Composite improvements and bug fixes

[digetx]

Here is a quite big update for the compositing using 3d: a big performance improvement (more to come) and various bug-fixes. The performance is greatly improved by using specifically-optimized shaders for the most common compositing operations and a bit smarter optimizations in the composite-preparation phase (texture reduction and coords wrapping checks), thus transparency and fading effects in plasma 5 are smooth now in comparison to a slideshow that it was before. Unpleasant bugs popped out due to the code changes, they are fixed now. It also turned out that the attributes-buffer set-up code was buggy and I was blaming HW before, that is fixed now and seems we can enable 3d compositing by default (I'll do more testing to be sure).

[digetx]

BTW, I'll probably add couple more patches to this PR soon'ish.

[digetx]

@kusma, I'm not going to add any new features to this PR. Will be nice if you could skim through the changes and tell if anything is ridiculous. In overall this work gives quite huge improvement to performance and fixes several serious bugs.

[Decatf]

The commit exa: composite-3d: support transformations breaks rendering on Galaxy Tab 10.1 (tegra 2). It happens when Option "AccelCompositing" "true". When this is false there is no rendering bug. This bug does not happen on Ouya.

Option "AccelCompositing" "false": https://i.imgur.com/5LJK2vC.png

Option "AccelCompositing" "true": https://i.imgur.com/11frFG3.png Xorg.0.log: https://paste.ubuntu.com/p/5rm4BzMpMm/

xorg conf:

Section "Device"
    Identifier "Tegra Graphics"
    Option "device" "/dev/dri/card0"
    Option "SWcursor" "false"
    Option "AccelCompositing" "true"
    Option "NoAccel" "false"
    Option "DisablePoolAllocator" "false"
    Option "DisablePixmapRefrigerator" "false"
    Option "DisableCompressionLZ4" "false"
    Option "DisableCompressionJPEG" "true"
    Option "JPEGCompressionQuality" "75"
    Option "DisableCompressionPNG" "false"
EndSection

[digetx]

@Decatf are you using the same kernel configuration on both devices?

[Decatf]

are you using the same kernel configuration on both devices?

Not exactly the same. They are similar aside from configs for peripheral drivers. GPU related configs should be the same though.

For Galaxy Tab 10.1. It is this kernel branch and defconfig https://github.com/Decatf/linux/tree/5.0-p4wifi_pmos https://gitlab.com/postmarketOS/pmaports/blob/d079a3d0/device/linux-samsung-p4wifi/config-samsung-p4wifi.armv7

For Ouya. It is this kernel branch and defconfig https://github.com/Decatf/linux/tree/5.0-next-20190305-ouya-V2 https://gitlab.com/ryang2678/pmaports/blob/66fc2a5340604f75d12f0de6b7708c2fccde3927/device/linux-ouya-ouya-mainline/config-ouya-ouya-mainline.armv7

The Galaxy Tab 10.1 kernel is based on upstream linux v5.0. The Ouya kernel is based on grate-driver/linux. For Ouya, I've reverted the commits related to the new host1x implementation.

XXX: drm/tegra: plane: Add custom CSC BLOB property 
WIP: drm/tegra: Introduce new Host1x driver and new v2 UAPI
drm/tegra: Remove staging UAPI and Host1x driver 

[digetx]

@Decatf could you please try this change:

diff --git a/src/shaders/vertex_common.asm b/src/shaders/vertex_common.asm
index de3e169..71606ba 100644
--- a/src/shaders/vertex_common.asm
+++ b/src/shaders/vertex_common.asm
@@ -53,6 +53,6 @@ EXEC
        RCPs r0.***w, c[4].wwww
 ;

-EXEC(export[1]=vector)
+EXEC_END(export[1]=vector)
        MULv r63.xyzw, r3.xyzw, r0.xyzw
 ;

[Decatf]

could you please try this change:

This change fixes the rendering bug.

[digetx]

@Decatf awesome, thank you!

[digetx]

Everything works fine, it all got enough testing and Erik is too busy for review. Hence merging!