Hi
i'm trying h21,one scene with much refraction is much slower than h20.5.
in 20.5 it takes 35 mins,but in h21 it takes 70 mins! and the optix(gpu) contribute less than before.

my gpu is 5070,and system is ubuntu,what can do?

and it occupation more vram, from 10G to 12G!

Did you use this Precompile Karma XPU Render Kernels by any chance? Under Render menu

not yet,i'm compiling now. thank you!

still sloooow!

luoqiulin
in 20.5 it takes 35 mins,but in h21 it takes 70 mins! and the optix(gpu) contribute less than before.

Is this in the viewport? Or offline render?

What happens if you try this envvar?
KARMA_XPU_OPTIX_FORCE_OPTIXIR=1

luoqiulin
and it occupation more vram, from 10G to 12G!

In what case are you seeing this?
If Houdini is running, then it could be the Vulcan viewport (or perhaps COPs) taking up this extra VRAM.

brians
is this in the viewport? Or offline render?

What happens if you try this envvar?
KARMA_XPU_OPTIX_FORCE_OPTIXIR=1

in the viewport

and it's still slow after add your envvar!
but i don't find this var in "hconfig"

luoqiulin
in the viewport

ah ok.

For offline rendering, performance should still be mostly the same for H20.5 vs H21. But for viewport rendering we take steps to enable faster interaction, which can unfortunately hurt performance a little.

One thing that is new in H21 is that we compile the shaders differently in viewport (for faster shader compilation/iteration).

So how about try this
- remove the KARMA_XPU_OPTIX_FORCE_OPTIXIR envvvar
- add in this envvar KARMA_XPU_OPTIX_FAST_COMPILE_OVERRIDE=0

luoqiulin
i don't find this var in "hconfig"

These are internal dev/debug envvars for XPU, and are not designed for common use (eg they change all the time, some can cause issues elsewhere, etc...). So typically we don't document them unless we see them being needed by users.

brians
So how about try this
- remove the KARMA_XPU_OPTIX_FORCE_OPTIXIR envvvar
- add in this envvar KARMA_XPU_OPTIX_FAST_COMPILE_OVERRIDE=0

thank you! by add this var, it's much faster now. but still slower than 20.5.
here is a compare pic!

There are a couple of new things that might be contributing to that

1) “Thinwall At Refract Limit”
This fixes black patches on refractive objects due to bounce limits. It means more rays are fired, which could explain your slowdown. But on the positive side it means a much lower bounce limit is needed now (ie defaults of 4 are fine), so we’ve found overall its a win in both look and performance.

Does your scene have refraction? Try setting your R+R limits lower (eg back to 4). Does that speed things up?

2)”indirect fake caustics” removes some black shadows cast from refractive objects. This may come at a slight performance cost, but improves look.

Again, does your scene have refraction? It might be because of this…

Both features are listed here
https://www.sidefx.com/docs/houdini/news/21/karma.html [www.sidefx.com]

brians
Again, does your scene have refraction?

oh yes,with too many refletion and refrection!

thank you for your reply!
i will feed back my result after i try again!

Hi, i'm back to report my result.
after i reduce reflection and refraction limit from 10,20 to 4,4.the speed is faster but still a bit slower than 20.5.
the good news is,even the limit is much lower,it looks better than 20.5.
so i think it's fair enough!

and one more thing:
the view port's Flat Shade mode does not work !

luoqiulin
the view port's Flat Shade mode does not work !

I think there is a bug-report about this, but in short...
- flat-shade mode in the vulkan viewport will work with regular lights (currently its disabled for the default headlamp light)
- flat-shade mode (and all those other modes) will never work with Karma