RFC: wined3d: Implementing depth/stencil buffer locking

[Octavian Voicu]

Hi,

Here is a first shot at implementing depth/stencil buffer locking in wined3d.

Currently depth/stencil buffer locations are onscreen and offscreen,
with transitions between the two states done in load_ds_location. I
added another location, sysmem, and implemented transitions from
offscreen to sysmem and from sysmem to offscreen. If buffer is
onscreen, it will be moved offscreen first, then to sysmem.

Some issues with this implementation:

1) make dsurface.ok in ddraw/tests crashes in context_acquire
(wined3d/context.c:2456) called from surface_load_location
(wined3d/surface.c:6323), I'm guessing  in because there is no
swapchain. How can I prevent this crash and still get a context?
Loading from offscreen to sysmem shouldn't require a swapchain.

2) When running my test app [1] I also get this error:

err:d3d_surface:surface_load_ds_location >>>>>>>>>>>>>>>>>
GL_INVALID_VALUE (0x501) from glTexSubImage2D @
../../../../wine/dlls/wined3d/surface.c / 5877

3) There are other issues with my test app (looks like it's related to
textures), so can't be sure if locking works correctly. There is an
improvement over the unpatched version, but can't be sure until the
other issue is resolved. Game uses pre-rendered backgrounds and
probably loads the Z-buffer by locking the surface, so it can render
3D characters correctly over the pre-rendered backgrounds.

It happens in the "if (surface->flags & SFLAG_DS_SYSMEM)" branch of
the surface_load_ds_location function. Uploading happens on each
frame, this error only happens for the first upload.

I'm quite new to wined3d code so I might have missed some obvious
things. Any comments and suggestions are welcome.

Thanks.

Octavian


[1] The demo of The Longest Journey, available for download here (to
reproduce, click on the Begin Journey, press ESC to skip intro movies;
notice how rendered characters look):
http://games.softpedia.com/get/Games-Demo/The-Longest-Journey-161-Demo.shtml

[Henri Verbeet]

On 2 November 2011 11:59, Octavian Voicu <octavian.voicu(a)gmail.com> wrote:
> 1) make dsurface.ok in ddraw/tests crashes in context_acquire
> (wined3d/context.c:2456) called from surface_load_location
> (wined3d/surface.c:6323), I'm guessing  in because there is no
> swapchain. How can I prevent this crash and still get a context?
> Loading from offscreen to sysmem shouldn't require a swapchain.

In practice it does. The reason is that to get a GL context you need a
window, or at least a Drawable, and currently in wined3d that means
you need a swapchain. There are a couple of things that need to be
fixed there. One is that wined3d should really be able to do rendering
without a swapchain for d3d10/11, the other is that ddraw should
create its swapchain much earlier, likely on SetCooperativeLevel().
I'm not sure all that matter a whole lot here though, the more
interesting question is probably how you managed to get the surface to
be in the offscreen location without a swapchain.

The sysmem <-> texture transfer code looks at first sight like it
should be using surface_upload_data() / surface_download_data().
Another issue is that SFLAG_DS_ONSCREEN and SFLAG_DS_OFFSCREEN should
actually go away and be replaced with SFLAG_INDRAWABLE /
SFLAG_INTEXTURE. These are separate flags for historical reasons, but
it should be possible to merge them back now.

[Octavian Voicu]

On Wed, Nov 2, 2011 at 3:41 PM, Henri Verbeet <hverbeet(a)gmail.com> wrote:
> I'm not sure all that matter a whole lot here though, the more
> interesting question is probably how you managed to get the surface to
> be in the offscreen location without a swapchain.

According to git blame, Stefan managed to do that :)

Crash happens in test_ddsd() in ddraw/tests/dsurface.c:4330, after
unlocking a newly created z-buffer. The surface is in video memory
when created, even if it contains junk, right?

It is strange though that there is no swapchain at that point, it is
created early on in the test.

Octavian

[Stefan Dösinger]

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1


Am 03.11.2011 um 11:35 schrieb Octavian Voicu:
> Crash happens in test_ddsd() in ddraw/tests/dsurface.c:4330, after
> unlocking a newly created z-buffer. The surface is in video memory
> when created, even if it contains junk, right?
> Surfaces are created with SFLAG_INSYSMEM(see dlls wined3d/surface.c, surface_private_setup), but due to the SFLAG_DS_ONSCREEN and SFLAG_DS_OFFSCREEN flags as well as your newly created SFLAG_DS_SYSMEM this doesn't apply for depth stencils. This problem should go away once depth stencils use the normal location flags.

-----BEGIN PGP SIGNATURE-----
Version: GnuPG/MacGPG2 v2.0.17 (Darwin)

iQIcBAEBAgAGBQJOsnEwAAoJEN0/YqbEcdMwzwoP/1zSKr2rGnug9rMWH79rGbYL
a3dSaNB4uVHtITM9QhJNU21aAJf5vtPIhNcY1DOusXduIGwIsvKqEvhl5xXdkLYf
96rJWRTLqS2XIqSMhah7jy6cmRYZoL16ZKQ5sW4ukY514EQuKfcZDB4pYfYqKPa7
Qt9UVNdcDwSFAX695i4KnVTlBT4va6LwndADInDKR0AaCX1PKRIAs1XK+qtJm1yH
R9DyRxsoYYWBRy/7RaQWEYvc3F/4DuxBGZhnKzWVrUaw7BPx7UPuI5M8k30dPhMZ
Wym6Rll30TuhNc5HbamPbAMu/YqusbGN+hdIRXotCtwAGFujcIhIXksesgibA0eO
MnQtOsfjF/xQtfcmuEm7NHhviGkO2Q5eKDPAPFxIZa49fthiW2rVSVEUJi3E3uXG
K4qX59831RooZfdJJPgpJ0PJHeyw9OSzENdVPNFqqArcv1yP4srzuIzsLsVbyM4B
DtbIrccBfr5DAeysuGxThLgyvN4R5IS7kBXezH3mbeBUME0wjkfRCfBYJf30hTuy
kER9FRHdGhNpXGdQBLNb6xAgIsBkdCIaqSmsxme8bJSsg7KYa0r78WPbCKegpT/i
H7fMVWv3iBSyycRqX9ZS3zT7EaCeV9SzThIYiaygMh4gLEuABmOEZqzmRgmQkA8p
QfYTqLr/wKaN3InfHnaB
=ypSa
-----END PGP SIGNATURE-----