Alexander Veselov

Computer Graphics & AI Algorithm Engineer @Intel, previously @AMD.

[AMD GPUOpen @GPUOpen]

#AMD Advanced #Rendering Research Group will be demonstrating our real-time global illumination research at SIGGRAPH exhibition! #ARR #raytracing (1/2)

[Mike Dereviannykh @Mishok2000]

Today I'm planning to present my new short paper "Real-Time Path-Guiding Based on Parametric Mixture Model" at the Eurographics 2022 at 9 am in Reims.
My technique helps to drastically decrease the variance of MC estimators for 1st bounce of light without sacrificing performance

[Alexander Veselov @TheKAJAJJJ]

It's unexpected graphics debugging behavior time!

[Yusuke Tokuyoshi @yusuketokuyoshi]

Tiled Reservoir Sampling for Many-Light Rendering (technical report)
https://t.co/adZNdJ2PEe

[Guillaume Boissé @guitio2002]

Thanks everyone for all the reaction and feedback, it means a lot! I’m a bit of a noob at this and hadn’t quite understood that preprint ≠ final version :) So here’s a preprint of the paper: https://t.co/Us2pzIGAzC

[Guillaume Boissé @guitio2002]

Got my first paper accepted at SIGGRAPH Asia: “World-space Spatiotemporal Reservoir Reuse for Ray-traced Global Illumination” 🎉

[Alexander Veselov @TheKAJAJJJ]

Verifying some materials in my toy GPU path tracer

[Mikkel Gjoel @pixelmager]

Reference GLSL-implementations of MaterialX shading nodes: https://t.co/reYbawQz31 (from https://t.co/NKlAq6yUBi )

[Alexander Veselov @TheKAJAJJJ]

Does anyone know about the Baader-Meinhof phenomenon? Not so long ago I learned about using Control Variates method in computer graphics. Now I'm finding mentions of this method in every second CG paper.

[Bartosz Ciechanowski @BCiechanowski]

Cameras and Lenses – https://t.co/RWp3NAicmW

[Alexander Veselov @TheKAJAJJJ]

@TooMuchVoltage @nholzschuch @pointinpolygon Great, now I fully realized your approach. Thank you for the inspiration!

[Alexander Veselov @TheKAJAJJJ]

@TooMuchVoltage @nholzschuch @pointinpolygon Yeah, I thought about spatial filtering of neighbor probes. But again, the visibility stops me: for example, if one probe is in a dark room and another is outside (or even one of probes is not computed). But now I'm thinking about screen-space averaging before injecting to grid.

[Alexander Veselov @TheKAJAJJJ]

@TooMuchVoltage @nholzschuch @pointinpolygon BTW, here you can see DDGI-like results I reached so far. It traces rays from probes and temporally accumulates into hierarchical SH9 grid around the camera. The number of rays is comparable to 1920*1080, but variance is still really noticeable.

[Alexander Veselov @TheKAJAJJJ]

@TooMuchVoltage @nholzschuch @pointinpolygon Wow, so you inject screen-space trace results into a world-space structure. Sounds promising to avoid leaks. Thanks, I have to try that!

[Alexander Veselov @TheKAJAJJJ]

@TooMuchVoltage @nholzschuch @pointinpolygon I spent a lot of time on a similar probe-based approach, but always ran into problems: Light/shadow leaks OR lack of temporal responsibility OR bad performance OR bad convergence... Your work looks like magic to me.

[Alexander Veselov @TheKAJAJJJ]

@TooMuchVoltage @nholzschuch @pointinpolygon Looks awesome! So, as far as I understand, your approach is similar to DDGI, but you use a kind of ambient cube to store directional info? How many rays do you trace per frame and do you handle visibility/leaks?

[inigo quilez @iquilezles]

Biplanar mapping - an alternative to Triplanar/RoundCube that only considers the two most relevant of the three projections at any given point. It has some singularities but in practice it works great and saves a lot of bandwidth:

https://t.co/PBbzki3E63

https://t.co/9yP9YYEA5P

[https://mastodon.social/@vitaut @vzverovich]

Just initialize a variable they said...

[Arseny Kapoulkine @zeuxcg]

@KostasAAA @SebLagarde A cute perf trick on PS3 for post-processing passes used to be to force an artificially high register count via a pragma in the shader.

This would decrease occupancy and, for simple post-fx like blur, could noticeably improve performance by reducing tex cache thrashing.