Mmap from /sys/devices/pci* on linux forces the cache-disable and
write-through bits, which turns our write-combining map into an
uncached-map, seriously impacting performance. It turns out that a bug in
mprotect allows us to fix this by disabling access to those pages and then
immediately re-enabling them.
The gen4_render_state is now always called "render_state" (i965_render.c
bookkeeping) and gen4_state_t is now always called "card_state" (the buffer
for state used by the chip).
This patch duplicates all of the surface sampler state needed to read from
the separate YUV surfaces and then has the planar sampler read from samplers
0, 2, 4 instead of 0, 1, 2. This appears to make things work, instead of
having the samplers break at random. I do not understand why this works.
We have a collection of wm_state objects for each ps kernel,
(one for each combination of src and mask extend and repeat
values).
Thanks to Dave Airlie for noticing an errant write through a
wild wm_state pointer in an early version of this commit.
(cherry picked from 7763706a93d3021907273f9b330750ba110e2fc3 commit)
This cherry-pick required more reformatting than most, due to the
projective texturing merge.
This will eventually allow for the elimination of sampler state
updates while compositing---and initializing everything in the
initialization function.
(cherry picked from commit d0874697be8086cd64740c24698df8cd4d31c76f)
This reverts commit f47486fab3.
Sorry for failing to push the #line processing changes to intel-gen4asm;
those are now pushed, so this change should no longer be necessary. Having
correct file name and line numbers in the error messages is awfully nice
though.
We need one for each possible combination of src and dst
blend_factors. Again, as with recent changes, this eliminates
state updates from prepare_composite and allows that function
to instead simply reference an existing object initialized
within gen4_state_init.
Thanks to Dave Airlie (and git-bisect) for pointing out that with
gnome-terminal all text was appearing as solid black with an early
version of this commit. As expected the bug was an alignment issue.
(cherry picked from 0c0ab52c2d100c47f38c7ef826ef585c8b9815e9 commit)
Performance is approximately equivalent on text tests, but may be
around +2%.
The YUV->RGB code was written to write directly to the dataport registers,
but that didn't work for the compositing functions (cause still unknown).
This change makes that code write RGB values to the src_sample registers as
with the other sample computation fragments.
This reverts commit 346cf57deabb4c336612df4c13650a87b5ef6775.
Mixing randr transforms and video caused screen corruption for Render
operations. No, I don't understand why.
The hardware has been marked as needing a sync, so the next video put will
block waiting for the previous one to complete. Adding a sync here just
stalls the video playback for no good reason.
Instead of leaving pixel values in src_sample registers, compute the pixel
values directl to the data port to save 8 moves. This cannot work when no
computation is done as there is both no way to wait for the sampler to
finish and because the sampler returns data in a different order from that
required by the data port (sigh).
This reduces the CPU overhead of memcpying them in every time, for a speedup
in aa24text of around 30%. This is based on work by Carl Worth which is
in the intel-batchbuffer branch.
The Intel driver appears to be coded to only work with displays
expecting 18 bit pixels. However I have an application using a LCD
display that expects pixel data in 24 bit format. The difference is
only 2 bits in a single GPU register. This patch implements that
change, controlled by a new driver option, "LVDS24Bit". The default
value is false, which is the previous behavior. When set to true,
then 24 bit panels should work (at least the one I'm testing here
does).
Fd.o bug #15201
Signed-off-by: Mike Isely <isely@pobox.com>
Using the updated factors even when BT709 conversion isn't available
(non-965) should still give us better color reproduction. Tested on a
945GM, examining the +/-5% of black bars of videotestsrc.
Keith Packard
Hong Liu
Eric Anholt
Carl Worth
Zhenyu Wang
Mike Isely
Jesse Barnes
David Schleef
Bryce Harrington