Replies: 7 comments 30 replies
-
|
Current WIP on generalized cluster shading: Level-opaque Landscape-alphacut Level-alphacut |
Beta Was this translation helpful? Give feedback.
-
|
This naturally requires |
Beta Was this translation helpful? Give feedback.
-
|
What about mesh shaders on Apple silicon? Right now the performance is pretty bad. Is there any possibility of improving it? |
Beta Was this translation helpful? Give feedback.
-
|
Apparently questions is not to me here. On M1/M2 (and equivalent) they went for full software emulation. On M3 is suppose to be fixed - yet can't be 100% sure. |
Beta Was this translation helpful? Give feedback.
-
|
oh that's a shame. Maybe in a few years on newer hardware it'll make sense. |
Beta Was this translation helpful? Give feedback.
-
|
Draw indirect on DX12 :( |
Beta Was this translation helpful? Give feedback.
-
|
@lmichaelis , @ezamelczyk , @Nindaleth , @thokkat Sorry guys for mass-tagging, yet if you have time, can you please build&test At this point most of game should work, except known issues:
|
Beta Was this translation helpful? Give feedback.
-
|
I'm getting GPU hangs at the moment on this branch (that includes both 39a5edd that worked previously and also the current 682af3a), wonder if it's caused by a Mesa update. |
Beta Was this translation helpful? Give feedback.
-
This is bad... hang you mean real hang? Asking as in Vulkan there is a system-defined timeout on gpu-execution (2sec usually). |
Beta Was this translation helpful? Give feedback.
-
|
Maybe my terminology is imprecise and it's a GPU reset? On the loading screen, the visual output freezes at a certain point, including the built-in FPS counter and the external MangoHUD overlay. After a few seconds my display resets (as if you disconnected and reconnected the display cable) and the same frame stays stuck on the display. Keyboard and mouse inputs don't have any effect since the initial freeze. Out of various attempts, my system partially recovered only once from the freeze and that happened basically by the X session dropping and starting again. Here's the log from that case (no logs were produced in the other cases): |
Beta Was this translation helpful? Give feedback.
-
|
A similar thing happens on my AMD rig. Whenever I start OG my window server crashes fully and does not let me log back in until I reboot. I was able to recover the following messages from syslog: Syslog ExtractWhen I try to log back in after the crash but before a reboot I get and then a bit later To verify I isn't a general system issue I launched BG3 which ran fine. The crash happens regardless of whether I enable or disable raytracing. |
Beta Was this translation helpful? Give feedback.
-
Important part here is Can be anything - infinite loop, invalid access to array, bad alignment, null, bad driver, etc. Today and yesterday been debugging similar error on IntelUHD - maybe it a shader/descriptor-set mismatch. Plan so far is to de-risk it a little and add new command-line key
Haven't traced BG3 :) Yet, probably it not as crazy on shading, and can work with draw-spam just fine. Here for GDR build we have 2 things that make life hard for hardware - GPU-driven visibility testing (lock-free + CS) and bindless (landscape is 2 draws pretty-much) |
Beta Was this translation helpful? Give feedback.
-
|
New testing version is available (178dd6c): Animated textures and water tesselation is back. Tested on my end RTX3070/IntelUHD/MacM1 - no validation errors, no crashes. DX12 still broken - if everything else is fine, probably it's OK to merge it as-is and solve DX problem later. Trouble shooting flags:
|
Beta Was this translation helpful? Give feedback.
-
|
Here's the results on my RX 6800 XT. I tested using a release build.
I did not notice any visual artifacts during testing. Apart from stutters when starting a new save file I didn't notice any. The FPS numbers seem very unreliable so perhaps its CPU bound now. The Gothic data files are stored on an SSD. Footnotes |
Beta Was this translation helpful? Give feedback.
-
|
what would be you opinion? |
Beta Was this translation helpful? Give feedback.
-
|
Looks good to merge I'd say. You should work out those stutters before the next release though so that new users don't immediately have a bad experience :) |
Beta Was this translation helpful? Give feedback.
-
|
Current branch state is also working for me (6700 XT) when loading a save game, no crashes or stutters on Mesa git. |
Beta Was this translation helpful? Give feedback.
-
|
Added quick fix for stutters: 0a4c1bf - not a 100% solution, but might be just good enough Do last testing before merge, GPU timeline: CPU flamegraph Wind animation on trees takes ~15%, what is absolutely delightful... Still more optimization work for future. |
Beta Was this translation helpful? Give feedback.
-
|
Hello!
Seems that it's too expensive, but i couldn't come up with better solution yet |
Beta Was this translation helpful? Give feedback.
-
|
What you proposing is well... normal MSAA based solution. Prepass is somewhat out of place (HiZ can be sourced from non-MSAA depth). I've mention my thought about AA sometime before. MSAA gbuffer is expensive memory-wise, for both store and drawing into. specially on 2k displays. I don't see great option to do AA today, so just do other stuff for now. Eventually we should expect Vulkan to support tile-shader, and we already got multi-sample to single-sample rendering. What we can do then:
Another option is Quincunx Antialiasing, - requires MSAA2 + custom sample locations + pain in fragment shader. And lastly MSAA visibility buffer can be an option for hi-end gpu's. We don't even have to use hardware-MSAA here, just 4x resolution, and write depth(32f) + clusterId(24bit) + primitiveId(7bit). Then compute-shader magic to resolve it to color and light-up. |
Beta Was this translation helpful? Give feedback.
-
|
Spent quite a time tuning sky/fog against path-traced reference: // 20:00 // 8:00 // 12:00 This change enables to remove hack for direct-light (it was artificially dimmed) and enables better GI |
Beta Was this translation helpful? Give feedback.
-
Prototype with draw-indirect (non-indexed clusters) looks promising. On Intel:
0.06ms2.34msPutting it together: 2-pass culling for landscape work well in bindless-hardware; can be scaled down to run "bindfull", with one indirect draw per material (338 draws for landscape).
Point from *1 are applicable here, yet less trivial - need to do depth re-projecting.
Tessellation is quite slow and [practically] not available on Apple. Can software rasterization (metal 3.2 afair), and tessellate water up to 1 triangle per-pixel.
Beta Was this translation helpful? Give feedback.
All reactions