@TheChowdhary That's simply not true at all. Simple projects used to be great to judge the level of someone.
What can easily be generated is indeed not impressive anymore, but dedicating time and effort to push a project far and beyond is valuable.
"Lone monk" scene by Carlo Bergonzini, rendered in albedo, my #rust#webgpu pathtracer (no RT cores).
Good result, but still a long way to go :)
(Unfortunately the roof tiles were made of a Blender modifier, and don't look that good on this export 😅)
I have a pretty weird #webgpu (#rust/#wgpu wasm codebase) issue only on Safari, where compute passes don't output anything to buffers.
Worse: I have a command encoder with [compute, compute..., debug_blit], the debug_blit is broken except if moved to another submit.
Anyone?
Not posting here much those days, but here is a Firefox vs Native comparison on my #webgpu#rustlang pathtracer.
Performance hit is clear in the web version. This scene is GPU bound, no CPU work and no CPU -> GPU flow.
Weirdly, Chrome is barely usable compared to Firefox.
@BrianBreiholz@sampavlovic You need an indirection lookup, contains LOD info and data start. Then based on the LOD you know how to perform the linear interpolation of the samples
@WebXREngine Shout-out to @pezcode and @noerihuisman for creating the original meshlet PoC that is the foundation of what will be available in 1.6.0!
Crazy to think that they prototyped that during a fun offsite in Austria :)
In order to start diversifying your timeline from AI-only content, here is a video of the next @WebXREngine version.
Both showcase meshlets with hierarchical culling, roughly 32*32*130K tris, nicely frustum / LOD culled.
Running on #webgl via Wonderland Engine
@ccallac7 With time you get better at understanding all the overloads it tried to resolve. I would also recommend to ditch the std and use better containers.
@ocornut Took a bit of setup, but VSCode + Cmake extension + LLDB extension is nice enough. I wrote some lldb provider for custom containers.
The only pain point is that sometimes the debugger view just freezes. Obviously doesn't happen on LLDB cli
@mrmaxm@WebXREngine Mix of both, we do build a voxel representation to efficiently place bricks mapping to the probes SH data.
The SH itself is however generated from camera rendering to cubemap.
@PlayingInCanvas@squareys This exact scene is 2.5mB, with 52K probes.
The sparser the scene is (think outdoor scene with props), the lighter the GI will be. Sponza is relatively dense, except in the middle corridor.