r->left && r->right) || abs(r->right - r->left) != surface->currentDesc.Width)
3216         return FALSE;
3217     if ((r->top && r->bottom) || abs(r->bottom - r->top) != surface->currentDesc.Height)
3218         return FALSE;
3219     return TRUE;
3220 }
3221
3222 /* blit between surface locations. onscreen on different swapchains is not supported.
3223  * depth / stencil is not supported. */
3224 static void surface_blt_fbo(IWineD3DDeviceImpl *device, const WINED3DTEXTUREFILTERTYPE filter,
3225         IWineD3DSurfaceImpl *src_surface, DWORD src_location, const RECT *src_rect_in,
3226         IWineD3DSurfaceImpl *dst_surface, DWORD dst_location, const RECT *dst_rect_in)
3227 {
3228     const struct wined3d_gl_info *gl_info;
3229     struct wined3d_context *context;
3230     RECT src_rect, dst_rect;
3231     GLenum gl_filter;
3232
3233     TRACE("device %p, filter %s,\n", device, debug_d3dtexturefiltertype(filter));
3234     TRACE("src_surface %p, src_location %s, src_rect %s,\n",
3235             src_surface, debug_surflocation(src_location), wine_dbgstr_rect(src_rect_in));
3236     TRACE("dst_surface %p, dst_location %s, dst_rect %s.\n",
3237             dst_surface, debug_surflocation(dst_location), wine_dbgstr_rect(dst_rect_in));
3238
3239     src_rect = *src_rect_in;
3240     dst_rect = *dst_rect_in;
3241
3242     switch (filter)
3243     {
3244         case WINED3DTEXF_LINEAR:
3245             gl_filter = GL_LINEAR;
3246             break;
3247
3248         default:
3249             FIXME("Unsupported filter mode %s (%#x).\n", debug_d3dtexturefiltertype(filter), filter);
3250         case WINED3DTEXF_NONE:
3251         case WINED3DTEXF_POINT:
3252             gl_filter = GL_NEAREST;
3253             break;
3254     }
3255
3256     if (src_location == SFLAG_INDRAWABLE && surface_is_offscreen(src_surface))
3257         src_location = SFLAG_INTEXTURE;
3258     if (dst_location == SFLAG_INDRAWABLE && surface_is_offscreen(dst_surface))
3259         dst_location = SFLAG_INTEXTURE;
3260
3261     /* Make sure the locations are up-to-date. Loading the destination
3262      * surface isn't required if the entire surface is overwritten. (And is
3263      * in fact harmful if we're being called by surface_load_location() with
3264      * the purpose of loading the destination surface.) */
3265     surface_load_location(src_surface, src_location, NULL);
3266     if (!surface_is_full_rect(dst_surface, &dst_rect))
3267         surface_load_location(dst_surface, dst_location, NULL);
3268
3269     if (src_location == SFLAG_INDRAWABLE) context = context_acquire(device, src_surface);
3270     else if (dst_location == SFLAG_INDRAWABLE) context = context_acquire(device, dst_surface);
3271     else context = context_acquire(device, NULL);
3272
3273     if (!context->valid)
3274     {
3275         context_release(context);
3276         WARN("Invalid context, skipping blit.\n");
3277         return;
3278     }
3279
3280     gl_info = context->gl_info;
3281
3282     if (src_location == SFLAG_INDRAWABLE)
3283     {
3284         GLenum buffer = surface_get_gl_buffer(src_surface);
3285
3286         TRACE("Source surface %p is onscreen.\n", src_surface);
3287
3288         if (buffer == GL_FRONT)
3289             surface_translate_frontbuffer_coords(src_surface, context->win_handle, &src_rect);
3290
3291         src_rect.top = src_surface->currentDesc.Height - src_rect.top;
3292         src_rect.bottom = src_surface->currentDesc.Height - src_rect.bottom;
3293
3294         ENTER_GL();
3295         context_bind_fbo(context, GL_READ_FRAMEBUFFER, NULL);
3296         glReadBuffer(buffer);
3297         checkGLcall("glReadBuffer()");
3298     }
3299     else
3300     {
3301         TRACE("Source surface %p is offscreen.\n", src_surface);
3302         ENTER_GL();
3303         context_apply_fbo_state_blit(context, GL_READ_FRAMEBUFFER, src_surface, NULL, src_location);
3304         glReadBuffer(GL_COLOR_ATTACHMENT0);
3305         checkGLcall("glReadBuffer()");
3306     }
3307     LEAVE_GL();
3308
3309     if (dst_location == SFLAG_INDRAWABLE)
3310     {
3311         GLenum buffer = surface_get_gl_buffer(dst_surface);
3312
3313         TRACE("Destination surface %p is onscreen.\n", dst_surface);
3314
3315         if (buffer == GL_FRONT)
3316             surface_translate_frontbuffer_coords(dst_surface, context->win_handle, &dst_rect);
3317
3318         dst_rect.top = dst_surface->currentDesc.Height - dst_rect.top;
3319         dst_rect.bottom = dst_surface->currentDesc.Height - dst_rect.bottom;
3320
3321         ENTER_GL();
3322         context_bind_fbo(context, GL_DRAW_FRAMEBUFFER, NULL);
3323         context_set_draw_buffer(context, buffer);
3324     }
3325     else
3326     {
3327         TRACE("Destination surface %p is offscreen.\n", dst_surface);
3328
3329         ENTER_GL();
3330         context_apply_fbo_state_blit(context, GL_DRAW_FRAMEBUFFER, dst_surface, NULL, dst_location);
3331         context_set_draw_buffer(context, GL_COLOR_ATTACHMENT0);
3332     }
3333
3334     glColorMask(GL_TRUE, GL_TRUE, GL_TRUE, GL_TRUE);
3335     IWineD3DDeviceImpl_MarkStateDirty(device, STATE_RENDER(WINED3DRS_COLORWRITEENABLE));
3336     IWineD3DDeviceImpl_MarkStateDirty(device, STATE_RENDER(WINED3DRS_COLORWRITEENABLE1));
3337     IWineD3DDeviceImpl_MarkStateDirty(device, STATE_RENDER(WINED3DRS_COLORWRITEENABLE2));
3338     IWineD3DDeviceImpl_MarkStateDirty(device, STATE_RENDER(WINED3DRS_COLORWRITEENABLE3));
3339
3340     glDisable(GL_SCISSOR_TEST);
3341     IWineD3DDeviceImpl_MarkStateDirty(device, STATE_RENDER(WINED3DRS_SCISSORTESTENABLE));
3342
3343     gl_info->fbo_ops.glBlitFramebuffer(src_rect.left, src_rect.top, src_rect.right, src_rect.bottom,
3344             dst_rect.left, dst_rect.top, dst_rect.right, dst_rect.bottom, GL_COLOR_BUFFER_BIT, gl_filter);
3345     checkGLcall("glBlitFramebuffer()");
3346
3347     LEAVE_GL();
3348
3349     if (wined3d_settings.strict_draw_ordering) wglFlush(); /* Flush to ensure ordering across contexts. */
3350
3351     context_release(context);
3352 }
3353
3354 /* Not called from the VTable */
3355 static HRESULT IWineD3DSurfaceImpl_BltOverride(IWineD3DSurfaceImpl *dst_surface, const RECT *DestRect,
3356         IWineD3DSurfaceImpl *src_surface, const RECT *SrcRect, DWORD Flags, const WINEDDBLTFX *DDBltFx,
3357         WINED3DTEXTUREFILTERTYPE Filter)
3358 {
3359     IWineD3DDeviceImpl *device = dst_surface->resource.device;
3360     IWineD3DSwapChainImpl *srcSwapchain = NULL, *dstSwapchain = NULL;
3361     RECT dst_rect, src_rect;
3362
3363     TRACE("dst_surface %p, dst_rect %s, src_surface %p, src_rect %s, flags %#x, blt_fx %p, filter %s.\n",
3364             dst_surface, wine_dbgstr_rect(DestRect), src_surface, wine_dbgstr_rect(SrcRect),
3365             Flags, DDBltFx, debug_d3dtexturefiltertype(Filter));
3366
3367     /* Get the swapchain. One of the surfaces has to be a primary surface */
3368     if (dst_surface->resource.pool == WINED3DPOOL_SYSTEMMEM)
3369     {
3370         WARN("Destination is in sysmem, rejecting gl blt\n");
3371         return WINED3DERR_INVALIDCALL;
3372     }
3373
3374     if (dst_surface->container.type == WINED3D_CONTAINER_SWAPCHAIN)
3375         dstSwapchain = dst_surface->container.u.swapchain;
3376
3377     if (src_surface)
3378     {
3379         if (src_surface->resource.pool == WINED3DPOOL_SYSTEMMEM)
3380         {
3381             WARN("Src is in sysmem, rejecting gl blt\n");
3382             return WINED3DERR_INVALIDCALL;
3383         }
3384
3385         if (src_surface->container.type == WINED3D_CONTAINER_SWAPCHAIN)
3386             srcSwapchain = src_surface->container.u.swapchain;
3387     }
3388
3389     /* Early sort out of cases where no render target is used */
3390     if (!dstSwapchain && !srcSwapchain
3391             && src_surface != device->render_targets[0]
3392             && dst_surface != device->render_targets[0])
3393     {
3394         TRACE("No surface is render target, not using hardware blit.\n");
3395         return WINED3DERR_INVALIDCALL;
3396     }
3397
3398     /* No destination color keying supported */
3399     if(Flags & (WINEDDBLT_KEYDEST | WINEDDBLT_KEYDESTOVERRIDE)) {
3400         /* Can we support that with glBlendFunc if blitting to the frame buffer? */
3401         TRACE("Destination color key not supported in accelerated Blit, falling back to software\n");
3402         return WINED3DERR_INVALIDCALL;
3403     }
3404
3405     surface_get_rect(dst_surface, DestRect, &dst_rect);
3406     if (src_surface) surface_get_rect(src_surface, SrcRect, &src_rect);
3407
3408     /* The only case where both surfaces on a swapchain are supported is a back buffer -> front buffer blit on the same swapchain */
3409     if (dstSwapchain && dstSwapchain == srcSwapchain && dstSwapchain->back_buffers
3410             && dst_surface == dstSwapchain->front_buffer
3411             && src_surface == dstSwapchain->back_buffers[0])
3412     {
3413         /* Half-Life does a Blt from the back buffer to the front buffer,
3414          * Full surface size, no flags... Use present instead
3415          *
3416          * This path will only be entered for d3d7 and ddraw apps, because d3d8/9 offer no way to blit TO the front buffer
3417          */
3418
3419         /* Check rects - IWineD3DDevice_Present doesn't handle them */
3420         while(1)
3421         {
3422             TRACE("Looking if a Present can be done...\n");
3423             /* Source Rectangle must be full surface */
3424             if (src_rect.left || src_rect.top
3425                     || src_rect.right != src_surface->currentDesc.Width
3426                     || src_rect.bottom != src_surface->currentDesc.Height)
3427             {
3428                 TRACE("No, Source rectangle doesn't match\n");
3429                 break;
3430             }
3431
3432             /* No stretching may occur */
3433             if(src_rect.right != dst_rect.right - dst_rect.left ||
3434                src_rect.bottom != dst_rect.bottom - dst_rect.top) {
3435                 TRACE("No, stretching is done\n");
3436                 break;
3437             }
3438
3439             /* Destination must be full surface or match the clipping rectangle */
3440             if (dst_surface->clipper && ((IWineD3DClipperImpl *)dst_surface->clipper)->hWnd)
3441             {
3442                 RECT cliprect;
3443                 POINT pos[2];
3444                 GetClientRect(((IWineD3DClipperImpl *)dst_surface->clipper)->hWnd, &cliprect);
3445                 pos[0].x = dst_rect.left;
3446                 pos[0].y = dst_rect.top;
3447                 pos[1].x = dst_rect.right;
3448                 pos[1].y = dst_rect.bottom;
3449                 MapWindowPoints(GetDesktopWindow(), ((IWineD3DClipperImpl *)dst_surface->clipper)->hWnd, pos, 2);
3450
3451                 if(pos[0].x != cliprect.left  || pos[0].y != cliprect.top   ||
3452                    pos[1].x != cliprect.right || pos[1].y != cliprect.bottom)
3453                 {
3454                     TRACE("No, dest rectangle doesn't match(clipper)\n");
3455                     TRACE("Clip rect at %s\n", wine_dbgstr_rect(&cliprect));
3456                     TRACE("Blt dest: %s\n", wine_dbgstr_rect(&dst_rect));
3457                     break;
3458                 }
3459             }
3460             else if (dst_rect.left || dst_rect.top
3461                     || dst_rect.right != dst_surface->currentDesc.Width
3462                     || dst_rect.bottom != dst_surface->currentDesc.Height)
3463             {
3464                 TRACE("No, dest rectangle doesn't match(surface size)\n");
3465                 break;
3466             }
3467
3468             TRACE("Yes\n");
3469
3470             /* These flags are unimportant for the flag check, remove them */
3471             if((Flags & ~(WINEDDBLT_DONOTWAIT | WINEDDBLT_WAIT)) == 0) {
3472                 WINED3DSWAPEFFECT orig_swap = dstSwapchain->presentParms.SwapEffect;
3473
3474                 /* The idea behind this is that a glReadPixels and a glDrawPixels call
3475                     * take very long, while a flip is fast.
3476                     * This applies to Half-Life, which does such Blts every time it finished
3477                     * a frame, and to Prince of Persia 3D, which uses this to draw at least the main
3478                     * menu. This is also used by all apps when they do windowed rendering
3479                     *
3480                     * The problem is that flipping is not really the same as copying. After a
3481                     * Blt the front buffer is a copy of the back buffer, and the back buffer is
3482                     * untouched. Therefore it's necessary to override the swap effect
3483                     * and to set it back after the flip.
3484                     *
3485                     * Windowed Direct3D < 7 apps do the same. The D3D7 sdk demos are nice
3486                     * testcases.
3487                     */
3488
3489                 dstSwapchain->presentParms.SwapEffect = WINED3DSWAPEFFECT_COPY;
3490                 dstSwapchain->presentParms.PresentationInterval = WINED3DPRESENT_INTERVAL_IMMEDIATE;
3491
3492                 TRACE("Full screen back buffer -> front buffer blt, performing a flip instead\n");
3493                 IWineD3DSwapChain_Present((IWineD3DSwapChain *)dstSwapchain,
3494                         NULL, NULL, dstSwapchain->win_handle, NULL, 0);
3495
3496                 dstSwapchain->presentParms.SwapEffect = orig_swap;
3497
3498                 return WINED3D_OK;
3499             }
3500             break;
3501         }
3502
3503         TRACE("Unsupported blit between buffers on the same swapchain\n");
3504         return WINED3DERR_INVALIDCALL;
3505     } else if(dstSwapchain && dstSwapchain == srcSwapchain) {
3506         FIXME("Implement hardware blit between two surfaces on the same swapchain\n");
3507         return WINED3DERR_INVALIDCALL;
3508     } else if(dstSwapchain && srcSwapchain) {
3509         FIXME("Implement hardware blit between two different swapchains\n");
3510         return WINED3DERR_INVALIDCALL;
3511     }
3512     else if (dstSwapchain)
3513     {
3514         /* Handled with regular texture -> swapchain blit */
3515         if (src_surface == device->render_targets[0])
3516             TRACE("Blit from active render target to a swapchain\n");
3517     }
3518     else if (srcSwapchain && dst_surface == device->render_targets[0])
3519     {
3520         FIXME("Implement blit from a swapchain to the active render target\n");
3521         return WINED3DERR_INVALIDCALL;
3522     }
3523
3524     if ((srcSwapchain || src_surface == device->render_targets[0]) && !dstSwapchain)
3525     {
3526         /* Blit from render target to texture */
3527         BOOL stretchx;
3528
3529         /* P8 read back is not implemented */
3530         if (src_surface->resource.format_desc->format == WINED3DFMT_P8_UINT
3531                 || dst_surface->resource.format_desc->format == WINED3DFMT_P8_UINT)
3532         {
3533             TRACE("P8 read back not supported by frame buffer to texture blit\n");
3534             return WINED3DERR_INVALIDCALL;
3535         }
3536
3537         if(Flags & (WINEDDBLT_KEYSRC | WINEDDBLT_KEYSRCOVERRIDE)) {
3538             TRACE("Color keying not supported by frame buffer to texture blit\n");
3539             return WINED3DERR_INVALIDCALL;
3540             /* Destination color key is checked above */
3541         }
3542
3543         if(dst_rect.right - dst_rect.left != src_rect.right - src_rect.left) {
3544             stretchx = TRUE;
3545         } else {
3546             stretchx = FALSE;
3547         }
3548
3549         /* Blt is a pretty powerful call, while glCopyTexSubImage2D is not. glCopyTexSubImage cannot
3550          * flip the image nor scale it.
3551          *
3552          * -> If the app asks for a unscaled, upside down copy, just perform one glCopyTexSubImage2D call
3553          * -> If the app wants a image width an unscaled width, copy it line per line
3554          * -> If the app wants a image that is scaled on the x axis, and the destination rectangle is smaller
3555          *    than the frame buffer, draw an upside down scaled image onto the fb, read it back and restore the
3556          *    back buffer. This is slower than reading line per line, thus not used for flipping
3557          * -> If the app wants a scaled image with a dest rect that is bigger than the fb, it has to be copied
3558          *    pixel by pixel
3559          *
3560          * If EXT_framebuffer_blit is supported that can be used instead. Note that EXT_framebuffer_blit implies
3561          * FBO support, so it doesn't really make sense to try and make it work with different offscreen rendering
3562          * backends.
3563          */
3564         if (fbo_blit_supported(&device->adapter->gl_info, BLIT_OP_BLIT,
3565                 &src_rect, src_surface->resource.usage, src_surface->resource.pool, src_surface->resource.format_desc,
3566                 &dst_rect, dst_surface->resource.usage, dst_surface->resource.pool, dst_surface->resource.format_desc))
3567         {
3568             surface_blt_fbo(device, Filter,
3569                     src_surface, SFLAG_INDRAWABLE, &src_rect,
3570                     dst_surface, SFLAG_INDRAWABLE, &dst_rect);
3571             surface_modify_location(dst_surface, SFLAG_INDRAWABLE, TRUE);
3572         }
3573         else if (!stretchx || dst_rect.right - dst_rect.left > src_surface->currentDesc.Width
3574                 || dst_rect.bottom - dst_rect.top > src_surface->currentDesc.Height)
3575         {
3576             TRACE("No stretching in x direction, using direct framebuffer -> texture copy\n");
3577             fb_copy_to_texture_direct(dst_surface, src_surface, &src_rect, &dst_rect, Filter);
3578         } else {
3579             TRACE("Using hardware stretching to flip / stretch the texture\n");
3580             fb_copy_to_texture_hwstretch(dst_surface, src_surface, &src_rect, &dst_rect, Filter);
3581         }
3582
3583         if (!(dst_surface->Flags & SFLAG_DONOTFREE))
3584         {
3585             HeapFree(GetProcessHeap(), 0, dst_surface->resource.heapMemory);
3586             dst_surface->resource.allocatedMemory = NULL;
3587             dst_surface->resource.heapMemory = NULL;
3588         }
3589         else
3590         {
3591             dst_surface->Flags &= ~SFLAG_INSYSMEM;
3592         }
3593
3594         return WINED3D_OK;
3595     }
3596     else if (src_surface)
3597     {
3598         /* Blit from offscreen surface to render target */
3599         DWORD oldCKeyFlags = src_surface->CKeyFlags;
3600         WINEDDCOLORKEY oldBltCKey = src_surface->SrcBltCKey;
3601         struct wined3d_context *context;
3602
3603         TRACE("Blt from surface %p to rendertarget %p\n", src_surface, dst_surface);
3604
3605         if (!(Flags & (WINEDDBLT_KEYSRC | WINEDDBLT_KEYSRCOVERRIDE))
3606                 && fbo_blit_supported(&device->adapter->gl_info, BLIT_OP_BLIT,
3607                         &src_rect, src_surface->resource.usage, src_surface->resource.pool,
3608                         src_surface->resource.format_desc,
3609                         &dst_rect, dst_surface->resource.usage, dst_surface->resource.pool,
3610                         dst_surface->resource.format_desc))
3611         {
3612             TRACE("Using surface_blt_fbo.\n");
3613             /* The source is always a texture, but never the currently active render target, and the texture
3614              * contents are never upside down. */
3615             surface_blt_fbo(device, Filter,
3616                     src_surface, SFLAG_INDRAWABLE, &src_rect,
3617                     dst_surface, SFLAG_INDRAWABLE, &dst_rect);
3618             surface_modify_location(dst_surface, SFLAG_INDRAWABLE, TRUE);
3619             return WINED3D_OK;
3620         }
3621
3622         if (!(Flags & (WINEDDBLT_KEYSRC | WINEDDBLT_KEYSRCOVERRIDE))
3623                 && arbfp_blit.blit_supported(&device->adapter->gl_info, BLIT_OP_BLIT,
3624                         &src_rect, src_surface->resource.usage, src_surface->resource.pool,
3625                         src_surface->resource.format_desc,
3626                         &dst_rect, dst_surface->resource.usage, dst_surface->resource.pool,
3627                         dst_surface->resource.format_desc))
3628         {
3629             return arbfp_blit_surface(device, src_surface, &src_rect, dst_surface, &dst_rect, BLIT_OP_BLIT, Filter);
3630         }
3631
3632         /* Color keying: Check if we have to do a color keyed blt,
3633          * and if not check if a color key is activated.
3634          *
3635          * Just modify the color keying parameters in the surface and restore them afterwards
3636          * The surface keeps track of the color key last used to load the opengl surface.
3637          * PreLoad will catch the change to the flags and color key and reload if necessary.
3638          */
3639         if(Flags & WINEDDBLT_KEYSRC) {
3640             /* Use color key from surface */
3641         } else if(Flags & WINEDDBLT_KEYSRCOVERRIDE) {
3642             /* Use color key from DDBltFx */
3643             src_surface->CKeyFlags |= WINEDDSD_CKSRCBLT;
3644             src_surface->SrcBltCKey = DDBltFx->ddckSrcColorkey;
3645         } else {
3646             /* Do not use color key */
3647             src_surface->CKeyFlags &= ~WINEDDSD_CKSRCBLT;
3648         }
3649
3650         /* Now load the surface */
3651         surface_internal_preload(src_surface, SRGB_RGB);
3652
3653         /* Activate the destination context, set it up for blitting */
3654         context = context_acquire(device, dst_surface);
3655         context_apply_blit_state(context, device);
3656
3657         if (dstSwapchain && dst_surface == dstSwapchain->front_buffer)
3658             surface_translate_frontbuffer_coords(dst_surface, context->win_handle, &dst_rect);
3659
3660         if (!device->blitter->blit_supported(&device->adapter->gl_info, BLIT_OP_BLIT,
3661                 &src_rect, src_surface->resource.usage, src_surface->resource.pool, src_surface->resource.format_desc,
3662                 &dst_rect, dst_surface->resource.usage, dst_surface->resource.pool, dst_surface->resource.format_desc))
3663         {
3664             FIXME("Unsupported blit operation falling back to software\n");
3665             return WINED3DERR_INVALIDCALL;
3666         }
3667
3668         device->blitter->set_shader((IWineD3DDevice *)device, src_surface);
3669
3670         ENTER_GL();
3671
3672         /* This is for color keying */
3673         if(Flags & (WINEDDBLT_KEYSRC | WINEDDBLT_KEYSRCOVERRIDE)) {
3674             glEnable(GL_ALPHA_TEST);
3675             checkGLcall("glEnable(GL_ALPHA_TEST)");
3676
3677             /* When the primary render target uses P8, the alpha component contains the palette index.
3678              * Which means that the colorkey is one of the palette entries. In other cases pixels that
3679              * should be masked away have alpha set to 0. */
3680             if (primary_render_target_is_p8(device))
3681                 glAlphaFunc(GL_NOTEQUAL, (float)src_surface->SrcBltCKey.dwColorSpaceLowValue / 256.0f);
3682             else
3683                 glAlphaFunc(GL_NOTEQUAL, 0.0f);
3684             checkGLcall("glAlphaFunc");
3685         } else {
3686             glDisable(GL_ALPHA_TEST);
3687             checkGLcall("glDisable(GL_ALPHA_TEST)");
3688         }
3689
3690         /* Draw a textured quad
3691          */
3692         draw_textured_quad(src_surface, &src_rect, &dst_rect, Filter);
3693
3694         if(Flags & (WINEDDBLT_KEYSRC | WINEDDBLT_KEYSRCOVERRIDE)) {
3695             glDisable(GL_ALPHA_TEST);
3696             checkGLcall("glDisable(GL_ALPHA_TEST)");
3697         }
3698
3699         /* Restore the color key parameters */
3700         src_surface->CKeyFlags = oldCKeyFlags;
3701         src_surface->SrcBltCKey = oldBltCKey;
3702
3703         LEAVE_GL();
3704
3705         /* Leave the opengl state valid for blitting */
3706         device->blitter->unset_shader((IWineD3DDevice *)device);
3707
3708         if (wined3d_settings.strict_draw_ordering || (dstSwapchain
3709                 && (dst_surface == dstSwapchain->front_buffer
3710                 || dstSwapchain->num_contexts > 1)))
3711             wglFlush(); /* Flush to ensure ordering across contexts. */
3712
3713         context_release(context);
3714
3715         /* TODO: If the surface is locked often, perform the Blt in software on the memory instead */
3716         /* The surface is now in the drawable. On onscreen surfaces or without fbos the texture
3717          * is outdated now
3718          */
3719         surface_modify_location(dst_surface, SFLAG_INDRAWABLE, TRUE);
3720
3721         return WINED3D_OK;
3722     }
3723     else
3724     {
3725         /* Source-Less Blit to render target */
3726         if (Flags & WINEDDBLT_COLORFILL)
3727         {
3728             WINED3DCOLORVALUE color;
3729
3730             TRACE("Colorfill\n");
3731
3732             /* The color as given in the Blt function is in the surface format. */
3733             if (!surface_convert_color_to_float(dst_surface, DDBltFx->u5.dwFillColor, &color))
3734                 return WINED3DERR_INVALIDCALL;
3735
3736             if (ffp_blit.blit_supported(&device->adapter->gl_info, BLIT_OP_COLOR_FILL,
3737                     NULL, 0, 0, NULL,
3738                     &dst_rect, dst_surface->resource.usage, dst_surface->resource.pool,
3739                     dst_surface->resource.format_desc))
3740             {
3741                 return ffp_blit.color_fill(device, dst_surface, &dst_rect, &color);
3742             }
3743             else if (cpu_blit.blit_supported(&device->adapter->gl_info, BLIT_OP_COLOR_FILL,
3744                     NULL, 0, 0, NULL,
3745                     &dst_rect, dst_surface->resource.usage, dst_surface->resource.pool,
3746                     dst_surface->resource.format_desc))
3747             {
3748                 return cpu_blit.color_fill(device, dst_surface, &dst_rect, &color);
3749             }
3750             return WINED3DERR_INVALIDCALL;
3751         }
3752     }
3753
3754     /* Default: Fall back to the generic blt. Not an error, a TRACE is enough */
3755     TRACE("Didn't find any usable render target setup for hw blit, falling back to software\n");
3756     return WINED3DERR_INVALIDCALL;
3757 }
3758
3759 static HRESULT IWineD3DSurfaceImpl_BltZ(IWineD3DSurfaceImpl *This, const RECT *DestRect,
3760         IWineD3DSurface *src_surface, const RECT *src_rect, DWORD Flags, const WINEDDBLTFX *DDBltFx)
3761 {
3762     IWineD3DDeviceImpl *device = This->resource.device;
3763     float depth;
3764
3765     if (Flags & WINEDDBLT_DEPTHFILL) {
3766         switch(This->resource.format_desc->format)
3767         {
3768             case WINED3DFMT_D16_UNORM:
3769                 depth = (float) DDBltFx->u5.dwFillDepth / (float) 0x0000ffff;
3770                 break;
3771             case WINED3DFMT_S1_UINT_D15_UNORM:
3772                 depth = (float) DDBltFx->u5.dwFillDepth / (float) 0x00007fff;
3773                 break;
3774             case WINED3DFMT_D24_UNORM_S8_UINT:
3775             case WINED3DFMT_X8D24_UNORM:
3776                 depth = (float) DDBltFx->u5.dwFillDepth / (float) 0x00ffffff;
3777                 break;
3778             case WINED3DFMT_D32_UNORM:
3779                 depth = (float) DDBltFx->u5.dwFillDepth / (float) 0xffffffff;
3780                 break;
3781             default:
3782                 depth = 0.0f;
3783                 ERR("Unexpected format for depth fill: %s\n", debug_d3dformat(This->resource.format_desc->format));
3784         }
3785
3786         return IWineD3DDevice_Clear((IWineD3DDevice *)device, DestRect ? 1 : 0, (const WINED3DRECT *)DestRect,
3787                 WINED3DCLEAR_ZBUFFER, 0x00000000, depth, 0x00000000);
3788     }
3789
3790     FIXME("(%p): Unsupp depthstencil blit\n", This);
3791     return WINED3DERR_INVALIDCALL;
3792 }
3793
3794 static HRESULT WINAPI IWineD3DSurfaceImpl_Blt(IWineD3DSurface *iface, const RECT *DestRect,
3795         IWineD3DSurface *src_surface, const RECT *SrcRect, DWORD Flags,
3796         const WINEDDBLTFX *DDBltFx, WINED3DTEXTUREFILTERTYPE Filter)
3797 {
3798     IWineD3DSurfaceImpl *This = (IWineD3DSurfaceImpl *)iface;
3799     IWineD3DSurfaceImpl *src = (IWineD3DSurfaceImpl *)src_surface;
3800     IWineD3DDeviceImpl *device = This->resource.device;
3801
3802     TRACE("iface %p, dst_rect %s, src_surface %p, src_rect %s, flags %#x, fx %p, filter %s.\n",
3803             iface, wine_dbgstr_rect(DestRect), src_surface, wine_dbgstr_rect(SrcRect),
3804             Flags, DDBltFx, debug_d3dtexturefiltertype(Filter));
3805     TRACE("Usage is %s.\n", debug_d3dusage(This->resource.usage));
3806
3807     if ((This->Flags & SFLAG_LOCKED) || (src && (src->Flags & SFLAG_LOCKED)))
3808     {
3809         WARN(" Surface is busy, returning DDERR_SURFACEBUSY\n");
3810         return WINEDDERR_SURFACEBUSY;
3811     }
3812
3813     /* Accessing the depth stencil is supposed to fail between a BeginScene and EndScene pair,
3814      * except depth blits, which seem to work
3815      */
3816     if (This == device->depth_stencil || (src && src == device->depth_stencil))
3817     {
3818         if (device->inScene && !(Flags & WINEDDBLT_DEPTHFILL))
3819         {
3820             TRACE("Attempt to access the depth stencil surface in a BeginScene / EndScene pair, returning WINED3DERR_INVALIDCALL\n");
3821             return WINED3DERR_INVALIDCALL;
3822         }
3823         else if (SUCCEEDED(IWineD3DSurfaceImpl_BltZ(This, DestRect, src_surface, SrcRect, Flags, DDBltFx)))
3824         {
3825             TRACE("Z Blit override handled the blit\n");
3826             return WINED3D_OK;
3827         }
3828     }
3829
3830     /* Special cases for RenderTargets */
3831     if ((This->resource.usage & WINED3DUSAGE_RENDERTARGET)
3832             || (src && (src->resource.usage & WINED3DUSAGE_RENDERTARGET)))
3833     {
3834         if (SUCCEEDED(IWineD3DSurfaceImpl_BltOverride(This, DestRect, src, SrcRect, Flags, DDBltFx, Filter)))
3835             return WINED3D_OK;
3836     }
3837
3838     /* For the rest call the X11 surface implementation.
3839      * For RenderTargets this should be implemented OpenGL accelerated in BltOverride,
3840      * other Blts are rather rare. */
3841     return IWineD3DBaseSurfaceImpl_Blt(iface, DestRect, src_surface, SrcRect, Flags, DDBltFx, Filter);
3842 }
3843
3844 static HRESULT WINAPI IWineD3DSurfaceImpl_BltFast(IWineD3DSurface *iface, DWORD dstx, DWORD dsty,
3845         IWineD3DSurface *src_surface, const RECT *rsrc, DWORD trans)
3846 {
3847     IWineD3DSurfaceImpl *This = (IWineD3DSurfaceImpl *)iface;
3848     IWineD3DSurfaceImpl *src = (IWineD3DSurfaceImpl *)src_surface;
3849     IWineD3DDeviceImpl *device = This->resource.device;
3850
3851     TRACE("iface %p, dst_x %u, dst_y %u, src_surface %p, src_rect %s, flags %#x.\n",
3852             iface, dstx, dsty, src_surface, wine_dbgstr_rect(rsrc), trans);
3853
3854     if ((This->Flags & SFLAG_LOCKED) || (src->Flags & SFLAG_LOCKED))
3855     {
3856         WARN(" Surface is busy, returning DDERR_SURFACEBUSY\n");
3857         return WINEDDERR_SURFACEBUSY;
3858     }
3859
3860     if (device->inScene && (This == device->depth_stencil || src == device->depth_stencil))
3861     {
3862         TRACE("Attempt to access the depth stencil surface in a BeginScene / EndScene pair, returning WINED3DERR_INVALIDCALL\n");
3863         return WINED3DERR_INVALIDCALL;
3864     }
3865
3866     /* Special cases for RenderTargets */
3867     if ((This->resource.usage & WINED3DUSAGE_RENDERTARGET)
3868             || (src->resource.usage & WINED3DUSAGE_RENDERTARGET))
3869     {
3870
3871         RECT SrcRect, DstRect;
3872         DWORD Flags=0;
3873
3874         surface_get_rect(src, rsrc, &SrcRect);
3875
3876         DstRect.left = dstx;
3877         DstRect.top=dsty;
3878         DstRect.right = dstx + SrcRect.right - SrcRect.left;
3879         DstRect.bottom = dsty + SrcRect.bottom - SrcRect.top;
3880
3881         /* Convert BltFast flags into Btl ones because it is called from SurfaceImpl_Blt as well */
3882         if(trans & WINEDDBLTFAST_SRCCOLORKEY)
3883             Flags |= WINEDDBLT_KEYSRC;
3884         if(trans & WINEDDBLTFAST_DESTCOLORKEY)
3885             Flags |= WINEDDBLT_KEYDEST;
3886         if(trans & WINEDDBLTFAST_WAIT)
3887             Flags |= WINEDDBLT_WAIT;
3888         if(trans & WINEDDBLTFAST_DONOTWAIT)
3889             Flags |= WINEDDBLT_DONOTWAIT;
3890
3891         if (SUCCEEDED(IWineD3DSurfaceImpl_BltOverride(This,
3892                 &DstRect, src, &SrcRect, Flags, NULL, WINED3DTEXF_POINT)))
3893             return WINED3D_OK;
3894     }
3895
3896     return IWineD3DBaseSurfaceImpl_BltFast(iface, dstx, dsty, src_surface, rsrc, trans);
3897 }
3898
3899 static HRESULT WINAPI IWineD3DSurfaceImpl_RealizePalette(IWineD3DSurface *iface)
3900 {
3901     IWineD3DSurfaceImpl *This = (IWineD3DSurfaceImpl *) iface;
3902     RGBQUAD col[256];
3903     IWineD3DPaletteImpl *pal = This->palette;
3904     unsigned int n;
3905     TRACE("(%p)\n", This);
3906
3907     if (!pal) return WINED3D_OK;
3908
3909     if (This->resource.format_desc->format == WINED3DFMT_P8_UINT
3910             || This->resource.format_desc->format == WINED3DFMT_P8_UINT_A8_UNORM)
3911     {
3912         if (This->resource.usage & WINED3DUSAGE_RENDERTARGET)
3913         {
3914             /* Make sure the texture is up to date. This call doesn't do
3915              * anything if the texture is already up to date. */
3916             surface_load_location(This, SFLAG_INTEXTURE, NULL);
3917
3918             /* We want to force a palette refresh, so mark the drawable as not being up to date */
3919             surface_modify_location(This, SFLAG_INDRAWABLE, FALSE);
3920         }
3921         else
3922         {
3923             if (!(This->Flags & SFLAG_INSYSMEM))
3924             {
3925                 TRACE("Palette changed with surface that does not have an up to date system memory copy.\n");
3926                 surface_load_location(This, SFLAG_INSYSMEM, NULL);
3927             }
3928             TRACE("Dirtifying surface\n");
3929             surface_modify_location(This, SFLAG_INSYSMEM, TRUE);
3930         }
3931     }
3932
3933     if(This->Flags & SFLAG_DIBSECTION) {
3934         TRACE("(%p): Updating the hdc's palette\n", This);
3935         for (n=0; n<256; n++) {
3936             col[n].rgbRed   = pal->palents[n].peRed;
3937             col[n].rgbGreen = pal->palents[n].peGreen;
3938             col[n].rgbBlue  = pal->palents[n].peBlue;
3939             col[n].rgbReserved = 0;
3940         }
3941         SetDIBColorTable(This->hDC, 0, 256, col);
3942     }
3943
3944     /* Propagate the changes to the drawable when we have a palette. */
3945     if (This->resource.usage & WINED3DUSAGE_RENDERTARGET)
3946         surface_load_location(This, SFLAG_INDRAWABLE, NULL);
3947
3948     return WINED3D_OK;
3949 }
3950
3951 static HRESULT WINAPI IWineD3DSurfaceImpl_PrivateSetup(IWineD3DSurface *iface) {
3952     /** Check against the maximum texture sizes supported by the video card **/
3953     IWineD3DSurfaceImpl *This = (IWineD3DSurfaceImpl *) iface;
3954     const struct wined3d_gl_info *gl_info = &This->resource.device->adapter->gl_info;
3955     unsigned int pow2Width, pow2Height;
3956
3957     This->texture_name = 0;
3958     This->texture_target = GL_TEXTURE_2D;
3959
3960     /* Non-power2 support */
3961     if (gl_info->supported[ARB_TEXTURE_NON_POWER_OF_TWO] || gl_info->supported[WINED3D_GL_NORMALIZED_TEXRECT])
3962     {
3963         pow2Width = This->currentDesc.Width;
3964         pow2Height = This->currentDesc.Height;
3965     }
3966     else
3967     {
3968         /* Find the nearest pow2 match */
3969         pow2Width = pow2Height = 1;
3970         while (pow2Width < This->currentDesc.Width) pow2Width <<= 1;
3971         while (pow2Height < This->currentDesc.Height) pow2Height <<= 1;
3972     }
3973     This->pow2Width  = pow2Width;
3974     This->pow2Height = pow2Height;
3975
3976     if (pow2Width > This->currentDesc.Width || pow2Height > This->currentDesc.Height) {
3977         /** TODO: add support for non power two compressed textures **/
3978         if (This->resource.format_desc->Flags & WINED3DFMT_FLAG_COMPRESSED)
3979         {
3980             FIXME("(%p) Compressed non-power-two textures are not supported w(%d) h(%d)\n",
3981                   This, This->currentDesc.Width, This->currentDesc.Height);
3982             return WINED3DERR_NOTAVAILABLE;
3983         }
3984     }
3985
3986     if(pow2Width != This->currentDesc.Width ||
3987        pow2Height != This->currentDesc.Height) {
3988         This->Flags |= SFLAG_NONPOW2;
3989     }
3990
3991     TRACE("%p\n", This);
3992     if ((This->pow2Width > gl_info->limits.texture_size || This->pow2Height > gl_info->limits.texture_size)
3993             && !(This->resource.usage & (WINED3DUSAGE_RENDERTARGET | WINED3DUSAGE_DEPTHSTENCIL)))
3994     {
3995         /* one of three options
3996         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)
3997         2: Set the texture to the maximum size (bad idea)
3998         3:    WARN and return WINED3DERR_NOTAVAILABLE;
3999         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.
4000         */
4001         if(This->resource.pool == WINED3DPOOL_DEFAULT || This->resource.pool == WINED3DPOOL_MANAGED)
4002         {
4003             WARN("(%p) Unable to allocate a surface which exceeds the maximum OpenGL texture size\n", This);
4004             return WINED3DERR_NOTAVAILABLE;
4005         }
4006
4007         /* We should never use this surface in combination with OpenGL! */
4008         TRACE("(%p) Creating an oversized surface: %ux%u\n", This, This->pow2Width, This->pow2Height);
4009     }
4010     else
4011     {
4012         /* Don't use ARB_TEXTURE_RECTANGLE in case the surface format is P8 and EXT_PALETTED_TEXTURE
4013            is used in combination with texture uploads (RTL_READTEX/RTL_TEXTEX). The reason is that EXT_PALETTED_TEXTURE
4014            doesn't work in combination with ARB_TEXTURE_RECTANGLE.
4015         */
4016         if (This->Flags & SFLAG_NONPOW2 && gl_info->supported[ARB_TEXTURE_RECTANGLE]
4017                 && !(This->resource.format_desc->format == WINED3DFMT_P8_UINT
4018                 && gl_info->supported[EXT_PALETTED_TEXTURE]
4019                 && wined3d_settings.rendertargetlock_mode == RTL_READTEX))
4020         {
4021             This->texture_target = GL_TEXTURE_RECTANGLE_ARB;
4022             This->pow2Width  = This->currentDesc.Width;
4023             This->pow2Height = This->currentDesc.Height;
4024             This->Flags &= ~(SFLAG_NONPOW2 | SFLAG_NORMCOORD);
4025         }
4026     }
4027
4028     switch (wined3d_settings.offscreen_rendering_mode)
4029     {
4030         case ORM_FBO:
4031             This->get_drawable_size = get_drawable_size_fbo;
4032             break;
4033
4034         case ORM_BACKBUFFER:
4035             This->get_drawable_size = get_drawable_size_backbuffer;
4036             break;
4037
4038         default:
4039             ERR("Unhandled offscreen rendering mode %#x.\n", wined3d_settings.offscreen_rendering_mode);
4040             return WINED3DERR_INVALIDCALL;
4041     }
4042
4043     This->Flags |= SFLAG_INSYSMEM;
4044
4045     return WINED3D_OK;
4046 }
4047
4048 /* GL locking is done by the caller */
4049 static void surface_depth_blt(IWineD3DSurfaceImpl *This, const struct wined3d_gl_info *gl_info,
4050         GLuint texture, GLsizei w, GLsizei h, GLenum target)
4051 {
4052     IWineD3DDeviceImpl *device = This->resource.device;
4053     GLint compare_mode = GL_NONE;
4054     struct blt_info info;
4055     GLint old_binding = 0;
4056
4057     glPushAttrib(GL_ENABLE_BIT | GL_DEPTH_BUFFER_BIT | GL_COLOR_BUFFER_BIT | GL_VIEWPORT_BIT);
4058
4059     glDisable(GL_CULL_FACE);
4060     glDisable(GL_BLEND);
4061     glDisable(GL_ALPHA_TEST);
4062     glDisable(GL_SCISSOR_TEST);
4063     glDisable(GL_STENCIL_TEST);
4064     glEnable(GL_DEPTH_TEST);
4065     glDepthFunc(GL_ALWAYS);
4066     glDepthMask(GL_TRUE);
4067     glColorMask(GL_FALSE, GL_FALSE, GL_FALSE, GL_FALSE);
4068     glViewport(0, 0, w, h);
4069
4070     surface_get_blt_info(target, NULL, w, h, &info);
4071     GL_EXTCALL(glActiveTextureARB(GL_TEXTURE0_ARB));
4072     glGetIntegerv(info.binding, &old_binding);
4073     glBindTexture(info.bind_target, texture);
4074     if (gl_info->supported[ARB_SHADOW])
4075     {
4076         glGetTexParameteriv(info.bind_target, GL_TEXTURE_COMPARE_MODE_ARB, &compare_mode);
4077         if (compare_mode != GL_NONE) glTexParameteri(info.bind_target, GL_TEXTURE_COMPARE_MODE_ARB, GL_NONE);
4078     }
4079
4080     device->shader_backend->shader_select_depth_blt((IWineD3DDevice *)device,
4081             info.tex_type, &This->ds_current_size);
4082
4083     glBegin(GL_TRIANGLE_STRIP);
4084     glTexCoord3fv(info.coords[0]);
4085     glVertex2f(-1.0f, -1.0f);
4086     glTexCoord3fv(info.coords[1]);
4087     glVertex2f(1.0f, -1.0f);
4088     glTexCoord3fv(info.coords[2]);
4089     glVertex2f(-1.0f, 1.0f);
4090     glTexCoord3fv(info.coords[3]);
4091     glVertex2f(1.0f, 1.0f);
4092     glEnd();
4093
4094     if (compare_mode != GL_NONE) glTexParameteri(info.bind_target, GL_TEXTURE_COMPARE_MODE_ARB, compare_mode);
4095     glBindTexture(info.bind_target, old_binding);
4096
4097     glPopAttrib();
4098
4099     device->shader_backend->shader_deselect_depth_blt((IWineD3DDevice *)device);
4100 }
4101
4102 void surface_modify_ds_location(IWineD3DSurfaceImpl *surface,
4103         DWORD location, UINT w, UINT h)
4104 {
4105     TRACE("surface %p, new location %#x, w %u, h %u.\n", surface, location, w, h);
4106
4107     if (location & ~SFLAG_DS_LOCATIONS)
4108         FIXME("Invalid location (%#x) specified.\n", location);
4109
4110     surface->ds_current_size.cx = w;
4111     surface->ds_current_size.cy = h;
4112     surface->Flags &= ~SFLAG_DS_LOCATIONS;
4113     surface->Flags |= location;
4114 }
4115
4116 /* Context activation is done by the caller. */
4117 void surface_load_ds_location(IWineD3DSurfaceImpl *surface, struct wined3d_context *context, DWORD location)
4118 {
4119     IWineD3DDeviceImpl *device = surface->resource.device;
4120     const struct wined3d_gl_info *gl_info = context->gl_info;
4121
4122     TRACE("surface %p, new location %#x.\n", surface, location);
4123
4124     /* TODO: Make this work for modes other than FBO */
4125     if (wined3d_settings.offscreen_rendering_mode != ORM_FBO) return;
4126
4127     if (!(surface->Flags & location))
4128     {
4129         surface->ds_current_size.cx = 0;
4130         surface->ds_current_size.cy = 0;
4131     }
4132
4133     if (surface->ds_current_size.cx == surface->currentDesc.Width
4134             && surface->ds_current_size.cy == surface->currentDesc.Height)
4135     {
4136         TRACE("Location (%#x) is already up to date.\n", location);
4137         return;
4138     }
4139
4140     if (surface->current_renderbuffer)
4141     {
4142         FIXME("Not supported with fixed up depth stencil.\n");
4143         return;
4144     }
4145
4146     if (!(surface->Flags & SFLAG_LOCATIONS))
4147     {
4148         FIXME("No up to date depth stencil location.\n");
4149         surface->Flags |= location;
4150         return;
4151     }
4152
4153     if (location == SFLAG_DS_OFFSCREEN)
4154     {
4155         GLint old_binding = 0;
4156         GLenum bind_target;
4157
4158         TRACE("Copying onscreen depth buffer to depth texture.\n");
4159
4160         ENTER_GL();
4161
4162         if (!device->depth_blt_texture)
4163         {
4164             glGenTextures(1, &device->depth_blt_texture);
4165         }
4166
4167         /* Note that we use depth_blt here as well, rather than glCopyTexImage2D
4168          * directly on the FBO texture. That's because we need to flip. */
4169         context_bind_fbo(context, GL_FRAMEBUFFER, NULL);
4170         if (surface->texture_target == GL_TEXTURE_RECTANGLE_ARB)
4171         {
4172             glGetIntegerv(GL_TEXTURE_BINDING_RECTANGLE_ARB, &old_binding);
4173             bind_target = GL_TEXTURE_RECTANGLE_ARB;
4174         }
4175         else
4176         {
4177             glGetIntegerv(GL_TEXTURE_BINDING_2D, &old_binding);
4178             bind_target = GL_TEXTURE_2D;
4179         }
4180         glBindTexture(bind_target, device->depth_blt_texture);
4181         glCopyTexImage2D(bind_target, surface->texture_level, surface->resource.format_desc->glInternal,
4182                 0, 0, surface->currentDesc.Width, surface->currentDesc.Height, 0);
4183         glTexParameteri(bind_target, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
4184         glTexParameteri(bind_target, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
4185         glTexParameteri(bind_target, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
4186         glTexParameteri(bind_target, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
4187         glTexParameteri(bind_target, GL_TEXTURE_WRAP_R, GL_CLAMP_TO_EDGE);
4188         glTexParameteri(bind_target, GL_DEPTH_TEXTURE_MODE_ARB, GL_LUMINANCE);
4189         glBindTexture(bind_target, old_binding);
4190
4191         /* Setup the destination */
4192         if (!device->depth_blt_rb)
4193         {
4194             gl_info->fbo_ops.glGenRenderbuffers(1, &device->depth_blt_rb);
4195             checkGLcall("glGenRenderbuffersEXT");
4196         }
4197         if (device->depth_blt_rb_w != surface->currentDesc.Width
4198                 || device->depth_blt_rb_h != surface->currentDesc.Height)
4199         {
4200             gl_info->fbo_ops.glBindRenderbuffer(GL_RENDERBUFFER, device->depth_blt_rb);
4201             checkGLcall("glBindRenderbufferEXT");
4202             gl_info->fbo_ops.glRenderbufferStorage(GL_RENDERBUFFER, GL_RGBA8,
4203                     surface->currentDesc.Width, surface->currentDesc.Height);
4204             checkGLcall("glRenderbufferStorageEXT");
4205             device->depth_blt_rb_w = surface->currentDesc.Width;
4206             device->depth_blt_rb_h = surface->currentDesc.Height;
4207         }
4208
4209         context_bind_fbo(context, GL_FRAMEBUFFER, &context->dst_fbo);
4210         gl_info->fbo_ops.glFramebufferRenderbuffer(GL_FRAMEBUFFER,
4211                 GL_COLOR_ATTACHMENT0, GL_RENDERBUFFER, device->depth_blt_rb);
4212         checkGLcall("glFramebufferRenderbufferEXT");
4213         context_attach_depth_stencil_fbo(context, GL_FRAMEBUFFER, surface, FALSE);
4214
4215         /* Do the actual blit */
4216         surface_depth_blt(surface, gl_info, device->depth_blt_texture,
4217                 surface->currentDesc.Width, surface->currentDesc.Height, bind_target);
4218         checkGLcall("depth_blt");
4219
4220         if (context->current_fbo) context_bind_fbo(context, GL_FRAMEBUFFER, &context->current_fbo->id);
4221         else context_bind_fbo(context, GL_FRAMEBUFFER, NULL);
4222
4223         LEAVE_GL();
4224
4225         if (wined3d_settings.strict_draw_ordering) wglFlush(); /* Flush to ensure ordering across contexts. */
4226     }
4227     else if (location == SFLAG_DS_ONSCREEN)
4228     {
4229         TRACE("Copying depth texture to onscreen depth buffer.\n");
4230
4231         ENTER_GL();
4232
4233         context_bind_fbo(context, GL_FRAMEBUFFER, NULL);
4234         surface_depth_blt(surface, gl_info, surface->texture_name,
4235                 surface->currentDesc.Width, surface->currentDesc.Height, surface->texture_target);
4236         checkGLcall("depth_blt");
4237
4238         if (context->current_fbo) context_bind_fbo(context, GL_FRAMEBUFFER, &context->current_fbo->id);
4239
4240         LEAVE_GL();
4241
4242         if (wined3d_settings.strict_draw_ordering) wglFlush(); /* Flush to ensure ordering across contexts. */
4243     }
4244     else
4245     {
4246         ERR("Invalid location (%#x) specified.\n", location);
4247     }
4248
4249     surface->Flags |= location;
4250     surface->ds_current_size.cx = surface->currentDesc.Width;
4251     surface->ds_current_size.cy = surface->currentDesc.Height;
4252 }
4253
4254 void surface_modify_location(IWineD3DSurfaceImpl *surface, DWORD flag, BOOL persistent)
4255 {
4256     IWineD3DSurfaceImpl *overlay;
4257
4258     TRACE("surface %p, location %s, persistent %#x.\n",
4259             surface, debug_surflocation(flag), persistent);
4260
4261     if (wined3d_settings.offscreen_rendering_mode == ORM_FBO)
4262     {
4263         if (surface_is_offscreen(surface))
4264         {
4265             /* With ORM_FBO, SFLAG_INTEXTURE and SFLAG_INDRAWABLE are the same for offscreen targets. */
4266             if (flag & (SFLAG_INTEXTURE | SFLAG_INDRAWABLE)) flag |= (SFLAG_INTEXTURE | SFLAG_INDRAWABLE);
4267         }
4268         else
4269         {
4270             TRACE("Surface %p is an onscreen surface.\n", surface);
4271         }
4272     }
4273
4274     if (persistent)
4275     {
4276         if (((surface->Flags & SFLAG_INTEXTURE) && !(flag & SFLAG_INTEXTURE))
4277                 || ((surface->Flags & SFLAG_INSRGBTEX) && !(flag & SFLAG_INSRGBTEX)))
4278         {
4279             if (surface->container.type == WINED3D_CONTAINER_TEXTURE)
4280             {
4281                 TRACE("Passing to container.\n");
4282                 IWineD3DBaseTexture_SetDirty((IWineD3DBaseTexture *)surface->container.u.texture, TRUE);
4283             }
4284         }
4285         surface->Flags &= ~SFLAG_LOCATIONS;
4286         surface->Flags |= flag;
4287
4288         /* Redraw emulated overlays, if any */
4289         if (flag & SFLAG_INDRAWABLE && !list_empty(&surface->overlays))
4290         {
4291             LIST_FOR_EACH_ENTRY(overlay, &surface->overlays, IWineD3DSurfaceImpl, overlay_entry)
4292             {
4293                 IWineD3DSurface_DrawOverlay((IWineD3DSurface *)overlay);
4294             }
4295         }
4296     }
4297     else
4298     {
4299         if ((surface->Flags & (SFLAG_INTEXTURE | SFLAG_INSRGBTEX)) && (flag & (SFLAG_INTEXTURE | SFLAG_INSRGBTEX)))
4300         {
4301             if (surface->container.type == WINED3D_CONTAINER_TEXTURE)
4302             {
4303                 TRACE("Passing to container\n");
4304                 IWineD3DBaseTexture_SetDirty((IWineD3DBaseTexture *)surface->container.u.texture, TRUE);
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     swapchain = This->container.type == WINED3D_CONTAINER_SWAPCHAIN ? This->container.u.swapchain : NULL;
4340     if (swapchain && This == swapchain->front_buffer)
4341         surface_translate_frontbuffer_coords(This, context->win_handle, &dst_rect);
4342
4343     device->blitter->set_shader((IWineD3DDevice *) device, This);
4344
4345     ENTER_GL();
4346     draw_textured_quad(This, &src_rect, &dst_rect, WINED3DTEXF_POINT);
4347     LEAVE_GL();
4348
4349     device->blitter->unset_shader((IWineD3DDevice *) device);
4350
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 WINED3DSURFTYPE WINAPI IWineD3DSurfaceImpl_GetImplType(IWineD3DSurface *iface) {
4661     return SURFACE_OPENGL;
4662 }
4663
4664 static HRESULT WINAPI IWineD3DSurfaceImpl_DrawOverlay(IWineD3DSurface *iface) {
4665     IWineD3DSurfaceImpl *This = (IWineD3DSurfaceImpl *) iface;
4666     HRESULT hr;
4667
4668     /* If there's no destination surface there is nothing to do */
4669     if(!This->overlay_dest) return WINED3D_OK;
4670
4671     /* Blt calls ModifyLocation on the dest surface, which in turn calls DrawOverlay to
4672      * update the overlay. Prevent an endless recursion
4673      */
4674     if(This->overlay_dest->Flags & SFLAG_INOVERLAYDRAW) {
4675         return WINED3D_OK;
4676     }
4677     This->overlay_dest->Flags |= SFLAG_INOVERLAYDRAW;
4678     hr = IWineD3DSurfaceImpl_Blt((IWineD3DSurface *) This->overlay_dest, &This->overlay_destrect,
4679                                  iface, &This->overlay_srcrect, WINEDDBLT_WAIT,
4680                                  NULL, WINED3DTEXF_LINEAR);
4681     This->overlay_dest->Flags &= ~SFLAG_INOVERLAYDRAW;
4682
4683     return hr;
4684 }
4685
4686 BOOL surface_is_offscreen(IWineD3DSurfaceImpl *surface)
4687 {
4688     IWineD3DSwapChainImpl *swapchain = surface->container.u.swapchain;
4689
4690     /* Not on a swapchain - must be offscreen */
4691     if (surface->container.type != WINED3D_CONTAINER_SWAPCHAIN) return TRUE;
4692
4693     /* The front buffer is always onscreen */
4694     if (surface == swapchain->front_buffer) return FALSE;
4695
4696     /* If the swapchain is rendered to an FBO, the backbuffer is
4697      * offscreen, otherwise onscreen */
4698     return swapchain->render_to_fbo;
4699 }
4700
4701 const IWineD3DSurfaceVtbl IWineD3DSurface_Vtbl =
4702 {
4703     /* IUnknown */
4704     IWineD3DBaseSurfaceImpl_QueryInterface,
4705     IWineD3DBaseSurfaceImpl_AddRef,
4706     IWineD3DSurfaceImpl_Release,
4707     /* IWineD3DResource */
4708     IWineD3DBaseSurfaceImpl_GetParent,
4709     IWineD3DBaseSurfaceImpl_SetPrivateData,
4710     IWineD3DBaseSurfaceImpl_GetPrivateData,
4711     IWineD3DBaseSurfaceImpl_FreePrivateData,
4712     IWineD3DBaseSurfaceImpl_SetPriority,
4713     IWineD3DBaseSurfaceImpl_GetPriority,
4714     IWineD3DSurfaceImpl_PreLoad,
4715     IWineD3DSurfaceImpl_UnLoad,
4716     IWineD3DBaseSurfaceImpl_GetType,
4717     /* IWineD3DSurface */
4718     IWineD3DBaseSurfaceImpl_GetContainer,
4719     IWineD3DBaseSurfaceImpl_GetDesc,
4720     IWineD3DSurfaceImpl_LockRect,
4721     IWineD3DSurfaceImpl_UnlockRect,
4722     IWineD3DSurfaceImpl_GetDC,
4723     IWineD3DSurfaceImpl_ReleaseDC,
4724     IWineD3DSurfaceImpl_Flip,
4725     IWineD3DSurfaceImpl_Blt,
4726     IWineD3DBaseSurfaceImpl_GetBltStatus,
4727     IWineD3DBaseSurfaceImpl_GetFlipStatus,
4728     IWineD3DBaseSurfaceImpl_IsLost,
4729     IWineD3DBaseSurfaceImpl_Restore,
4730     IWineD3DSurfaceImpl_BltFast,
4731     IWineD3DBaseSurfaceImpl_GetPalette,
4732     IWineD3DBaseSurfaceImpl_SetPalette,
4733     IWineD3DSurfaceImpl_RealizePalette,
4734     IWineD3DBaseSurfaceImpl_SetColorKey,
4735     IWineD3DBaseSurfaceImpl_GetPitch,
4736     IWineD3DSurfaceImpl_SetMem,
4737     IWineD3DBaseSurfaceImpl_SetOverlayPosition,
4738     IWineD3DBaseSurfaceImpl_GetOverlayPosition,
4739     IWineD3DBaseSurfaceImpl_UpdateOverlayZOrder,
4740     IWineD3DBaseSurfaceImpl_UpdateOverlay,
4741     IWineD3DBaseSurfaceImpl_SetClipper,
4742     IWineD3DBaseSurfaceImpl_GetClipper,
4743     /* Internal use: */
4744     IWineD3DSurfaceImpl_LoadTexture,
4745     IWineD3DSurfaceImpl_BindTexture,
4746     IWineD3DBaseSurfaceImpl_GetData,
4747     IWineD3DSurfaceImpl_SetFormat,
4748     IWineD3DSurfaceImpl_PrivateSetup,
4749     IWineD3DSurfaceImpl_GetImplType,
4750     IWineD3DSurfaceImpl_DrawOverlay
4751 };
4752
4753 static HRESULT ffp_blit_alloc(IWineD3DDevice *iface) { return WINED3D_OK; }
4754 /* Context activation is done by the caller. */
4755 static void ffp_blit_free(IWineD3DDevice *iface) { }
4756
4757 /* This function is used in case of 8bit paletted textures using GL_EXT_paletted_texture */
4758 /* Context activation is done by the caller. */
4759 static void ffp_blit_p8_upload_palette(IWineD3DSurfaceImpl *surface, const struct wined3d_gl_info *gl_info)
4760 {
4761     BYTE table[256][4];
4762     BOOL colorkey_active = (surface->CKeyFlags & WINEDDSD_CKSRCBLT) ? TRUE : FALSE;
4763
4764     d3dfmt_p8_init_palette(surface, table, colorkey_active);
4765
4766     TRACE("Using GL_EXT_PALETTED_TEXTURE for 8-bit paletted texture support\n");
4767     ENTER_GL();
4768     GL_EXTCALL(glColorTableEXT(surface->texture_target, GL_RGBA, 256, GL_RGBA, GL_UNSIGNED_BYTE, table));
4769     LEAVE_GL();
4770 }
4771
4772 /* Context activation is done by the caller. */
4773 static HRESULT ffp_blit_set(IWineD3DDevice *iface, IWineD3DSurfaceImpl *surface)
4774 {
4775     IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *) iface;
4776     const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
4777     enum complex_fixup fixup = get_complex_fixup(surface->resource.format_desc->color_fixup);
4778
4779     /* When EXT_PALETTED_TEXTURE is around, palette conversion is done by the GPU
4780      * else the surface is converted in software at upload time in LoadLocation.
4781      */
4782     if(fixup == COMPLEX_FIXUP_P8 && gl_info->supported[EXT_PALETTED_TEXTURE])
4783         ffp_blit_p8_upload_palette(surface, gl_info);
4784
4785     ENTER_GL();
4786     glEnable(surface->texture_target);
4787     checkGLcall("glEnable(surface->texture_target)");
4788     LEAVE_GL();
4789     return WINED3D_OK;
4790 }
4791
4792 /* Context activation is done by the caller. */
4793 static void ffp_blit_unset(IWineD3DDevice *iface)
4794 {
4795     IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *) iface;
4796     const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
4797
4798     ENTER_GL();
4799     glDisable(GL_TEXTURE_2D);
4800     checkGLcall("glDisable(GL_TEXTURE_2D)");
4801     if (gl_info->supported[ARB_TEXTURE_CUBE_MAP])
4802     {
4803         glDisable(GL_TEXTURE_CUBE_MAP_ARB);
4804         checkGLcall("glDisable(GL_TEXTURE_CUBE_MAP_ARB)");
4805     }
4806     if (gl_info->supported[ARB_TEXTURE_RECTANGLE])
4807     {
4808         glDisable(GL_TEXTURE_RECTANGLE_ARB);
4809         checkGLcall("glDisable(GL_TEXTURE_RECTANGLE_ARB)");
4810     }
4811     LEAVE_GL();
4812 }
4813
4814 static BOOL ffp_blit_supported(const struct wined3d_gl_info *gl_info, enum blit_operation blit_op,
4815                                const RECT *src_rect, DWORD src_usage, WINED3DPOOL src_pool,
4816                                const struct wined3d_format_desc *src_format_desc,
4817                                const RECT *dst_rect, DWORD dst_usage, WINED3DPOOL dst_pool,
4818                                const struct wined3d_format_desc *dst_format_desc)
4819 {
4820     enum complex_fixup src_fixup;
4821
4822     if (blit_op == BLIT_OP_COLOR_FILL)
4823     {
4824         if (!(dst_usage & WINED3DUSAGE_RENDERTARGET))
4825         {
4826             TRACE("Color fill not supported\n");
4827             return FALSE;
4828         }
4829
4830         return TRUE;
4831     }
4832
4833     src_fixup = get_complex_fixup(src_format_desc->color_fixup);
4834     if (TRACE_ON(d3d_surface) && TRACE_ON(d3d))
4835     {
4836         TRACE("Checking support for fixup:\n");
4837         dump_color_fixup_desc(src_format_desc->color_fixup);
4838     }
4839
4840     if (blit_op != BLIT_OP_BLIT)
4841     {
4842         TRACE("Unsupported blit_op=%d\n", blit_op);
4843         return FALSE;
4844      }
4845
4846     if (!is_identity_fixup(dst_format_desc->color_fixup))
4847     {
4848         TRACE("Destination fixups are not supported\n");
4849         return FALSE;
4850     }
4851
4852     if (src_fixup == COMPLEX_FIXUP_P8 && gl_info->supported[EXT_PALETTED_TEXTURE])
4853     {
4854         TRACE("P8 fixup supported\n");
4855         return TRUE;
4856     }
4857
4858     /* We only support identity conversions. */
4859     if (is_identity_fixup(src_format_desc->color_fixup))
4860     {
4861         TRACE("[OK]\n");
4862         return TRUE;
4863     }
4864
4865     TRACE("[FAILED]\n");
4866     return FALSE;
4867 }
4868
4869 static HRESULT ffp_blit_color_fill(IWineD3DDeviceImpl *device, IWineD3DSurfaceImpl *dst_surface,
4870         const RECT *dst_rect, const WINED3DCOLORVALUE *color)
4871 {
4872     const RECT draw_rect = {0, 0, dst_surface->currentDesc.Width, dst_surface->currentDesc.Height};
4873
4874     return device_clear_render_targets(device, 1 /* rt_count */, &dst_surface, 1 /* rect_count */,
4875             dst_rect, &draw_rect, WINED3DCLEAR_TARGET, color, 0.0f /* depth */, 0 /* stencil */);
4876 }
4877
4878 const struct blit_shader ffp_blit =  {
4879     ffp_blit_alloc,
4880     ffp_blit_free,
4881     ffp_blit_set,
4882     ffp_blit_unset,
4883     ffp_blit_supported,
4884     ffp_blit_color_fill
4885 };
4886
4887 static HRESULT cpu_blit_alloc(IWineD3DDevice *iface)
4888 {
4889     return WINED3D_OK;
4890 }
4891
4892 /* Context activation is done by the caller. */
4893 static void cpu_blit_free(IWineD3DDevice *iface)
4894 {
4895 }
4896
4897 /* Context activation is done by the caller. */
4898 static HRESULT cpu_blit_set(IWineD3DDevice *iface, IWineD3DSurfaceImpl *surface)
4899 {
4900     return WINED3D_OK;
4901 }
4902
4903 /* Context activation is done by the caller. */
4904 static void cpu_blit_unset(IWineD3DDevice *iface)
4905 {
4906 }
4907
4908 static BOOL cpu_blit_supported(const struct wined3d_gl_info *gl_info, enum blit_operation blit_op,
4909                                const RECT *src_rect, DWORD src_usage, WINED3DPOOL src_pool,
4910                                const struct wined3d_format_desc *src_format_desc,
4911                                const RECT *dst_rect, DWORD dst_usage, WINED3DPOOL dst_pool,
4912                                const struct wined3d_format_desc *dst_format_desc)
4913 {
4914     if (blit_op == BLIT_OP_COLOR_FILL)
4915     {
4916         return TRUE;
4917     }
4918
4919     return FALSE;
4920 }
4921
4922 static HRESULT cpu_blit_color_fill(IWineD3DDeviceImpl *device, IWineD3DSurfaceImpl *dst_surface,
4923         const RECT *dst_rect, const WINED3DCOLORVALUE *color)
4924 {
4925     WINEDDBLTFX BltFx;
4926
4927     memset(&BltFx, 0, sizeof(BltFx));
4928     BltFx.dwSize = sizeof(BltFx);
4929     BltFx.u5.dwFillColor = wined3d_format_convert_from_float(dst_surface->resource.format_desc, color);
4930     return IWineD3DBaseSurfaceImpl_Blt((IWineD3DSurface*)dst_surface, dst_rect,
4931             NULL, NULL, WINEDDBLT_COLORFILL, &BltFx, WINED3DTEXF_POINT);
4932 }
4933
4934 const struct blit_shader cpu_blit =  {
4935     cpu_blit_alloc,
4936     cpu_blit_free,
4937     cpu_blit_set,
4938     cpu_blit_unset,
4939     cpu_blit_supported,
4940     cpu_blit_color_fill
4941 };
4942
4943 static BOOL fbo_blit_supported(const struct wined3d_gl_info *gl_info, enum blit_operation blit_op,
4944                                const RECT *src_rect, DWORD src_usage, WINED3DPOOL src_pool,
4945                                const struct wined3d_format_desc *src_format_desc,
4946                                const RECT *dst_rect, DWORD dst_usage, WINED3DPOOL dst_pool,
4947                                const struct wined3d_format_desc *dst_format_desc)
4948 {
4949     if ((wined3d_settings.offscreen_rendering_mode != ORM_FBO) || !gl_info->fbo_ops.glBlitFramebuffer)
4950         return FALSE;
4951
4952     /* We only support blitting. Things like color keying / color fill should
4953      * be handled by other blitters.
4954      */
4955     if (blit_op != BLIT_OP_BLIT)
4956         return FALSE;
4957
4958     /* Source and/or destination need to be on the GL side */
4959     if (src_pool == WINED3DPOOL_SYSTEMMEM || dst_pool == WINED3DPOOL_SYSTEMMEM)
4960         return FALSE;
4961
4962     if(!((src_format_desc->Flags & WINED3DFMT_FLAG_FBO_ATTACHABLE) || (src_usage & WINED3DUSAGE_RENDERTARGET))
4963         && ((dst_format_desc->Flags & WINED3DFMT_FLAG_FBO_ATTACHABLE) || (dst_usage & WINED3DUSAGE_RENDERTARGET)))
4964         return FALSE;
4965
4966     if (!is_identity_fixup(src_format_desc->color_fixup) ||
4967         !is_identity_fixup(dst_format_desc->color_fixup))
4968         return FALSE;
4969
4970     if (!(src_format_desc->format == dst_format_desc->format
4971         || (is_identity_fixup(src_format_desc->color_fixup)
4972         && is_identity_fixup(dst_format_desc->color_fixup))))
4973         return FALSE;
4974
4975     return TRUE;
4976 }