daisytuner / docc / 26433735230

push

github

Activate Einsum pipeline for Python/PyTorch frontend (#720)

- Added Einsum step to compilation (after which the SDFG is also dumped)
- Generate tasklets and cmath library nodes directly instead of tensor nodes if all tensor types are scalar
- Merged the whole "lifting" part of Einsum nodes into one pass: the EinsumDetectionPass that runs in linear time
- Renamed the EinsumExpand transformation to the EinsumPromotion transformation because the name was confusing with expanding Einsum/library nodes
- Fixed a bug where the lifting failed for multiple input edges of the same access node on a tasklet and added unit tests
- Moved Einsum node source files into tensor folder (but the Einsum node is not a tensor node yet)

85 of 115 new or added lines in 13 files covered. (73.91%)

35027 of 57560 relevant lines covered (60.85%)

11105.41 hits per line