Am 10.12.2009 um 23:15 schrieb Henri Verbeet:
2009/12/10 Stefan Dösinger stefan@codeweavers.com:
@@ -6304,7 +6304,7 @@ HRESULT create_primary_opengl_context(IWineD3DDevice *iface, IWineD3DSwapChain *
...
- swapchain->context[0]->render_offscreen = swapchain->render_to_fbo;
- swapchain->context[0]->render_offscreen = surface_is_offscreen(target);
This will introduce a warning, surface_is_offscreen() takes an IWineD3DSurface pointer.
Woops, I guess I should not use my own gcc all the time which chokes on the OSX headers and writes warning spam :-/
- Setting "render_offscreen" is really the responsibility of the
context_create() call, similar to create_primary_opengl_context().
context_create() doesn't take a swapchain as parameter, only a window. This is correct, because GL contexts are linked to the drawable, not the swapchain. Thus context_create doesn't have the needed information to set the offscreen flag. (I could just pass in a TRUE or FALSE parameter, but there's something else)
The reason why this line is needed however is because the context has a last render target set, and if the app never changes the render target the sorta-cached render-offscreen value never gets set properly. Maybe we should make sure that the full FindContext() runs at least once for a newly created context, then we can drop this line and the one in create_primary_opengl_context(). Should keeping the last_rt set to NULL after context creation work correctly?
As a sidenode, maybe create_primary_opengl_context is named badly. It's not part of the context management, its a device helper function that happens to use the context management to set up a GL context, among other things. Maybe device_restart_opengl or something similar is better.
- The line is actually wrong, the relevant context has been setup to
render to the front buffer, so "render_offscreen" should always be FALSE, even if there is a back buffer.
Well, this code is in the backbuffer path. Render_offscreen is false if we don't set it due to the alloc HEAP_ZERO_MEMORY. The point of the line is to set it to true if we render to an FBO. We never set a swapchain to render to the front buffer at creation time if there is a back buffer.
- It doesn't make sense for "render_to_fbo" to be TRUE for a
swapchain without back buffer.
Correct, but this needs another patch to fix.
First of all, render_to_fbo shouldn't matter when we're dealing with the front buffer, even when a back buffer is there. surface_is_offscreen takes care of that, but we have to use it. This is important when we're forcing render_to_fbo on for driver bug workarounds, or if we have a double buffered environment with render_to_fbo on for a real(non-driver) reason and access the front buffer due to some ddraw/d3d9ex tricks or wined3d internal stuff.
In the ddraw windowed mode rendering situation rendering the backbuffer to a FBO is actually the correct thing to do, because from the application's point of view the backbuffer is an offscreen plain surface. We just construct a double-buffered GL setup to be able to use windowed ddraw rendering without FBOs on older drivers. Currently this triggers a few other bugs(e.g. in GetDC() + Client_storage) that I want to fix first.
So the next step after fixing the GL errors and the GetDC bug is to ignore the back buffer size in CreateSwapchain and Reset if there is no backbuffer. Then render_to_fbo will be false for single buffered setups.
The academic question afterwards is if we render windowed d3d7 setups to an FBO if available or not. If we have FBOs, the difference is small. I haven't seen an app that depends on this yet, and unless we are on MacOS 10.5 or earlier which cannot write to the front buffer I don't expect driver issues either. This whole thing is done do please drivers without FBOs.
So once the other bugs are fixed I'll add a check to CreateSwapchain and reset to ignore the backbuffer size if backbuffercount == 0