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 2009-2010 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 (!container && type != WINED3D_CONTAINER_NONE)
 100         ERR("Setting NULL container of type %#x.\n", type);
 101
 102     if (type == WINED3D_CONTAINER_SWAPCHAIN)
 103     {
 104         surface->get_drawable_size = get_drawable_size_swapchain;
 105     }
 106     else
 107     {
 108         switch (wined3d_settings.offscreen_rendering_mode)
 109         {
 110             case ORM_FBO:
 111                 surface->get_drawable_size = get_drawable_size_fbo;
 112                 break;
 113
 114             case ORM_BACKBUFFER:
 115                 surface->get_drawable_size = get_drawable_size_backbuffer;
 116                 break;
 117
 118             default:
 119                 ERR("Unhandled offscreen rendering mode %#x.\n", wined3d_settings.offscreen_rendering_mode);
 120                 return;
 121         }
 122     }
 123
 124     surface->container.type = type;
 125     surface->container.u.base = container;
 126 }
 127
 128 struct blt_info
 129 {
 130     GLenum binding;
 131     GLenum bind_target;
 132     enum tex_types tex_type;
 133     GLfloat coords[4][3];
 134 };
 135
 136 struct float_rect
 137 {
 138     float l;
 139     float t;
 140     float r;
 141     float b;
 142 };
 143
 144 static inline void cube_coords_float(const RECT *r, UINT w, UINT h, struct float_rect *f)
 145 {
 146     f->l = ((r->left * 2.0f) / w) - 1.0f;
 147     f->t = ((r->top * 2.0f) / h) - 1.0f;
 148     f->r = ((r->right * 2.0f) / w) - 1.0f;
 149     f->b = ((r->bottom * 2.0f) / h) - 1.0f;
 150 }
 151
 152 static void surface_get_blt_info(GLenum target, const RECT *rect, GLsizei w, GLsizei h, struct blt_info *info)
 153 {
 154     GLfloat (*coords)[3] = info->coords;
 155     struct float_rect f;
 156
 157     switch (target)
 158     {
 159         default:
 160             FIXME("Unsupported texture target %#x\n", target);
 161             /* Fall back to GL_TEXTURE_2D */
 162         case GL_TEXTURE_2D:
 163             info->binding = GL_TEXTURE_BINDING_2D;
 164             info->bind_target = GL_TEXTURE_2D;
 165             info->tex_type = tex_2d;
 166             coords[0][0] = (float)rect->left / w;
 167             coords[0][1] = (float)rect->top / h;
 168             coords[0][2] = 0.0f;
 169
 170             coords[1][0] = (float)rect->right / w;
 171             coords[1][1] = (float)rect->top / h;
 172             coords[1][2] = 0.0f;
 173
 174             coords[2][0] = (float)rect->left / w;
 175             coords[2][1] = (float)rect->bottom / h;
 176             coords[2][2] = 0.0f;
 177
 178             coords[3][0] = (float)rect->right / w;
 179             coords[3][1] = (float)rect->bottom / h;
 180             coords[3][2] = 0.0f;
 181             break;
 182
 183         case GL_TEXTURE_RECTANGLE_ARB:
 184             info->binding = GL_TEXTURE_BINDING_RECTANGLE_ARB;
 185             info->bind_target = GL_TEXTURE_RECTANGLE_ARB;
 186             info->tex_type = tex_rect;
 187             coords[0][0] = rect->left;  coords[0][1] = rect->top;       coords[0][2] = 0.0f;
 188             coords[1][0] = rect->right; coords[1][1] = rect->top;       coords[1][2] = 0.0f;
 189             coords[2][0] = rect->left;  coords[2][1] = rect->bottom;    coords[2][2] = 0.0f;
 190             coords[3][0] = rect->right; coords[3][1] = rect->bottom;    coords[3][2] = 0.0f;
 191             break;
 192
 193         case GL_TEXTURE_CUBE_MAP_POSITIVE_X:
 194             info->binding = GL_TEXTURE_BINDING_CUBE_MAP_ARB;
 195             info->bind_target = GL_TEXTURE_CUBE_MAP_ARB;
 196             info->tex_type = tex_cube;
 197             cube_coords_float(rect, w, h, &f);
 198
 199             coords[0][0] =  1.0f;   coords[0][1] = -f.t;   coords[0][2] = -f.l;
 200             coords[1][0] =  1.0f;   coords[1][1] = -f.t;   coords[1][2] = -f.r;
 201             coords[2][0] =  1.0f;   coords[2][1] = -f.b;   coords[2][2] = -f.l;
 202             coords[3][0] =  1.0f;   coords[3][1] = -f.b;   coords[3][2] = -f.r;
 203             break;
 204
 205         case GL_TEXTURE_CUBE_MAP_NEGATIVE_X:
 206             info->binding = GL_TEXTURE_BINDING_CUBE_MAP_ARB;
 207             info->bind_target = GL_TEXTURE_CUBE_MAP_ARB;
 208             info->tex_type = tex_cube;
 209             cube_coords_float(rect, w, h, &f);
 210
 211             coords[0][0] = -1.0f;   coords[0][1] = -f.t;   coords[0][2] = f.l;
 212             coords[1][0] = -1.0f;   coords[1][1] = -f.t;   coords[1][2] = f.r;
 213             coords[2][0] = -1.0f;   coords[2][1] = -f.b;   coords[2][2] = f.l;
 214             coords[3][0] = -1.0f;   coords[3][1] = -f.b;   coords[3][2] = f.r;
 215             break;
 216
 217         case GL_TEXTURE_CUBE_MAP_POSITIVE_Y:
 218             info->binding = GL_TEXTURE_BINDING_CUBE_MAP_ARB;
 219             info->bind_target = GL_TEXTURE_CUBE_MAP_ARB;
 220             info->tex_type = tex_cube;
 221             cube_coords_float(rect, w, h, &f);
 222
 223             coords[0][0] = f.l;   coords[0][1] =  1.0f;   coords[0][2] = f.t;
 224             coords[1][0] = f.r;   coords[1][1] =  1.0f;   coords[1][2] = f.t;
 225             coords[2][0] = f.l;   coords[2][1] =  1.0f;   coords[2][2] = f.b;
 226             coords[3][0] = f.r;   coords[3][1] =  1.0f;   coords[3][2] = f.b;
 227             break;
 228
 229         case GL_TEXTURE_CUBE_MAP_NEGATIVE_Y:
 230             info->binding = GL_TEXTURE_BINDING_CUBE_MAP_ARB;
 231             info->bind_target = GL_TEXTURE_CUBE_MAP_ARB;
 232             info->tex_type = tex_cube;
 233             cube_coords_float(rect, w, h, &f);
 234
 235             coords[0][0] = f.l;   coords[0][1] = -1.0f;   coords[0][2] = -f.t;
 236             coords[1][0] = f.r;   coords[1][1] = -1.0f;   coords[1][2] = -f.t;
 237             coords[2][0] = f.l;   coords[2][1] = -1.0f;   coords[2][2] = -f.b;
 238             coords[3][0] = f.r;   coords[3][1] = -1.0f;   coords[3][2] = -f.b;
 239             break;
 240
 241         case GL_TEXTURE_CUBE_MAP_POSITIVE_Z:
 242             info->binding = GL_TEXTURE_BINDING_CUBE_MAP_ARB;
 243             info->bind_target = GL_TEXTURE_CUBE_MAP_ARB;
 244             info->tex_type = tex_cube;
 245             cube_coords_float(rect, w, h, &f);
 246
 247             coords[0][0] = f.l;   coords[0][1] = -f.t;   coords[0][2] =  1.0f;
 248             coords[1][0] = f.r;   coords[1][1] = -f.t;   coords[1][2] =  1.0f;
 249             coords[2][0] = f.l;   coords[2][1] = -f.b;   coords[2][2] =  1.0f;
 250             coords[3][0] = f.r;   coords[3][1] = -f.b;   coords[3][2] =  1.0f;
 251             break;
 252
 253         case GL_TEXTURE_CUBE_MAP_NEGATIVE_Z:
 254             info->binding = GL_TEXTURE_BINDING_CUBE_MAP_ARB;
 255             info->bind_target = GL_TEXTURE_CUBE_MAP_ARB;
 256             info->tex_type = tex_cube;
 257             cube_coords_float(rect, w, h, &f);
 258
 259             coords[0][0] = -f.l;   coords[0][1] = -f.t;   coords[0][2] = -1.0f;
 260             coords[1][0] = -f.r;   coords[1][1] = -f.t;   coords[1][2] = -1.0f;
 261             coords[2][0] = -f.l;   coords[2][1] = -f.b;   coords[2][2] = -1.0f;
 262             coords[3][0] = -f.r;   coords[3][1] = -f.b;   coords[3][2] = -1.0f;
 263             break;
 264     }
 265 }
 266
 267 static inline void surface_get_rect(IWineD3DSurfaceImpl *This, const RECT *rect_in, RECT *rect_out)
 268 {
 269     if (rect_in)
 270         *rect_out = *rect_in;
 271     else
 272     {
 273         rect_out->left = 0;
 274         rect_out->top = 0;
 275         rect_out->right = This->currentDesc.Width;
 276         rect_out->bottom = This->currentDesc.Height;
 277     }
 278 }
 279
 280 /* GL locking and context activation is done by the caller */
 281 void draw_textured_quad(IWineD3DSurfaceImpl *src_surface, const RECT *src_rect, const RECT *dst_rect, WINED3DTEXTUREFILTERTYPE Filter)
 282 {
 283     struct blt_info info;
 284
 285     surface_get_blt_info(src_surface->texture_target, src_rect, src_surface->pow2Width, src_surface->pow2Height, &info);
 286
 287     glEnable(info.bind_target);
 288     checkGLcall("glEnable(bind_target)");
 289
 290     /* Bind the texture */
 291     glBindTexture(info.bind_target, src_surface->texture_name);
 292     checkGLcall("glBindTexture");
 293
 294     /* Filtering for StretchRect */
 295     glTexParameteri(info.bind_target, GL_TEXTURE_MAG_FILTER,
 296             wined3d_gl_mag_filter(magLookup, Filter));
 297     checkGLcall("glTexParameteri");
 298     glTexParameteri(info.bind_target, GL_TEXTURE_MIN_FILTER,
 299             wined3d_gl_min_mip_filter(minMipLookup, Filter, WINED3DTEXF_NONE));
 300     checkGLcall("glTexParameteri");
 301     glTexParameteri(info.bind_target, GL_TEXTURE_WRAP_S, GL_CLAMP);
 302     glTexParameteri(info.bind_target, GL_TEXTURE_WRAP_T, GL_CLAMP);
 303     glTexEnvi(GL_TEXTURE_ENV, GL_TEXTURE_ENV_MODE, GL_REPLACE);
 304     checkGLcall("glTexEnvi");
 305
 306     /* Draw a quad */
 307     glBegin(GL_TRIANGLE_STRIP);
 308     glTexCoord3fv(info.coords[0]);
 309     glVertex2i(dst_rect->left, dst_rect->top);
 310
 311     glTexCoord3fv(info.coords[1]);
 312     glVertex2i(dst_rect->right, dst_rect->top);
 313
 314     glTexCoord3fv(info.coords[2]);
 315     glVertex2i(dst_rect->left, dst_rect->bottom);
 316
 317     glTexCoord3fv(info.coords[3]);
 318     glVertex2i(dst_rect->right, dst_rect->bottom);
 319     glEnd();
 320
 321     /* Unbind the texture */
 322     glBindTexture(info.bind_target, 0);
 323     checkGLcall("glBindTexture(info->bind_target, 0)");
 324
 325     /* We changed the filtering settings on the texture. Inform the
 326      * container about this to get the filters reset properly next draw. */
 327     if (src_surface->container.type == WINED3D_CONTAINER_TEXTURE)
 328     {
 329         IWineD3DBaseTextureImpl *texture = src_surface->container.u.texture;
 330         texture->baseTexture.texture_rgb.states[WINED3DTEXSTA_MAGFILTER] = WINED3DTEXF_POINT;
 331         texture->baseTexture.texture_rgb.states[WINED3DTEXSTA_MINFILTER] = WINED3DTEXF_POINT;
 332         texture->baseTexture.texture_rgb.states[WINED3DTEXSTA_MIPFILTER] = WINED3DTEXF_NONE;
 333     }
 334 }
 335
 336 HRESULT surface_init(IWineD3DSurfaceImpl *surface, WINED3DSURFTYPE surface_type, UINT alignment,
 337         UINT width, UINT height, UINT level, BOOL lockable, BOOL discard, WINED3DMULTISAMPLE_TYPE multisample_type,
 338         UINT multisample_quality, IWineD3DDeviceImpl *device, DWORD usage, enum wined3d_format_id format_id,
 339         WINED3DPOOL pool, void *parent, const struct wined3d_parent_ops *parent_ops)
 340 {
 341     const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
 342     const struct wined3d_format *format = wined3d_get_format(gl_info, format_id);
 343     void (*cleanup)(IWineD3DSurfaceImpl *This);
 344     unsigned int resource_size;
 345     HRESULT hr;
 346
 347     if (multisample_quality > 0)
 348     {
 349         FIXME("multisample_quality set to %u, substituting 0\n", multisample_quality);
 350         multisample_quality = 0;
 351     }
 352
 353     /* FIXME: Check that the format is supported by the device. */
 354
 355     resource_size = wined3d_format_calculate_size(format, alignment, width, height);
 356     if (!resource_size) return WINED3DERR_INVALIDCALL;
 357
 358     /* Look at the implementation and set the correct Vtable. */
 359     switch (surface_type)
 360     {
 361         case SURFACE_OPENGL:
 362             surface->lpVtbl = &IWineD3DSurface_Vtbl;
 363             cleanup = surface_cleanup;
 364             break;
 365
 366         case SURFACE_GDI:
 367             surface->lpVtbl = &IWineGDISurface_Vtbl;
 368             cleanup = surface_gdi_cleanup;
 369             break;
 370
 371         default:
 372             ERR("Requested unknown surface implementation %#x.\n", surface_type);
 373             return WINED3DERR_INVALIDCALL;
 374     }
 375
 376     hr = resource_init((IWineD3DResource *)surface, WINED3DRTYPE_SURFACE,
 377             device, resource_size, usage, format, pool, parent, parent_ops);
 378     if (FAILED(hr))
 379     {
 380         WARN("Failed to initialize resource, returning %#x.\n", hr);
 381         return hr;
 382     }
 383
 384     /* "Standalone" surface. */
 385     surface_set_container(surface, WINED3D_CONTAINER_NONE, NULL);
 386
 387     surface->currentDesc.Width = width;
 388     surface->currentDesc.Height = height;
 389     surface->currentDesc.MultiSampleType = multisample_type;
 390     surface->currentDesc.MultiSampleQuality = multisample_quality;
 391     surface->texture_level = level;
 392     list_init(&surface->overlays);
 393
 394     /* Flags */
 395     surface->flags = SFLAG_NORMCOORD; /* Default to normalized coords. */
 396     if (discard) surface->flags |= SFLAG_DISCARD;
 397     if (lockable || format_id == WINED3DFMT_D16_LOCKABLE) surface->flags |= SFLAG_LOCKABLE;
 398
 399     /* Quick lockable sanity check.
 400      * TODO: remove this after surfaces, usage and lockability have been debugged properly
 401      * this function is too deep to need to care about things like this.
 402      * Levels need to be checked too, since they all affect what can be done. */
 403     switch (pool)
 404     {
 405         case WINED3DPOOL_SCRATCH:
 406             if(!lockable)
 407             {
 408                 FIXME("Called with a pool of SCRATCH and a lockable of FALSE "
 409                         "which are mutually exclusive, setting lockable to TRUE.\n");
 410                 lockable = TRUE;
 411             }
 412             break;
 413
 414         case WINED3DPOOL_SYSTEMMEM:
 415             if (!lockable)
 416                 FIXME("Called with a pool of SYSTEMMEM and a lockable of FALSE, this is acceptable but unexpected.\n");
 417             break;
 418
 419         case WINED3DPOOL_MANAGED:
 420             if (usage & WINED3DUSAGE_DYNAMIC)
 421                 FIXME("Called with a pool of MANAGED and a usage of DYNAMIC which are mutually exclusive.\n");
 422             break;
 423
 424         case WINED3DPOOL_DEFAULT:
 425             if (lockable && !(usage & (WINED3DUSAGE_DYNAMIC | WINED3DUSAGE_RENDERTARGET | WINED3DUSAGE_DEPTHSTENCIL)))
 426                 WARN("Creating a lockable surface with a POOL of DEFAULT, that doesn't specify DYNAMIC usage.\n");
 427             break;
 428
 429         default:
 430             FIXME("Unknown pool %#x.\n", pool);
 431             break;
 432     };
 433
 434     if (usage & WINED3DUSAGE_RENDERTARGET && pool != WINED3DPOOL_DEFAULT)
 435     {
 436         FIXME("Trying to create a render target that isn't in the default pool.\n");
 437     }
 438
 439     /* Mark the texture as dirty so that it gets loaded first time around. */
 440     surface_add_dirty_rect(surface, NULL);
 441     list_init(&surface->renderbuffers);
 442
 443     TRACE("surface %p, memory %p, size %u\n", surface, surface->resource.allocatedMemory, surface->resource.size);
 444
 445     /* Call the private setup routine */
 446     hr = IWineD3DSurface_PrivateSetup((IWineD3DSurface *)surface);
 447     if (FAILED(hr))
 448     {
 449         ERR("Private setup failed, returning %#x\n", hr);
 450         cleanup(surface);
 451         return hr;
 452     }
 453
 454     return hr;
 455 }
 456
 457 static void surface_force_reload(IWineD3DSurfaceImpl *surface)
 458 {
 459     surface->flags &= ~(SFLAG_ALLOCATED | SFLAG_SRGBALLOCATED);
 460 }
 461
 462 void surface_set_texture_name(IWineD3DSurfaceImpl *surface, GLuint new_name, BOOL srgb)
 463 {
 464     GLuint *name;
 465     DWORD flag;
 466
 467     TRACE("surface %p, new_name %u, srgb %#x.\n", surface, new_name, srgb);
 468
 469     if(srgb)
 470     {
 471         name = &surface->texture_name_srgb;
 472         flag = SFLAG_INSRGBTEX;
 473     }
 474     else
 475     {
 476         name = &surface->texture_name;
 477         flag = SFLAG_INTEXTURE;
 478     }
 479
 480     if (!*name && new_name)
 481     {
 482         /* FIXME: We shouldn't need to remove SFLAG_INTEXTURE if the
 483          * surface has no texture name yet. See if we can get rid of this. */
 484         if (surface->flags & flag)
 485             ERR("Surface has %s set, but no texture name.\n", debug_surflocation(flag));
 486         surface_modify_location(surface, flag, FALSE);
 487     }
 488
 489     *name = new_name;
 490     surface_force_reload(surface);
 491 }
 492
 493 void surface_set_texture_target(IWineD3DSurfaceImpl *surface, GLenum target)
 494 {
 495     TRACE("surface %p, target %#x.\n", surface, target);
 496
 497     if (surface->texture_target != target)
 498     {
 499         if (target == GL_TEXTURE_RECTANGLE_ARB)
 500         {
 501             surface->flags &= ~SFLAG_NORMCOORD;
 502         }
 503         else if (surface->texture_target == GL_TEXTURE_RECTANGLE_ARB)
 504         {
 505             surface->flags |= SFLAG_NORMCOORD;
 506         }
 507     }
 508     surface->texture_target = target;
 509     surface_force_reload(surface);
 510 }
 511
 512 /* Context activation is done by the caller. */
 513 static void surface_bind_and_dirtify(IWineD3DSurfaceImpl *This, BOOL srgb) {
 514     DWORD active_sampler;
 515
 516     /* We don't need a specific texture unit, but after binding the texture the current unit is dirty.
 517      * Read the unit back instead of switching to 0, this avoids messing around with the state manager's
 518      * gl states. The current texture unit should always be a valid one.
 519      *
 520      * To be more specific, this is tricky because we can implicitly be called
 521      * from sampler() in state.c. This means we can't touch anything other than
 522      * whatever happens to be the currently active texture, or we would risk
 523      * marking already applied sampler states dirty again.
 524      *
 525      * TODO: Track the current active texture per GL context instead of using glGet
 526      */
 527     GLint active_texture;
 528     ENTER_GL();
 529     glGetIntegerv(GL_ACTIVE_TEXTURE, &active_texture);
 530     LEAVE_GL();
 531     active_sampler = This->resource.device->rev_tex_unit_map[active_texture - GL_TEXTURE0_ARB];
 532
 533     if (active_sampler != WINED3D_UNMAPPED_STAGE)
 534     {
 535         IWineD3DDeviceImpl_MarkStateDirty(This->resource.device, STATE_SAMPLER(active_sampler));
 536     }
 537     IWineD3DSurface_BindTexture((IWineD3DSurface *)This, srgb);
 538 }
 539
 540 /* This function checks if the primary render target uses the 8bit paletted format. */
 541 static BOOL primary_render_target_is_p8(IWineD3DDeviceImpl *device)
 542 {
 543     if (device->render_targets && device->render_targets[0])
 544     {
 545         IWineD3DSurfaceImpl *render_target = device->render_targets[0];
 546         if ((render_target->resource.usage & WINED3DUSAGE_RENDERTARGET)
 547                 && (render_target->resource.format->id == WINED3DFMT_P8_UINT))
 548             return TRUE;
 549     }
 550     return FALSE;
 551 }
 552
 553 /* This call just downloads data, the caller is responsible for binding the
 554  * correct texture. */
 555 /* Context activation is done by the caller. */
 556 static void surface_download_data(IWineD3DSurfaceImpl *This, const struct wined3d_gl_info *gl_info)
 557 {
 558     const struct wined3d_format *format = This->resource.format;
 559
 560     /* Only support read back of converted P8 surfaces */
 561     if (This->flags & SFLAG_CONVERTED && format->id != WINED3DFMT_P8_UINT)
 562     {
 563         FIXME("Readback conversion not supported for format %s.\n", debug_d3dformat(format->id));
 564         return;
 565     }
 566
 567     ENTER_GL();
 568
 569     if (format->flags & WINED3DFMT_FLAG_COMPRESSED)
 570     {
 571         TRACE("(%p) : Calling glGetCompressedTexImageARB level %d, format %#x, type %#x, data %p.\n",
 572                 This, This->texture_level, format->glFormat, format->glType,
 573                 This->resource.allocatedMemory);
 574
 575         if (This->flags & SFLAG_PBO)
 576         {
 577             GL_EXTCALL(glBindBufferARB(GL_PIXEL_PACK_BUFFER_ARB, This->pbo));
 578             checkGLcall("glBindBufferARB");
 579             GL_EXTCALL(glGetCompressedTexImageARB(This->texture_target, This->texture_level, NULL));
 580             checkGLcall("glGetCompressedTexImageARB");
 581             GL_EXTCALL(glBindBufferARB(GL_PIXEL_PACK_BUFFER_ARB, 0));
 582             checkGLcall("glBindBufferARB");
 583         }
 584         else
 585         {
 586             GL_EXTCALL(glGetCompressedTexImageARB(This->texture_target,
 587                     This->texture_level, This->resource.allocatedMemory));
 588             checkGLcall("glGetCompressedTexImageARB");
 589         }
 590
 591         LEAVE_GL();
 592     } else {
 593         void *mem;
 594         GLenum gl_format = format->glFormat;
 595         GLenum gl_type = format->glType;
 596         int src_pitch = 0;
 597         int dst_pitch = 0;
 598
 599         /* In case of P8 the index is stored in the alpha component if the primary render target uses P8 */
 600         if (format->id == WINED3DFMT_P8_UINT && primary_render_target_is_p8(This->resource.device))
 601         {
 602             gl_format = GL_ALPHA;
 603             gl_type = GL_UNSIGNED_BYTE;
 604         }
 605
 606         if (This->flags & SFLAG_NONPOW2)
 607         {
 608             unsigned char alignment = This->resource.device->surface_alignment;
 609             src_pitch = format->byte_count * This->pow2Width;
 610             dst_pitch = IWineD3DSurface_GetPitch((IWineD3DSurface *) This);
 611             src_pitch = (src_pitch + alignment - 1) & ~(alignment - 1);
 612             mem = HeapAlloc(GetProcessHeap(), 0, src_pitch * This->pow2Height);
 613         } else {
 614             mem = This->resource.allocatedMemory;
 615         }
 616
 617         TRACE("(%p) : Calling glGetTexImage level %d, format %#x, type %#x, data %p\n",
 618                 This, This->texture_level, gl_format, gl_type, mem);
 619
 620         if (This->flags & SFLAG_PBO)
 621         {
 622             GL_EXTCALL(glBindBufferARB(GL_PIXEL_PACK_BUFFER_ARB, This->pbo));
 623             checkGLcall("glBindBufferARB");
 624
 625             glGetTexImage(This->texture_target, This->texture_level, gl_format, gl_type, NULL);
 626             checkGLcall("glGetTexImage");
 627
 628             GL_EXTCALL(glBindBufferARB(GL_PIXEL_PACK_BUFFER_ARB, 0));
 629             checkGLcall("glBindBufferARB");
 630         } else {
 631             glGetTexImage(This->texture_target, This->texture_level, gl_format, gl_type, mem);
 632             checkGLcall("glGetTexImage");
 633         }
 634         LEAVE_GL();
 635
 636         if (This->flags & SFLAG_NONPOW2)
 637         {
 638             const BYTE *src_data;
 639             BYTE *dst_data;
 640             UINT y;
 641             /*
 642              * Some games (e.g. warhammer 40k) don't work properly with the odd pitches, preventing
 643              * the surface pitch from being used to box non-power2 textures. Instead we have to use a hack to
 644              * repack the texture so that the bpp * width pitch can be used instead of bpp * pow2width.
 645              *
 646              * We're doing this...
 647              *
 648              * instead of boxing the texture :
 649              * |<-texture width ->|  -->pow2width|   /\
 650              * |111111111111111111|              |   |
 651              * |222 Texture 222222| boxed empty  | texture height
 652              * |3333 Data 33333333|              |   |
 653              * |444444444444444444|              |   \/
 654              * -----------------------------------   |
 655              * |     boxed  empty | boxed empty  | pow2height
 656              * |                  |              |   \/
 657              * -----------------------------------
 658              *
 659              *
 660              * we're repacking the data to the expected texture width
 661              *
 662              * |<-texture width ->|  -->pow2width|   /\
 663              * |111111111111111111222222222222222|   |
 664              * |222333333333333333333444444444444| texture height
 665              * |444444                           |   |
 666              * |                                 |   \/
 667              * |                                 |   |
 668              * |            empty                | pow2height
 669              * |                                 |   \/
 670              * -----------------------------------
 671              *
 672              * == is the same as
 673              *
 674              * |<-texture width ->|    /\
 675              * |111111111111111111|
 676              * |222222222222222222|texture height
 677              * |333333333333333333|
 678              * |444444444444444444|    \/
 679              * --------------------
 680              *
 681              * this also means that any references to allocatedMemory should work with the data as if were a
 682              * standard texture with a non-power2 width instead of texture boxed up to be a power2 texture.
 683              *
 684              * internally the texture is still stored in a boxed format so any references to textureName will
 685              * get a boxed texture with width pow2width and not a texture of width currentDesc.Width.
 686              *
 687              * Performance should not be an issue, because applications normally do not lock the surfaces when
 688              * rendering. If an app does, the SFLAG_DYNLOCK flag will kick in and the memory copy won't be released,
 689              * and doesn't have to be re-read.
 690              */
 691             src_data = mem;
 692             dst_data = This->resource.allocatedMemory;
 693             TRACE("(%p) : Repacking the surface data from pitch %d to pitch %d\n", This, src_pitch, dst_pitch);
 694             for (y = 1 ; y < This->currentDesc.Height; y++) {
 695                 /* skip the first row */
 696                 src_data += src_pitch;
 697                 dst_data += dst_pitch;
 698                 memcpy(dst_data, src_data, dst_pitch);
 699             }
 700
 701             HeapFree(GetProcessHeap(), 0, mem);
 702         }
 703     }
 704
 705     /* Surface has now been downloaded */
 706     This->flags |= SFLAG_INSYSMEM;
 707 }
 708
 709 /* This call just uploads data, the caller is responsible for binding the
 710  * correct texture. */
 711 /* Context activation is done by the caller. */
 712 static void surface_upload_data(IWineD3DSurfaceImpl *This, const struct wined3d_gl_info *gl_info,
 713         const struct wined3d_format *format, BOOL srgb, const GLvoid *data)
 714 {
 715     GLsizei width = This->currentDesc.Width;
 716     GLsizei height = This->currentDesc.Height;
 717     GLenum internal;
 718
 719     if (srgb)
 720     {
 721         internal = format->glGammaInternal;
 722     }
 723     else if (This->resource.usage & WINED3DUSAGE_RENDERTARGET && surface_is_offscreen(This))
 724     {
 725         internal = format->rtInternal;
 726     }
 727     else
 728     {
 729         internal = format->glInternal;
 730     }
 731
 732     TRACE("This %p, internal %#x, width %d, height %d, format %#x, type %#x, data %p.\n",
 733             This, internal, width, height, format->glFormat, format->glType, data);
 734     TRACE("target %#x, level %u, resource size %u.\n",
 735             This->texture_target, This->texture_level, This->resource.size);
 736
 737     if (format->heightscale != 1.0f && format->heightscale != 0.0f) height *= format->heightscale;
 738
 739     ENTER_GL();
 740
 741     if (This->flags & SFLAG_PBO)
 742     {
 743         GL_EXTCALL(glBindBufferARB(GL_PIXEL_UNPACK_BUFFER_ARB, This->pbo));
 744         checkGLcall("glBindBufferARB");
 745
 746         TRACE("(%p) pbo: %#x, data: %p.\n", This, This->pbo, data);
 747         data = NULL;
 748     }
 749
 750     if (format->flags & WINED3DFMT_FLAG_COMPRESSED)
 751     {
 752         TRACE("Calling glCompressedTexSubImage2DARB.\n");
 753
 754         GL_EXTCALL(glCompressedTexSubImage2DARB(This->texture_target, This->texture_level,
 755                 0, 0, width, height, internal, This->resource.size, data));
 756         checkGLcall("glCompressedTexSubImage2DARB");
 757     }
 758     else
 759     {
 760         TRACE("Calling glTexSubImage2D.\n");
 761
 762         glTexSubImage2D(This->texture_target, This->texture_level,
 763                 0, 0, width, height, format->glFormat, format->glType, data);
 764         checkGLcall("glTexSubImage2D");
 765     }
 766
 767     if (This->flags & SFLAG_PBO)
 768     {
 769         GL_EXTCALL(glBindBufferARB(GL_PIXEL_UNPACK_BUFFER_ARB, 0));
 770         checkGLcall("glBindBufferARB");
 771     }
 772
 773     LEAVE_GL();
 774
 775     if (gl_info->quirks & WINED3D_QUIRK_FBO_TEX_UPDATE)
 776     {
 777         IWineD3DDeviceImpl *device = This->resource.device;
 778         unsigned int i;
 779
 780         for (i = 0; i < device->numContexts; ++i)
 781         {
 782             context_surface_update(device->contexts[i], This);
 783         }
 784     }
 785 }
 786
 787 /* This call just allocates the texture, the caller is responsible for binding
 788  * the correct texture. */
 789 /* Context activation is done by the caller. */
 790 static void surface_allocate_surface(IWineD3DSurfaceImpl *This, const struct wined3d_gl_info *gl_info,
 791         const struct wined3d_format *format, BOOL srgb)
 792 {
 793     BOOL enable_client_storage = FALSE;
 794     GLsizei width = This->pow2Width;
 795     GLsizei height = This->pow2Height;
 796     const BYTE *mem = NULL;
 797     GLenum internal;
 798
 799     if (srgb)
 800     {
 801         internal = format->glGammaInternal;
 802     }
 803     else if (This->resource.usage & WINED3DUSAGE_RENDERTARGET && surface_is_offscreen(This))
 804     {
 805         internal = format->rtInternal;
 806     }
 807     else
 808     {
 809         internal = format->glInternal;
 810     }
 811
 812     if (format->heightscale != 1.0f && format->heightscale != 0.0f) height *= format->heightscale;
 813
 814     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",
 815             This, This->texture_target, This->texture_level, debug_d3dformat(format->id),
 816             internal, width, height, format->glFormat, format->glType);
 817
 818     ENTER_GL();
 819
 820     if (gl_info->supported[APPLE_CLIENT_STORAGE])
 821     {
 822         if (This->flags & (SFLAG_NONPOW2 | SFLAG_DIBSECTION | SFLAG_CONVERTED)
 823                 || !This->resource.allocatedMemory)
 824         {
 825             /* In some cases we want to disable client storage.
 826              * SFLAG_NONPOW2 has a bigger opengl texture than the client memory, and different pitches
 827              * SFLAG_DIBSECTION: Dibsections may have read / write protections on the memory. Avoid issues...
 828              * SFLAG_CONVERTED: The conversion destination memory is freed after loading the surface
 829              * allocatedMemory == NULL: Not defined in the extension. Seems to disable client storage effectively
 830              */
 831             glPixelStorei(GL_UNPACK_CLIENT_STORAGE_APPLE, GL_FALSE);
 832             checkGLcall("glPixelStorei(GL_UNPACK_CLIENT_STORAGE_APPLE, GL_FALSE)");
 833             This->flags &= ~SFLAG_CLIENT;
 834             enable_client_storage = TRUE;
 835         }
 836         else
 837         {
 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->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->glFormat, format->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->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 *format = surface->resource.format;
 990     IWineD3DDeviceImpl *device = surface->resource.device;
 991
 992     switch (format->id)
 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->id));
1034             return FALSE;
1035     }
1036
1037     return TRUE;
1038 }
1039
1040 /* Do not call while under the GL lock. */
1041 static ULONG WINAPI IWineD3DSurfaceImpl_Release(IWineD3DSurface *iface)
1042 {
1043     IWineD3DSurfaceImpl *This = (IWineD3DSurfaceImpl *)iface;
1044     ULONG ref = InterlockedDecrement(&This->resource.ref);
1045     TRACE("(%p) : Releasing from %d\n", This, ref + 1);
1046
1047     if (!ref)
1048     {
1049         surface_cleanup(This);
1050         This->resource.parent_ops->wined3d_object_destroyed(This->resource.parent);
1051
1052         TRACE("(%p) Released.\n", This);
1053         HeapFree(GetProcessHeap(), 0, This);
1054     }
1055
1056     return ref;
1057 }
1058
1059 /* ****************************************************
1060    IWineD3DSurface IWineD3DResource parts follow
1061    **************************************************** */
1062
1063 /* Do not call while under the GL lock. */
1064 void surface_internal_preload(IWineD3DSurfaceImpl *surface, enum WINED3DSRGB srgb)
1065 {
1066     IWineD3DDeviceImpl *device = surface->resource.device;
1067
1068     TRACE("iface %p, srgb %#x.\n", surface, srgb);
1069
1070     if (surface->container.type == WINED3D_CONTAINER_TEXTURE)
1071     {
1072         IWineD3DBaseTextureImpl *texture = surface->container.u.texture;
1073
1074         TRACE("Passing to container.\n");
1075         texture->baseTexture.internal_preload((IWineD3DBaseTexture *)texture, srgb);
1076     }
1077     else
1078     {
1079         struct wined3d_context *context = NULL;
1080
1081         TRACE("(%p) : About to load surface\n", surface);
1082
1083         if (!device->isInDraw) context = context_acquire(device, NULL);
1084
1085         if (surface->resource.format->id == WINED3DFMT_P8_UINT
1086                 || surface->resource.format->id == WINED3DFMT_P8_UINT_A8_UNORM)
1087         {
1088             if (palette9_changed(surface))
1089             {
1090                 TRACE("Reloading surface because the d3d8/9 palette was changed\n");
1091                 /* TODO: This is not necessarily needed with hw palettized texture support */
1092                 surface_load_location(surface, SFLAG_INSYSMEM, NULL);
1093                 /* Make sure the texture is reloaded because of the palette change, this kills performance though :( */
1094                 surface_modify_location(surface, SFLAG_INTEXTURE, FALSE);
1095             }
1096         }
1097
1098         IWineD3DSurface_LoadTexture((IWineD3DSurface *)surface, srgb == SRGB_SRGB ? TRUE : FALSE);
1099
1100         if (surface->resource.pool == WINED3DPOOL_DEFAULT)
1101         {
1102             /* Tell opengl to try and keep this texture in video ram (well mostly) */
1103             GLclampf tmp;
1104             tmp = 0.9f;
1105             ENTER_GL();
1106             glPrioritizeTextures(1, &surface->texture_name, &tmp);
1107             LEAVE_GL();
1108         }
1109
1110         if (context) context_release(context);
1111     }
1112 }
1113
1114 static void WINAPI IWineD3DSurfaceImpl_PreLoad(IWineD3DSurface *iface)
1115 {
1116     surface_internal_preload((IWineD3DSurfaceImpl *)iface, SRGB_ANY);
1117 }
1118
1119 /* Context activation is done by the caller. */
1120 static void surface_remove_pbo(IWineD3DSurfaceImpl *This, const struct wined3d_gl_info *gl_info)
1121 {
1122     This->resource.heapMemory = HeapAlloc(GetProcessHeap() ,0 , This->resource.size + RESOURCE_ALIGNMENT);
1123     This->resource.allocatedMemory =
1124             (BYTE *)(((ULONG_PTR) This->resource.heapMemory + (RESOURCE_ALIGNMENT - 1)) & ~(RESOURCE_ALIGNMENT - 1));
1125
1126     ENTER_GL();
1127     GL_EXTCALL(glBindBufferARB(GL_PIXEL_UNPACK_BUFFER_ARB, This->pbo));
1128     checkGLcall("glBindBufferARB(GL_PIXEL_UNPACK_BUFFER, This->pbo)");
1129     GL_EXTCALL(glGetBufferSubDataARB(GL_PIXEL_UNPACK_BUFFER_ARB, 0, This->resource.size, This->resource.allocatedMemory));
1130     checkGLcall("glGetBufferSubDataARB");
1131     GL_EXTCALL(glDeleteBuffersARB(1, &This->pbo));
1132     checkGLcall("glDeleteBuffersARB");
1133     LEAVE_GL();
1134
1135     This->pbo = 0;
1136     This->flags &= ~SFLAG_PBO;
1137 }
1138
1139 BOOL surface_init_sysmem(IWineD3DSurfaceImpl *surface)
1140 {
1141     if (!surface->resource.allocatedMemory)
1142     {
1143         surface->resource.heapMemory = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY,
1144                 surface->resource.size + RESOURCE_ALIGNMENT);
1145         if (!surface->resource.heapMemory)
1146         {
1147             ERR("Out of memory\n");
1148             return FALSE;
1149         }
1150         surface->resource.allocatedMemory =
1151             (BYTE *)(((ULONG_PTR)surface->resource.heapMemory + (RESOURCE_ALIGNMENT - 1)) & ~(RESOURCE_ALIGNMENT - 1));
1152     }
1153     else
1154     {
1155         memset(surface->resource.allocatedMemory, 0, surface->resource.size);
1156     }
1157
1158     surface_modify_location(surface, SFLAG_INSYSMEM, TRUE);
1159
1160     return TRUE;
1161 }
1162
1163 /* Do not call while under the GL lock. */
1164 static void WINAPI IWineD3DSurfaceImpl_UnLoad(IWineD3DSurface *iface)
1165 {
1166     IWineD3DSurfaceImpl *This = (IWineD3DSurfaceImpl *) iface;
1167     IWineD3DDeviceImpl *device = This->resource.device;
1168     const struct wined3d_gl_info *gl_info;
1169     renderbuffer_entry_t *entry, *entry2;
1170     struct wined3d_context *context;
1171
1172     TRACE("(%p)\n", iface);
1173
1174     if(This->resource.pool == WINED3DPOOL_DEFAULT) {
1175         /* Default pool resources are supposed to be destroyed before Reset is called.
1176          * Implicit resources stay however. So this means we have an implicit render target
1177          * or depth stencil. The content may be destroyed, but we still have to tear down
1178          * opengl resources, so we cannot leave early.
1179          *
1180          * Put the surfaces into sysmem, and reset the content. The D3D content is undefined,
1181          * but we can't set the sysmem INDRAWABLE because when we're rendering the swapchain
1182          * or the depth stencil into an FBO the texture or render buffer will be removed
1183          * and all flags get lost
1184          */
1185         surface_init_sysmem(This);
1186     }
1187     else
1188     {
1189         /* Load the surface into system memory */
1190         surface_load_location(This, SFLAG_INSYSMEM, NULL);
1191         surface_modify_location(This, SFLAG_INDRAWABLE, FALSE);
1192     }
1193     surface_modify_location(This, SFLAG_INTEXTURE, FALSE);
1194     surface_modify_location(This, SFLAG_INSRGBTEX, FALSE);
1195     This->flags &= ~(SFLAG_ALLOCATED | SFLAG_SRGBALLOCATED);
1196
1197     context = context_acquire(device, NULL);
1198     gl_info = context->gl_info;
1199
1200     /* Destroy PBOs, but load them into real sysmem before */
1201     if (This->flags & SFLAG_PBO)
1202         surface_remove_pbo(This, gl_info);
1203
1204     /* Destroy fbo render buffers. This is needed for implicit render targets, for
1205      * all application-created targets the application has to release the surface
1206      * before calling _Reset
1207      */
1208     LIST_FOR_EACH_ENTRY_SAFE(entry, entry2, &This->renderbuffers, renderbuffer_entry_t, entry) {
1209         ENTER_GL();
1210         gl_info->fbo_ops.glDeleteRenderbuffers(1, &entry->id);
1211         LEAVE_GL();
1212         list_remove(&entry->entry);
1213         HeapFree(GetProcessHeap(), 0, entry);
1214     }
1215     list_init(&This->renderbuffers);
1216     This->current_renderbuffer = NULL;
1217
1218     /* If we're in a texture, the texture name belongs to the texture.
1219      * Otherwise, destroy it. */
1220     if (This->container.type != WINED3D_CONTAINER_TEXTURE)
1221     {
1222         ENTER_GL();
1223         glDeleteTextures(1, &This->texture_name);
1224         This->texture_name = 0;
1225         glDeleteTextures(1, &This->texture_name_srgb);
1226         This->texture_name_srgb = 0;
1227         LEAVE_GL();
1228     }
1229
1230     context_release(context);
1231
1232     resource_unload((IWineD3DResourceImpl *)This);
1233 }
1234
1235 /* ******************************************************
1236    IWineD3DSurface IWineD3DSurface parts follow
1237    ****************************************************** */
1238
1239 /* Read the framebuffer back into the surface */
1240 static void read_from_framebuffer(IWineD3DSurfaceImpl *This, const RECT *rect, void *dest, UINT pitch)
1241 {
1242     IWineD3DDeviceImpl *device = This->resource.device;
1243     const struct wined3d_gl_info *gl_info;
1244     struct wined3d_context *context;
1245     BYTE *mem;
1246     GLint fmt;
1247     GLint type;
1248     BYTE *row, *top, *bottom;
1249     int i;
1250     BOOL bpp;
1251     RECT local_rect;
1252     BOOL srcIsUpsideDown;
1253     GLint rowLen = 0;
1254     GLint skipPix = 0;
1255     GLint skipRow = 0;
1256
1257     if(wined3d_settings.rendertargetlock_mode == RTL_DISABLE) {
1258         static BOOL warned = FALSE;
1259         if(!warned) {
1260             ERR("The application tries to lock the render target, but render target locking is disabled\n");
1261             warned = TRUE;
1262         }
1263         return;
1264     }
1265
1266     context = context_acquire(device, This);
1267     context_apply_blit_state(context, device);
1268     gl_info = context->gl_info;
1269
1270     ENTER_GL();
1271
1272     /* Select the correct read buffer, and give some debug output.
1273      * There is no need to keep track of the current read buffer or reset it, every part of the code
1274      * that reads sets the read buffer as desired.
1275      */
1276     if (surface_is_offscreen(This))
1277     {
1278         /* Mapping the primary render target which is not on a swapchain.
1279          * Read from the back buffer. */
1280         TRACE("Mapping offscreen render target.\n");
1281         glReadBuffer(device->offscreenBuffer);
1282         srcIsUpsideDown = TRUE;
1283     }
1284     else
1285     {
1286         /* Onscreen surfaces are always part of a swapchain */
1287         GLenum buffer = surface_get_gl_buffer(This);
1288         TRACE("Mapping %#x buffer.\n", buffer);
1289         glReadBuffer(buffer);
1290         checkGLcall("glReadBuffer");
1291         srcIsUpsideDown = FALSE;
1292     }
1293
1294     /* TODO: Get rid of the extra rectangle comparison and construction of a full surface rectangle */
1295     if(!rect) {
1296         local_rect.left = 0;
1297         local_rect.top = 0;
1298         local_rect.right = This->currentDesc.Width;
1299         local_rect.bottom = This->currentDesc.Height;
1300     } else {
1301         local_rect = *rect;
1302     }
1303     /* TODO: Get rid of the extra GetPitch call, LockRect does that too. Cache the pitch */
1304
1305     switch (This->resource.format->id)
1306     {
1307         case WINED3DFMT_P8_UINT:
1308         {
1309             if (primary_render_target_is_p8(device))
1310             {
1311                 /* In case of P8 render targets the index is stored in the alpha component */
1312                 fmt = GL_ALPHA;
1313                 type = GL_UNSIGNED_BYTE;
1314                 mem = dest;
1315                 bpp = This->resource.format->byte_count;
1316             } else {
1317                 /* GL can't return palettized data, so read ARGB pixels into a
1318                  * separate block of memory and convert them into palettized format
1319                  * in software. Slow, but if the app means to use palettized render
1320                  * targets and locks it...
1321                  *
1322                  * Use GL_RGB, GL_UNSIGNED_BYTE to read the surface for performance reasons
1323                  * Don't use GL_BGR as in the WINED3DFMT_R8G8B8 case, instead watch out
1324                  * for the color channels when palettizing the colors.
1325                  */
1326                 fmt = GL_RGB;
1327                 type = GL_UNSIGNED_BYTE;
1328                 pitch *= 3;
1329                 mem = HeapAlloc(GetProcessHeap(), 0, This->resource.size * 3);
1330                 if(!mem) {
1331                     ERR("Out of memory\n");
1332                     LEAVE_GL();
1333                     return;
1334                 }
1335                 bpp = This->resource.format->byte_count * 3;
1336             }
1337         }
1338         break;
1339
1340         default:
1341             mem = dest;
1342             fmt = This->resource.format->glFormat;
1343             type = This->resource.format->glType;
1344             bpp = This->resource.format->byte_count;
1345     }
1346
1347     if (This->flags & SFLAG_PBO)
1348     {
1349         GL_EXTCALL(glBindBufferARB(GL_PIXEL_PACK_BUFFER_ARB, This->pbo));
1350         checkGLcall("glBindBufferARB");
1351         if (mem)
1352         {
1353             ERR("mem not null for pbo -- unexpected\n");
1354             mem = NULL;
1355         }
1356     }
1357
1358     /* Save old pixel store pack state */
1359     glGetIntegerv(GL_PACK_ROW_LENGTH, &rowLen);
1360     checkGLcall("glGetIntegerv");
1361     glGetIntegerv(GL_PACK_SKIP_PIXELS, &skipPix);
1362     checkGLcall("glGetIntegerv");
1363     glGetIntegerv(GL_PACK_SKIP_ROWS, &skipRow);
1364     checkGLcall("glGetIntegerv");
1365
1366     /* Setup pixel store pack state -- to glReadPixels into the correct place */
1367     glPixelStorei(GL_PACK_ROW_LENGTH, This->currentDesc.Width);
1368     checkGLcall("glPixelStorei");
1369     glPixelStorei(GL_PACK_SKIP_PIXELS, local_rect.left);
1370     checkGLcall("glPixelStorei");
1371     glPixelStorei(GL_PACK_SKIP_ROWS, local_rect.top);
1372     checkGLcall("glPixelStorei");
1373
1374     glReadPixels(local_rect.left, (!srcIsUpsideDown) ? (This->currentDesc.Height - local_rect.bottom) : local_rect.top ,
1375                  local_rect.right - local_rect.left,
1376                  local_rect.bottom - local_rect.top,
1377                  fmt, type, mem);
1378     checkGLcall("glReadPixels");
1379
1380     /* Reset previous pixel store pack state */
1381     glPixelStorei(GL_PACK_ROW_LENGTH, rowLen);
1382     checkGLcall("glPixelStorei");
1383     glPixelStorei(GL_PACK_SKIP_PIXELS, skipPix);
1384     checkGLcall("glPixelStorei");
1385     glPixelStorei(GL_PACK_SKIP_ROWS, skipRow);
1386     checkGLcall("glPixelStorei");
1387
1388     if (This->flags & SFLAG_PBO)
1389     {
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->id == 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         {
1436             GL_EXTCALL(glUnmapBufferARB(GL_PIXEL_UNPACK_BUFFER_ARB));
1437             GL_EXTCALL(glBindBufferARB(GL_PIXEL_UNPACK_BUFFER_ARB, 0));
1438         }
1439     }
1440
1441     LEAVE_GL();
1442     context_release(context);
1443
1444     /* For P8 textures we need to perform an inverse palette lookup. This is done by searching for a palette
1445      * index which matches the RGB value. Note this isn't guaranteed to work when there are multiple entries for
1446      * the same color but we have no choice.
1447      * In case of P8 render targets, the index is stored in the alpha component so no conversion is needed.
1448      */
1449     if (This->resource.format->id == WINED3DFMT_P8_UINT && !primary_render_target_is_p8(device))
1450     {
1451         const PALETTEENTRY *pal = NULL;
1452         DWORD width = pitch / 3;
1453         int x, y, c;
1454
1455         if(This->palette) {
1456             pal = This->palette->palents;
1457         } else {
1458             ERR("Palette is missing, cannot perform inverse palette lookup\n");
1459             HeapFree(GetProcessHeap(), 0, mem);
1460             return ;
1461         }
1462
1463         for(y = local_rect.top; y < local_rect.bottom; y++) {
1464             for(x = local_rect.left; x < local_rect.right; x++) {
1465                 /*                      start              lines            pixels      */
1466                 const BYTE *blue = mem + y * pitch + x * (sizeof(BYTE) * 3);
1467                 const BYTE *green = blue  + 1;
1468                 const BYTE *red = green + 1;
1469
1470                 for(c = 0; c < 256; c++) {
1471                     if(*red   == pal[c].peRed   &&
1472                        *green == pal[c].peGreen &&
1473                        *blue  == pal[c].peBlue)
1474                     {
1475                         *((BYTE *) dest + y * width + x) = c;
1476                         break;
1477                     }
1478                 }
1479             }
1480         }
1481         HeapFree(GetProcessHeap(), 0, mem);
1482     }
1483 }
1484
1485 /* Read the framebuffer contents into a texture */
1486 static void read_from_framebuffer_texture(IWineD3DSurfaceImpl *This, BOOL srgb)
1487 {
1488     IWineD3DDeviceImpl *device = This->resource.device;
1489     const struct wined3d_gl_info *gl_info;
1490     struct wined3d_context *context;
1491
1492     if (!surface_is_offscreen(This))
1493     {
1494         /* We would need to flip onscreen surfaces, but there's no efficient
1495          * way to do that here. It makes more sense for the caller to
1496          * explicitly go through sysmem. */
1497         ERR("Not supported for onscreen targets.\n");
1498         return;
1499     }
1500
1501     /* Activate the surface to read from. In some situations it isn't the currently active target(e.g. backbuffer
1502      * locking during offscreen rendering). RESOURCELOAD is ok because glCopyTexSubImage2D isn't affected by any
1503      * states in the stateblock, and no driver was found yet that had bugs in that regard.
1504      */
1505     context = context_acquire(device, This);
1506     gl_info = context->gl_info;
1507
1508     surface_prepare_texture(This, gl_info, srgb);
1509     surface_bind_and_dirtify(This, srgb);
1510
1511     TRACE("Reading back offscreen render target %p.\n", This);
1512
1513     ENTER_GL();
1514
1515     glReadBuffer(device->offscreenBuffer);
1516     checkGLcall("glReadBuffer");
1517
1518     glCopyTexSubImage2D(This->texture_target, This->texture_level,
1519             0, 0, 0, 0, This->currentDesc.Width, This->currentDesc.Height);
1520     checkGLcall("glCopyTexSubImage2D");
1521
1522     LEAVE_GL();
1523
1524     context_release(context);
1525 }
1526
1527 /* Context activation is done by the caller. */
1528 static void surface_prepare_texture_internal(IWineD3DSurfaceImpl *surface,
1529         const struct wined3d_gl_info *gl_info, BOOL srgb)
1530 {
1531     DWORD alloc_flag = srgb ? SFLAG_SRGBALLOCATED : SFLAG_ALLOCATED;
1532     CONVERT_TYPES convert;
1533     struct wined3d_format format;
1534
1535     if (surface->flags & alloc_flag) return;
1536
1537     d3dfmt_get_conv(surface, TRUE, TRUE, &format, &convert);
1538     if (convert != NO_CONVERSION || format.convert) surface->flags |= SFLAG_CONVERTED;
1539     else surface->flags &= ~SFLAG_CONVERTED;
1540
1541     surface_bind_and_dirtify(surface, srgb);
1542     surface_allocate_surface(surface, gl_info, &format, srgb);
1543     surface->flags |= alloc_flag;
1544 }
1545
1546 /* Context activation is done by the caller. */
1547 void surface_prepare_texture(IWineD3DSurfaceImpl *surface, const struct wined3d_gl_info *gl_info, BOOL srgb)
1548 {
1549     if (surface->container.type == WINED3D_CONTAINER_TEXTURE)
1550     {
1551         IWineD3DBaseTextureImpl *texture = surface->container.u.texture;
1552         UINT sub_count = texture->baseTexture.level_count * texture->baseTexture.layer_count;
1553         UINT i;
1554
1555         TRACE("surface %p is a subresource of texture %p.\n", surface, texture);
1556
1557         for (i = 0; i < sub_count; ++i)
1558         {
1559             IWineD3DSurfaceImpl *s = (IWineD3DSurfaceImpl *)texture->baseTexture.sub_resources[i];
1560             surface_prepare_texture_internal(s, gl_info, srgb);
1561         }
1562
1563         return;
1564     }
1565
1566     surface_prepare_texture_internal(surface, gl_info, srgb);
1567 }
1568
1569 static void surface_prepare_system_memory(IWineD3DSurfaceImpl *This)
1570 {
1571     IWineD3DDeviceImpl *device = This->resource.device;
1572     const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
1573
1574     /* Performance optimization: Count how often a surface is locked, if it is locked regularly do not throw away the system memory copy.
1575      * This avoids the need to download the surface from opengl all the time. The surface is still downloaded if the opengl texture is
1576      * changed
1577      */
1578     if (!(This->flags & SFLAG_DYNLOCK))
1579     {
1580         This->lockCount++;
1581         /* MAXLOCKCOUNT is defined in wined3d_private.h */
1582         if(This->lockCount > MAXLOCKCOUNT) {
1583             TRACE("Surface is locked regularly, not freeing the system memory copy any more\n");
1584             This->flags |= SFLAG_DYNLOCK;
1585         }
1586     }
1587
1588     /* Create a PBO for dynamically locked surfaces but don't do it for converted or non-pow2 surfaces.
1589      * Also don't create a PBO for systemmem surfaces.
1590      */
1591     if (gl_info->supported[ARB_PIXEL_BUFFER_OBJECT] && (This->flags & SFLAG_DYNLOCK)
1592             && !(This->flags & (SFLAG_PBO | SFLAG_CONVERTED | SFLAG_NONPOW2))
1593             && (This->resource.pool != WINED3DPOOL_SYSTEMMEM))
1594     {
1595         GLenum error;
1596         struct wined3d_context *context;
1597
1598         context = context_acquire(device, NULL);
1599         ENTER_GL();
1600
1601         GL_EXTCALL(glGenBuffersARB(1, &This->pbo));
1602         error = glGetError();
1603         if (!This->pbo || error != GL_NO_ERROR)
1604             ERR("Failed to bind the PBO with error %s (%#x)\n", debug_glerror(error), error);
1605
1606         TRACE("Attaching pbo=%#x to (%p)\n", This->pbo, This);
1607
1608         GL_EXTCALL(glBindBufferARB(GL_PIXEL_UNPACK_BUFFER_ARB, This->pbo));
1609         checkGLcall("glBindBufferARB");
1610
1611         GL_EXTCALL(glBufferDataARB(GL_PIXEL_UNPACK_BUFFER_ARB, This->resource.size + 4, This->resource.allocatedMemory, GL_STREAM_DRAW_ARB));
1612         checkGLcall("glBufferDataARB");
1613
1614         GL_EXTCALL(glBindBufferARB(GL_PIXEL_UNPACK_BUFFER_ARB, 0));
1615         checkGLcall("glBindBufferARB");
1616
1617         /* We don't need the system memory anymore and we can't even use it for PBOs */
1618         if (!(This->flags & SFLAG_CLIENT))
1619         {
1620             HeapFree(GetProcessHeap(), 0, This->resource.heapMemory);
1621             This->resource.heapMemory = NULL;
1622         }
1623         This->resource.allocatedMemory = NULL;
1624         This->flags |= SFLAG_PBO;
1625         LEAVE_GL();
1626         context_release(context);
1627     }
1628     else if (!(This->resource.allocatedMemory || This->flags & SFLAG_PBO))
1629     {
1630         /* Whatever surface we have, make sure that there is memory allocated for the downloaded copy,
1631          * or a pbo to map
1632          */
1633         if(!This->resource.heapMemory) {
1634             This->resource.heapMemory = HeapAlloc(GetProcessHeap() ,0 , This->resource.size + RESOURCE_ALIGNMENT);
1635         }
1636         This->resource.allocatedMemory =
1637                 (BYTE *)(((ULONG_PTR) This->resource.heapMemory + (RESOURCE_ALIGNMENT - 1)) & ~(RESOURCE_ALIGNMENT - 1));
1638         if (This->flags & SFLAG_INSYSMEM)
1639         {
1640             ERR("Surface without memory or pbo has SFLAG_INSYSMEM set!\n");
1641         }
1642     }
1643 }
1644
1645 static HRESULT WINAPI IWineD3DSurfaceImpl_Map(IWineD3DSurface *iface,
1646         WINED3DLOCKED_RECT *pLockedRect, const RECT *pRect, DWORD Flags)
1647 {
1648     IWineD3DSurfaceImpl *This = (IWineD3DSurfaceImpl *)iface;
1649     IWineD3DDeviceImpl *device = This->resource.device;
1650     const RECT *pass_rect = pRect;
1651
1652     TRACE("iface %p, locked_rect %p, rect %s, flags %#x.\n",
1653             iface, pLockedRect, wine_dbgstr_rect(pRect), Flags);
1654
1655     /* This is also done in the base class, but we have to verify this before loading any data from
1656      * gl into the sysmem copy. The PBO may be mapped, a different rectangle locked, the discard flag
1657      * may interfere, and all other bad things may happen
1658      */
1659     if (This->flags & SFLAG_LOCKED)
1660     {
1661         WARN("Surface is already locked, returning D3DERR_INVALIDCALL\n");
1662         return WINED3DERR_INVALIDCALL;
1663     }
1664     This->flags |= SFLAG_LOCKED;
1665
1666     if (!(This->flags & SFLAG_LOCKABLE))
1667     {
1668         TRACE("Warning: trying to lock unlockable surf@%p\n", This);
1669     }
1670
1671     if (Flags & WINED3DLOCK_DISCARD)
1672     {
1673         TRACE("WINED3DLOCK_DISCARD flag passed, marking SYSMEM as up to date.\n");
1674         surface_prepare_system_memory(This);
1675         surface_modify_location(This, SFLAG_INSYSMEM, TRUE);
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 && !pRect->left
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_Map(iface, pLockedRect, pRect, Flags);
1742 }
1743
1744 static void flush_to_framebuffer_drawpixels(IWineD3DSurfaceImpl *This,
1745         GLenum fmt, GLenum type, UINT bpp, const BYTE *mem)
1746 {
1747     UINT pitch = IWineD3DSurface_GetPitch((IWineD3DSurface *) This);    /* target is argb, 4 byte */
1748     IWineD3DDeviceImpl *device = This->resource.device;
1749     const struct wined3d_gl_info *gl_info;
1750     struct wined3d_context *context;
1751     RECT rect;
1752     UINT w, h;
1753
1754     if (This->flags & SFLAG_LOCKED)
1755         rect = This->lockedRect;
1756     else
1757         SetRect(&rect, 0, 0, This->currentDesc.Width, This->currentDesc.Height);
1758
1759     mem += rect.top * pitch + rect.left * bpp;
1760     w = rect.right - rect.left;
1761     h = rect.bottom - rect.top;
1762
1763     /* Activate the correct context for the render target */
1764     context = context_acquire(device, This);
1765     context_apply_blit_state(context, device);
1766     gl_info = context->gl_info;
1767
1768     ENTER_GL();
1769
1770     if (!surface_is_offscreen(This))
1771     {
1772         GLenum buffer = surface_get_gl_buffer(This);
1773         TRACE("Unlocking %#x buffer.\n", buffer);
1774         context_set_draw_buffer(context, buffer);
1775
1776         surface_translate_drawable_coords(This, context->win_handle, &rect);
1777         glPixelZoom(1.0f, -1.0f);
1778     }
1779     else
1780     {
1781         /* Primary offscreen render target */
1782         TRACE("Offscreen render target.\n");
1783         context_set_draw_buffer(context, device->offscreenBuffer);
1784
1785         glPixelZoom(1.0f, 1.0f);
1786     }
1787
1788     glRasterPos3i(rect.left, rect.top, 1);
1789     checkGLcall("glRasterPos3i");
1790
1791     /* Some drivers(radeon dri, others?) don't like exceptions during
1792      * glDrawPixels. If the surface is a DIB section, it might be in GDIMode
1793      * after ReleaseDC. Reading it will cause an exception, which x11drv will
1794      * catch to put the dib section in InSync mode, which leads to a crash
1795      * and a blocked x server on my radeon card.
1796      *
1797      * The following lines read the dib section so it is put in InSync mode
1798      * before glDrawPixels is called and the crash is prevented. There won't
1799      * be any interfering gdi accesses, because UnlockRect is called from
1800      * ReleaseDC, and the app won't use the dc any more afterwards.
1801      */
1802     if ((This->flags & SFLAG_DIBSECTION) && !(This->flags & SFLAG_PBO))
1803     {
1804         volatile BYTE read;
1805         read = This->resource.allocatedMemory[0];
1806     }
1807
1808     /* If not fullscreen, we need to skip a number of bytes to find the next row of data */
1809     glPixelStorei(GL_UNPACK_ROW_LENGTH, This->currentDesc.Width);
1810
1811     if (This->flags & SFLAG_PBO)
1812     {
1813         GL_EXTCALL(glBindBufferARB(GL_PIXEL_UNPACK_BUFFER_ARB, This->pbo));
1814         checkGLcall("glBindBufferARB");
1815     }
1816
1817     glDrawPixels(w, h, fmt, type, mem);
1818     checkGLcall("glDrawPixels");
1819
1820     if (This->flags & SFLAG_PBO)
1821     {
1822         GL_EXTCALL(glBindBufferARB(GL_PIXEL_UNPACK_BUFFER_ARB, 0));
1823         checkGLcall("glBindBufferARB");
1824     }
1825
1826     glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
1827     checkGLcall("glPixelStorei(GL_UNPACK_ROW_LENGTH, 0)");
1828
1829     LEAVE_GL();
1830     context_release(context);
1831 }
1832
1833 static HRESULT WINAPI IWineD3DSurfaceImpl_Unmap(IWineD3DSurface *iface)
1834 {
1835     IWineD3DSurfaceImpl *This = (IWineD3DSurfaceImpl *)iface;
1836     IWineD3DDeviceImpl *device = This->resource.device;
1837     BOOL fullsurface;
1838
1839     if (!(This->flags & SFLAG_LOCKED))
1840     {
1841         WARN("trying to Unlock an unlocked surf@%p\n", This);
1842         return WINEDDERR_NOTLOCKED;
1843     }
1844
1845     if (This->flags & SFLAG_PBO)
1846     {
1847         const struct wined3d_gl_info *gl_info;
1848         struct wined3d_context *context;
1849
1850         TRACE("Freeing PBO memory\n");
1851
1852         context = context_acquire(device, NULL);
1853         gl_info = context->gl_info;
1854
1855         ENTER_GL();
1856         GL_EXTCALL(glBindBufferARB(GL_PIXEL_UNPACK_BUFFER_ARB, This->pbo));
1857         GL_EXTCALL(glUnmapBufferARB(GL_PIXEL_UNPACK_BUFFER_ARB));
1858         GL_EXTCALL(glBindBufferARB(GL_PIXEL_UNPACK_BUFFER_ARB, 0));
1859         checkGLcall("glUnmapBufferARB");
1860         LEAVE_GL();
1861         context_release(context);
1862
1863         This->resource.allocatedMemory = NULL;
1864     }
1865
1866     TRACE("(%p) : dirtyfied(%d)\n", This, This->flags & (SFLAG_INDRAWABLE | SFLAG_INTEXTURE) ? 0 : 1);
1867
1868     if (This->flags & (SFLAG_INDRAWABLE | SFLAG_INTEXTURE))
1869     {
1870         TRACE("(%p) : Not Dirtified so nothing to do, return now\n", This);
1871         goto unlock_end;
1872     }
1873
1874     if (This->container.type == WINED3D_CONTAINER_SWAPCHAIN
1875             || (device->render_targets && This == device->render_targets[0]))
1876     {
1877         if(wined3d_settings.rendertargetlock_mode == RTL_DISABLE) {
1878             static BOOL warned = FALSE;
1879             if(!warned) {
1880                 ERR("The application tries to write to the render target, but render target locking is disabled\n");
1881                 warned = TRUE;
1882             }
1883             goto unlock_end;
1884         }
1885
1886         if (!This->dirtyRect.left && !This->dirtyRect.top
1887                 && This->dirtyRect.right == This->currentDesc.Width
1888                 && This->dirtyRect.bottom == This->currentDesc.Height)
1889         {
1890             fullsurface = TRUE;
1891         } else {
1892             /* TODO: Proper partial rectangle tracking */
1893             fullsurface = FALSE;
1894             This->flags |= SFLAG_INSYSMEM;
1895         }
1896
1897         surface_load_location(This, SFLAG_INDRAWABLE, fullsurface ? NULL : &This->dirtyRect);
1898
1899         if (!fullsurface)
1900         {
1901             /* Partial rectangle tracking is not commonly implemented, it is
1902              * only done for render targets. Overwrite the flags to bring
1903              * them back into a sane state. INSYSMEM was set before to tell
1904              * surface_load_location() where to read the rectangle from.
1905              * Indrawable is set because all modifications from the partial
1906              * sysmem copy are written back to the drawable, thus the surface
1907              * is merged again in the drawable. The sysmem copy is not fully
1908              * up to date because only a subrectangle was read in Map(). */
1909             This->flags &= ~SFLAG_INSYSMEM;
1910             This->flags |= SFLAG_INDRAWABLE;
1911         }
1912
1913         This->dirtyRect.left   = This->currentDesc.Width;
1914         This->dirtyRect.top    = This->currentDesc.Height;
1915         This->dirtyRect.right  = 0;
1916         This->dirtyRect.bottom = 0;
1917     }
1918     else if (This->resource.format->flags & (WINED3DFMT_FLAG_DEPTH | WINED3DFMT_FLAG_STENCIL))
1919     {
1920         FIXME("Depth Stencil buffer locking is not implemented\n");
1921     }
1922
1923     unlock_end:
1924     This->flags &= ~SFLAG_LOCKED;
1925     memset(&This->lockedRect, 0, sizeof(RECT));
1926
1927     /* Overlays have to be redrawn manually after changes with the GL implementation */
1928     if(This->overlay_dest) {
1929         IWineD3DSurface_DrawOverlay(iface);
1930     }
1931     return WINED3D_OK;
1932 }
1933
1934 static void surface_release_client_storage(IWineD3DSurfaceImpl *surface)
1935 {
1936     struct wined3d_context *context;
1937
1938     context = context_acquire(surface->resource.device, NULL);
1939
1940     ENTER_GL();
1941     glPixelStorei(GL_UNPACK_CLIENT_STORAGE_APPLE, GL_FALSE);
1942     if (surface->texture_name)
1943     {
1944         surface_bind_and_dirtify(surface, FALSE);
1945         glTexImage2D(surface->texture_target, surface->texture_level,
1946                 GL_RGB, 1, 1, 0, GL_RGB, GL_UNSIGNED_BYTE, NULL);
1947     }
1948     if (surface->texture_name_srgb)
1949     {
1950         surface_bind_and_dirtify(surface, TRUE);
1951         glTexImage2D(surface->texture_target, surface->texture_level,
1952                 GL_RGB, 1, 1, 0, GL_RGB, GL_UNSIGNED_BYTE, NULL);
1953     }
1954     glPixelStorei(GL_UNPACK_CLIENT_STORAGE_APPLE, GL_TRUE);
1955
1956     LEAVE_GL();
1957     context_release(context);
1958
1959     surface_modify_location(surface, SFLAG_INSRGBTEX, FALSE);
1960     surface_modify_location(surface, SFLAG_INTEXTURE, FALSE);
1961     surface_force_reload(surface);
1962 }
1963
1964 static HRESULT WINAPI IWineD3DSurfaceImpl_GetDC(IWineD3DSurface *iface, HDC *pHDC)
1965 {
1966     IWineD3DSurfaceImpl *This = (IWineD3DSurfaceImpl *)iface;
1967     WINED3DLOCKED_RECT lock;
1968     HRESULT hr;
1969     RGBQUAD col[256];
1970
1971     TRACE("(%p)->(%p)\n",This,pHDC);
1972
1973     if (This->flags & SFLAG_USERPTR)
1974     {
1975         ERR("Not supported on surfaces with an application-provided surfaces\n");
1976         return WINEDDERR_NODC;
1977     }
1978
1979     /* Give more detailed info for ddraw */
1980     if (This->flags & SFLAG_DCINUSE)
1981         return WINEDDERR_DCALREADYCREATED;
1982
1983     /* Can't GetDC if the surface is locked */
1984     if (This->flags & SFLAG_LOCKED)
1985         return WINED3DERR_INVALIDCALL;
1986
1987     memset(&lock, 0, sizeof(lock)); /* To be sure */
1988
1989     /* Create a DIB section if there isn't a hdc yet */
1990     if (!This->hDC)
1991     {
1992         if (This->flags & SFLAG_CLIENT)
1993         {
1994             surface_load_location(This, SFLAG_INSYSMEM, NULL);
1995             surface_release_client_storage(This);
1996         }
1997         hr = IWineD3DBaseSurfaceImpl_CreateDIBSection(iface);
1998         if(FAILED(hr)) return WINED3DERR_INVALIDCALL;
1999
2000         /* Use the dib section from now on if we are not using a PBO */
2001         if (!(This->flags & SFLAG_PBO))
2002             This->resource.allocatedMemory = This->dib.bitmap_data;
2003     }
2004
2005     /* Map the surface */
2006     hr = IWineD3DSurface_Map(iface, &lock, NULL, 0);
2007
2008     /* Sync the DIB with the PBO. This can't be done earlier because Map()
2009      * activates the allocatedMemory. */
2010     if (This->flags & SFLAG_PBO)
2011         memcpy(This->dib.bitmap_data, This->resource.allocatedMemory, This->dib.bitmap_size);
2012
2013     if (FAILED(hr))
2014     {
2015         ERR("IWineD3DSurface_Map failed, hr %#x.\n", hr);
2016         /* keep the dib section */
2017         return hr;
2018     }
2019
2020     if (This->resource.format->id == WINED3DFMT_P8_UINT
2021             || This->resource.format->id == WINED3DFMT_P8_UINT_A8_UNORM)
2022     {
2023         /* GetDC on palettized formats is unsupported in D3D9, and the method is missing in
2024             D3D8, so this should only be used for DX <=7 surfaces (with non-device palettes) */
2025         unsigned int n;
2026         const PALETTEENTRY *pal = NULL;
2027
2028         if(This->palette) {
2029             pal = This->palette->palents;
2030         } else {
2031             IWineD3DSurfaceImpl *dds_primary;
2032             IWineD3DSwapChainImpl *swapchain;
2033             swapchain = (IWineD3DSwapChainImpl *)This->resource.device->swapchains[0];
2034             dds_primary = swapchain->front_buffer;
2035             if (dds_primary && dds_primary->palette)
2036                 pal = dds_primary->palette->palents;
2037         }
2038
2039         if (pal) {
2040             for (n=0; n<256; n++) {
2041                 col[n].rgbRed   = pal[n].peRed;
2042                 col[n].rgbGreen = pal[n].peGreen;
2043                 col[n].rgbBlue  = pal[n].peBlue;
2044                 col[n].rgbReserved = 0;
2045             }
2046             SetDIBColorTable(This->hDC, 0, 256, col);
2047         }
2048     }
2049
2050     *pHDC = This->hDC;
2051     TRACE("returning %p\n",*pHDC);
2052     This->flags |= SFLAG_DCINUSE;
2053
2054     return WINED3D_OK;
2055 }
2056
2057 static HRESULT WINAPI IWineD3DSurfaceImpl_ReleaseDC(IWineD3DSurface *iface, HDC hDC)
2058 {
2059     IWineD3DSurfaceImpl *This = (IWineD3DSurfaceImpl *)iface;
2060
2061     TRACE("(%p)->(%p)\n",This,hDC);
2062
2063     if (!(This->flags & SFLAG_DCINUSE))
2064         return WINEDDERR_NODC;
2065
2066     if (This->hDC !=hDC) {
2067         WARN("Application tries to release an invalid DC(%p), surface dc is %p\n", hDC, This->hDC);
2068         return WINEDDERR_NODC;
2069     }
2070
2071     if ((This->flags & SFLAG_PBO) && This->resource.allocatedMemory)
2072     {
2073         /* Copy the contents of the DIB over to the PBO */
2074         memcpy(This->resource.allocatedMemory, This->dib.bitmap_data, This->dib.bitmap_size);
2075     }
2076
2077     /* we locked first, so unlock now */
2078     IWineD3DSurface_Unmap(iface);
2079
2080     This->flags &= ~SFLAG_DCINUSE;
2081
2082     return WINED3D_OK;
2083 }
2084
2085 /* ******************************************************
2086    IWineD3DSurface Internal (No mapping to directx api) parts follow
2087    ****************************************************** */
2088
2089 HRESULT d3dfmt_get_conv(IWineD3DSurfaceImpl *This, BOOL need_alpha_ck,
2090         BOOL use_texturing, struct wined3d_format *format, CONVERT_TYPES *convert)
2091 {
2092     BOOL colorkey_active = need_alpha_ck && (This->CKeyFlags & WINEDDSD_CKSRCBLT);
2093     IWineD3DDeviceImpl *device = This->resource.device;
2094     const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
2095     BOOL blit_supported = FALSE;
2096
2097     /* Copy the default values from the surface. Below we might perform fixups */
2098     /* TODO: get rid of color keying desc fixups by using e.g. a table. */
2099     *format = *This->resource.format;
2100     *convert = NO_CONVERSION;
2101
2102     /* Ok, now look if we have to do any conversion */
2103     switch (This->resource.format->id)
2104     {
2105         case WINED3DFMT_P8_UINT:
2106             /* Below the call to blit_supported is disabled for Wine 1.2
2107              * because the function isn't operating correctly yet. At the
2108              * moment 8-bit blits are handled in software and if certain GL
2109              * extensions are around, surface conversion is performed at
2110              * upload time. The blit_supported call recognizes it as a
2111              * destination fixup. This type of upload 'fixup' and 8-bit to
2112              * 8-bit blits need to be handled by the blit_shader.
2113              * TODO: get rid of this #if 0. */
2114 #if 0
2115             blit_supported = device->blitter->blit_supported(&device->adapter->gl_info, BLIT_OP_BLIT,
2116                     &rect, This->resource.usage, This->resource.pool, This->resource.format,
2117                     &rect, This->resource.usage, This->resource.pool, This->resource.format);
2118 #endif
2119             blit_supported = gl_info->supported[EXT_PALETTED_TEXTURE] || gl_info->supported[ARB_FRAGMENT_PROGRAM];
2120
2121             /* Use conversion when the blit_shader backend supports it. It only supports this in case of
2122              * texturing. Further also use conversion in case of color keying.
2123              * Paletted textures can be emulated using shaders but only do that for 2D purposes e.g. situations
2124              * in which the main render target uses p8. Some games like GTA Vice City use P8 for texturing which
2125              * conflicts with this.
2126              */
2127             if (!((blit_supported && device->render_targets && This == device->render_targets[0]))
2128                     || colorkey_active || !use_texturing)
2129             {
2130                 format->glFormat = GL_RGBA;
2131                 format->glInternal = GL_RGBA;
2132                 format->glType = GL_UNSIGNED_BYTE;
2133                 format->conv_byte_count = 4;
2134                 if (colorkey_active)
2135                     *convert = CONVERT_PALETTED_CK;
2136                 else
2137                     *convert = CONVERT_PALETTED;
2138             }
2139             break;
2140
2141         case WINED3DFMT_B2G3R3_UNORM:
2142             /* **********************
2143                 GL_UNSIGNED_BYTE_3_3_2
2144                 ********************** */
2145             if (colorkey_active) {
2146                 /* This texture format will never be used.. So do not care about color keying
2147                     up until the point in time it will be needed :-) */
2148                 FIXME(" ColorKeying not supported in the RGB 332 format !\n");
2149             }
2150             break;
2151
2152         case WINED3DFMT_B5G6R5_UNORM:
2153             if (colorkey_active)
2154             {
2155                 *convert = CONVERT_CK_565;
2156                 format->glFormat = GL_RGBA;
2157                 format->glInternal = GL_RGB5_A1;
2158                 format->glType = GL_UNSIGNED_SHORT_5_5_5_1;
2159                 format->conv_byte_count = 2;
2160             }
2161             break;
2162
2163         case WINED3DFMT_B5G5R5X1_UNORM:
2164             if (colorkey_active)
2165             {
2166                 *convert = CONVERT_CK_5551;
2167                 format->glFormat = GL_BGRA;
2168                 format->glInternal = GL_RGB5_A1;
2169                 format->glType = GL_UNSIGNED_SHORT_1_5_5_5_REV;
2170                 format->conv_byte_count = 2;
2171             }
2172             break;
2173
2174         case WINED3DFMT_B8G8R8_UNORM:
2175             if (colorkey_active)
2176             {
2177                 *convert = CONVERT_CK_RGB24;
2178                 format->glFormat = GL_RGBA;
2179                 format->glInternal = GL_RGBA8;
2180                 format->glType = GL_UNSIGNED_INT_8_8_8_8;
2181                 format->conv_byte_count = 4;
2182             }
2183             break;
2184
2185         case WINED3DFMT_B8G8R8X8_UNORM:
2186             if (colorkey_active)
2187             {
2188                 *convert = CONVERT_RGB32_888;
2189                 format->glFormat = GL_RGBA;
2190                 format->glInternal = GL_RGBA8;
2191                 format->glType = GL_UNSIGNED_INT_8_8_8_8;
2192                 format->conv_byte_count = 4;
2193             }
2194             break;
2195
2196         default:
2197             break;
2198     }
2199
2200     return WINED3D_OK;
2201 }
2202
2203 void d3dfmt_p8_init_palette(IWineD3DSurfaceImpl *This, BYTE table[256][4], BOOL colorkey)
2204 {
2205     IWineD3DDeviceImpl *device = This->resource.device;
2206     IWineD3DPaletteImpl *pal = This->palette;
2207     BOOL index_in_alpha = FALSE;
2208     unsigned int i;
2209
2210     /* Old games like StarCraft, C&C, Red Alert and others use P8 render targets.
2211      * Reading back the RGB output each lockrect (each frame as they lock the whole screen)
2212      * is slow. Further RGB->P8 conversion is not possible because palettes can have
2213      * duplicate entries. Store the color key in the unused alpha component to speed the
2214      * download up and to make conversion unneeded. */
2215     index_in_alpha = primary_render_target_is_p8(device);
2216
2217     if (!pal)
2218     {
2219         UINT dxVersion = ((IWineD3DImpl *)device->wined3d)->dxVersion;
2220
2221         /* In DirectDraw the palette is a property of the surface, there are no such things as device palettes. */
2222         if (dxVersion <= 7)
2223         {
2224             ERR("This code should never get entered for DirectDraw!, expect problems\n");
2225             if (index_in_alpha)
2226             {
2227                 /* Guarantees that memory representation remains correct after sysmem<->texture transfers even if
2228                  * there's no palette at this time. */
2229                 for (i = 0; i < 256; i++) table[i][3] = i;
2230             }
2231         }
2232         else
2233         {
2234             /* Direct3D >= 8 palette usage style: P8 textures use device palettes, palette entry format is A8R8G8B8,
2235              * alpha is stored in peFlags and may be used by the app if D3DPTEXTURECAPS_ALPHAPALETTE device
2236              * capability flag is present (wine does advertise this capability) */
2237             for (i = 0; i < 256; ++i)
2238             {
2239                 table[i][0] = device->palettes[device->currentPalette][i].peRed;
2240                 table[i][1] = device->palettes[device->currentPalette][i].peGreen;
2241                 table[i][2] = device->palettes[device->currentPalette][i].peBlue;
2242                 table[i][3] = device->palettes[device->currentPalette][i].peFlags;
2243             }
2244         }
2245     }
2246     else
2247     {
2248         TRACE("Using surface palette %p\n", pal);
2249         /* Get the surface's palette */
2250         for (i = 0; i < 256; ++i)
2251         {
2252             table[i][0] = pal->palents[i].peRed;
2253             table[i][1] = pal->palents[i].peGreen;
2254             table[i][2] = pal->palents[i].peBlue;
2255
2256             /* When index_in_alpha is set the palette index is stored in the
2257              * alpha component. In case of a readback we can then read
2258              * GL_ALPHA. Color keying is handled in BltOverride using a
2259              * GL_ALPHA_TEST using GL_NOT_EQUAL. In case of index_in_alpha the
2260              * color key itself is passed to glAlphaFunc in other cases the
2261              * alpha component of pixels that should be masked away is set to 0. */
2262             if (index_in_alpha)
2263             {
2264                 table[i][3] = i;
2265             }
2266             else if (colorkey && (i >= This->SrcBltCKey.dwColorSpaceLowValue)
2267                     && (i <= This->SrcBltCKey.dwColorSpaceHighValue))
2268             {
2269                 table[i][3] = 0x00;
2270             }
2271             else if (pal->flags & WINEDDPCAPS_ALPHA)
2272             {
2273                 table[i][3] = pal->palents[i].peFlags;
2274             }
2275             else
2276             {
2277                 table[i][3] = 0xFF;
2278             }
2279         }
2280     }
2281 }
2282
2283 static HRESULT d3dfmt_convert_surface(const BYTE *src, BYTE *dst, UINT pitch, UINT width,
2284         UINT height, UINT outpitch, CONVERT_TYPES convert, IWineD3DSurfaceImpl *This)
2285 {
2286     const BYTE *source;
2287     BYTE *dest;
2288     TRACE("(%p)->(%p),(%d,%d,%d,%d,%p)\n", src, dst, pitch, height, outpitch, convert,This);
2289
2290     switch (convert) {
2291         case NO_CONVERSION:
2292         {
2293             memcpy(dst, src, pitch * height);
2294             break;
2295         }
2296         case CONVERT_PALETTED:
2297         case CONVERT_PALETTED_CK:
2298         {
2299             BYTE table[256][4];
2300             unsigned int x, y;
2301
2302             d3dfmt_p8_init_palette(This, table, (convert == CONVERT_PALETTED_CK));
2303
2304             for (y = 0; y < height; y++)
2305             {
2306                 source = src + pitch * y;
2307                 dest = dst + outpitch * y;
2308                 /* This is an 1 bpp format, using the width here is fine */
2309                 for (x = 0; x < width; x++) {
2310                     BYTE color = *source++;
2311                     *dest++ = table[color][0];
2312                     *dest++ = table[color][1];
2313                     *dest++ = table[color][2];
2314                     *dest++ = table[color][3];
2315                 }
2316             }
2317         }
2318         break;
2319
2320         case CONVERT_CK_565:
2321         {
2322             /* Converting the 565 format in 5551 packed to emulate color-keying.
2323
2324               Note : in all these conversion, it would be best to average the averaging
2325                       pixels to get the color of the pixel that will be color-keyed to
2326                       prevent 'color bleeding'. This will be done later on if ever it is
2327                       too visible.
2328
2329               Note2: Nvidia documents say that their driver does not support alpha + color keying
2330                      on the same surface and disables color keying in such a case
2331             */
2332             unsigned int x, y;
2333             const WORD *Source;
2334             WORD *Dest;
2335
2336             TRACE("Color keyed 565\n");
2337
2338             for (y = 0; y < height; y++) {
2339                 Source = (const WORD *)(src + y * pitch);
2340                 Dest = (WORD *) (dst + y * outpitch);
2341                 for (x = 0; x < width; x++ ) {
2342                     WORD color = *Source++;
2343                     *Dest = ((color & 0xFFC0) | ((color & 0x1F) << 1));
2344                     if ((color < This->SrcBltCKey.dwColorSpaceLowValue) ||
2345                         (color > This->SrcBltCKey.dwColorSpaceHighValue)) {
2346                         *Dest |= 0x0001;
2347                     }
2348                     Dest++;
2349                 }
2350             }
2351         }
2352         break;
2353
2354         case CONVERT_CK_5551:
2355         {
2356             /* Converting X1R5G5B5 format to R5G5B5A1 to emulate color-keying. */
2357             unsigned int x, y;
2358             const WORD *Source;
2359             WORD *Dest;
2360             TRACE("Color keyed 5551\n");
2361             for (y = 0; y < height; y++) {
2362                 Source = (const WORD *)(src + y * pitch);
2363                 Dest = (WORD *) (dst + y * outpitch);
2364                 for (x = 0; x < width; x++ ) {
2365                     WORD color = *Source++;
2366                     *Dest = color;
2367                     if ((color < This->SrcBltCKey.dwColorSpaceLowValue) ||
2368                         (color > This->SrcBltCKey.dwColorSpaceHighValue)) {
2369                         *Dest |= (1 << 15);
2370                     }
2371                     else {
2372                         *Dest &= ~(1 << 15);
2373                     }
2374                     Dest++;
2375                 }
2376             }
2377         }
2378         break;
2379
2380         case CONVERT_CK_RGB24:
2381         {
2382             /* Converting R8G8B8 format to R8G8B8A8 with color-keying. */
2383             unsigned int x, y;
2384             for (y = 0; y < height; y++)
2385             {
2386                 source = src + pitch * y;
2387                 dest = dst + outpitch * y;
2388                 for (x = 0; x < width; x++) {
2389                     DWORD color = ((DWORD)source[0] << 16) + ((DWORD)source[1] << 8) + (DWORD)source[2] ;
2390                     DWORD dstcolor = color << 8;
2391                     if ((color < This->SrcBltCKey.dwColorSpaceLowValue) ||
2392                         (color > This->SrcBltCKey.dwColorSpaceHighValue)) {
2393                         dstcolor |= 0xff;
2394                     }
2395                     *(DWORD*)dest = dstcolor;
2396                     source += 3;
2397                     dest += 4;
2398                 }
2399             }
2400         }
2401         break;
2402
2403         case CONVERT_RGB32_888:
2404         {
2405             /* Converting X8R8G8B8 format to R8G8B8A8 with color-keying. */
2406             unsigned int x, y;
2407             for (y = 0; y < height; y++)
2408             {
2409                 source = src + pitch * y;
2410                 dest = dst + outpitch * y;
2411                 for (x = 0; x < width; x++) {
2412                     DWORD color = 0xffffff & *(const DWORD*)source;
2413                     DWORD dstcolor = color << 8;
2414                     if ((color < This->SrcBltCKey.dwColorSpaceLowValue) ||
2415                         (color > This->SrcBltCKey.dwColorSpaceHighValue)) {
2416                         dstcolor |= 0xff;
2417                     }
2418                     *(DWORD*)dest = dstcolor;
2419                     source += 4;
2420                     dest += 4;
2421                 }
2422             }
2423         }
2424         break;
2425
2426         default:
2427             ERR("Unsupported conversion type %#x.\n", convert);
2428     }
2429     return WINED3D_OK;
2430 }
2431
2432 BOOL palette9_changed(IWineD3DSurfaceImpl *This)
2433 {
2434     IWineD3DDeviceImpl *device = This->resource.device;
2435
2436     if (This->palette || (This->resource.format->id != WINED3DFMT_P8_UINT
2437             && This->resource.format->id != WINED3DFMT_P8_UINT_A8_UNORM))
2438     {
2439         /* If a ddraw-style palette is attached assume no d3d9 palette change.
2440          * Also the palette isn't interesting if the surface format isn't P8 or A8P8
2441          */
2442         return FALSE;
2443     }
2444
2445     if (This->palette9)
2446     {
2447         if (!memcmp(This->palette9, device->palettes[device->currentPalette], sizeof(PALETTEENTRY) * 256))
2448         {
2449             return FALSE;
2450         }
2451     } else {
2452         This->palette9 = HeapAlloc(GetProcessHeap(), 0, sizeof(PALETTEENTRY) * 256);
2453     }
2454     memcpy(This->palette9, device->palettes[device->currentPalette], sizeof(PALETTEENTRY) * 256);
2455     return TRUE;
2456 }
2457
2458 static HRESULT WINAPI IWineD3DSurfaceImpl_LoadTexture(IWineD3DSurface *iface, BOOL srgb_mode) {
2459     IWineD3DSurfaceImpl *This = (IWineD3DSurfaceImpl *)iface;
2460     DWORD flag = srgb_mode ? SFLAG_INSRGBTEX : SFLAG_INTEXTURE;
2461
2462     TRACE("iface %p, srgb %#x.\n", iface, srgb_mode);
2463
2464     if (!(This->flags & flag))
2465     {
2466         TRACE("Reloading because surface is dirty\n");
2467     }
2468     /* Reload if either the texture and sysmem have different ideas about the
2469      * color key, or the actual key values changed. */
2470     else if (!(This->flags & SFLAG_GLCKEY) != !(This->CKeyFlags & WINEDDSD_CKSRCBLT)
2471             || ((This->CKeyFlags & WINEDDSD_CKSRCBLT)
2472             && (This->glCKey.dwColorSpaceLowValue != This->SrcBltCKey.dwColorSpaceLowValue
2473             || This->glCKey.dwColorSpaceHighValue != This->SrcBltCKey.dwColorSpaceHighValue)))
2474     {
2475         TRACE("Reloading because of color keying\n");
2476         /* To perform the color key conversion we need a sysmem copy of
2477          * the surface. Make sure we have it
2478          */
2479
2480         surface_load_location(This, SFLAG_INSYSMEM, NULL);
2481         /* Make sure the texture is reloaded because of the color key change, this kills performance though :( */
2482         /* TODO: This is not necessarily needed with hw palettized texture support */
2483         surface_modify_location(This, SFLAG_INSYSMEM, TRUE);
2484     } else {
2485         TRACE("surface is already in texture\n");
2486         return WINED3D_OK;
2487     }
2488
2489     /* Resources are placed in system RAM and do not need to be recreated when a device is lost.
2490      *  These resources are not bound by device size or format restrictions. Because of this,
2491      *  these resources cannot be accessed by the Direct3D device nor set as textures or render targets.
2492      *  However, these resources can always be created, locked, and copied.
2493      */
2494     if (This->resource.pool == WINED3DPOOL_SCRATCH )
2495     {
2496         FIXME("(%p) Operation not supported for scratch textures\n",This);
2497         return WINED3DERR_INVALIDCALL;
2498     }
2499
2500     surface_load_location(This, flag, NULL /* no partial locking for textures yet */);
2501
2502     if (!(This->flags & SFLAG_DONOTFREE))
2503     {
2504         HeapFree(GetProcessHeap(), 0, This->resource.heapMemory);
2505         This->resource.allocatedMemory = NULL;
2506         This->resource.heapMemory = NULL;
2507         surface_modify_location(This, SFLAG_INSYSMEM, FALSE);
2508     }
2509
2510     return WINED3D_OK;
2511 }
2512
2513 /* Context activation is done by the caller. */
2514 static void WINAPI IWineD3DSurfaceImpl_BindTexture(IWineD3DSurface *iface, BOOL srgb)
2515 {
2516     IWineD3DSurfaceImpl *This = (IWineD3DSurfaceImpl *)iface;
2517
2518     TRACE("iface %p, srgb %#x.\n", iface, srgb);
2519
2520     if (This->container.type == WINED3D_CONTAINER_TEXTURE)
2521     {
2522         TRACE("Passing to container.\n");
2523         IWineD3DBaseTexture_BindTexture((IWineD3DBaseTexture *)This->container.u.texture, srgb);
2524     }
2525     else
2526     {
2527         GLuint *name;
2528
2529         TRACE("(%p) : Binding surface\n", This);
2530
2531         name = srgb ? &This->texture_name_srgb : &This->texture_name;
2532
2533         ENTER_GL();
2534
2535         if (!This->texture_level)
2536         {
2537             if (!*name) {
2538                 glGenTextures(1, name);
2539                 checkGLcall("glGenTextures");
2540                 TRACE("Surface %p given name %d\n", This, *name);
2541
2542                 glBindTexture(This->texture_target, *name);
2543                 checkGLcall("glBindTexture");
2544                 glTexParameteri(This->texture_target, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
2545                 checkGLcall("glTexParameteri(dimension, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE)");
2546                 glTexParameteri(This->texture_target, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
2547                 checkGLcall("glTexParameteri(dimension, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE)");
2548                 glTexParameteri(This->texture_target, GL_TEXTURE_WRAP_R, GL_CLAMP_TO_EDGE);
2549                 checkGLcall("glTexParameteri(dimension, GL_TEXTURE_WRAP_R, GL_CLAMP_TO_EDGE)");
2550                 glTexParameteri(This->texture_target, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
2551                 checkGLcall("glTexParameteri(dimension, GL_TEXTURE_MIN_FILTER, GL_NEAREST)");
2552                 glTexParameteri(This->texture_target, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
2553                 checkGLcall("glTexParameteri(dimension, GL_TEXTURE_MAG_FILTER, GL_NEAREST)");
2554             }
2555             /* This is where we should be reducing the amount of GLMemoryUsed */
2556         } else if (*name) {
2557             /* Mipmap surfaces should have a base texture container */
2558             ERR("Mipmap surface has a glTexture bound to it!\n");
2559         }
2560
2561         glBindTexture(This->texture_target, *name);
2562         checkGLcall("glBindTexture");
2563
2564         LEAVE_GL();
2565     }
2566 }
2567
2568 static HRESULT WINAPI IWineD3DSurfaceImpl_SetFormat(IWineD3DSurface *iface, enum wined3d_format_id format)
2569 {
2570     IWineD3DSurfaceImpl *This = (IWineD3DSurfaceImpl *)iface;
2571     HRESULT hr;
2572
2573     TRACE("(%p) : Calling base function first\n", This);
2574     hr = IWineD3DBaseSurfaceImpl_SetFormat(iface, format);
2575     if (SUCCEEDED(hr))
2576     {
2577         This->flags &= ~(SFLAG_ALLOCATED | SFLAG_SRGBALLOCATED);
2578         TRACE("(%p) : glFormat %d, glFormatInternal %d, glType %d\n", This, This->resource.format->glFormat,
2579                 This->resource.format->glInternal, This->resource.format->glType);
2580     }
2581     return hr;
2582 }
2583
2584 static HRESULT WINAPI IWineD3DSurfaceImpl_SetMem(IWineD3DSurface *iface, void *Mem) {
2585     IWineD3DSurfaceImpl *This = (IWineD3DSurfaceImpl *) iface;
2586
2587     TRACE("iface %p, mem %p.\n", iface, Mem);
2588
2589     if (This->flags & (SFLAG_LOCKED | SFLAG_DCINUSE))
2590     {
2591         WARN("Surface is locked or the HDC is in use\n");
2592         return WINED3DERR_INVALIDCALL;
2593     }
2594
2595     if(Mem && Mem != This->resource.allocatedMemory) {
2596         void *release = NULL;
2597
2598         /* Do I have to copy the old surface content? */
2599         if (This->flags & SFLAG_DIBSECTION)
2600         {
2601             SelectObject(This->hDC, This->dib.holdbitmap);
2602             DeleteDC(This->hDC);
2603             /* Release the DIB section */
2604             DeleteObject(This->dib.DIBsection);
2605             This->dib.bitmap_data = NULL;
2606             This->resource.allocatedMemory = NULL;
2607             This->hDC = NULL;
2608             This->flags &= ~SFLAG_DIBSECTION;
2609         }
2610         else if (!(This->flags & SFLAG_USERPTR))
2611         {
2612             release = This->resource.heapMemory;
2613             This->resource.heapMemory = NULL;
2614         }
2615         This->resource.allocatedMemory = Mem;
2616         This->flags |= SFLAG_USERPTR;
2617
2618         /* Now the surface memory is most up do date. Invalidate drawable and texture */
2619         surface_modify_location(This, SFLAG_INSYSMEM, TRUE);
2620
2621         /* For client textures opengl has to be notified */
2622         if (This->flags & SFLAG_CLIENT)
2623             surface_release_client_storage(This);
2624
2625         /* Now free the old memory if any */
2626         HeapFree(GetProcessHeap(), 0, release);
2627     }
2628     else if (This->flags & SFLAG_USERPTR)
2629     {
2630         /* Map and GetDC will re-create the dib section and allocated memory. */
2631         This->resource.allocatedMemory = NULL;
2632         /* HeapMemory should be NULL already */
2633         if (This->resource.heapMemory)
2634             ERR("User pointer surface has heap memory allocated.\n");
2635         This->flags &= ~(SFLAG_USERPTR | SFLAG_INSYSMEM);
2636
2637         if (This->flags & SFLAG_CLIENT)
2638             surface_release_client_storage(This);
2639
2640         surface_prepare_system_memory(This);
2641         surface_modify_location(This, SFLAG_INSYSMEM, TRUE);
2642     }
2643     return WINED3D_OK;
2644 }
2645
2646 void flip_surface(IWineD3DSurfaceImpl *front, IWineD3DSurfaceImpl *back) {
2647
2648     /* Flip the surface contents */
2649     /* Flip the DC */
2650     {
2651         HDC tmp;
2652         tmp = front->hDC;
2653         front->hDC = back->hDC;
2654         back->hDC = tmp;
2655     }
2656
2657     /* Flip the DIBsection */
2658     {
2659         HBITMAP tmp;
2660         BOOL hasDib = front->flags & SFLAG_DIBSECTION;
2661         tmp = front->dib.DIBsection;
2662         front->dib.DIBsection = back->dib.DIBsection;
2663         back->dib.DIBsection = tmp;
2664
2665         if (back->flags & SFLAG_DIBSECTION) front->flags |= SFLAG_DIBSECTION;
2666         else front->flags &= ~SFLAG_DIBSECTION;
2667         if (hasDib) back->flags |= SFLAG_DIBSECTION;
2668         else back->flags &= ~SFLAG_DIBSECTION;
2669     }
2670
2671     /* Flip the surface data */
2672     {
2673         void* tmp;
2674
2675         tmp = front->dib.bitmap_data;
2676         front->dib.bitmap_data = back->dib.bitmap_data;
2677         back->dib.bitmap_data = tmp;
2678
2679         tmp = front->resource.allocatedMemory;
2680         front->resource.allocatedMemory = back->resource.allocatedMemory;
2681         back->resource.allocatedMemory = tmp;
2682
2683         tmp = front->resource.heapMemory;
2684         front->resource.heapMemory = back->resource.heapMemory;
2685         back->resource.heapMemory = tmp;
2686     }
2687
2688     /* Flip the PBO */
2689     {
2690         GLuint tmp_pbo = front->pbo;
2691         front->pbo = back->pbo;
2692         back->pbo = tmp_pbo;
2693     }
2694
2695     /* client_memory should not be different, but just in case */
2696     {
2697         BOOL tmp;
2698         tmp = front->dib.client_memory;
2699         front->dib.client_memory = back->dib.client_memory;
2700         back->dib.client_memory = tmp;
2701     }
2702
2703     /* Flip the opengl texture */
2704     {
2705         GLuint tmp;
2706
2707         tmp = back->texture_name;
2708         back->texture_name = front->texture_name;
2709         front->texture_name = tmp;
2710
2711         tmp = back->texture_name_srgb;
2712         back->texture_name_srgb = front->texture_name_srgb;
2713         front->texture_name_srgb = tmp;
2714
2715         resource_unload((IWineD3DResourceImpl *)back);
2716         resource_unload((IWineD3DResourceImpl *)front);
2717     }
2718
2719     {
2720         DWORD tmp_flags = back->flags;
2721         back->flags = front->flags;
2722         front->flags = tmp_flags;
2723     }
2724 }
2725
2726 static HRESULT WINAPI IWineD3DSurfaceImpl_Flip(IWineD3DSurface *iface, IWineD3DSurface *override, DWORD Flags) {
2727     IWineD3DSurfaceImpl *This = (IWineD3DSurfaceImpl *)iface;
2728     IWineD3DSwapChainImpl *swapchain = NULL;
2729
2730     TRACE("(%p)->(%p,%x)\n", This, override, Flags);
2731
2732     /* Flipping is only supported on RenderTargets and overlays*/
2733     if( !(This->resource.usage & (WINED3DUSAGE_RENDERTARGET | WINED3DUSAGE_OVERLAY)) ) {
2734         WARN("Tried to flip a non-render target, non-overlay surface\n");
2735         return WINEDDERR_NOTFLIPPABLE;
2736     }
2737
2738     if(This->resource.usage & WINED3DUSAGE_OVERLAY) {
2739         flip_surface(This, (IWineD3DSurfaceImpl *) override);
2740
2741         /* Update the overlay if it is visible */
2742         if(This->overlay_dest) {
2743             return IWineD3DSurface_DrawOverlay((IWineD3DSurface *) This);
2744         } else {
2745             return WINED3D_OK;
2746         }
2747     }
2748
2749     if(override) {
2750         /* DDraw sets this for the X11 surfaces, so don't confuse the user
2751          * FIXME("(%p) Target override is not supported by now\n", This);
2752          * Additionally, it isn't really possible to support triple-buffering
2753          * properly on opengl at all
2754          */
2755     }
2756
2757     if (This->container.type != WINED3D_CONTAINER_SWAPCHAIN)
2758     {
2759         ERR("Flipped surface is not on a swapchain\n");
2760         return WINEDDERR_NOTFLIPPABLE;
2761     }
2762     swapchain = This->container.u.swapchain;
2763
2764     /* Just overwrite the swapchain presentation interval. This is ok because only ddraw apps can call Flip,
2765      * and only d3d8 and d3d9 apps specify the presentation interval
2766      */
2767     if (!(Flags & (WINEDDFLIP_NOVSYNC | WINEDDFLIP_INTERVAL2 | WINEDDFLIP_INTERVAL3 | WINEDDFLIP_INTERVAL4)))
2768     {
2769         /* Most common case first to avoid wasting time on all the other cases */
2770         swapchain->presentParms.PresentationInterval = WINED3DPRESENT_INTERVAL_ONE;
2771     } else if(Flags & WINEDDFLIP_NOVSYNC) {
2772         swapchain->presentParms.PresentationInterval = WINED3DPRESENT_INTERVAL_IMMEDIATE;
2773     } else if(Flags & WINEDDFLIP_INTERVAL2) {
2774         swapchain->presentParms.PresentationInterval = WINED3DPRESENT_INTERVAL_TWO;
2775     } else if(Flags & WINEDDFLIP_INTERVAL3) {
2776         swapchain->presentParms.PresentationInterval = WINED3DPRESENT_INTERVAL_THREE;
2777     } else {
2778         swapchain->presentParms.PresentationInterval = WINED3DPRESENT_INTERVAL_FOUR;
2779     }
2780
2781     /* Flipping a OpenGL surface -> Use WineD3DDevice::Present */
2782     return IWineD3DSwapChain_Present((IWineD3DSwapChain *)swapchain,
2783             NULL, NULL, swapchain->win_handle, NULL, 0);
2784 }
2785
2786 /* Does a direct frame buffer -> texture copy. Stretching is done
2787  * with single pixel copy calls
2788  */
2789 static void fb_copy_to_texture_direct(IWineD3DSurfaceImpl *dst_surface, IWineD3DSurfaceImpl *src_surface,
2790         const RECT *src_rect, const RECT *dst_rect_in, WINED3DTEXTUREFILTERTYPE Filter)
2791 {
2792     IWineD3DDeviceImpl *device = dst_surface->resource.device;
2793     float xrel, yrel;
2794     UINT row;
2795     struct wined3d_context *context;
2796     BOOL upsidedown = FALSE;
2797     RECT dst_rect = *dst_rect_in;
2798
2799     /* Make sure that the top pixel is always above the bottom pixel, and keep a separate upside down flag
2800      * glCopyTexSubImage is a bit picky about the parameters we pass to it
2801      */
2802     if(dst_rect.top > dst_rect.bottom) {
2803         UINT tmp = dst_rect.bottom;
2804         dst_rect.bottom = dst_rect.top;
2805         dst_rect.top = tmp;
2806         upsidedown = TRUE;
2807     }
2808
2809     context = context_acquire(device, src_surface);
2810     context_apply_blit_state(context, device);
2811     surface_internal_preload(dst_surface, SRGB_RGB);
2812     ENTER_GL();
2813
2814     /* Bind the target texture */
2815     glBindTexture(dst_surface->texture_target, dst_surface->texture_name);
2816     checkGLcall("glBindTexture");
2817     if (surface_is_offscreen(src_surface))
2818     {
2819         TRACE("Reading from an offscreen target\n");
2820         upsidedown = !upsidedown;
2821         glReadBuffer(device->offscreenBuffer);
2822     }
2823     else
2824     {
2825         glReadBuffer(surface_get_gl_buffer(src_surface));
2826     }
2827     checkGLcall("glReadBuffer");
2828
2829     xrel = (float) (src_rect->right - src_rect->left) / (float) (dst_rect.right - dst_rect.left);
2830     yrel = (float) (src_rect->bottom - src_rect->top) / (float) (dst_rect.bottom - dst_rect.top);
2831
2832     if ((xrel - 1.0f < -eps) || (xrel - 1.0f > eps))
2833     {
2834         FIXME("Doing a pixel by pixel copy from the framebuffer to a texture, expect major performance issues\n");
2835
2836         if(Filter != WINED3DTEXF_NONE && Filter != WINED3DTEXF_POINT) {
2837             ERR("Texture filtering not supported in direct blit\n");
2838         }
2839     }
2840     else if ((Filter != WINED3DTEXF_NONE && Filter != WINED3DTEXF_POINT)
2841             && ((yrel - 1.0f < -eps) || (yrel - 1.0f > eps)))
2842     {
2843         ERR("Texture filtering not supported in direct blit\n");
2844     }
2845
2846     if (upsidedown
2847             && !((xrel - 1.0f < -eps) || (xrel - 1.0f > eps))
2848             && !((yrel - 1.0f < -eps) || (yrel - 1.0f > eps)))
2849     {
2850         /* Upside down copy without stretching is nice, one glCopyTexSubImage call will do */
2851
2852         glCopyTexSubImage2D(dst_surface->texture_target, dst_surface->texture_level,
2853                 dst_rect.left /*xoffset */, dst_rect.top /* y offset */,
2854                 src_rect->left, src_surface->currentDesc.Height - src_rect->bottom,
2855                 dst_rect.right - dst_rect.left, dst_rect.bottom - dst_rect.top);
2856     } else {
2857         UINT yoffset = src_surface->currentDesc.Height - src_rect->top + dst_rect.top - 1;
2858         /* I have to process this row by row to swap the image,
2859          * otherwise it would be upside down, so stretching in y direction
2860          * doesn't cost extra time
2861          *
2862          * However, stretching in x direction can be avoided if not necessary
2863          */
2864         for(row = dst_rect.top; row < dst_rect.bottom; row++) {
2865             if ((xrel - 1.0f < -eps) || (xrel - 1.0f > eps))
2866             {
2867                 /* Well, that stuff works, but it's very slow.
2868                  * find a better way instead
2869                  */
2870                 UINT col;
2871
2872                 for (col = dst_rect.left; col < dst_rect.right; ++col)
2873                 {
2874                     glCopyTexSubImage2D(dst_surface->texture_target, dst_surface->texture_level,
2875                             dst_rect.left + col /* x offset */, row /* y offset */,
2876                             src_rect->left + col * xrel, yoffset - (int) (row * yrel), 1, 1);
2877                 }
2878             }
2879             else
2880             {
2881                 glCopyTexSubImage2D(dst_surface->texture_target, dst_surface->texture_level,
2882                         dst_rect.left /* x offset */, row /* y offset */,
2883                         src_rect->left, yoffset - (int) (row * yrel), dst_rect.right - dst_rect.left, 1);
2884             }
2885         }
2886     }
2887     checkGLcall("glCopyTexSubImage2D");
2888
2889     LEAVE_GL();
2890     context_release(context);
2891
2892     /* The texture is now most up to date - If the surface is a render target and has a drawable, this
2893      * path is never entered
2894      */
2895     surface_modify_location(dst_surface, SFLAG_INTEXTURE, TRUE);
2896 }
2897
2898 /* Uses the hardware to stretch and flip the image */
2899 static void fb_copy_to_texture_hwstretch(IWineD3DSurfaceImpl *dst_surface, IWineD3DSurfaceImpl *src_surface,
2900         const RECT *src_rect, const RECT *dst_rect_in, WINED3DTEXTUREFILTERTYPE Filter)
2901 {
2902     IWineD3DDeviceImpl *device = dst_surface->resource.device;
2903     GLuint src, backup = 0;
2904     IWineD3DSwapChainImpl *src_swapchain = NULL;
2905     float left, right, top, bottom; /* Texture coordinates */
2906     UINT fbwidth = src_surface->currentDesc.Width;
2907     UINT fbheight = src_surface->currentDesc.Height;
2908     struct wined3d_context *context;
2909     GLenum drawBuffer = GL_BACK;
2910     GLenum texture_target;
2911     BOOL noBackBufferBackup;
2912     BOOL src_offscreen;
2913     BOOL upsidedown = FALSE;
2914     RECT dst_rect = *dst_rect_in;
2915
2916     TRACE("Using hwstretch blit\n");
2917     /* Activate the Proper context for reading from the source surface, set it up for blitting */
2918     context = context_acquire(device, src_surface);
2919     context_apply_blit_state(context, device);
2920     surface_internal_preload(dst_surface, SRGB_RGB);
2921
2922     src_offscreen = surface_is_offscreen(src_surface);
2923     noBackBufferBackup = src_offscreen && wined3d_settings.offscreen_rendering_mode == ORM_FBO;
2924     if (!noBackBufferBackup && !src_surface->texture_name)
2925     {
2926         /* Get it a description */
2927         surface_internal_preload(src_surface, SRGB_RGB);
2928     }
2929     ENTER_GL();
2930
2931     /* Try to use an aux buffer for drawing the rectangle. This way it doesn't need restoring.
2932      * This way we don't have to wait for the 2nd readback to finish to leave this function.
2933      */
2934     if (context->aux_buffers >= 2)
2935     {
2936         /* Got more than one aux buffer? Use the 2nd aux buffer */
2937         drawBuffer = GL_AUX1;
2938     }
2939     else if ((!src_offscreen || device->offscreenBuffer == GL_BACK) && context->aux_buffers >= 1)
2940     {
2941         /* Only one aux buffer, but it isn't used (Onscreen rendering, or non-aux orm)? Use it! */
2942         drawBuffer = GL_AUX0;
2943     }
2944
2945     if(noBackBufferBackup) {
2946         glGenTextures(1, &backup);
2947         checkGLcall("glGenTextures");
2948         glBindTexture(GL_TEXTURE_2D, backup);
2949         checkGLcall("glBindTexture(GL_TEXTURE_2D, backup)");
2950         texture_target = GL_TEXTURE_2D;
2951     } else {
2952         /* Backup the back buffer and copy the source buffer into a texture to draw an upside down stretched quad. If
2953          * we are reading from the back buffer, the backup can be used as source texture
2954          */
2955         texture_target = src_surface->texture_target;
2956         glBindTexture(texture_target, src_surface->texture_name);
2957         checkGLcall("glBindTexture(texture_target, src_surface->texture_name)");
2958         glEnable(texture_target);
2959         checkGLcall("glEnable(texture_target)");
2960
2961         /* For now invalidate the texture copy of the back buffer. Drawable and sysmem copy are untouched */
2962         src_surface->flags &= ~SFLAG_INTEXTURE;
2963     }
2964
2965     /* Make sure that the top pixel is always above the bottom pixel, and keep a separate upside down flag
2966      * glCopyTexSubImage is a bit picky about the parameters we pass to it
2967      */
2968     if(dst_rect.top > dst_rect.bottom) {
2969         UINT tmp = dst_rect.bottom;
2970         dst_rect.bottom = dst_rect.top;
2971         dst_rect.top = tmp;
2972         upsidedown = TRUE;
2973     }
2974
2975     if (src_offscreen)
2976     {
2977         TRACE("Reading from an offscreen target\n");
2978         upsidedown = !upsidedown;
2979         glReadBuffer(device->offscreenBuffer);
2980     }
2981     else
2982     {
2983         glReadBuffer(surface_get_gl_buffer(src_surface));
2984     }
2985
2986     /* TODO: Only back up the part that will be overwritten */
2987     glCopyTexSubImage2D(texture_target, 0,
2988                         0, 0 /* read offsets */,
2989                         0, 0,
2990                         fbwidth,
2991                         fbheight);
2992
2993     checkGLcall("glCopyTexSubImage2D");
2994
2995     /* No issue with overriding these - the sampler is dirty due to blit usage */
2996     glTexParameteri(texture_target, GL_TEXTURE_MAG_FILTER,
2997             wined3d_gl_mag_filter(magLookup, Filter));
2998     checkGLcall("glTexParameteri");
2999     glTexParameteri(texture_target, GL_TEXTURE_MIN_FILTER,
3000             wined3d_gl_min_mip_filter(minMipLookup, Filter, WINED3DTEXF_NONE));
3001     checkGLcall("glTexParameteri");
3002
3003     if (src_surface->container.type == WINED3D_CONTAINER_SWAPCHAIN)
3004         src_swapchain = src_surface->container.u.swapchain;
3005     if (!src_swapchain || src_surface == src_swapchain->back_buffers[0])
3006     {
3007         src = backup ? backup : src_surface->texture_name;
3008     }
3009     else
3010     {
3011         glReadBuffer(GL_FRONT);
3012         checkGLcall("glReadBuffer(GL_FRONT)");
3013
3014         glGenTextures(1, &src);
3015         checkGLcall("glGenTextures(1, &src)");
3016         glBindTexture(GL_TEXTURE_2D, src);
3017         checkGLcall("glBindTexture(GL_TEXTURE_2D, src)");
3018
3019         /* TODO: Only copy the part that will be read. Use src_rect->left, src_rect->bottom as origin, but with the width watch
3020          * out for power of 2 sizes
3021          */
3022         glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA, src_surface->pow2Width,
3023                 src_surface->pow2Height, 0, GL_RGBA, GL_UNSIGNED_BYTE, NULL);
3024         checkGLcall("glTexImage2D");
3025         glCopyTexSubImage2D(GL_TEXTURE_2D, 0,
3026                             0, 0 /* read offsets */,
3027                             0, 0,
3028                             fbwidth,
3029                             fbheight);
3030
3031         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
3032         checkGLcall("glTexParameteri");
3033         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
3034         checkGLcall("glTexParameteri");
3035
3036         glReadBuffer(GL_BACK);
3037         checkGLcall("glReadBuffer(GL_BACK)");
3038
3039         if(texture_target != GL_TEXTURE_2D) {
3040             glDisable(texture_target);
3041             glEnable(GL_TEXTURE_2D);
3042             texture_target = GL_TEXTURE_2D;
3043         }
3044     }
3045     checkGLcall("glEnd and previous");
3046
3047     left = src_rect->left;
3048     right = src_rect->right;
3049
3050     if (!upsidedown)
3051     {
3052         top = src_surface->currentDesc.Height - src_rect->top;
3053         bottom = src_surface->currentDesc.Height - src_rect->bottom;
3054     }
3055     else
3056     {
3057         top = src_surface->currentDesc.Height - src_rect->bottom;
3058         bottom = src_surface->currentDesc.Height - src_rect->top;
3059     }
3060
3061     if (src_surface->flags & SFLAG_NORMCOORD)
3062     {
3063         left /= src_surface->pow2Width;
3064         right /= src_surface->pow2Width;
3065         top /= src_surface->pow2Height;
3066         bottom /= src_surface->pow2Height;
3067     }
3068
3069     /* draw the source texture stretched and upside down. The correct surface is bound already */
3070     glTexParameteri(texture_target, GL_TEXTURE_WRAP_S, GL_CLAMP);
3071     glTexParameteri(texture_target, GL_TEXTURE_WRAP_T, GL_CLAMP);
3072
3073     context_set_draw_buffer(context, drawBuffer);
3074     glReadBuffer(drawBuffer);
3075
3076     glBegin(GL_QUADS);
3077         /* bottom left */
3078         glTexCoord2f(left, bottom);
3079         glVertex2i(0, 0);
3080
3081         /* top left */
3082         glTexCoord2f(left, top);
3083         glVertex2i(0, dst_rect.bottom - dst_rect.top);
3084
3085         /* top right */
3086         glTexCoord2f(right, top);
3087         glVertex2i(dst_rect.right - dst_rect.left, dst_rect.bottom - dst_rect.top);
3088
3089         /* bottom right */
3090         glTexCoord2f(right, bottom);
3091         glVertex2i(dst_rect.right - dst_rect.left, 0);
3092     glEnd();
3093     checkGLcall("glEnd and previous");
3094
3095     if (texture_target != dst_surface->texture_target)
3096     {
3097         glDisable(texture_target);
3098         glEnable(dst_surface->texture_target);
3099         texture_target = dst_surface->texture_target;
3100     }
3101
3102     /* Now read the stretched and upside down image into the destination texture */
3103     glBindTexture(texture_target, dst_surface->texture_name);
3104     checkGLcall("glBindTexture");
3105     glCopyTexSubImage2D(texture_target,
3106                         0,
3107                         dst_rect.left, dst_rect.top, /* xoffset, yoffset */
3108                         0, 0, /* We blitted the image to the origin */
3109                         dst_rect.right - dst_rect.left, dst_rect.bottom - dst_rect.top);
3110     checkGLcall("glCopyTexSubImage2D");
3111
3112     if(drawBuffer == GL_BACK) {
3113         /* Write the back buffer backup back */
3114         if(backup) {
3115             if(texture_target != GL_TEXTURE_2D) {
3116                 glDisable(texture_target);
3117                 glEnable(GL_TEXTURE_2D);
3118                 texture_target = GL_TEXTURE_2D;
3119             }
3120             glBindTexture(GL_TEXTURE_2D, backup);
3121             checkGLcall("glBindTexture(GL_TEXTURE_2D, backup)");
3122         }
3123         else
3124         {
3125             if (texture_target != src_surface->texture_target)
3126             {
3127                 glDisable(texture_target);
3128                 glEnable(src_surface->texture_target);
3129                 texture_target = src_surface->texture_target;
3130             }
3131             glBindTexture(src_surface->texture_target, src_surface->texture_name);
3132             checkGLcall("glBindTexture(src_surface->texture_target, src_surface->texture_name)");
3133         }
3134
3135         glBegin(GL_QUADS);
3136             /* top left */
3137             glTexCoord2f(0.0f, 0.0f);
3138             glVertex2i(0, fbheight);
3139
3140             /* bottom left */
3141             glTexCoord2f(0.0f, (float)fbheight / (float)src_surface->pow2Height);
3142             glVertex2i(0, 0);
3143
3144             /* bottom right */
3145             glTexCoord2f((float)fbwidth / (float)src_surface->pow2Width,
3146                     (float)fbheight / (float)src_surface->pow2Height);
3147             glVertex2i(fbwidth, 0);
3148
3149             /* top right */
3150             glTexCoord2f((float)fbwidth / (float)src_surface->pow2Width, 0.0f);
3151             glVertex2i(fbwidth, fbheight);
3152         glEnd();
3153     }
3154     glDisable(texture_target);
3155     checkGLcall("glDisable(texture_target)");
3156
3157     /* Cleanup */
3158     if (src != src_surface->texture_name && src != backup)
3159     {
3160         glDeleteTextures(1, &src);
3161         checkGLcall("glDeleteTextures(1, &src)");
3162     }
3163     if(backup) {
3164         glDeleteTextures(1, &backup);
3165         checkGLcall("glDeleteTextures(1, &backup)");
3166     }
3167
3168     LEAVE_GL();
3169
3170     if (wined3d_settings.strict_draw_ordering) wglFlush(); /* Flush to ensure ordering across contexts. */
3171
3172     context_release(context);
3173
3174     /* The texture is now most up to date - If the surface is a render target and has a drawable, this
3175      * path is never entered
3176      */
3177     surface_modify_location(dst_surface, SFLAG_INTEXTURE, TRUE);
3178 }
3179
3180 /* Until the blit_shader is ready, define some prototypes here. */
3181 static BOOL fbo_blit_supported(const struct wined3d_gl_info *gl_info, enum blit_operation blit_op,
3182         const RECT *src_rect, DWORD src_usage, WINED3DPOOL src_pool, const struct wined3d_format *src_format,
3183         const RECT *dst_rect, DWORD dst_usage, WINED3DPOOL dst_pool, const struct wined3d_format *dst_format);
3184
3185 /* Front buffer coordinates are always full screen coordinates, but our GL
3186  * drawable is limited to the window's client area. The sysmem and texture
3187  * copies do have the full screen size. Note that GL has a bottom-left
3188  * origin, while D3D has a top-left origin. */
3189 void surface_translate_drawable_coords(IWineD3DSurfaceImpl *surface, HWND window, RECT *rect)
3190 {
3191     UINT drawable_height;
3192
3193     if (surface->container.type == WINED3D_CONTAINER_SWAPCHAIN
3194             && surface == surface->container.u.swapchain->front_buffer)
3195     {
3196         POINT offset = {0, 0};
3197         RECT windowsize;
3198
3199         ScreenToClient(window, &offset);
3200         OffsetRect(rect, offset.x, offset.y);
3201
3202         GetClientRect(window, &windowsize);
3203         drawable_height = windowsize.bottom - windowsize.top;
3204     }
3205     else
3206     {
3207         drawable_height = surface->currentDesc.Height;
3208     }
3209
3210     rect->top = drawable_height - rect->top;
3211     rect->bottom = drawable_height - rect->bottom;
3212 }
3213
3214 static BOOL surface_is_full_rect(IWineD3DSurfaceImpl *surface, const RECT *r)
3215 {
3216     if ((r->left && r->right) || abs(r->right - r->left) != surface->currentDesc.Width)
3217         return FALSE;
3218     if ((r->top && r->bottom) || abs(r->bottom - r->top) != surface->currentDesc.Height)
3219         return FALSE;
3220     return TRUE;
3221 }
3222
3223 /* blit between surface locations. onscreen on different swapchains is not supported.
3224  * depth / stencil is not supported. */
3225 static void surface_blt_fbo(IWineD3DDeviceImpl *device, const WINED3DTEXTUREFILTERTYPE filter,
3226         IWineD3DSurfaceImpl *src_surface, DWORD src_location, const RECT *src_rect_in,
3227         IWineD3DSurfaceImpl *dst_surface, DWORD dst_location, const RECT *dst_rect_in)
3228 {
3229     const struct wined3d_gl_info *gl_info;
3230     struct wined3d_context *context;
3231     RECT src_rect, dst_rect;
3232     GLenum gl_filter;
3233
3234     TRACE("device %p, filter %s,\n", device, debug_d3dtexturefiltertype(filter));
3235     TRACE("src_surface %p, src_location %s, src_rect %s,\n",
3236             src_surface, debug_surflocation(src_location), wine_dbgstr_rect(src_rect_in));
3237     TRACE("dst_surface %p, dst_location %s, dst_rect %s.\n",
3238             dst_surface, debug_surflocation(dst_location), wine_dbgstr_rect(dst_rect_in));
3239
3240     src_rect = *src_rect_in;
3241     dst_rect = *dst_rect_in;
3242
3243     switch (filter)
3244     {
3245         case WINED3DTEXF_LINEAR:
3246             gl_filter = GL_LINEAR;
3247             break;
3248
3249         default:
3250             FIXME("Unsupported filter mode %s (%#x).\n", debug_d3dtexturefiltertype(filter), filter);
3251         case WINED3DTEXF_NONE:
3252         case WINED3DTEXF_POINT:
3253             gl_filter = GL_NEAREST;
3254             break;
3255     }
3256
3257     if (src_location == SFLAG_INDRAWABLE && surface_is_offscreen(src_surface))
3258         src_location = SFLAG_INTEXTURE;
3259     if (dst_location == SFLAG_INDRAWABLE && surface_is_offscreen(dst_surface))
3260         dst_location = SFLAG_INTEXTURE;
3261
3262     /* Make sure the locations are up-to-date. Loading the destination
3263      * surface isn't required if the entire surface is overwritten. (And is
3264      * in fact harmful if we're being called by surface_load_location() with
3265      * the purpose of loading the destination surface.) */
3266     surface_load_location(src_surface, src_location, NULL);
3267     if (!surface_is_full_rect(dst_surface, &dst_rect))
3268         surface_load_location(dst_surface, dst_location, NULL);
3269
3270     if (src_location == SFLAG_INDRAWABLE) context = context_acquire(device, src_surface);
3271     else if (dst_location == SFLAG_INDRAWABLE) context = context_acquire(device, dst_surface);
3272     else context = context_acquire(device, NULL);
3273
3274     if (!context->valid)
3275     {
3276         context_release(context);
3277         WARN("Invalid context, skipping blit.\n");
3278         return;
3279     }
3280
3281     gl_info = context->gl_info;
3282
3283     if (src_location == SFLAG_INDRAWABLE)
3284     {
3285         GLenum buffer = surface_get_gl_buffer(src_surface);
3286
3287         TRACE("Source surface %p is onscreen.\n", src_surface);
3288
3289         surface_translate_drawable_coords(src_surface, context->win_handle, &src_rect);
3290
3291         ENTER_GL();
3292         context_bind_fbo(context, GL_READ_FRAMEBUFFER, NULL);
3293         glReadBuffer(buffer);
3294         checkGLcall("glReadBuffer()");
3295     }
3296     else
3297     {
3298         TRACE("Source surface %p is offscreen.\n", src_surface);
3299         ENTER_GL();
3300         context_apply_fbo_state_blit(context, GL_READ_FRAMEBUFFER, src_surface, NULL, src_location);
3301         glReadBuffer(GL_COLOR_ATTACHMENT0);
3302         checkGLcall("glReadBuffer()");
3303     }
3304     LEAVE_GL();
3305
3306     if (dst_location == SFLAG_INDRAWABLE)
3307     {
3308         GLenum buffer = surface_get_gl_buffer(dst_surface);
3309
3310         TRACE("Destination surface %p is onscreen.\n", dst_surface);
3311
3312         surface_translate_drawable_coords(dst_surface, context->win_handle, &dst_rect);
3313
3314         ENTER_GL();
3315         context_bind_fbo(context, GL_DRAW_FRAMEBUFFER, NULL);
3316         context_set_draw_buffer(context, buffer);
3317     }
3318     else
3319     {
3320         TRACE("Destination surface %p is offscreen.\n", dst_surface);
3321
3322         ENTER_GL();
3323         context_apply_fbo_state_blit(context, GL_DRAW_FRAMEBUFFER, dst_surface, NULL, dst_location);
3324         context_set_draw_buffer(context, GL_COLOR_ATTACHMENT0);
3325     }
3326
3327     glColorMask(GL_TRUE, GL_TRUE, GL_TRUE, GL_TRUE);
3328     IWineD3DDeviceImpl_MarkStateDirty(device, STATE_RENDER(WINED3DRS_COLORWRITEENABLE));
3329     IWineD3DDeviceImpl_MarkStateDirty(device, STATE_RENDER(WINED3DRS_COLORWRITEENABLE1));
3330     IWineD3DDeviceImpl_MarkStateDirty(device, STATE_RENDER(WINED3DRS_COLORWRITEENABLE2));
3331     IWineD3DDeviceImpl_MarkStateDirty(device, STATE_RENDER(WINED3DRS_COLORWRITEENABLE3));
3332
3333     glDisable(GL_SCISSOR_TEST);
3334     IWineD3DDeviceImpl_MarkStateDirty(device, STATE_RENDER(WINED3DRS_SCISSORTESTENABLE));
3335
3336     gl_info->fbo_ops.glBlitFramebuffer(src_rect.left, src_rect.top, src_rect.right, src_rect.bottom,
3337             dst_rect.left, dst_rect.top, dst_rect.right, dst_rect.bottom, GL_COLOR_BUFFER_BIT, gl_filter);
3338     checkGLcall("glBlitFramebuffer()");
3339
3340     LEAVE_GL();
3341
3342     if (wined3d_settings.strict_draw_ordering) wglFlush(); /* Flush to ensure ordering across contexts. */
3343
3344     context_release(context);
3345 }
3346
3347 static void surface_blt_to_drawable(IWineD3DDeviceImpl *device,
3348         WINED3DTEXTUREFILTERTYPE filter, BOOL color_key,
3349         IWineD3DSurfaceImpl *src_surface, const RECT *src_rect_in,
3350         IWineD3DSurfaceImpl *dst_surface, const RECT *dst_rect_in)
3351 {
3352     IWineD3DSwapChainImpl *swapchain = NULL;
3353     struct wined3d_context *context;
3354     RECT src_rect, dst_rect;
3355
3356     src_rect = *src_rect_in;
3357     dst_rect = *dst_rect_in;
3358
3359     if (dst_surface->container.type == WINED3D_CONTAINER_SWAPCHAIN)
3360         swapchain = dst_surface->container.u.swapchain;
3361
3362     /* Make sure the surface is up-to-date. This should probably use
3363      * surface_load_location() and worry about the destination surface too,
3364      * unless we're overwriting it completely. */
3365     surface_internal_preload(src_surface, SRGB_RGB);
3366
3367     /* Activate the destination context, set it up for blitting */
3368     context = context_acquire(device, dst_surface);
3369     context_apply_blit_state(context, device);
3370
3371     if (!surface_is_offscreen(dst_surface))
3372         surface_translate_drawable_coords(dst_surface, context->win_handle, &dst_rect);
3373
3374     device->blitter->set_shader((IWineD3DDevice *)device, src_surface);
3375
3376     ENTER_GL();
3377
3378     if (color_key)
3379     {
3380         glEnable(GL_ALPHA_TEST);
3381         checkGLcall("glEnable(GL_ALPHA_TEST)");
3382
3383         /* When the primary render target uses P8, the alpha component
3384          * contains the palette index. Which means that the colorkey is one of
3385          * the palette entries. In other cases pixels that should be masked
3386          * away have alpha set to 0. */
3387         if (primary_render_target_is_p8(device))
3388             glAlphaFunc(GL_NOTEQUAL, (float)src_surface->SrcBltCKey.dwColorSpaceLowValue / 256.0f);
3389         else
3390             glAlphaFunc(GL_NOTEQUAL, 0.0f);
3391         checkGLcall("glAlphaFunc");
3392     }
3393     else
3394     {
3395         glDisable(GL_ALPHA_TEST);
3396         checkGLcall("glDisable(GL_ALPHA_TEST)");
3397     }
3398
3399     draw_textured_quad(src_surface, &src_rect, &dst_rect, filter);
3400
3401     if (color_key)
3402     {
3403         glDisable(GL_ALPHA_TEST);
3404         checkGLcall("glDisable(GL_ALPHA_TEST)");
3405     }
3406
3407     LEAVE_GL();
3408
3409     /* Leave the opengl state valid for blitting */
3410     device->blitter->unset_shader((IWineD3DDevice *)device);
3411
3412     if (wined3d_settings.strict_draw_ordering || (swapchain
3413             && (dst_surface == swapchain->front_buffer || swapchain->num_contexts > 1)))
3414         wglFlush(); /* Flush to ensure ordering across contexts. */
3415
3416     context_release(context);
3417 }
3418
3419 /* Do not call while under the GL lock. */
3420 HRESULT surface_color_fill(IWineD3DSurfaceImpl *s, const RECT *rect, const WINED3DCOLORVALUE *color)
3421 {
3422     IWineD3DDeviceImpl *device = s->resource.device;
3423     const struct blit_shader *blitter;
3424
3425     blitter = wined3d_select_blitter(&device->adapter->gl_info, BLIT_OP_COLOR_FILL,
3426             NULL, 0, 0, NULL, rect, s->resource.usage, s->resource.pool, s->resource.format);
3427     if (!blitter)
3428     {
3429         FIXME("No blitter is capable of performing the requested color fill operation.\n");
3430         return WINED3DERR_INVALIDCALL;
3431     }
3432
3433     return blitter->color_fill(device, s, rect, color);
3434 }
3435
3436 /* Not called from the VTable */
3437 /* Do not call while under the GL lock. */
3438 static HRESULT IWineD3DSurfaceImpl_BltOverride(IWineD3DSurfaceImpl *dst_surface, const RECT *DestRect,
3439         IWineD3DSurfaceImpl *src_surface, const RECT *SrcRect, DWORD Flags, const WINEDDBLTFX *DDBltFx,
3440         WINED3DTEXTUREFILTERTYPE Filter)
3441 {
3442     IWineD3DDeviceImpl *device = dst_surface->resource.device;
3443     const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
3444     IWineD3DSwapChainImpl *srcSwapchain = NULL, *dstSwapchain = NULL;
3445     RECT dst_rect, src_rect;
3446
3447     TRACE("dst_surface %p, dst_rect %s, src_surface %p, src_rect %s, flags %#x, blt_fx %p, filter %s.\n",
3448             dst_surface, wine_dbgstr_rect(DestRect), src_surface, wine_dbgstr_rect(SrcRect),
3449             Flags, DDBltFx, debug_d3dtexturefiltertype(Filter));
3450
3451     /* Get the swapchain. One of the surfaces has to be a primary surface */
3452     if (dst_surface->resource.pool == WINED3DPOOL_SYSTEMMEM)
3453     {
3454         WARN("Destination is in sysmem, rejecting gl blt\n");
3455         return WINED3DERR_INVALIDCALL;
3456     }
3457
3458     if (dst_surface->container.type == WINED3D_CONTAINER_SWAPCHAIN)
3459         dstSwapchain = dst_surface->container.u.swapchain;
3460
3461     if (src_surface)
3462     {
3463         if (src_surface->resource.pool == WINED3DPOOL_SYSTEMMEM)
3464         {
3465             WARN("Src is in sysmem, rejecting gl blt\n");
3466             return WINED3DERR_INVALIDCALL;
3467         }
3468
3469         if (src_surface->container.type == WINED3D_CONTAINER_SWAPCHAIN)
3470             srcSwapchain = src_surface->container.u.swapchain;
3471     }
3472
3473     /* Early sort out of cases where no render target is used */
3474     if (!dstSwapchain && !srcSwapchain
3475             && src_surface != device->render_targets[0]
3476             && dst_surface != device->render_targets[0])
3477     {
3478         TRACE("No surface is render target, not using hardware blit.\n");
3479         return WINED3DERR_INVALIDCALL;
3480     }
3481
3482     /* No destination color keying supported */
3483     if(Flags & (WINEDDBLT_KEYDEST | WINEDDBLT_KEYDESTOVERRIDE)) {
3484         /* Can we support that with glBlendFunc if blitting to the frame buffer? */
3485         TRACE("Destination color key not supported in accelerated Blit, falling back to software\n");
3486         return WINED3DERR_INVALIDCALL;
3487     }
3488
3489     surface_get_rect(dst_surface, DestRect, &dst_rect);
3490     if (src_surface) surface_get_rect(src_surface, SrcRect, &src_rect);
3491
3492     /* The only case where both surfaces on a swapchain are supported is a back buffer -> front buffer blit on the same swapchain */
3493     if (dstSwapchain && dstSwapchain == srcSwapchain && dstSwapchain->back_buffers
3494             && dst_surface == dstSwapchain->front_buffer
3495             && src_surface == dstSwapchain->back_buffers[0])
3496     {
3497         /* Half-Life does a Blt from the back buffer to the front buffer,
3498          * Full surface size, no flags... Use present instead
3499          *
3500          * This path will only be entered for d3d7 and ddraw apps, because d3d8/9 offer no way to blit TO the front buffer
3501          */
3502
3503         /* Check rects - IWineD3DDevice_Present doesn't handle them */
3504         while(1)
3505         {
3506             TRACE("Looking if a Present can be done...\n");
3507             /* Source Rectangle must be full surface */
3508             if (src_rect.left || src_rect.top
3509                     || src_rect.right != src_surface->currentDesc.Width
3510                     || src_rect.bottom != src_surface->currentDesc.Height)
3511             {
3512                 TRACE("No, Source rectangle doesn't match\n");
3513                 break;
3514             }
3515
3516             /* No stretching may occur */
3517             if(src_rect.right != dst_rect.right - dst_rect.left ||
3518                src_rect.bottom != dst_rect.bottom - dst_rect.top) {
3519                 TRACE("No, stretching is done\n");
3520                 break;
3521             }
3522
3523             /* Destination must be full surface or match the clipping rectangle */
3524             if (dst_surface->clipper && ((IWineD3DClipperImpl *)dst_surface->clipper)->hWnd)
3525             {
3526                 RECT cliprect;
3527                 POINT pos[2];
3528                 GetClientRect(((IWineD3DClipperImpl *)dst_surface->clipper)->hWnd, &cliprect);
3529                 pos[0].x = dst_rect.left;
3530                 pos[0].y = dst_rect.top;
3531                 pos[1].x = dst_rect.right;
3532                 pos[1].y = dst_rect.bottom;
3533                 MapWindowPoints(GetDesktopWindow(), ((IWineD3DClipperImpl *)dst_surface->clipper)->hWnd, pos, 2);
3534
3535                 if(pos[0].x != cliprect.left  || pos[0].y != cliprect.top   ||
3536                    pos[1].x != cliprect.right || pos[1].y != cliprect.bottom)
3537                 {
3538                     TRACE("No, dest rectangle doesn't match(clipper)\n");
3539                     TRACE("Clip rect at %s\n", wine_dbgstr_rect(&cliprect));
3540                     TRACE("Blt dest: %s\n", wine_dbgstr_rect(&dst_rect));
3541                     break;
3542                 }
3543             }
3544             else if (dst_rect.left || dst_rect.top
3545                     || dst_rect.right != dst_surface->currentDesc.Width
3546                     || dst_rect.bottom != dst_surface->currentDesc.Height)
3547             {
3548                 TRACE("No, dest rectangle doesn't match(surface size)\n");
3549                 break;
3550             }
3551
3552             TRACE("Yes\n");
3553
3554             /* These flags are unimportant for the flag check, remove them */
3555             if (!(Flags & ~(WINEDDBLT_DONOTWAIT | WINEDDBLT_WAIT)))
3556             {
3557                 WINED3DSWAPEFFECT orig_swap = dstSwapchain->presentParms.SwapEffect;
3558
3559                 /* The idea behind this is that a glReadPixels and a glDrawPixels call
3560                     * take very long, while a flip is fast.
3561                     * This applies to Half-Life, which does such Blts every time it finished
3562                     * a frame, and to Prince of Persia 3D, which uses this to draw at least the main
3563                     * menu. This is also used by all apps when they do windowed rendering
3564                     *
3565                     * The problem is that flipping is not really the same as copying. After a
3566                     * Blt the front buffer is a copy of the back buffer, and the back buffer is
3567                     * untouched. Therefore it's necessary to override the swap effect
3568                     * and to set it back after the flip.
3569                     *
3570                     * Windowed Direct3D < 7 apps do the same. The D3D7 sdk demos are nice
3571                     * testcases.
3572                     */
3573
3574                 dstSwapchain->presentParms.SwapEffect = WINED3DSWAPEFFECT_COPY;
3575                 dstSwapchain->presentParms.PresentationInterval = WINED3DPRESENT_INTERVAL_IMMEDIATE;
3576
3577                 TRACE("Full screen back buffer -> front buffer blt, performing a flip instead\n");
3578                 IWineD3DSwapChain_Present((IWineD3DSwapChain *)dstSwapchain,
3579                         NULL, NULL, dstSwapchain->win_handle, NULL, 0);
3580
3581                 dstSwapchain->presentParms.SwapEffect = orig_swap;
3582
3583                 return WINED3D_OK;
3584             }
3585             break;
3586         }
3587
3588         TRACE("Unsupported blit between buffers on the same swapchain\n");
3589         return WINED3DERR_INVALIDCALL;
3590     } else if(dstSwapchain && dstSwapchain == srcSwapchain) {
3591         FIXME("Implement hardware blit between two surfaces on the same swapchain\n");
3592         return WINED3DERR_INVALIDCALL;
3593     } else if(dstSwapchain && srcSwapchain) {
3594         FIXME("Implement hardware blit between two different swapchains\n");
3595         return WINED3DERR_INVALIDCALL;
3596     }
3597     else if (dstSwapchain)
3598     {
3599         /* Handled with regular texture -> swapchain blit */
3600         if (src_surface == device->render_targets[0])
3601             TRACE("Blit from active render target to a swapchain\n");
3602     }
3603     else if (srcSwapchain && dst_surface == device->render_targets[0])
3604     {
3605         FIXME("Implement blit from a swapchain to the active render target\n");
3606         return WINED3DERR_INVALIDCALL;
3607     }
3608
3609     if ((srcSwapchain || src_surface == device->render_targets[0]) && !dstSwapchain)
3610     {
3611         /* Blit from render target to texture */
3612         BOOL stretchx;
3613
3614         /* P8 read back is not implemented */
3615         if (src_surface->resource.format->id == WINED3DFMT_P8_UINT
3616                 || dst_surface->resource.format->id == WINED3DFMT_P8_UINT)
3617         {
3618             TRACE("P8 read back not supported by frame buffer to texture blit\n");
3619             return WINED3DERR_INVALIDCALL;
3620         }
3621
3622         if(Flags & (WINEDDBLT_KEYSRC | WINEDDBLT_KEYSRCOVERRIDE)) {
3623             TRACE("Color keying not supported by frame buffer to texture blit\n");
3624             return WINED3DERR_INVALIDCALL;
3625             /* Destination color key is checked above */
3626         }
3627
3628         if(dst_rect.right - dst_rect.left != src_rect.right - src_rect.left) {
3629             stretchx = TRUE;
3630         } else {
3631             stretchx = FALSE;
3632         }
3633
3634         /* Blt is a pretty powerful call, while glCopyTexSubImage2D is not. glCopyTexSubImage cannot
3635          * flip the image nor scale it.
3636          *
3637          * -> If the app asks for a unscaled, upside down copy, just perform one glCopyTexSubImage2D call
3638          * -> If the app wants a image width an unscaled width, copy it line per line
3639          * -> If the app wants a image that is scaled on the x axis, and the destination rectangle is smaller
3640          *    than the frame buffer, draw an upside down scaled image onto the fb, read it back and restore the
3641          *    back buffer. This is slower than reading line per line, thus not used for flipping
3642          * -> If the app wants a scaled image with a dest rect that is bigger than the fb, it has to be copied
3643          *    pixel by pixel
3644          *
3645          * If EXT_framebuffer_blit is supported that can be used instead. Note that EXT_framebuffer_blit implies
3646          * FBO support, so it doesn't really make sense to try and make it work with different offscreen rendering
3647          * backends.
3648          */
3649         if (fbo_blit_supported(gl_info, BLIT_OP_BLIT,
3650                 &src_rect, src_surface->resource.usage, src_surface->resource.pool, src_surface->resource.format,
3651                 &dst_rect, dst_surface->resource.usage, dst_surface->resource.pool, dst_surface->resource.format))
3652         {
3653             surface_blt_fbo(device, Filter,
3654                     src_surface, SFLAG_INDRAWABLE, &src_rect,
3655                     dst_surface, SFLAG_INDRAWABLE, &dst_rect);
3656             surface_modify_location(dst_surface, SFLAG_INDRAWABLE, TRUE);
3657         }
3658         else if (!stretchx || dst_rect.right - dst_rect.left > src_surface->currentDesc.Width
3659                 || dst_rect.bottom - dst_rect.top > src_surface->currentDesc.Height)
3660         {
3661             TRACE("No stretching in x direction, using direct framebuffer -> texture copy\n");
3662             fb_copy_to_texture_direct(dst_surface, src_surface, &src_rect, &dst_rect, Filter);
3663         } else {
3664             TRACE("Using hardware stretching to flip / stretch the texture\n");
3665             fb_copy_to_texture_hwstretch(dst_surface, src_surface, &src_rect, &dst_rect, Filter);
3666         }
3667
3668         if (!(dst_surface->flags & SFLAG_DONOTFREE))
3669         {
3670             HeapFree(GetProcessHeap(), 0, dst_surface->resource.heapMemory);
3671             dst_surface->resource.allocatedMemory = NULL;
3672             dst_surface->resource.heapMemory = NULL;
3673         }
3674         else
3675         {
3676             dst_surface->flags &= ~SFLAG_INSYSMEM;
3677         }
3678
3679         return WINED3D_OK;
3680     }
3681     else if (src_surface)
3682     {
3683         /* Blit from offscreen surface to render target */
3684         DWORD oldCKeyFlags = src_surface->CKeyFlags;
3685         WINEDDCOLORKEY oldBltCKey = src_surface->SrcBltCKey;
3686
3687         TRACE("Blt from surface %p to rendertarget %p\n", src_surface, dst_surface);
3688
3689         if (!(Flags & (WINEDDBLT_KEYSRC | WINEDDBLT_KEYSRCOVERRIDE))
3690                 && fbo_blit_supported(gl_info, BLIT_OP_BLIT,
3691                         &src_rect, src_surface->resource.usage, src_surface->resource.pool,
3692                         src_surface->resource.format,
3693                         &dst_rect, dst_surface->resource.usage, dst_surface->resource.pool,
3694                         dst_surface->resource.format))
3695         {
3696             TRACE("Using surface_blt_fbo.\n");
3697             /* The source is always a texture, but never the currently active render target, and the texture
3698              * contents are never upside down. */
3699             surface_blt_fbo(device, Filter,
3700                     src_surface, SFLAG_INDRAWABLE, &src_rect,
3701                     dst_surface, SFLAG_INDRAWABLE, &dst_rect);
3702             surface_modify_location(dst_surface, SFLAG_INDRAWABLE, TRUE);
3703             return WINED3D_OK;
3704         }
3705
3706         if (!(Flags & (WINEDDBLT_KEYSRC | WINEDDBLT_KEYSRCOVERRIDE))
3707                 && arbfp_blit.blit_supported(gl_info, BLIT_OP_BLIT,
3708                         &src_rect, src_surface->resource.usage, src_surface->resource.pool,
3709                         src_surface->resource.format,
3710                         &dst_rect, dst_surface->resource.usage, dst_surface->resource.pool,
3711                         dst_surface->resource.format))
3712         {
3713             return arbfp_blit_surface(device, src_surface, &src_rect, dst_surface, &dst_rect, BLIT_OP_BLIT, Filter);
3714         }
3715
3716         if (!device->blitter->blit_supported(gl_info, BLIT_OP_BLIT,
3717                 &src_rect, src_surface->resource.usage, src_surface->resource.pool, src_surface->resource.format,
3718                 &dst_rect, dst_surface->resource.usage, dst_surface->resource.pool, dst_surface->resource.format))
3719         {
3720             FIXME("Unsupported blit operation falling back to software\n");
3721             return WINED3DERR_INVALIDCALL;
3722         }
3723
3724         /* Color keying: Check if we have to do a color keyed blt,
3725          * and if not check if a color key is activated.
3726          *
3727          * Just modify the color keying parameters in the surface and restore them afterwards
3728          * The surface keeps track of the color key last used to load the opengl surface.
3729          * PreLoad will catch the change to the flags and color key and reload if necessary.
3730          */
3731         if(Flags & WINEDDBLT_KEYSRC) {
3732             /* Use color key from surface */
3733         } else if(Flags & WINEDDBLT_KEYSRCOVERRIDE) {
3734             /* Use color key from DDBltFx */
3735             src_surface->CKeyFlags |= WINEDDSD_CKSRCBLT;
3736             src_surface->SrcBltCKey = DDBltFx->ddckSrcColorkey;
3737         } else {
3738             /* Do not use color key */
3739             src_surface->CKeyFlags &= ~WINEDDSD_CKSRCBLT;
3740         }
3741
3742         surface_blt_to_drawable(device, Filter, Flags & (WINEDDBLT_KEYSRC | WINEDDBLT_KEYSRCOVERRIDE),
3743                 src_surface, &src_rect, dst_surface, &dst_rect);
3744
3745         /* Restore the color key parameters */
3746         src_surface->CKeyFlags = oldCKeyFlags;
3747         src_surface->SrcBltCKey = oldBltCKey;
3748
3749         surface_modify_location(dst_surface, SFLAG_INDRAWABLE, TRUE);
3750
3751         return WINED3D_OK;
3752     }
3753     else
3754     {
3755         /* Source-Less Blit to render target */
3756         if (Flags & WINEDDBLT_COLORFILL)
3757         {
3758             WINED3DCOLORVALUE color;
3759
3760             TRACE("Colorfill\n");
3761
3762             /* The color as given in the Blt function is in the surface format. */
3763             if (!surface_convert_color_to_float(dst_surface, DDBltFx->u5.dwFillColor, &color))
3764                 return WINED3DERR_INVALIDCALL;
3765
3766             return surface_color_fill(dst_surface, &dst_rect, &color);
3767         }
3768     }
3769
3770     /* Default: Fall back to the generic blt. Not an error, a TRACE is enough */
3771     TRACE("Didn't find any usable render target setup for hw blit, falling back to software\n");
3772     return WINED3DERR_INVALIDCALL;
3773 }
3774
3775 static HRESULT IWineD3DSurfaceImpl_BltZ(IWineD3DSurfaceImpl *This, const RECT *DestRect,
3776         IWineD3DSurface *src_surface, const RECT *src_rect, DWORD Flags, const WINEDDBLTFX *DDBltFx)
3777 {
3778     IWineD3DDeviceImpl *device = This->resource.device;
3779     float depth;
3780
3781     if (Flags & WINEDDBLT_DEPTHFILL)
3782     {
3783         switch (This->resource.format->id)
3784         {
3785             case WINED3DFMT_D16_UNORM:
3786                 depth = (float) DDBltFx->u5.dwFillDepth / (float) 0x0000ffff;
3787                 break;
3788             case WINED3DFMT_S1_UINT_D15_UNORM:
3789                 depth = (float) DDBltFx->u5.dwFillDepth / (float) 0x00007fff;
3790                 break;
3791             case WINED3DFMT_D24_UNORM_S8_UINT:
3792             case WINED3DFMT_X8D24_UNORM:
3793                 depth = (float) DDBltFx->u5.dwFillDepth / (float) 0x00ffffff;
3794                 break;
3795             case WINED3DFMT_D32_UNORM:
3796                 depth = (float) DDBltFx->u5.dwFillDepth / (float) 0xffffffff;
3797                 break;
3798             default:
3799                 depth = 0.0f;
3800                 ERR("Unexpected format for depth fill: %s.\n", debug_d3dformat(This->resource.format->id));
3801         }
3802
3803         return IWineD3DDevice_Clear((IWineD3DDevice *)device, DestRect ? 1 : 0, DestRect,
3804                 WINED3DCLEAR_ZBUFFER, 0x00000000, depth, 0x00000000);
3805     }
3806
3807     FIXME("(%p): Unsupp depthstencil blit\n", This);
3808     return WINED3DERR_INVALIDCALL;
3809 }
3810
3811 static HRESULT WINAPI IWineD3DSurfaceImpl_Blt(IWineD3DSurface *iface, const RECT *DestRect,
3812         IWineD3DSurface *src_surface, const RECT *SrcRect, DWORD Flags,
3813         const WINEDDBLTFX *DDBltFx, WINED3DTEXTUREFILTERTYPE Filter)
3814 {
3815     IWineD3DSurfaceImpl *This = (IWineD3DSurfaceImpl *)iface;
3816     IWineD3DSurfaceImpl *src = (IWineD3DSurfaceImpl *)src_surface;
3817     IWineD3DDeviceImpl *device = This->resource.device;
3818
3819     TRACE("iface %p, dst_rect %s, src_surface %p, src_rect %s, flags %#x, fx %p, filter %s.\n",
3820             iface, wine_dbgstr_rect(DestRect), src_surface, wine_dbgstr_rect(SrcRect),
3821             Flags, DDBltFx, debug_d3dtexturefiltertype(Filter));
3822     TRACE("Usage is %s.\n", debug_d3dusage(This->resource.usage));
3823
3824     if ((This->flags & SFLAG_LOCKED) || (src && (src->flags & SFLAG_LOCKED)))
3825     {
3826         WARN(" Surface is busy, returning DDERR_SURFACEBUSY\n");
3827         return WINEDDERR_SURFACEBUSY;
3828     }
3829
3830     /* Accessing the depth stencil is supposed to fail between a BeginScene and EndScene pair,
3831      * except depth blits, which seem to work
3832      */
3833     if (This == device->depth_stencil || (src && src == device->depth_stencil))
3834     {
3835         if (device->inScene && !(Flags & WINEDDBLT_DEPTHFILL))
3836         {
3837             TRACE("Attempt to access the depth stencil surface in a BeginScene / EndScene pair, returning WINED3DERR_INVALIDCALL\n");
3838             return WINED3DERR_INVALIDCALL;
3839         }
3840         else if (SUCCEEDED(IWineD3DSurfaceImpl_BltZ(This, DestRect, src_surface, SrcRect, Flags, DDBltFx)))
3841         {
3842             TRACE("Z Blit override handled the blit\n");
3843             return WINED3D_OK;
3844         }
3845     }
3846
3847     /* Special cases for RenderTargets */
3848     if ((This->resource.usage & WINED3DUSAGE_RENDERTARGET)
3849             || (src && (src->resource.usage & WINED3DUSAGE_RENDERTARGET)))
3850     {
3851         if (SUCCEEDED(IWineD3DSurfaceImpl_BltOverride(This, DestRect, src, SrcRect, Flags, DDBltFx, Filter)))
3852             return WINED3D_OK;
3853     }
3854
3855     /* For the rest call the X11 surface implementation.
3856      * For RenderTargets this should be implemented OpenGL accelerated in BltOverride,
3857      * other Blts are rather rare. */
3858     return IWineD3DBaseSurfaceImpl_Blt(iface, DestRect, src_surface, SrcRect, Flags, DDBltFx, Filter);
3859 }
3860
3861 static HRESULT WINAPI IWineD3DSurfaceImpl_BltFast(IWineD3DSurface *iface, DWORD dstx, DWORD dsty,
3862         IWineD3DSurface *src_surface, const RECT *rsrc, DWORD trans)
3863 {
3864     IWineD3DSurfaceImpl *This = (IWineD3DSurfaceImpl *)iface;
3865     IWineD3DSurfaceImpl *src = (IWineD3DSurfaceImpl *)src_surface;
3866     IWineD3DDeviceImpl *device = This->resource.device;
3867
3868     TRACE("iface %p, dst_x %u, dst_y %u, src_surface %p, src_rect %s, flags %#x.\n",
3869             iface, dstx, dsty, src_surface, wine_dbgstr_rect(rsrc), trans);
3870
3871     if ((This->flags & SFLAG_LOCKED) || (src->flags & SFLAG_LOCKED))
3872     {
3873         WARN(" Surface is busy, returning DDERR_SURFACEBUSY\n");
3874         return WINEDDERR_SURFACEBUSY;
3875     }
3876
3877     if (device->inScene && (This == device->depth_stencil || src == device->depth_stencil))
3878     {
3879         TRACE("Attempt to access the depth stencil surface in a BeginScene / EndScene pair, returning WINED3DERR_INVALIDCALL\n");
3880         return WINED3DERR_INVALIDCALL;
3881     }
3882
3883     /* Special cases for RenderTargets */
3884     if ((This->resource.usage & WINED3DUSAGE_RENDERTARGET)
3885             || (src->resource.usage & WINED3DUSAGE_RENDERTARGET))
3886     {
3887
3888         RECT SrcRect, DstRect;
3889         DWORD Flags=0;
3890
3891         surface_get_rect(src, rsrc, &SrcRect);
3892
3893         DstRect.left = dstx;
3894         DstRect.top=dsty;
3895         DstRect.right = dstx + SrcRect.right - SrcRect.left;
3896         DstRect.bottom = dsty + SrcRect.bottom - SrcRect.top;
3897
3898         /* Convert BltFast flags into Btl ones because it is called from SurfaceImpl_Blt as well */
3899         if(trans & WINEDDBLTFAST_SRCCOLORKEY)
3900             Flags |= WINEDDBLT_KEYSRC;
3901         if(trans & WINEDDBLTFAST_DESTCOLORKEY)
3902             Flags |= WINEDDBLT_KEYDEST;
3903         if(trans & WINEDDBLTFAST_WAIT)
3904             Flags |= WINEDDBLT_WAIT;
3905         if(trans & WINEDDBLTFAST_DONOTWAIT)
3906             Flags |= WINEDDBLT_DONOTWAIT;
3907
3908         if (SUCCEEDED(IWineD3DSurfaceImpl_BltOverride(This,
3909                 &DstRect, src, &SrcRect, Flags, NULL, WINED3DTEXF_POINT)))
3910             return WINED3D_OK;
3911     }
3912
3913     return IWineD3DBaseSurfaceImpl_BltFast(iface, dstx, dsty, src_surface, rsrc, trans);
3914 }
3915
3916 static HRESULT WINAPI IWineD3DSurfaceImpl_RealizePalette(IWineD3DSurface *iface)
3917 {
3918     IWineD3DSurfaceImpl *This = (IWineD3DSurfaceImpl *) iface;
3919     RGBQUAD col[256];
3920     IWineD3DPaletteImpl *pal = This->palette;
3921     unsigned int n;
3922     TRACE("(%p)\n", This);
3923
3924     if (!pal) return WINED3D_OK;
3925
3926     if (This->resource.format->id == WINED3DFMT_P8_UINT
3927             || This->resource.format->id == WINED3DFMT_P8_UINT_A8_UNORM)
3928     {
3929         if (This->resource.usage & WINED3DUSAGE_RENDERTARGET)
3930         {
3931             /* Make sure the texture is up to date. This call doesn't do
3932              * anything if the texture is already up to date. */
3933             surface_load_location(This, SFLAG_INTEXTURE, NULL);
3934
3935             /* We want to force a palette refresh, so mark the drawable as not being up to date */
3936             surface_modify_location(This, SFLAG_INDRAWABLE, FALSE);
3937         }
3938         else
3939         {
3940             if (!(This->flags & SFLAG_INSYSMEM))
3941             {
3942                 TRACE("Palette changed with surface that does not have an up to date system memory copy.\n");
3943                 surface_load_location(This, SFLAG_INSYSMEM, NULL);
3944             }
3945             TRACE("Dirtifying surface\n");
3946             surface_modify_location(This, SFLAG_INSYSMEM, TRUE);
3947         }
3948     }
3949
3950     if (This->flags & SFLAG_DIBSECTION)
3951     {
3952         TRACE("(%p): Updating the hdc's palette\n", This);
3953         for (n=0; n<256; n++) {
3954             col[n].rgbRed   = pal->palents[n].peRed;
3955             col[n].rgbGreen = pal->palents[n].peGreen;
3956             col[n].rgbBlue  = pal->palents[n].peBlue;
3957             col[n].rgbReserved = 0;
3958         }
3959         SetDIBColorTable(This->hDC, 0, 256, col);
3960     }
3961
3962     /* Propagate the changes to the drawable when we have a palette. */
3963     if (This->resource.usage & WINED3DUSAGE_RENDERTARGET)
3964         surface_load_location(This, SFLAG_INDRAWABLE, NULL);
3965
3966     return WINED3D_OK;
3967 }
3968
3969 static HRESULT WINAPI IWineD3DSurfaceImpl_PrivateSetup(IWineD3DSurface *iface) {
3970     /** Check against the maximum texture sizes supported by the video card **/
3971     IWineD3DSurfaceImpl *This = (IWineD3DSurfaceImpl *) iface;
3972     const struct wined3d_gl_info *gl_info = &This->resource.device->adapter->gl_info;
3973     unsigned int pow2Width, pow2Height;
3974
3975     This->texture_name = 0;
3976     This->texture_target = GL_TEXTURE_2D;
3977
3978     /* Non-power2 support */
3979     if (gl_info->supported[ARB_TEXTURE_NON_POWER_OF_TWO] || gl_info->supported[WINED3D_GL_NORMALIZED_TEXRECT])
3980     {
3981         pow2Width = This->currentDesc.Width;
3982         pow2Height = This->currentDesc.Height;
3983     }
3984     else
3985     {
3986         /* Find the nearest pow2 match */
3987         pow2Width = pow2Height = 1;
3988         while (pow2Width < This->currentDesc.Width) pow2Width <<= 1;
3989         while (pow2Height < This->currentDesc.Height) pow2Height <<= 1;
3990     }
3991     This->pow2Width  = pow2Width;
3992     This->pow2Height = pow2Height;
3993
3994     if (pow2Width > This->currentDesc.Width || pow2Height > This->currentDesc.Height)
3995     {
3996         /* TODO: Add support for non power two compressed textures. */
3997         if (This->resource.format->flags & WINED3DFMT_FLAG_COMPRESSED)
3998         {
3999             FIXME("(%p) Compressed non-power-two textures are not supported w(%d) h(%d)\n",
4000                   This, This->currentDesc.Width, This->currentDesc.Height);
4001             return WINED3DERR_NOTAVAILABLE;
4002         }
4003     }
4004
4005     if (pow2Width != This->currentDesc.Width
4006             || pow2Height != This->currentDesc.Height)
4007     {
4008         This->flags |= SFLAG_NONPOW2;
4009     }
4010
4011     TRACE("%p\n", This);
4012     if ((This->pow2Width > gl_info->limits.texture_size || This->pow2Height > gl_info->limits.texture_size)
4013             && !(This->resource.usage & (WINED3DUSAGE_RENDERTARGET | WINED3DUSAGE_DEPTHSTENCIL)))
4014     {
4015         /* one of three options
4016         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)
4017         2: Set the texture to the maximum size (bad idea)
4018         3:    WARN and return WINED3DERR_NOTAVAILABLE;
4019         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.
4020         */
4021         if(This->resource.pool == WINED3DPOOL_DEFAULT || This->resource.pool == WINED3DPOOL_MANAGED)
4022         {
4023             WARN("(%p) Unable to allocate a surface which exceeds the maximum OpenGL texture size\n", This);
4024             return WINED3DERR_NOTAVAILABLE;
4025         }
4026
4027         /* We should never use this surface in combination with OpenGL! */
4028         TRACE("(%p) Creating an oversized surface: %ux%u\n", This, This->pow2Width, This->pow2Height);
4029     }
4030     else
4031     {
4032         /* Don't use ARB_TEXTURE_RECTANGLE in case the surface format is P8 and EXT_PALETTED_TEXTURE
4033            is used in combination with texture uploads (RTL_READTEX/RTL_TEXTEX). The reason is that EXT_PALETTED_TEXTURE
4034            doesn't work in combination with ARB_TEXTURE_RECTANGLE.
4035         */
4036         if (This->flags & SFLAG_NONPOW2 && gl_info->supported[ARB_TEXTURE_RECTANGLE]
4037                 && !(This->resource.format->id == WINED3DFMT_P8_UINT
4038                 && gl_info->supported[EXT_PALETTED_TEXTURE]
4039                 && wined3d_settings.rendertargetlock_mode == RTL_READTEX))
4040         {
4041             This->texture_target = GL_TEXTURE_RECTANGLE_ARB;
4042             This->pow2Width  = This->currentDesc.Width;
4043             This->pow2Height = This->currentDesc.Height;
4044             This->flags &= ~(SFLAG_NONPOW2 | SFLAG_NORMCOORD);
4045         }
4046     }
4047
4048     switch (wined3d_settings.offscreen_rendering_mode)
4049     {
4050         case ORM_FBO:
4051             This->get_drawable_size = get_drawable_size_fbo;
4052             break;
4053
4054         case ORM_BACKBUFFER:
4055             This->get_drawable_size = get_drawable_size_backbuffer;
4056             break;
4057
4058         default:
4059             ERR("Unhandled offscreen rendering mode %#x.\n", wined3d_settings.offscreen_rendering_mode);
4060             return WINED3DERR_INVALIDCALL;
4061     }
4062
4063     This->flags |= SFLAG_INSYSMEM;
4064
4065     return WINED3D_OK;
4066 }
4067
4068 /* GL locking is done by the caller */
4069 static void surface_depth_blt(IWineD3DSurfaceImpl *This, const struct wined3d_gl_info *gl_info,
4070         GLuint texture, GLsizei w, GLsizei h, GLenum target)
4071 {
4072     IWineD3DDeviceImpl *device = This->resource.device;
4073     GLint compare_mode = GL_NONE;
4074     struct blt_info info;
4075     GLint old_binding = 0;
4076     RECT rect;
4077
4078     glPushAttrib(GL_ENABLE_BIT | GL_DEPTH_BUFFER_BIT | GL_COLOR_BUFFER_BIT | GL_VIEWPORT_BIT);
4079
4080     glDisable(GL_CULL_FACE);
4081     glDisable(GL_BLEND);
4082     glDisable(GL_ALPHA_TEST);
4083     glDisable(GL_SCISSOR_TEST);
4084     glDisable(GL_STENCIL_TEST);
4085     glEnable(GL_DEPTH_TEST);
4086     glDepthFunc(GL_ALWAYS);
4087     glDepthMask(GL_TRUE);
4088     glColorMask(GL_FALSE, GL_FALSE, GL_FALSE, GL_FALSE);
4089     glViewport(0, 0, w, h);
4090
4091     SetRect(&rect, 0, h, w, 0);
4092     surface_get_blt_info(target, &rect, w, h, &info);
4093     GL_EXTCALL(glActiveTextureARB(GL_TEXTURE0_ARB));
4094     glGetIntegerv(info.binding, &old_binding);
4095     glBindTexture(info.bind_target, texture);
4096     if (gl_info->supported[ARB_SHADOW])
4097     {
4098         glGetTexParameteriv(info.bind_target, GL_TEXTURE_COMPARE_MODE_ARB, &compare_mode);
4099         if (compare_mode != GL_NONE) glTexParameteri(info.bind_target, GL_TEXTURE_COMPARE_MODE_ARB, GL_NONE);
4100     }
4101
4102     device->shader_backend->shader_select_depth_blt((IWineD3DDevice *)device,
4103             info.tex_type, &This->ds_current_size);
4104
4105     glBegin(GL_TRIANGLE_STRIP);
4106     glTexCoord3fv(info.coords[0]);
4107     glVertex2f(-1.0f, -1.0f);
4108     glTexCoord3fv(info.coords[1]);
4109     glVertex2f(1.0f, -1.0f);
4110     glTexCoord3fv(info.coords[2]);
4111     glVertex2f(-1.0f, 1.0f);
4112     glTexCoord3fv(info.coords[3]);
4113     glVertex2f(1.0f, 1.0f);
4114     glEnd();
4115
4116     if (compare_mode != GL_NONE) glTexParameteri(info.bind_target, GL_TEXTURE_COMPARE_MODE_ARB, compare_mode);
4117     glBindTexture(info.bind_target, old_binding);
4118
4119     glPopAttrib();
4120
4121     device->shader_backend->shader_deselect_depth_blt((IWineD3DDevice *)device);
4122 }
4123
4124 void surface_modify_ds_location(IWineD3DSurfaceImpl *surface,
4125         DWORD location, UINT w, UINT h)
4126 {
4127     TRACE("surface %p, new location %#x, w %u, h %u.\n", surface, location, w, h);
4128
4129     if (location & ~SFLAG_DS_LOCATIONS)
4130         FIXME("Invalid location (%#x) specified.\n", location);
4131
4132     surface->ds_current_size.cx = w;
4133     surface->ds_current_size.cy = h;
4134     surface->flags &= ~SFLAG_DS_LOCATIONS;
4135     surface->flags |= location;
4136 }
4137
4138 /* Context activation is done by the caller. */
4139 void surface_load_ds_location(IWineD3DSurfaceImpl *surface, struct wined3d_context *context, DWORD location)
4140 {
4141     IWineD3DDeviceImpl *device = surface->resource.device;
4142     const struct wined3d_gl_info *gl_info = context->gl_info;
4143
4144     TRACE("surface %p, new location %#x.\n", surface, location);
4145
4146     /* TODO: Make this work for modes other than FBO */
4147     if (wined3d_settings.offscreen_rendering_mode != ORM_FBO) return;
4148
4149     if (!(surface->flags & location))
4150     {
4151         surface->ds_current_size.cx = 0;
4152         surface->ds_current_size.cy = 0;
4153     }
4154
4155     if (surface->ds_current_size.cx == surface->currentDesc.Width
4156             && surface->ds_current_size.cy == surface->currentDesc.Height)
4157     {
4158         TRACE("Location (%#x) is already up to date.\n", location);
4159         return;
4160     }
4161
4162     if (surface->current_renderbuffer)
4163     {
4164         FIXME("Not supported with fixed up depth stencil.\n");
4165         return;
4166     }
4167
4168     if (!(surface->flags & SFLAG_LOCATIONS))
4169     {
4170         FIXME("No up to date depth stencil location.\n");
4171         surface->flags |= location;
4172         return;
4173     }
4174
4175     if (location == SFLAG_DS_OFFSCREEN)
4176     {
4177         GLint old_binding = 0;
4178         GLenum bind_target;
4179         GLsizei w, h;
4180
4181         /* The render target is allowed to be smaller than the depth/stencil
4182          * buffer, so the onscreen depth/stencil buffer is potentially smaller
4183          * than the offscreen surface. Don't overwrite the offscreen surface
4184          * with undefined data. */
4185         w = min(surface->currentDesc.Width, context->swapchain->presentParms.BackBufferWidth);
4186         h = min(surface->currentDesc.Height, context->swapchain->presentParms.BackBufferHeight);
4187
4188         TRACE("Copying onscreen depth buffer to depth texture.\n");
4189
4190         ENTER_GL();
4191
4192         if (!device->depth_blt_texture)
4193         {
4194             glGenTextures(1, &device->depth_blt_texture);
4195         }
4196
4197         /* Note that we use depth_blt here as well, rather than glCopyTexImage2D
4198          * directly on the FBO texture. That's because we need to flip. */
4199         context_bind_fbo(context, GL_FRAMEBUFFER, NULL);
4200         if (surface->texture_target == GL_TEXTURE_RECTANGLE_ARB)
4201         {
4202             glGetIntegerv(GL_TEXTURE_BINDING_RECTANGLE_ARB, &old_binding);
4203             bind_target = GL_TEXTURE_RECTANGLE_ARB;
4204         }
4205         else
4206         {
4207             glGetIntegerv(GL_TEXTURE_BINDING_2D, &old_binding);
4208             bind_target = GL_TEXTURE_2D;
4209         }
4210         glBindTexture(bind_target, device->depth_blt_texture);
4211         glCopyTexImage2D(bind_target, surface->texture_level, surface->resource.format->glInternal, 0, 0, w, h, 0);
4212         glTexParameteri(bind_target, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
4213         glTexParameteri(bind_target, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
4214         glTexParameteri(bind_target, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
4215         glTexParameteri(bind_target, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
4216         glTexParameteri(bind_target, GL_TEXTURE_WRAP_R, GL_CLAMP_TO_EDGE);
4217         glTexParameteri(bind_target, GL_DEPTH_TEXTURE_MODE_ARB, GL_LUMINANCE);
4218         glBindTexture(bind_target, old_binding);
4219
4220         /* Setup the destination */
4221         if (!device->depth_blt_rb)
4222         {
4223             gl_info->fbo_ops.glGenRenderbuffers(1, &device->depth_blt_rb);
4224             checkGLcall("glGenRenderbuffersEXT");
4225         }
4226         if (device->depth_blt_rb_w != w || device->depth_blt_rb_h != h)
4227         {
4228             gl_info->fbo_ops.glBindRenderbuffer(GL_RENDERBUFFER, device->depth_blt_rb);
4229             checkGLcall("glBindRenderbufferEXT");
4230             gl_info->fbo_ops.glRenderbufferStorage(GL_RENDERBUFFER, GL_RGBA8, w, h);
4231             checkGLcall("glRenderbufferStorageEXT");
4232             device->depth_blt_rb_w = w;
4233             device->depth_blt_rb_h = h;
4234         }
4235
4236         context_bind_fbo(context, GL_FRAMEBUFFER, &context->dst_fbo);
4237         gl_info->fbo_ops.glFramebufferRenderbuffer(GL_FRAMEBUFFER,
4238                 GL_COLOR_ATTACHMENT0, GL_RENDERBUFFER, device->depth_blt_rb);
4239         checkGLcall("glFramebufferRenderbufferEXT");
4240         context_attach_depth_stencil_fbo(context, GL_FRAMEBUFFER, surface, FALSE);
4241
4242         /* Do the actual blit */
4243         surface_depth_blt(surface, gl_info, device->depth_blt_texture, w, h, bind_target);
4244         checkGLcall("depth_blt");
4245
4246         if (context->current_fbo) context_bind_fbo(context, GL_FRAMEBUFFER, &context->current_fbo->id);
4247         else context_bind_fbo(context, GL_FRAMEBUFFER, NULL);
4248
4249         LEAVE_GL();
4250
4251         if (wined3d_settings.strict_draw_ordering) wglFlush(); /* Flush to ensure ordering across contexts. */
4252     }
4253     else if (location == SFLAG_DS_ONSCREEN)
4254     {
4255         TRACE("Copying depth texture to onscreen depth buffer.\n");
4256
4257         ENTER_GL();
4258
4259         context_bind_fbo(context, GL_FRAMEBUFFER, NULL);
4260         surface_depth_blt(surface, gl_info, surface->texture_name,
4261                 surface->currentDesc.Width, surface->currentDesc.Height, surface->texture_target);
4262         checkGLcall("depth_blt");
4263
4264         if (context->current_fbo) context_bind_fbo(context, GL_FRAMEBUFFER, &context->current_fbo->id);
4265
4266         LEAVE_GL();
4267
4268         if (wined3d_settings.strict_draw_ordering) wglFlush(); /* Flush to ensure ordering across contexts. */
4269     }
4270     else
4271     {
4272         ERR("Invalid location (%#x) specified.\n", location);
4273     }
4274
4275     surface->flags |= location;
4276     surface->ds_current_size.cx = surface->currentDesc.Width;
4277     surface->ds_current_size.cy = surface->currentDesc.Height;
4278 }
4279
4280 void surface_modify_location(IWineD3DSurfaceImpl *surface, DWORD flag, BOOL persistent)
4281 {
4282     IWineD3DSurfaceImpl *overlay;
4283
4284     TRACE("surface %p, location %s, persistent %#x.\n",
4285             surface, debug_surflocation(flag), persistent);
4286
4287     if (wined3d_settings.offscreen_rendering_mode == ORM_FBO)
4288     {
4289         if (surface_is_offscreen(surface))
4290         {
4291             /* With ORM_FBO, SFLAG_INTEXTURE and SFLAG_INDRAWABLE are the same for offscreen targets. */
4292             if (flag & (SFLAG_INTEXTURE | SFLAG_INDRAWABLE)) flag |= (SFLAG_INTEXTURE | SFLAG_INDRAWABLE);
4293         }
4294         else
4295         {
4296             TRACE("Surface %p is an onscreen surface.\n", surface);
4297         }
4298     }
4299
4300     if (persistent)
4301     {
4302         if (((surface->flags & SFLAG_INTEXTURE) && !(flag & SFLAG_INTEXTURE))
4303                 || ((surface->flags & SFLAG_INSRGBTEX) && !(flag & SFLAG_INSRGBTEX)))
4304         {
4305             if (surface->container.type == WINED3D_CONTAINER_TEXTURE)
4306             {
4307                 TRACE("Passing to container.\n");
4308                 IWineD3DBaseTexture_SetDirty((IWineD3DBaseTexture *)surface->container.u.texture, TRUE);
4309             }
4310         }
4311         surface->flags &= ~SFLAG_LOCATIONS;
4312         surface->flags |= flag;
4313
4314         /* Redraw emulated overlays, if any */
4315         if (flag & SFLAG_INDRAWABLE && !list_empty(&surface->overlays))
4316         {
4317             LIST_FOR_EACH_ENTRY(overlay, &surface->overlays, IWineD3DSurfaceImpl, overlay_entry)
4318             {
4319                 IWineD3DSurface_DrawOverlay((IWineD3DSurface *)overlay);
4320             }
4321         }
4322     }
4323     else
4324     {
4325         if ((surface->flags & (SFLAG_INTEXTURE | SFLAG_INSRGBTEX)) && (flag & (SFLAG_INTEXTURE | SFLAG_INSRGBTEX)))
4326         {
4327             if (surface->container.type == WINED3D_CONTAINER_TEXTURE)
4328             {
4329                 TRACE("Passing to container\n");
4330                 IWineD3DBaseTexture_SetDirty((IWineD3DBaseTexture *)surface->container.u.texture, TRUE);
4331             }
4332         }
4333         surface->flags &= ~flag;
4334     }
4335
4336     if (!(surface->flags & SFLAG_LOCATIONS))
4337     {
4338         ERR("Surface %p does not have any up to date location.\n", surface);
4339     }
4340 }
4341
4342 HRESULT surface_load_location(IWineD3DSurfaceImpl *surface, DWORD flag, const RECT *rect)
4343 {
4344     IWineD3DDeviceImpl *device = surface->resource.device;
4345     const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
4346     BOOL drawable_read_ok = surface_is_offscreen(surface);
4347     struct wined3d_format format;
4348     CONVERT_TYPES convert;
4349     int width, pitch, outpitch;
4350     BYTE *mem;
4351     BOOL in_fbo = FALSE;
4352
4353     TRACE("surface %p, location %s, rect %s.\n", surface, debug_surflocation(flag), wine_dbgstr_rect(rect));
4354
4355     if (surface->resource.usage & WINED3DUSAGE_DEPTHSTENCIL)
4356     {
4357         if (flag == SFLAG_INTEXTURE)
4358         {
4359             struct wined3d_context *context = context_acquire(device, NULL);
4360             surface_load_ds_location(surface, context, SFLAG_DS_OFFSCREEN);
4361             context_release(context);
4362             return WINED3D_OK;
4363         }
4364         else
4365         {
4366             FIXME("Unimplemented location %s for depth/stencil buffers.\n", debug_surflocation(flag));
4367             return WINED3DERR_INVALIDCALL;
4368         }
4369     }
4370
4371     if (wined3d_settings.offscreen_rendering_mode == ORM_FBO)
4372     {
4373         if (surface_is_offscreen(surface))
4374         {
4375             /* With ORM_FBO, SFLAG_INTEXTURE and SFLAG_INDRAWABLE are the same for offscreen targets.
4376              * Prefer SFLAG_INTEXTURE. */
4377             if (flag == SFLAG_INDRAWABLE) flag = SFLAG_INTEXTURE;
4378             drawable_read_ok = FALSE;
4379             in_fbo = TRUE;
4380         }
4381         else
4382         {
4383             TRACE("Surface %p is an onscreen surface.\n", surface);
4384         }
4385     }
4386
4387     if (surface->flags & flag)
4388     {
4389         TRACE("Location already up to date\n");
4390         return WINED3D_OK;
4391     }
4392
4393     if (!(surface->flags & SFLAG_LOCATIONS))
4394     {
4395         ERR("Surface %p does not have any up to date location.\n", surface);
4396         surface->flags |= SFLAG_LOST;
4397         return WINED3DERR_DEVICELOST;
4398     }
4399
4400     if (flag == SFLAG_INSYSMEM)
4401     {
4402         surface_prepare_system_memory(surface);
4403
4404         /* Download the surface to system memory */
4405         if (surface->flags & (SFLAG_INTEXTURE | SFLAG_INSRGBTEX))
4406         {
4407             struct wined3d_context *context = NULL;
4408
4409             if (!device->isInDraw) context = context_acquire(device, NULL);
4410
4411             surface_bind_and_dirtify(surface, !(surface->flags & SFLAG_INTEXTURE));
4412             surface_download_data(surface, gl_info);
4413
4414             if (context) context_release(context);
4415         }
4416         else
4417         {
4418             /* Note: It might be faster to download into a texture first. */
4419             read_from_framebuffer(surface, rect, surface->resource.allocatedMemory,
4420                     IWineD3DSurface_GetPitch((IWineD3DSurface *)surface));
4421         }
4422     }
4423     else if (flag == SFLAG_INDRAWABLE)
4424     {
4425         if (wined3d_settings.rendertargetlock_mode == RTL_READTEX)
4426             surface_load_location(surface, SFLAG_INTEXTURE, NULL);
4427
4428         if (surface->flags & SFLAG_INTEXTURE)
4429         {
4430             RECT r;
4431
4432             surface_get_rect(surface, rect, &r);
4433             surface_blt_to_drawable(device, WINED3DTEXF_POINT, FALSE, surface, &r, surface, &r);
4434         }
4435         else
4436         {
4437             int byte_count;
4438             if ((surface->flags & SFLAG_LOCATIONS) == SFLAG_INSRGBTEX)
4439             {
4440                 /* This needs a shader to convert the srgb data sampled from the GL texture into RGB
4441                  * values, otherwise we get incorrect values in the target. For now go the slow way
4442                  * via a system memory copy
4443                  */
4444                 surface_load_location(surface, SFLAG_INSYSMEM, rect);
4445             }
4446
4447             d3dfmt_get_conv(surface, FALSE /* We need color keying */,
4448                     FALSE /* We won't use textures */, &format, &convert);
4449
4450             /* The width is in 'length' not in bytes */
4451             width = surface->currentDesc.Width;
4452             pitch = IWineD3DSurface_GetPitch((IWineD3DSurface *)surface);
4453
4454             /* Don't use PBOs for converted surfaces. During PBO conversion we look at SFLAG_CONVERTED
4455              * but it isn't set (yet) in all cases it is getting called. */
4456             if ((convert != NO_CONVERSION) && (surface->flags & SFLAG_PBO))
4457             {
4458                 struct wined3d_context *context = NULL;
4459
4460                 TRACE("Removing the pbo attached to surface %p.\n", surface);
4461
4462                 if (!device->isInDraw) context = context_acquire(device, NULL);
4463                 surface_remove_pbo(surface, gl_info);
4464                 if (context) context_release(context);
4465             }
4466
4467             if ((convert != NO_CONVERSION) && surface->resource.allocatedMemory)
4468             {
4469                 int height = surface->currentDesc.Height;
4470                 byte_count = format.conv_byte_count;
4471
4472                 /* Stick to the alignment for the converted surface too, makes it easier to load the surface */
4473                 outpitch = width * byte_count;
4474                 outpitch = (outpitch + device->surface_alignment - 1) & ~(device->surface_alignment - 1);
4475
4476                 mem = HeapAlloc(GetProcessHeap(), 0, outpitch * height);
4477                 if(!mem) {
4478                     ERR("Out of memory %d, %d!\n", outpitch, height);
4479                     return WINED3DERR_OUTOFVIDEOMEMORY;
4480                 }
4481                 d3dfmt_convert_surface(surface->resource.allocatedMemory, mem, pitch,
4482                         width, height, outpitch, convert, surface);
4483
4484                 surface->flags |= SFLAG_CONVERTED;
4485             }
4486             else
4487             {
4488                 surface->flags &= ~SFLAG_CONVERTED;
4489                 mem = surface->resource.allocatedMemory;
4490                 byte_count = format.byte_count;
4491             }
4492
4493             flush_to_framebuffer_drawpixels(surface, format.glFormat, format.glType, byte_count, mem);
4494
4495             /* Don't delete PBO memory */
4496             if ((mem != surface->resource.allocatedMemory) && !(surface->flags & SFLAG_PBO))
4497                 HeapFree(GetProcessHeap(), 0, mem);
4498         }
4499     }
4500     else /* if(flag & (SFLAG_INTEXTURE | SFLAG_INSRGBTEX)) */
4501     {
4502         const DWORD attach_flags = WINED3DFMT_FLAG_FBO_ATTACHABLE | WINED3DFMT_FLAG_FBO_ATTACHABLE_SRGB;
4503
4504         if (drawable_read_ok && (surface->flags & SFLAG_INDRAWABLE))
4505         {
4506             read_from_framebuffer_texture(surface, flag == SFLAG_INSRGBTEX);
4507         }
4508         else if (surface->flags & (SFLAG_INSRGBTEX | SFLAG_INTEXTURE)
4509                 && (surface->resource.format->flags & attach_flags) == attach_flags
4510                 && fbo_blit_supported(gl_info, BLIT_OP_BLIT,
4511                         NULL, surface->resource.usage, surface->resource.pool, surface->resource.format,
4512                         NULL, surface->resource.usage, surface->resource.pool, surface->resource.format))
4513         {
4514             DWORD src_location = flag == SFLAG_INSRGBTEX ? SFLAG_INTEXTURE : SFLAG_INSRGBTEX;
4515             RECT rect = {0, 0, surface->currentDesc.Width, surface->currentDesc.Height};
4516
4517             surface_blt_fbo(surface->resource.device, WINED3DTEXF_POINT,
4518                     surface, src_location, &rect, surface, flag, &rect);
4519         }
4520         else
4521         {
4522             /* Upload from system memory */
4523             BOOL srgb = flag == SFLAG_INSRGBTEX;
4524             struct wined3d_context *context = NULL;
4525
4526             d3dfmt_get_conv(surface, TRUE /* We need color keying */,
4527                     TRUE /* We will use textures */, &format, &convert);
4528
4529             if (srgb)
4530             {
4531                 if ((surface->flags & (SFLAG_INTEXTURE | SFLAG_INSYSMEM)) == SFLAG_INTEXTURE)
4532                 {
4533                     /* Performance warning... */
4534                     FIXME("Downloading RGB surface %p to reload it as sRGB.\n", surface);
4535                     surface_load_location(surface, SFLAG_INSYSMEM, rect);
4536                 }
4537             }
4538             else
4539             {
4540                 if ((surface->flags & (SFLAG_INSRGBTEX | SFLAG_INSYSMEM)) == SFLAG_INSRGBTEX)
4541                 {
4542                     /* Performance warning... */
4543                     FIXME("Downloading sRGB surface %p to reload it as RGB.\n", surface);
4544                     surface_load_location(surface, SFLAG_INSYSMEM, rect);
4545                 }
4546             }
4547             if (!(surface->flags & SFLAG_INSYSMEM))
4548             {
4549                 WARN("Trying to load a texture from sysmem, but SFLAG_INSYSMEM is not set.\n");
4550                 /* Lets hope we get it from somewhere... */
4551                 surface_load_location(surface, SFLAG_INSYSMEM, rect);
4552             }
4553
4554             if (!device->isInDraw) context = context_acquire(device, NULL);
4555
4556             surface_prepare_texture(surface, gl_info, srgb);
4557             surface_bind_and_dirtify(surface, srgb);
4558
4559             if (surface->CKeyFlags & WINEDDSD_CKSRCBLT)
4560             {
4561                 surface->flags |= SFLAG_GLCKEY;
4562                 surface->glCKey = surface->SrcBltCKey;
4563             }
4564             else surface->flags &= ~SFLAG_GLCKEY;
4565
4566             /* The width is in 'length' not in bytes */
4567             width = surface->currentDesc.Width;
4568             pitch = IWineD3DSurface_GetPitch((IWineD3DSurface *)surface);
4569
4570             /* Don't use PBOs for converted surfaces. During PBO conversion we look at SFLAG_CONVERTED
4571              * but it isn't set (yet) in all cases it is getting called. */
4572             if ((convert != NO_CONVERSION || format.convert) && (surface->flags & SFLAG_PBO))
4573             {
4574                 TRACE("Removing the pbo attached to surface %p.\n", surface);
4575                 surface_remove_pbo(surface, gl_info);
4576             }
4577
4578             if (format.convert)
4579             {
4580                 /* This code is entered for texture formats which need a fixup. */
4581                 int height = surface->currentDesc.Height;
4582
4583                 /* Stick to the alignment for the converted surface too, makes it easier to load the surface */
4584                 outpitch = width * format.conv_byte_count;
4585                 outpitch = (outpitch + device->surface_alignment - 1) & ~(device->surface_alignment - 1);
4586
4587                 mem = HeapAlloc(GetProcessHeap(), 0, outpitch * height);
4588                 if(!mem) {
4589                     ERR("Out of memory %d, %d!\n", outpitch, height);
4590                     if (context) context_release(context);
4591                     return WINED3DERR_OUTOFVIDEOMEMORY;
4592                 }
4593                 format.convert(surface->resource.allocatedMemory, mem, pitch, width, height);
4594             }
4595             else if (convert != NO_CONVERSION && surface->resource.allocatedMemory)
4596             {
4597                 /* This code is only entered for color keying fixups */
4598                 int height = surface->currentDesc.Height;
4599
4600                 /* Stick to the alignment for the converted surface too, makes it easier to load the surface */
4601                 outpitch = width * format.conv_byte_count;
4602                 outpitch = (outpitch + device->surface_alignment - 1) & ~(device->surface_alignment - 1);
4603
4604                 mem = HeapAlloc(GetProcessHeap(), 0, outpitch * height);
4605                 if(!mem) {
4606                     ERR("Out of memory %d, %d!\n", outpitch, height);
4607                     if (context) context_release(context);
4608                     return WINED3DERR_OUTOFVIDEOMEMORY;
4609                 }
4610                 d3dfmt_convert_surface(surface->resource.allocatedMemory, mem, pitch,
4611                         width, height, outpitch, convert, surface);
4612             }
4613             else
4614             {
4615                 mem = surface->resource.allocatedMemory;
4616             }
4617
4618             /* Make sure the correct pitch is used */
4619             ENTER_GL();
4620             glPixelStorei(GL_UNPACK_ROW_LENGTH, width);
4621             LEAVE_GL();
4622
4623             if (mem || (surface->flags & SFLAG_PBO))
4624                 surface_upload_data(surface, gl_info, &format, srgb, mem);
4625
4626             /* Restore the default pitch */
4627             ENTER_GL();
4628             glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
4629             LEAVE_GL();
4630
4631             if (context) context_release(context);
4632
4633             /* Don't delete PBO memory */
4634             if ((mem != surface->resource.allocatedMemory) && !(surface->flags & SFLAG_PBO))
4635                 HeapFree(GetProcessHeap(), 0, mem);
4636         }
4637     }
4638
4639     if (!rect) surface->flags |= flag;
4640
4641     if (in_fbo && (surface->flags & (SFLAG_INTEXTURE | SFLAG_INDRAWABLE)))
4642     {
4643         /* With ORM_FBO, SFLAG_INTEXTURE and SFLAG_INDRAWABLE are the same for offscreen targets. */
4644         surface->flags |= (SFLAG_INTEXTURE | SFLAG_INDRAWABLE);
4645     }
4646
4647     return WINED3D_OK;
4648 }
4649
4650 static WINED3DSURFTYPE WINAPI IWineD3DSurfaceImpl_GetImplType(IWineD3DSurface *iface) {
4651     return SURFACE_OPENGL;
4652 }
4653
4654 static HRESULT WINAPI IWineD3DSurfaceImpl_DrawOverlay(IWineD3DSurface *iface) {
4655     IWineD3DSurfaceImpl *This = (IWineD3DSurfaceImpl *) iface;
4656     HRESULT hr;
4657
4658     /* If there's no destination surface there is nothing to do */
4659     if(!This->overlay_dest) return WINED3D_OK;
4660
4661     /* Blt calls ModifyLocation on the dest surface, which in turn calls DrawOverlay to
4662      * update the overlay. Prevent an endless recursion. */
4663     if (This->overlay_dest->flags & SFLAG_INOVERLAYDRAW)
4664         return WINED3D_OK;
4665
4666     This->overlay_dest->flags |= SFLAG_INOVERLAYDRAW;
4667     hr = IWineD3DSurfaceImpl_Blt((IWineD3DSurface *)This->overlay_dest,
4668             &This->overlay_destrect, iface, &This->overlay_srcrect,
4669             WINEDDBLT_WAIT, NULL, WINED3DTEXF_LINEAR);
4670     This->overlay_dest->flags &= ~SFLAG_INOVERLAYDRAW;
4671
4672     return hr;
4673 }
4674
4675 BOOL surface_is_offscreen(IWineD3DSurfaceImpl *surface)
4676 {
4677     IWineD3DSwapChainImpl *swapchain = surface->container.u.swapchain;
4678
4679     /* Not on a swapchain - must be offscreen */
4680     if (surface->container.type != WINED3D_CONTAINER_SWAPCHAIN) return TRUE;
4681
4682     /* The front buffer is always onscreen */
4683     if (surface == swapchain->front_buffer) return FALSE;
4684
4685     /* If the swapchain is rendered to an FBO, the backbuffer is
4686      * offscreen, otherwise onscreen */
4687     return swapchain->render_to_fbo;
4688 }
4689
4690 const IWineD3DSurfaceVtbl IWineD3DSurface_Vtbl =
4691 {
4692     /* IUnknown */
4693     IWineD3DBaseSurfaceImpl_QueryInterface,
4694     IWineD3DBaseSurfaceImpl_AddRef,
4695     IWineD3DSurfaceImpl_Release,
4696     /* IWineD3DResource */
4697     IWineD3DBaseSurfaceImpl_GetParent,
4698     IWineD3DBaseSurfaceImpl_SetPrivateData,
4699     IWineD3DBaseSurfaceImpl_GetPrivateData,
4700     IWineD3DBaseSurfaceImpl_FreePrivateData,
4701     IWineD3DBaseSurfaceImpl_SetPriority,
4702     IWineD3DBaseSurfaceImpl_GetPriority,
4703     IWineD3DSurfaceImpl_PreLoad,
4704     IWineD3DSurfaceImpl_UnLoad,
4705     IWineD3DBaseSurfaceImpl_GetType,
4706     /* IWineD3DSurface */
4707     IWineD3DBaseSurfaceImpl_GetDesc,
4708     IWineD3DSurfaceImpl_Map,
4709     IWineD3DSurfaceImpl_Unmap,
4710     IWineD3DSurfaceImpl_GetDC,
4711     IWineD3DSurfaceImpl_ReleaseDC,
4712     IWineD3DSurfaceImpl_Flip,
4713     IWineD3DSurfaceImpl_Blt,
4714     IWineD3DBaseSurfaceImpl_GetBltStatus,
4715     IWineD3DBaseSurfaceImpl_GetFlipStatus,
4716     IWineD3DBaseSurfaceImpl_IsLost,
4717     IWineD3DBaseSurfaceImpl_Restore,
4718     IWineD3DSurfaceImpl_BltFast,
4719     IWineD3DBaseSurfaceImpl_GetPalette,
4720     IWineD3DBaseSurfaceImpl_SetPalette,
4721     IWineD3DSurfaceImpl_RealizePalette,
4722     IWineD3DBaseSurfaceImpl_SetColorKey,
4723     IWineD3DBaseSurfaceImpl_GetPitch,
4724     IWineD3DSurfaceImpl_SetMem,
4725     IWineD3DBaseSurfaceImpl_SetOverlayPosition,
4726     IWineD3DBaseSurfaceImpl_GetOverlayPosition,
4727     IWineD3DBaseSurfaceImpl_UpdateOverlayZOrder,
4728     IWineD3DBaseSurfaceImpl_UpdateOverlay,
4729     IWineD3DBaseSurfaceImpl_SetClipper,
4730     IWineD3DBaseSurfaceImpl_GetClipper,
4731     /* Internal use: */
4732     IWineD3DSurfaceImpl_LoadTexture,
4733     IWineD3DSurfaceImpl_BindTexture,
4734     IWineD3DBaseSurfaceImpl_GetData,
4735     IWineD3DSurfaceImpl_SetFormat,
4736     IWineD3DSurfaceImpl_PrivateSetup,
4737     IWineD3DSurfaceImpl_GetImplType,
4738     IWineD3DSurfaceImpl_DrawOverlay
4739 };
4740
4741 static HRESULT ffp_blit_alloc(IWineD3DDevice *iface) { return WINED3D_OK; }
4742 /* Context activation is done by the caller. */
4743 static void ffp_blit_free(IWineD3DDevice *iface) { }
4744
4745 /* This function is used in case of 8bit paletted textures using GL_EXT_paletted_texture */
4746 /* Context activation is done by the caller. */
4747 static void ffp_blit_p8_upload_palette(IWineD3DSurfaceImpl *surface, const struct wined3d_gl_info *gl_info)
4748 {
4749     BYTE table[256][4];
4750     BOOL colorkey_active = (surface->CKeyFlags & WINEDDSD_CKSRCBLT) ? TRUE : FALSE;
4751
4752     d3dfmt_p8_init_palette(surface, table, colorkey_active);
4753
4754     TRACE("Using GL_EXT_PALETTED_TEXTURE for 8-bit paletted texture support\n");
4755     ENTER_GL();
4756     GL_EXTCALL(glColorTableEXT(surface->texture_target, GL_RGBA, 256, GL_RGBA, GL_UNSIGNED_BYTE, table));
4757     LEAVE_GL();
4758 }
4759
4760 /* Context activation is done by the caller. */
4761 static HRESULT ffp_blit_set(IWineD3DDevice *iface, IWineD3DSurfaceImpl *surface)
4762 {
4763     IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *) iface;
4764     const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
4765     enum complex_fixup fixup = get_complex_fixup(surface->resource.format->color_fixup);
4766
4767     /* When EXT_PALETTED_TEXTURE is around, palette conversion is done by the GPU
4768      * else the surface is converted in software at upload time in LoadLocation.
4769      */
4770     if(fixup == COMPLEX_FIXUP_P8 && gl_info->supported[EXT_PALETTED_TEXTURE])
4771         ffp_blit_p8_upload_palette(surface, gl_info);
4772
4773     ENTER_GL();
4774     glEnable(surface->texture_target);
4775     checkGLcall("glEnable(surface->texture_target)");
4776     LEAVE_GL();
4777     return WINED3D_OK;
4778 }
4779
4780 /* Context activation is done by the caller. */
4781 static void ffp_blit_unset(IWineD3DDevice *iface)
4782 {
4783     IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *) iface;
4784     const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
4785
4786     ENTER_GL();
4787     glDisable(GL_TEXTURE_2D);
4788     checkGLcall("glDisable(GL_TEXTURE_2D)");
4789     if (gl_info->supported[ARB_TEXTURE_CUBE_MAP])
4790     {
4791         glDisable(GL_TEXTURE_CUBE_MAP_ARB);
4792         checkGLcall("glDisable(GL_TEXTURE_CUBE_MAP_ARB)");
4793     }
4794     if (gl_info->supported[ARB_TEXTURE_RECTANGLE])
4795     {
4796         glDisable(GL_TEXTURE_RECTANGLE_ARB);
4797         checkGLcall("glDisable(GL_TEXTURE_RECTANGLE_ARB)");
4798     }
4799     LEAVE_GL();
4800 }
4801
4802 static BOOL ffp_blit_supported(const struct wined3d_gl_info *gl_info, enum blit_operation blit_op,
4803         const RECT *src_rect, DWORD src_usage, WINED3DPOOL src_pool, const struct wined3d_format *src_format,
4804         const RECT *dst_rect, DWORD dst_usage, WINED3DPOOL dst_pool, const struct wined3d_format *dst_format)
4805 {
4806     enum complex_fixup src_fixup;
4807
4808     if (blit_op == BLIT_OP_COLOR_FILL)
4809     {
4810         if (!(dst_usage & WINED3DUSAGE_RENDERTARGET))
4811         {
4812             TRACE("Color fill not supported\n");
4813             return FALSE;
4814         }
4815
4816         return TRUE;
4817     }
4818
4819     src_fixup = get_complex_fixup(src_format->color_fixup);
4820     if (TRACE_ON(d3d_surface) && TRACE_ON(d3d))
4821     {
4822         TRACE("Checking support for fixup:\n");
4823         dump_color_fixup_desc(src_format->color_fixup);
4824     }
4825
4826     if (blit_op != BLIT_OP_BLIT)
4827     {
4828         TRACE("Unsupported blit_op=%d\n", blit_op);
4829         return FALSE;
4830      }
4831
4832     if (!is_identity_fixup(dst_format->color_fixup))
4833     {
4834         TRACE("Destination fixups are not supported\n");
4835         return FALSE;
4836     }
4837
4838     if (src_fixup == COMPLEX_FIXUP_P8 && gl_info->supported[EXT_PALETTED_TEXTURE])
4839     {
4840         TRACE("P8 fixup supported\n");
4841         return TRUE;
4842     }
4843
4844     /* We only support identity conversions. */
4845     if (is_identity_fixup(src_format->color_fixup))
4846     {
4847         TRACE("[OK]\n");
4848         return TRUE;
4849     }
4850
4851     TRACE("[FAILED]\n");
4852     return FALSE;
4853 }
4854
4855 /* Do not call while under the GL lock. */
4856 static HRESULT ffp_blit_color_fill(IWineD3DDeviceImpl *device, IWineD3DSurfaceImpl *dst_surface,
4857         const RECT *dst_rect, const WINED3DCOLORVALUE *color)
4858 {
4859     const RECT draw_rect = {0, 0, dst_surface->currentDesc.Width, dst_surface->currentDesc.Height};
4860
4861     return device_clear_render_targets(device, 1 /* rt_count */, &dst_surface, 1 /* rect_count */,
4862             dst_rect, &draw_rect, WINED3DCLEAR_TARGET, color, 0.0f /* depth */, 0 /* stencil */);
4863 }
4864
4865 const struct blit_shader ffp_blit =  {
4866     ffp_blit_alloc,
4867     ffp_blit_free,
4868     ffp_blit_set,
4869     ffp_blit_unset,
4870     ffp_blit_supported,
4871     ffp_blit_color_fill
4872 };
4873
4874 static HRESULT cpu_blit_alloc(IWineD3DDevice *iface)
4875 {
4876     return WINED3D_OK;
4877 }
4878
4879 /* Context activation is done by the caller. */
4880 static void cpu_blit_free(IWineD3DDevice *iface)
4881 {
4882 }
4883
4884 /* Context activation is done by the caller. */
4885 static HRESULT cpu_blit_set(IWineD3DDevice *iface, IWineD3DSurfaceImpl *surface)
4886 {
4887     return WINED3D_OK;
4888 }
4889
4890 /* Context activation is done by the caller. */
4891 static void cpu_blit_unset(IWineD3DDevice *iface)
4892 {
4893 }
4894
4895 static BOOL cpu_blit_supported(const struct wined3d_gl_info *gl_info, enum blit_operation blit_op,
4896         const RECT *src_rect, DWORD src_usage, WINED3DPOOL src_pool, const struct wined3d_format *src_format,
4897         const RECT *dst_rect, DWORD dst_usage, WINED3DPOOL dst_pool, const struct wined3d_format *dst_format)
4898 {
4899     if (blit_op == BLIT_OP_COLOR_FILL)
4900     {
4901         return TRUE;
4902     }
4903
4904     return FALSE;
4905 }
4906
4907 /* Do not call while under the GL lock. */
4908 static HRESULT cpu_blit_color_fill(IWineD3DDeviceImpl *device, IWineD3DSurfaceImpl *dst_surface,
4909         const RECT *dst_rect, const WINED3DCOLORVALUE *color)
4910 {
4911     WINEDDBLTFX BltFx;
4912
4913     memset(&BltFx, 0, sizeof(BltFx));
4914     BltFx.dwSize = sizeof(BltFx);
4915     BltFx.u5.dwFillColor = wined3d_format_convert_from_float(dst_surface->resource.format, color);
4916     return IWineD3DBaseSurfaceImpl_Blt((IWineD3DSurface*)dst_surface, dst_rect,
4917             NULL, NULL, WINEDDBLT_COLORFILL, &BltFx, WINED3DTEXF_POINT);
4918 }
4919
4920 const struct blit_shader cpu_blit =  {
4921     cpu_blit_alloc,
4922     cpu_blit_free,
4923     cpu_blit_set,
4924     cpu_blit_unset,
4925     cpu_blit_supported,
4926     cpu_blit_color_fill
4927 };
4928
4929 static BOOL fbo_blit_supported(const struct wined3d_gl_info *gl_info, enum blit_operation blit_op,
4930         const RECT *src_rect, DWORD src_usage, WINED3DPOOL src_pool, const struct wined3d_format *src_format,
4931         const RECT *dst_rect, DWORD dst_usage, WINED3DPOOL dst_pool, const struct wined3d_format *dst_format)
4932 {
4933     if ((wined3d_settings.offscreen_rendering_mode != ORM_FBO) || !gl_info->fbo_ops.glBlitFramebuffer)
4934         return FALSE;
4935
4936     /* We only support blitting. Things like color keying / color fill should
4937      * be handled by other blitters.
4938      */
4939     if (blit_op != BLIT_OP_BLIT)
4940         return FALSE;
4941
4942     /* Source and/or destination need to be on the GL side */
4943     if (src_pool == WINED3DPOOL_SYSTEMMEM || dst_pool == WINED3DPOOL_SYSTEMMEM)
4944         return FALSE;
4945
4946     if (!((src_format->flags & WINED3DFMT_FLAG_FBO_ATTACHABLE) || (src_usage & WINED3DUSAGE_RENDERTARGET))
4947             && ((dst_format->flags & WINED3DFMT_FLAG_FBO_ATTACHABLE) || (dst_usage & WINED3DUSAGE_RENDERTARGET)))
4948         return FALSE;
4949
4950     if (!(src_format->id == dst_format->id
4951             || (is_identity_fixup(src_format->color_fixup)
4952             && is_identity_fixup(dst_format->color_fixup))))
4953         return FALSE;
4954
4955     return TRUE;
4956 }