BufferizedNDArray inherits from EagerTensor

[sreevickrant]

#351

[willow-ahrens]

Can bufferizedndarray get moved to be imported after interface? I think it should be imported around the same time we import FIberTensor

[sreevickrant]

I was going to ask. I didn't know where to move it so I resorted to update in place.

[willow-ahrens]

Why is count nonzeroes in this pr?

[sreevickrant]

tests/test_notation_interpreter.py::test_count_nonfill_vector needs count_non_zeros in EagerTensor. I can revert it back.

[willow-ahrens]

does it? It looks like it's just asking for np.count_nonzeros. The count_nonfill op is implemented in FinchNotation. I think all that's needed there is to convert a to a ndarray just before we call np.count_nonzeros, is that right?

[sreevickrant]

Can bufferizedndarray get moved to be imported after interface? I think it should be imported around the same time we import FIberTensor

I think the problem here is I need to fix dc_stats.py to get rid of the circular import because it uses BufferizedNDArray.

[willow-ahrens]

We're trying to avoid dynamic imports in the library, but I'm not sure what to do about this one.

[sreevickrant]

I tried of bunch of stuff but I think this was the cleanest fix. We can always force the linter to not sort init and we can avoid a bunch of the dynamic imports.

[willow-ahrens]

@kylebd99 @mtsokol do you see a way to avoid this dynamic import? The summary is that DCStats uses bufferizedNDArray to accumulate results in finchnotation, but DCStats is defined in autoscheduling module, after the definition of interface. I see two easy fixes:

1. Embrace chaos and allow relative imports (strongly against but can be convinced)
2. Define a set of "advanced" autoschedulers after we define interface functions, so that they can reuse the interface functions. This is just moving complex statistics stuff to a different folder that gets loaded after interface gets loaded.

[willow-ahrens]

Is option 2 above a possibility?

[sreevickrant]

I agree with option 2. It would also fix the other dynamic import that I had to make:

autoschedule/compiler.py — from ..compile.lower import make_extent inside _lower_query_of_reorder and _lower_query_of_aggregate.