Oliver Rausch

Custom scale transformations

1

I'd like to be able to make a custom scale. My particular use-case is transforming unit magnitudes, for example, by having a scale that multiplies values by 1e6. In plotnine,...

Lower subarray patterns to new grid mapped array nodes

1

Previously they lowered to mpi.BlockScatter and mpi.BlockGather from dace.

Add grid-mapped-array communication nodes

1

These nodes abstract mapping an array onto a process grid. They fill a similar role to BlockScatter and BlockGather in dace, but with a more array-programming focused API. This means...

Add DistributedMemlet node and scheduling function

1

This change adds the DistributedMemlet library node and the scheduling function for distributed computation. This allows you to distribute the work in the top-level map of the SDFG by specifying...

Support nested SDFGs in distributed lowering

1

This allows us to support reductions with their intialization states. The idea is that nested SDFG are required to be schedule such that there is no communication within them. The...

Since Im2col has a ONNX node within a map, the schedule of the onnx node is inferred to be Sequential (since the schedule of the outer map is Sequential). Then...

Both the ORT expansion and the pure expansions ignore the input/output subsets and always operate on the full array specified by the memlet. For the pure expansions, fixing this should...

Something like ```python C[:] = ONNXAdd(A, B) ``` instead of ```python ONNXAdd(A=A, B=B, C=C) ``` This would require renaming the current ONNX nodes to something else, maybe like ONNXAddNode. (so...