4 ms·
Hugh Hoppe, from Microsoft, wrote some fundamental progressive-mesh code that he showed at SIGGRAPH '97 using this model. Instead of the code having to choose f
by SavantIdiot 5y ago
Hugh Hoppe, from Microsoft, wrote some fundamental progressive-mesh code that he showed at SIGGRAPH '97 using this model. Instead of the code having to choose from different model resolutions (like the original four bunny models), the algorithm would allow DirectX to reduce the vertex count based on the rendered size of the object (1 pixel only needs 1 vertex!) to save compute cycles.
I wonder if geometry compression is still a thing given the performance increases of GPUs. I haven't done 3D driver work since DirectX 5....
- wetmore 5y agoVery much still a thing, e.g. for reducing the poly count of a mesh before running it through a procedure to spit out gcode for 3d printing or milling.
- Causality1 5y agoIt pleases me that someone named Hoppe picked a rabbit for the demo.
- SavantIdiot 5y agoHa! In 24 years I never made that connection.
- Philip-J-Fry 5y agoIt's funny you say that because that's effectively what Unreal Engine 5 is doing with it's Nanite technology. Also released in early access today. I do wonder what the problem was that needed to be solved before we got Nanite. Everyone knew that automatically optimising the geometry for the size on the screen was doable. Maybe it was a storage limitation?
- rasz 5y agoThat whole UE5 demo runs close to 1 triangle per pixel, its pure graphical masturbation. The problem is GPU power.
- flakiness 5y agoFor someone who's moderately interested: http://hhoppe.com/lapped_stonebunny_siggraph_cover.html http://hhoppe.com/lapped_stonebunny_siggraph_cover.html Glad to see he's still "the mesh guy", wondering what he's doing at Google. http://hhoppe.com/#publications http://hhoppe.com/#publications
- dahart 5y agoYeah, geometry compression is definitely still a thing. The main reason is that the relative gap between compute speed and memory bandwidth is still growing. It's pretty common now to bump into situations where redundant computation is faster than reading saved results from memory. Compression is perhaps ironically or surprisingly getting more important as GPU perf increases...
- corysama 5y agoContinuous LOD research dropped off dramatically when GPUs got vertex shaders. At that point, having the CPU touching the vertex data at all became more expensive than it could be worth. Instead, 3 or 4 static LODs have been the way to go. Those static meshes can be individually well optimized. There are surprisingly few good options for auto generating static LODs even today... Mesh Shaders are opening up whole new opportunities that are only starting to be explored. You get far more general purpose, compute shader like features. But, you have to work in small packets. Funny enough, they are a lot like PlayStation 2 vector unit programming made modern. That design really was ahead of it’s time. But, the fundamental physics advantages of SIMD working on a small chunk of SRAM are as valid as ever.
- plafl 5y agoNot for GPUs AFAIK but still related: https://www.blosc.org/pages/blosc-in-depth/ https://www.blosc.org/pages/blosc-in-depth/
- boulos 5y ago> Hugh Hoppe Hugues (though most people pronounce it fairly close to “Hugh’s”, so maybe that’s what stuck for you?)