compile *.comp with glslang optimizations enabled (-Os)

[sharkautarch]

It seems that glslang doesn't actually do any optimizations by default.
I compared the cs_composite_*.h file sizes between gamescope built from ValveSoftware/gamescope/master, and gamescope built with a modification to src/meson.build where I added -Od to glslang's args
And the file sizes with or without -Od turned out to be the same...

Oddly enough, glslang only lets you specify -Os or -Od, not sure why...

I exclude one shader from optimization: cs_composite_rcas.comp
because I had noticed that if optimization was enabled for cs_composite_rcas.comp, it would take longer for gamescope to startup when running w/ fsr (also, it seems like that specific shader's .h output file would balloon in size w/ -Os)

[misyltoad]

FWIU, glslang optimizations are pretty worthless, the majority of the good stuff comes from the driver.

[sharkautarch]

FWIU, glslang optimizations are pretty worthless, the majority of the good stuff comes from the driver.

From some testing with running gamescope on my integrate intel igpu (Intel Alder Lake-P GT2 [Iris Xe Graphics]), it seems like glslang seems to be able to make some optimizations that the driver might be lacking...
I ran a gpuvis-trace of a gamescope nested session w/

• ValveSoftware/gamescope/master
• glslang -Os optimizations

where I ran the same exact command for both: ./gamescope -W 1920 -H 1200 -w 992 -h 620 -F pixel -- vkcube
and found that the minimum paintall time for ValveSoftware/gamescope/master was around 3.25ms,
whereas the minimum paintall time for glslang -Os was around 3.1ms.
And overall, it seemed like paintall for glslang -Os was about 0.1 to 0.2 ms faster.

Maybe some drivers implement more shader optimizations than others?
Also maybe it might only be noticeable when using gamescope to do upscaling, maybe even where it is only noticeable with specific upscaling algorithms like the pixel one???

w/o -Os :

w/ -Os:

also more specs for my laptop I tested this on:

CPU: 14-core (6-mt/8-st) 12th Gen Intel Core i7-12700H (-MST AMCP-)
speed/min/max: 426/400/4600:4700:3500 MHz Kernel: 6.6.10-arch1-1 x86_64
vulkan-intel 1:23.3.3-1
mesa 1:23.3.3-1

[misyltoad]

3.1ms seems extraordinarily slow for the composite shader, but what you are reporting seems kinda margin of error given it looks like n = 1 in your testing.

If you disable the 1D + 3D LUT stuff going on, does it make it faster on Intel?

(Just comment it out)

[sharkautarch]

3.1ms seems extraordinarily slow for the composite shader, but what you are reporting seems kinda margin of error given it looks like n = 1 in your testing.

If you disable the 1D + 3D LUT stuff going on, does it make it faster on Intel?

(Just comment it out)

Ok:
I've retested gamescope w/ one trial run with latest upstream gamescope/master and one trial run with the 1D + 3D LUT stuff commented out

Here's the branch I'm running for testing running gamescope w/ 1D + 3D LUT commented out (in case I messed something up) : https://github.com/sharkautarch/gamescope/tree/test_w_no_lut
Same command for running gamescope as before, and for all of these runs, I've been running gamescope with cap_sys_nice=eip perms.

gamescope/master:

1D + 3D LUT commented out:

EDIT: I did another test where I removed a conditional from apply_layer_color_mgmt():
https://github.com/sharkautarch/gamescope/tree/color-mgmt-removed-condition
gpuvis:

[misyltoad]

Where's the GPU submit on your graph?

[sharkautarch]

Where's the GPU submit on your graph?

If by GPU submit, you mean the i915_* and intel_* perf stats, they're not being recorded, because those stats are only available for recording if your kernel is built w/ CONFIG_EXPERT=y and CONFIG_DRM_I915_LOW_LEVEL_TRACEPOINTS=y

btw, here's another gpuvis trace I did w/ just perform_3dlut commented out + a slight modification to perform_1dlut():
(branch: https://github.com/sharkautarch/gamescope/tree/gamescope_no_lut3d_opt_lut1d)

[sharkautarch]

closing this, since it seems that glslang -Os doesn't make any difference for me when color management stuff in the shader is commented out