Quality of Life Pass/Pipeline Features

[AlexanderViand-Intel]

Small fixes/add-ons that I noticed while drawing up a pass/pipeline overview figure:

• --secretize's entry-function parameter should be optional and, if omitted, the pass should apply to all functions in the input (module?). This should then also be reflected in any pipelines that use --secretize, making their entry-function optional, too.
• In order to better support mixed ctxt/ptxt examples, and for better integration with frontends, create --annotated-mlir-to-.... that are the same as the current --mlir-to-... pipelines but without the --secretize pass.
• The ciphertext-degree parameter in the various passes & pipelines that use it should be optional, with the compiler using the size of (the right-most-non-one-dimension of) the first secret tensor in the function signature if the parameter is omitted (and erroring out if no such secret tensor function parameter exists)
  See: Encode (R)LWE Parameters into the IR #1197
• --optimize-relinearization should be added to the RLWE pipeline builder? (EDIT: Not until --optimize-relinearization crashing #1148, and ideally also Generalize OptimizeRelinearization to support schemes besides BGV #1032 are fixed)
• --operation-balancer shuld be added to the mlir-to-secret-arithmetic pipeline
• Investigate if the --bgv-to-openfhe and --ckks-to-openfhe passes can be merged into a single --rlwe-to-openfhe pass combining the patterns from both passes.
  Also: Unify generic "convert binary op" patterns #1150
• (Optional/TBD): Split the library-specific parts of the pipeline into separate pipelines. While that might require doing something like --mlir-to-ckks --rlwe-to-openfhe, it might be nicer to keep backend-specific stuff separate/modular.
• Better error messages ( Improve Error/Warning Messages in HEIR through more meaningful Op "location" #1143)

[AlexanderViand-Intel]

Based on discussion from today's office hours:

For the ciphertext-degree parameter, we should decouple that into the actual ctxt degree (which can be optional and default to something reasoanble such as 16k) and something like --assume-native-tensor-size that does what was mentioned above.

In general, these things should not be just stored in a single pass's paramaters but encoded into the IR (at the module level)

[ZenithalHourlyRate]

Also I'm wondering why --operation-balancer did not get into the mlir-to-secret-arithmetic pipeline, though its absense makes me happy when constructing deep circuit for debugging.

[AlexanderViand-Intel]

Also I'm wondering why --operation-balancer did not get into the mlir-to-secret-arithmetic pipeline, though its absense makes me happy when constructing deep circuit for debugging.

Added it to the list! Do you happen to know which tests you expect to break?

[ZenithalHourlyRate]

Do you happen to know which tests you expect to break?

No, haven't tried it.

[AlexanderViand-Intel]

With #1196 addressing the last "clean up" part of this and the ciphertext degree thing + error messages being quite significant work packages in their own right with their own issues to track them, I've set this to close when #1196 gets merged.