| Age | Commit message (Collapse) | Author |
|---|
|
Fencing was used in two places: ensuring that we didn't get too many frames
ahead of ourselves, and glFinish. glFinish will be satisfied by waiting on
buffers like we would do for CPU access on them. The "don't get too far ahead"
is now the responsibility of the execution manager (kernel).
|
|
|
|
|
|
|
|
Makes state emission into a 2 phase, prepare sets things up and accounts
the size of all referenced buffer objects. The emit stage then actually
does the batchbuffer touching for emitting the objects.
There is an assert in dri_emit_reloc if a reloc occurs for a buffer
that hasn't been accounted yet.
|
|
|
|
batchbuffer > aperture size.
So with compiz on Intel hw with fake bufmgr, opening 4 firefox windows at 1680x1050 and hitting alt-tab, could cause the batchbuffer to try and reference more than the 32MB of RAM allocated.
Fix 1:
Fix 1 is to pre-verify the list of buffers against the current batchbuffer and if it can't possibly fit in the aperture to flush the batchbuffer to the hardware
and try again. If the buffers still can't fit well then you are hosed as I'm not sure there is a nice way to tell anyone.
Fix 2:
Next problem was that even with a simple check for total < aperture, we ran
into fragmentation issues, this meant that half way down a set of buffers,
we would fail as no blocks were available. Fix this by nuking the memory
manager from orbit and letting it start again and relayout the blocks in a
manner that fits.
Fix 3:
Finally the initial problem we were seeing was a memcpy to a NULL backing store.
We seem to end up with a texture at some point that never gets mapped but ends up with data in it. compiz al-tab icons have this property. So I created a card dirty bit that memcpy's any buffer that is !static and is written to back to memory. This probably is wrong but it makes compiz work for now.
Caveats:
965 support is still fail.
|
|
|
|
Add entrypoints to glapi XML file and regenerate files.
Implement glStencilOpSeparateATI().
Consolidate some code in stencil.c
|
|
|
|
Add DRI2 direct rendering support to libGL and add DRI2 client side
protocol code. Extend the GLX 1.3 create drawable functions in
glx_pbuffer.c to call into the DRI driver when possible.
Introduce __DRIconfig, opaque struct that represents a DRI driver
configuration. Get's rid of the open coded __GLcontextModes in the
DRI driver interface and the context modes create and destroy
functions that the loader was requires to provide. glcore.h is no
longer part of the DRI driver interface. The DRI config is GL binding
agnostic, that is, not specific to GLX, EGL or other bindings.
The core API is now also an extension, and the driver exports a list
of extensions as the symbol __driDriverExtensions, which the loader
must dlsym() for. The list of extension will always include the DRI
core extension, which allows creating and manipulating DRI screens,
drawables and contexts. The DRI legacy extension, when available,
provides alternative entry points for creating the DRI objects that
work with the XF86DRI infrastructure.
Change DRI2 client code to not use drm drawables or contexts. We
never used drm_drawable_t's and the only use for drm_context_t was as
a unique identifier when taking the lock. We now just allocate a
unique lock ID out of the DRILock sarea block. Once we get rid of the
lock entirely, we can drop this hack.
Change the interface between dri_util.c and the drivers, so that the
drivers now export the DriverAPI struct as driDriverAPI instead of the
InitScreen entry point. This lets us avoid dlsym()'ing for the DRI2
init screen function to see if DRI2 is supported by the driver.
|
|
|
|
__dri2ParseEvents() would determine the kind of event, but then call
UpdateBuffer() in either case, and UpdateBuffer() would then have to
figure that out again to dispatch to HandleBufferAttach() or
HandleDrawableConfig(). Pretty pointless.
|
|
Makes a lot more sense, since the screen is always implicit in the
DRI drawable, but it may not be possible to track down a context from
just a drawable.
|
|
The DRI driver needs to know where in the buffer to start reading.
|
|
When the DRI doesn't parse the event buffer for a while, the X server
may overwrite data that the driver didn't get a chance to look at. The
reemitDrawableInfo callback requests that the X server reemit all info
for the specified drawable. To make use of this, the drive needs to know
the new tail pointer so it know where to start reading from.
|
|
This also fixes the problem where the X server does multiple resizes before
the DRI driver gets the events. The obsolete buffer attach events then
reference already destroyed buffer objects.
|
|
Instead of passing in a fixed struct, the loader now passes in a list
of __DRIextension structs, to advertise the functionality it can provide
to the driver. Each extension is individually versioned and can be
extended or phased out as the interface develops.
|
|
Right now the DRI2 screen constructor takes 3 different versions:
DRI, DDX and DRM. This is mostly useless, though:
DRI: The DRI driver doesn't actually care about the DRI protocol,
it only talks to the loader, which in turn speaks DRI protocol. Thus,
the DRI protocol version is of not interest to the DRI driver, but it
needs to know what functionality the loader provides. At this point
that's reflected in the __DRIinterfaceMethods struct and the
internal_version integer.
DDX: The DDX version number is essentially used to track extensions
to the SAREA. With DRI2 the SAREA consists of a number of versioned,
self-describing blocks, so the DDX version is no longer interesting.
DRM: We have the fd, lets just ask the kernel ourselves.
|
|
|
|
Argh, seriously, I did test that, I just forgot to amend the commit
before pushing.
|
|
|
|
Sort-of a stop-gap workaround. There are a couple of nicer ways to
do this that doesn't require dlfcn.h, but they're more invasive.
|
|
The DriverAPI is internal to the DRI drivers and GetDrawableMSC
obsoletes GetMSC. Also, since the DRI driver interface has not yet
been released, just drop the getMSC function from the DRI interface
instead using the ABI preserving version mechanism.
Finally, using void pointer privates in the DRI interface is not allowed,
always pass the actual types around (__DRIdrawable in this case) to
enhance type safety and readability of the code.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
We have two consumers of relocations. One is static state buffers, which
want the same relocation every time. The other is the batchbuffer, which gets
thrown out immediately after submit. This lets us reduce repeated computation
for static state buffers, and clean up the code by moving relocations nearer
to where the state buffer is computed.
|
|
|
|
|
|
Now that the dri_bufmgr is stored in the context rather than the screen, all
access to one is single-threaded anyway.
|
|
|
|
bufmgr_fake doesn't care about it, but with ttm we would end up with the
buffer remaining referenced until application exit.
|
|
|
|
This is currently believed to work but be a significant performance loss.
Performance recovery should be soon to follow.
The dri_bo_fake_disable_backing_store() call was added to allow backing store
disable like bufmgr_fake.c did, which is a significant performance win (though
it's missing the no-fence-subdata part).
This commit is a squash merge of the 965-ttm branch, which had some history
I wanted to avoid pulling due to noisiness and brokenness at many points
for git-bisecting.
|
|
by another window.
|
|
|
|
The uint64_t flags (as defined by drm.h) were being used as unsigned ints in
many places.
|
|
|
|
|
|
This is required for 965 performance, as it avoids a lot of repeated data
uploads of the state caches due to surface offsets in them.
|
|
They shouldn't be created, and this often helps catch stupid issues.
|
|
This is required for 965 support, which has relocations in other places than
just the batchbuffer.
|
|
|
|
This should restore gears speed on 9xx hardware
|
