Thu, Sep 19
Ok both Linux + GTX 960M (driver 430.40) and Win 10 x64 + GTX 960 with latest drivers are working correctly. So I'm guessing it's an issue with the RTX card driver.
I cannot reproduce on any of my setups (even linux + nv430.40). I'm trying to figure out what's happening.
Wed, Sep 18
Tue, Sep 17
- Fix Gpencil matrix
- Fix crash due to drawing with buffers not big enough
Mon, Sep 16
I think it's more related to the de-correlation factor I applied on soft shadows sampling. The factor is ~1000 so maybe changing this factor may reduce convergence time. But it's a bit black magic at this point, it can work in some case and fail in others.
Sun, Sep 15
When we debugged with LazyDodo yesterday we found out that probe_display pass was the issue. Maybe there is other passes that crash. Our tests lead to the modelBlock ubo. It seems to be an issue with the way it is setup or queried as the crash does not occur if the ubo is renamed (and not used).
Sat, Sep 14
Fri, Sep 13
With D4997 applied, the slowdown.blend file is ~32fps which is about the same performance as 2.79 [which is around 30fps (hard to tell because the fps is really not stable)].
I did some exploration of the idea. (see P1099)
Implemented in rBffd5e1e6acd2
But the performance is quite noticeable with a lot of textures.
Is it technically possible, to give the user the ability, to deactivate certain calculations eevee is performing?
Implemented by rBcaf494444cd7
Thu, Sep 12
Wed, Sep 11
Tue, Sep 10
I'm not sure if that indicates a bug in the non-optimized code, or if it's intentional the optimized code is more accurate as well?
- Fix intel performance regression
Ok just adding back ModelMatrix triggers the legacy path and just draw without instancing, bringing back old performance for intel.
- Fix previous fix
A test with the spaceship test file show that the slowdown is not acceptable:
Workbench time 9.7ms > 15.5ms
- DRW: Fix crash with old intel GPU driver
- GPencil: Fix wrong stencil ID clearing
- GPU: Fix shader crashing on some intel gpu
Sun, Sep 8
Sat, Sep 7
After trying Sobol, I settled on using a multiplicator to halton as the points are better distributed than Sobol (in my opinion).
Fri, Sep 6
Thanks for raising awareness. After investigating, it seems to be a correlation artifacts. The same random numbers are used for AA and soft shadows jitter.
About the intel performance regression. I would really like some help in testing if the difference is noticeable in normal scene on other drivers/GPU.
This is why I would rather give full control over the shading and shadowing with a shader script node than trying to alienate EEVEE into something that is half usable for both NPR and PBR
A shader script node would probably more than make up the difference, and I think in the long term it's better for EEVEE to focus on PBR and matching Cycles. A more dedicated NPR system (like LANPR) might be the best thing to do.
Is there anywhere on devtalk or developer.blender.org that NPR is being roadmapped or discussed specifically? I'd like to list out some of the scenarios I use EEVEE for, now, and the limitations they have. Maybe I could also make an RCS proposal, like the one Marius Oberholster did a few years back.
Thanks for the reply and for all the work you do!
Commited as rBd8aaf25c23fa
Thu, Sep 5
I've also noticed that SSS translucency isn't being preserved by Shader-To-RGB node anymore. I think the changes to SSS in this commit are responsible for this.
Ha right! This is because translucency is now done as a separate pass in post and is not computed before the shader to RGB. I will have to add it as a limitation.
The regression were fixed but the tests hair_instancer_uv and visibility_particles were actually broken in 2.80 (scale affecting normals). Will update tests references after committing.