Radeon RADV LBVH Patches Take One Vulkan Ray-Tracing Test From ~13 To ~250 FPS

[phoronix]

Phoronix: Radeon RADV LBVH Patches Take One Vulkan Ray-Tracing Test From ~13 To ~250 FPS

On top of other recent RADV ray-tracing improvements, there is another recently opened merge request pending that can dramatically help some Vulkan ray-tracing workloads with much better performance...

https://www.phoronix.com/scan.php?page=news_item&px=RADV-LBVH-Patches

[CochainComplex]

Very nice improvements Btw are there still some efforts to make raytracing workable on pre RDNA2 Hardware with mesa?

[Danny3]

This is great, if everything is rendered correctly!

[Linuxxx]

Very nice improvements Btw are there still some efforts to make raytracing workable on pre RDNA2 Hardware with mesa?

What's the point when the performance will always be abysmal without dedicated RT hardware?

And I believe I should take this moment to sincerely thank AMD for delaying the open-sourcing of AMDVLK just enough so that RADV could be created in the mean time;
else I would be in a parallel universe right now where my R9 380 is serving me well as my most expensive paper-weight...

[arti]

Very nice improvements Btw are there still some efforts to make raytracing workable on pre RDNA2 Hardware with mesa?

raytracing works with my rx480 on stock mesa on arch linux https://www.youtube.com/watch?v=BC_QMOfFg_c

Launch command used was

RADV_PERFTEST=rt RADV_DEBUG=nocache ./q2rtx.sh

https://blog.froggi.es/bringing-vulk...-amd-hardware/

[DrYak]

What's the point when the performance will always be abysmal without dedicated RT hardware?

As anyone who has played around with the source of POV-Ray will tell you, there's no inherent magic in ray-tracing, just tons upon tons of linear algebra.
And GPUs are already pretty good at repeating the same maths operation over and over on very large data sets.

So for the "what's the point" part: well it's a low hanging-fruit, hardware is already able to perform the kind of maths required for it. Would be a shame not to try it.

"performance will always be abysmal without dedicated RT hardware": Well, not necessarily. Granted, you'll not be able to play Control at 4k on old cards, but I wouldn't be surprised if it turns out that a Vega 64 could kind of run ray traced games at 480p.

[coder]

What's the point when the performance will always be abysmal without dedicated RT hardware?

The best argument that comes to mind is to enable developers to fiddle with it before they can afford to upgrade their GPUs. For software development, you don't necessarily need very playable framerates or resolutions.

Another use case would be for non-realtime rendering, for which a shader-based GPU implementation should still be much faster than a desktop CPU renderer.

[anarki2]

Did they remove some of the sleep()'s?

[coder]

Did they remove some of the sleep()'s?

Joking aside, BVH is an acceleration structure used to reduce the number of ray/poly collision tests needed per-pixel. Constructing BVH's tends to be costly, so I gather the patch implemented some techniques like those mentioned in this paper:

https://research.nvidia.com/sites/de...2hpg_paper.pdf

Indeed, the patch confirms:

"The 'new' bvh building code uses a combination of (link) and (link) to sort the nodes along a space filling curve before creating the internal nodes."

See the news article, for the MR link, if you're interested in learning more.