Andrew Adams
Andrew Adams
Given that the tests already passed without the polyhedral support, should there be new tests that exercise it?
A(i) would have to be injective and surjective for that scatter to define a total function. We need to decide on the semantics for repeated values on A(i) and values...
My inclination would be to make O(A(i)) = B(i) work if A(i) has known bounds (e.g. it's a uint8).
If they knew the inverse of A, they'd probably already be done. The code that inverts A is something like: invA(A(i)) = i
Those inputs are 1-D, not scalar. I think you want: ``` def avgpool(float(B, C, H, W) input, float kH, float kW, float sH, float sW) -> (output) { output(b, c,...
By pytorch plumbing changes I meant changes within TC's flow from inputs into the kernel more generally. Even with these plumbing changes, I'd see there being a type-system difference between...
In the following signature: ``` def avgpool(float(N) input, int K, int(2) shape) -> (output) { ``` There's very little distinction between the values of N, K, and shape(0) in the...
Note that it's straightforward to take a set of Halide Exprs in terms of some tensors and scalar parameters and jit-compile some code that evaluates them, either using our own...
Duplicate of #9 ?
I started working on this a while ago, but got stuck for the reason described in there.