When allocating objects, we need to check the size of the full fenced
regions against the mappable limits in order to be able to mmap the
object later.
References: https://bugs.freedesktop.org/show_bug.cgi?id=42813
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
PutImage rarely uploads the entire image inside a single request,
instead breaking up into scanline segments to fit within the protocol
limits. A few optimisations are based on detecting when we can discard
the GPU bo for an all-dirty CPU pixmap, which are useful in this case so
check for an entirely dirty pixmap following an incremental PutImage.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
For a long lived pixmap which we are repeatedly using for upload and
copying to other pixmaps, we don't want to keep to maintain a GPU
buffer. So instead, reset the source counter if we discard the GPU bo
and treat the pixmap as purely CPU.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
This is the stub of the decoder, sufficient to give details of the ops
within the batch and to keep the debugger happy.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Otherwise gnome-shell forgets to update. Eventually, I'll get the async
pageflipping bits merged into the Xserver and this path and its extra
bw wastage will be history! But still I'll be undermined by the
compositor, grrr.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
A large part of the function preamble overhead is the multi-indirection
lookup for retrieving the sna pointer. We can eliminate most of these by
storing a pointer on the Pixmap as well, which we often need to lookup
anyway.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
This is to keep the sanity checks upon the caches happy by keeping a
potential flush out of the inactive cache.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Yes, writing 3 channels is slower than writing 4. But it's okay we
simply ignore the alpha value anyway.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Although the original precison need only 16-bits to store the offsets,
after projecting on to the sample grid we need a few more bits of
precision and so need a larger integer type to avoid overflow and render
glitches.
Reported-by: Clemens Eisserer <linuxhippy@gmail.com>
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=42680
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
The principle behind the opertator reduction of WHITE * maskca is valid,
except that we failed to account for the src/mask transposition when
emitting the vertices - garbage ensued.
Given that we agressively reduce the shader required for WHITE * maskca,
it does not seem worthwhile to special case the primitive emitter as
well.
Reported-by: Clemens Eisserer <linuxhippy@gmail.com>
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=42676
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Chris Wilson