There is currently no lowering out of `ml_program` in the LLVM
repository. This change adds a lowering to `memref` so that it can be
lowered all the way to LLVM. This lowering was taken from the [reference
backend in
torch-mlir](f416953600
).
I had tried implementing the `BufferizableOpInterface` for `ml_program`
instead of adding a new pass but that did not work because
`OneShotBufferize` does not visit module-level ops like
`ml_program.global`.
Added pass optimizes MLProgram global operations by reducing to only
the minimal load/store operations for global tensors. This avoids
unnecessary global operations throughout a program and potentially
improves operation gusion.
Reviewed By: jpienaar
Differential Revision: https://reviews.llvm.org/D159228
* Split ops into X_graph variants as discussed;
* Remove tokens from non-Graph region variants and rely on side-effect
modelling there while removing side-effect modelling from Graph
variants and relying on explicit ordering there;
* Make tokens required to be produced by Graph variants - but kept
explicit token type specification given previous discussion on this
potentially being configurable in future;
This results in duplicating some code. I considered adding helper
functions but decided against adding an abstraction there early given
size of duplication and creating accidental coupling.
Differential Revision: https://reviews.llvm.org/D127813
The approach I took was to define a dialect 'extern' attribute that a GlobalOp can take as a value to signify external linkage. I think this approach should compose well and should also work with wherever the OpaqueElements work goes in the future (since that is just another kind of attribute). I special cased the GlobalOp parser/printer for this case because it is significantly easier on the eyes.
In the discussion, Jeff Niu had proposed an alternative syntax for GlobalOp that I ended up not taking. I did try to implement it but a) I don't think it made anything easier to read in the common case, and b) it made the parsing/printing logic a lot more complicated (I think I would need a completely custom parser/printer to do it well). Please have a look at the common cases where the global type and initial value type match: I don't think how I have it is too bad. The less common cases seem ok to me.
I chose to only implement the direct, constant load op since that is non side effecting and there was still discussion pending on that.
Differential Revision: https://reviews.llvm.org/D124318
