Un-prioritized list of things that generally should be done:

• Support inter-block reductions (similar template function approach)
• Reduction to a scalar does not work as we no longer have a tensor axis. Need to figure out how to fix this. Likely want to implement a zero-dim tensor which is just a scalar (this is how PyTorch does it).
• Fusion printer that only prints math exprs from outputs. Rework the ir_printer class.
• SetRandom on fusion is unlikely necessary, lets see if we can pull this out of so much of the logic in the codebase.
• Remove TensorView Reorder code, use tensor domain.
• Cross thread reduction, predicate blocks of code not using threads (i.e. downstream of reduction)
• Remove active view from lower2device
• Move logic out of lower2device, so lower2device is just a wrapper around lowering passes
• Reduction op can only be run on tensor views, we should restrict the IR node to TensorView input/output.
• Rework predicates, per loop, include thread guards (if thread dim doesn't participate, predicate out on threads (i.e. threadIdx.y>0). This can be done at the highest for-loop that doesn’t use that thread dim.
• Remove predicate logic (besides that required for unrolling) out of unrolling.
• Get an external compilation working with torchlib like we do with test_gpu. I'd like to be able to create tutorials that can be individually compiled and run.
• Cleanup decltype in loops in favor of size_t
• Work through graph viz to update usage for TensorDomain::rootDomain() TensorDomain::rfactorDomain() and TensorDomain::domain() #1396
• Explicitly deleting just the copy operations is sufficient. See: https://abseil.io/tips/143 Transform replay refactor #53 (comment)
• Work back through graph viz and maybe add labels to the Domains we have/want.
• Change split semantics to directly take a parallel type. Maybe call splitOnThreadDim.
• Make threadDim and threadIndices special values we don't keep re-generating. Maybe attach to the fusion?