Ailing Zhang

Current ndarray implementation doesn't respect python. For example if you do `del ndarray1` only the ndarray object is deleted but the underlying memory is not freed. (managed by Program) Ideally...

as titled

Support ndarray deepcopy in C++ wrapper of AOT api

3

There're cases where users would want a deepcopy of an ndarray, so we can probably provide this as ``` auto arr_copy = arr.clone() ``` in our C++ wrapper. cc: @YuCrazing

Validate ndarray dtype in C++ wrapper.

2

In deployment scenarios, we may accidentally pass in an ndarray with wrong dtype compared to the one required by the compiled kernel, we should throw proper error to users when...

Ndarray type annotations & error message refinement

1

We had a discussion about refining ndarray definition and type annotation in https://github.com/taichi-dev/taichi/discussions/4085 earlier this year and it's now a hard blocker if we want to properly release ndarray to...

This is a long-term feature request that eventually we should decouple the compilation from runtime. In other words AOT users should be able to generate aot modules for vulkan backend...

[Tracking issue] Pain points when interacting with external framework (Numpy/PyTorch)

3

This issue serves tracking purpose for all issues related to user experience when interacting with external framework like Numpy and PyTorch. - [ ] Extra copy if a non-contiguous torch...

Custom C++ and CUDA Extensions tutorial need to be updated to use dispatcher API

4

As discussed in the sync, we need to rework this tutorial so that it's registering ops to dispatcher instead of only create python binding using `PYBIND11_MODULE`. cc: @ezyang @bhosmer @bdhirsh

as titled

Tune android system resource scheduler for performance measurement

5

While we developed [Android AOT demo for v1.0 release](https://github.com/taichi-dev/taichi/releases/tag/v1.0.0) we noticed an interesting (or confusing) behavior in our demo app. It runs super fast upon startup, and then slows down...