Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

UE5 Nanite in WebGPU

github.com

Post reply on HN
Print view View original

111–117 of 117 posts

Re: UE5 Nanite in WebGPU

#111

Post by 01HNNWZ0MV43FF »

jsheard wrote: post #11
throwaway17_17 wrote: post #3

Earlier quoted context omitted.

What is the use case for atomics in the rasterizer? I can’t figure out what the atomic operations do inside the rendering pipeline. I looked at the GitHub, but couldn’t find the place the atomic were hoped for.

With traditional hardware rasterization there are specialized hardware blocks which handle atomically updating the framebuffer to whichever sample is currently the closest to the camera, and discarding anything behind that. Nanite does software rasterization instead, and one of their insights was figuring out a practical way to cram all of the data needed for each pixel into just 64 bits (depth in the high bits and e…

Software rasterization, in a GPGPU compute shader, in hardware.

Re: UE5 Nanite in WebGPU

#112

Post by 01HNNWZ0MV43FF »

kllrnohj wrote: post #76
NotGMan wrote: post #18

Earlier quoted context omitted.

I'm a bit out of the GPU game but so this might be slightly wrong in some places: the issue is in small triangles because you end up paying a huge cost. GPUs ALWAYS shade in 2x2 blocks of pixels, not 1x1 pixels. So if you have a very small triangle (small as in how many pixels on the screen it covers) that covers 1 pixel you will still pay the price of a 2x2 block (4 pixels instead of 1), so you just wasted 300% of y…

> So even if you do it in software the point is that if you can get rid of that 2x2 block penalty as much as possible you could be faster than GPU doing 2x2 blocks in hardware since pixel shaders can be very expensive. Of course the obvious problem with that is if you don't have most of the screen covered in such small triangles then you're paying a large cost for nanite vs traditional means.

Nanite has an heuristic to decide between pixel-sized compute shader rasterizing and fixed-function rasterizing. You can have screen-sized quads in Nanite and it's fine

Re: UE5 Nanite in WebGPU

#113

Post by password4321 »

soulofmischief wrote: post #108
password4321 wrote: post #86

Earlier quoted context omitted.

Your ideas are intriguing to me and I wish to subscribe to your newsletter.

I appreciate that, anything in particular catch your interest?

I am most interested in speech recognition including diarization.

Re: UE5 Nanite in WebGPU

#114

Post by Pannoniae »

andybak wrote: post #99
TinkersW wrote: post #57

Earlier quoted context omitted.

A couple reasons 1. HW does 2x2 blocks of pixels always so it can have derivatives, even if you don't use them.. 2. Accessing SV_PrimitiveID is surprisingly slow on Nvidia/AMD, by writing it out in the PS you will take a huge perf hit in HW. There are ways to work around this, but they aren't trivial and differ between vendors, and you have to be aware of the issue it in the first place! I think some of the "software…

What's PS? Pixel shader? I'm guessing here.

Yes, correct

Re: UE5 Nanite in WebGPU

#115

Post by soulofmischief »

password4321 wrote: post #113
soulofmischief wrote: post #108

Earlier quoted context omitted.

I appreciate that, anything in particular catch your interest?

I am most interested in speech recognition including diarization.

I'll try to find time to write about it, but in the meantime if you just want to try something that works, Xenova published some tools and examples about two months ago which I'm sure will give you a good start.

https://github.com/xenova/whisper-web/tree/experimental-webg...

https://huggingface.co/spaces/Xenova/whisper-speaker-diariza...

https://huggingface.co/onnx-community/pyannote-segmentation-...

Speaker diarization is quite difficult as you know, especially in loud or crowded environments, and the model is only part of the story. A lot of tooling needs to be built out for things like natural interruption, speaker memory, context-switching, etc. in order to create a believable experience.

Re: UE5 Nanite in WebGPU

#116

Post by turtledragonfly »

runevault wrote: post #91
DaoVeles wrote: post #80

Earlier quoted context omitted.

In a past life (2000's) I was doing some dev stuff on Ps3, trying to figure out so decent uses for Cell's mass of compute and working around RSX's limited memory bandwidth while having the luxury of Blu-ray storage to burn through. One such thing I did get a fair way into was something like Nanite - I called it compressive meshing. It is the typical case of misguided engineering hubris at work. The initial work looke…

Keep in mind, it took Epic a long time to get it sorted. I think I saw the primary creator say it took him a decade of research and work to come to the initial implementation of Nanite that shipped in Unreal.

And it's crazy to think that during that decade, GPU functionality was changing pretty rapidly. So, you really are aiming for future GPUs that don't exist yet as you develop the tech.

Of course, the trajectory of GPU advancements is somewhat predictable, and settled down a little bit relative to the not-too-distant past. Perhaps some luck involved, too (:

Re: UE5 Nanite in WebGPU

#117

Post by password4321 »

soulofmischief wrote: post #115
password4321 wrote: post #113

Earlier quoted context omitted.

I am most interested in speech recognition including diarization.

I'll try to find time to write about it, but in the meantime if you just want to try something that works, Xenova published some tools and examples about two months ago which I'm sure will give you a good start. https://github.com/xenova/whisper-web/tree/experimental-webg... https://huggingface.co/spaces/Xenova/whisper-speaker-diariza... https://huggingface.co/onnx-community/pyannote-segmentation-... Speaker diarizat…

Thanks for the links! Most of the fined tuned diarization models are locked up behind SaaS paywalls, not open weights usable directly.
Post reply on HN