How to record the order of kernel and automatically launch them?

[Hanke98]

Hi, everyone!

We are currently working on implementing our simulator using Taichi AOT. We can dump the shaders according to the demo code of implicit_fem.py. However, it seems that we cannot record the calling order of each kernel and we have to manually launch the kernels, which is not very productive. Is it possible to dump the launch order of the kernels and automatically launch them?

Thanks in advance!

[bobcao3]

You can check out the compute graph work we have been doing. We decided to go with a graph API instead of a tracing API, as capturing a series of kernels is very hard to implement correctly. @ailzhang knows more about this

[ailzhang]

@Hanke98 You can checkout a cgraph version of implicit fem here https://github.com/taichi-dev/Taichi-UnityExample/blob/main/scripts/implicit_fem.cgraph.py. With this you'll be able to encode the calling order of kernels and simplify the calling code inside deployment environment. https://github.com/taichi-dev/Taichi-UnityExample/blob/main/Assets/Scripts/ImplicitFem.cs#L218-L244 Please note this feature is WIP so API/implementation details might change, but we'll keep the demos in UnityExample repo up to date.