[differentiable] make_adjoint pass should support mutable local variables and complex branching

[yuanming-hu]

(need to come up with a systematic solution.)

[archibate]

Btw, what does differentiable mean? I guess it's related to kernel.grad()? Was this a unique feature of taichi but no others like tvm? Sounds interesting! A recent issue #571 says if is not allowed in grad kernels, and I told him/her to use max/min instead. Expected to be resolved by this issue?

[yuanming-hu]

Btw, what does differentiable mean? I guess it's related to kernel.grad()?

Exactly!

Was this a unique feature of taichi but no others like tvm?

Yes, because Taichi has an imperative mega-kernel design. This is different from coarser-grained systems such as tensorflow/Halide/tvm.

A recent issue #571 says if is not allowed in grad kernels, and I told him/her to use max/min instead. Expected to be resolved by this issue?

True.

I just came up with a solution and will try to implement it tomorrow.