Our Vulkan regression test suite includes a Vulkan trace of the Sasha Willems shadowmapping demo. We found that the fps when playing back that trace file had decreased by about 20%. I narrowed down the performance drop to commit 888b7fc.
I duplicated the performance drop with the demo itself. Without commit 888b7fc, the demo runs at an average of 3125 fps. With commit 888b7fc, it runs at 2675 fps. This is a performance regression of about 14%. I ran this on a system with an Intel i5-6600K CPU @ 3.50GHz and an AMD Radeon RX470 graphics card.
Since the referenced sha is for the 18.1 rather than the master branch I'm copying here for completeness.
Author: Samuel Pitoiset <email@example.com>
Date: Wed Jun 13 20:19:23 2018 +0200
radv: don't fast clear HTILE for 16-bit depth surfaces on GFX8
This causes rendering issues in Shadow Warrior 2 with DXVK.
Fixes: ccc64f3133 ("radv: enable TC-compat HTILE for 16-bit depth surfaces on GFX8")
Signed-off-by: Samuel Pitoiset <firstname.lastname@example.org>
Reviewed-by: Bas Nieuwenhuizen <email@example.com>
(cherry picked from commit 51e23d34190076159129dd7b449b95a1ac3d4949)
Thanks for reporting this.
Unfortunately the original performance tweak doesn't really work on GFX8, as TC compatible HTILE for 16-bit depth surfaces is only natively supported on GFX9.
I know that this decreases performance for the shadowmapping demo but the impact is really tiny for real apps and we can't restore it without introducing rendering problems in some situations (as explained in the commit message).
It's nice to see that your CI system Vulkan is working. :)