This is slower than falling back to swrast for x11perf (up to 4x slower
on SNB), it is still faster than doing that rasterisation through a
WC-mapping and much faster in ordinary usage due to avoiding the
readback hit.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
By constructing the batch buffer directly for PolyRectangle, rather than
via miPolyRectangle->(PolyFillRectangle/PolyLine), we dramatically
reduce the CPU overhead and can saturate the GPU.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
We are hitting this path upon retiring the old scanout buffer following
a page-flip. We want to treat this as being hot and available for reuse,
so mark it as such until the next retirement pass.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
For many of the core drawing routines, passing a BoxRec for the fill is
more convenient since they already have one generated by the clip
intersection.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
As damage accumulation is handled modally, we do not need to track the
mode per elt and so attempt to simplify the code.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
I translated the region to copy by the composite pixmap offset, only
failed to use the translated region for the actual copy command (using
instead the original boxes). Fix that mistake by avoiding the temporary
region entirely and applying the translation inplace.
We also have to be careful in the case of copying between two composited
windows that have different offsets into the same screen pixmap.
This fixes the regression introduced with a3466c8b69 (sna/accel:
Implement a simpler path for CopyArea between the same pixmaps).
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
As the bo is outside of our control whilst it is under the influence of
an external renderer, we try to maintain it on the gpu so as to avoid
unnecessary ping-pong. But once it is wholly back under our control, we
want to stop paying the penalty for sharing it.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Take advantage of any available temporary buffer that we reuse for
readback knowing that it is the last operation in the batch.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
We track the last ring used when active so as to avoid stalling between
batches. Once the GPU has retired all the pending requests, we can use
whichever ring is preferrable for the next operation without any danger
of stalling upon submission.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
This is to workaround a ping-pong issue involving small icons. The
horrible sequence of operations appears to use a tiled FillRect to copy
from the scanout onto to a temporary pixmap, which causes us to
readback from the scanout. We are destined to hit the fallback path there
anyway until we implement stippling...
References: https://bugs.freedesktop.org/show_bug.cgi?id=41718
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
We can only emit state between primitives, ergo we need only check for
state updates if we've finished the vbo or are starting a new operation.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Chris Wilson