at 7:15 in this video
He says it's grid free, you can download the demo to play with the balls and in some worlds you can move the balls anywhere you want, no walls
how Nvidia does the balls collision detection in Flex without using a 3D grid?
Instead of storing a full grid of cells (whether or not they are empty), they store a flat representation of just the active cells.
Because fluid simulations need to constantly update neighbouring cells, the grid representation is a lot easier/faster to update, but you are constrained by the maximum size of grid (i.e. 3D texture) that you can store on your GPU, and if the majority of cells are empty, then you are just plain wasting memory.
The secret sauce is how they update potentially millions of neighbour interactions efficiently on a GPU with a sparse representation - that's traditionally been a prohibitively difficult thing to accomplish.
Because fluid simulations need to constantly update neighbouring cells, the grid representation is a lot easier/faster to update, but you are constrained by the maximum size of grid (i.e. 3D texture) that you can store on your GPU, and if the majority of cells are empty, then you are just plain wasting memory.
The secret sauce is how they update potentially millions of neighbour interactions efficiently on a GPU with a sparse representation - that's traditionally been a prohibitively difficult thing to accomplish.
Volume Tiled Resources. Same as in this demo:
Here's a slideshow: http://developer.download.nvidia.com/assets/events/GDC15/Dunn_Alex_SparseFluidSimulation.pdf
Volume Tiled Resources. Same as in this demo
Well that's a whole lot less exciting than I was imagining.
Looks like it boils down to in-driver support for sparse virtual textures?
Or the memory controller hardware: 64bit address space, several different choices in page size, ability to gracefully handle page-not-mapped HW exceptions, ability for texture samplers to return a page-miss boolean, etc, which allows drivers to implement "tiled resources", which allows games to implement HW-accelerated SVT :)Looks like it boils down to in-driver support for sparse virtual textures?
Allocating/deallocating virtual pages is a CPU side operation though. So if your GPU particleshaders created a list of voxels that must be resident, you've got to wait a frame for the CPU to read back that list and actualy map the appropriate tiles... This is still a complex system to construct!
This is still a complex system to construct!
Yeah. Looks like the only piece of it that is truly new is the ability to sample out of bounds and get a zero value?
Everything else it seems you could emulate (albeit probably not with sufficient performance) with SVT and mixing texture dimensions.
Isn't SVT still logically a grid? It's just that empty parts aren't resident. I'd still consider that a grid though.
This topic is closed to new replies.
Advertisement
Popular Topics
Advertisement
