Introduce Pulse IR skeleton #11767
Introduce Pulse IR skeleton #11767
Conversation
|
One or more of the the following people are requested to review this:
|
nkanazawa1989
added
experimental
nkanazawa1989
changed the title
There was a problem hiding this comment.
Thanks @TsafrirA this is great starting point of discussion. I like the use of composite pattern. This really fits in with our pulse model. However, I prefer simpler implementation and want to delegate all complexity to passes unless the method is used by multiple passes (so I prefer starting with minimum implementation and add methods later if we find it's necessary).
| pass | ||
|
|
||
| @abstractmethod | ||
| def shift_initial_time(self, value: int): |
There was a problem hiding this comment.
I think this must be implemented by a pass.
There was a problem hiding this comment.
Let's leave it in at the moment, and remove it later if we decide it's better to have it as a pass. I think during scheduling this is something we might have to do often, so it might make sense to have it as a method.
| @property | ||
| def initial_time(self) -> int | None: | ||
| """Return the initial time ``initial_time`` of the object""" | ||
| elements_initial_times = [element.initial_time for element in self._elements] |
There was a problem hiding this comment.
This is O(N) lookup and expensive. I'd prefer turning the data structure into (rustworkx based) DAG and then return the first node to calculate the initial time. Alternatively we can use dict keyed on the initial time, but this doesn't work because None shouldn't be a key.
| @property | ||
| def final_time(self) -> int | None: | ||
| """Return the final time of the ``IrBlock``object""" | ||
| elements_final_times = [element.final_time for element in self._elements] |
There was a problem hiding this comment.
With the same reason I prefer DAG.
| """Return the length of the IR, defined as the number of elements in it""" | ||
| return len(self._elements) | ||
|
|
||
| def __repr__(self) -> str: |
There was a problem hiding this comment.
In my experience repr for sequence (e.g. ScheduleBlock) is not really useful, because it's super lengthy. If we prefer LLVM IR like, probably we can support .dump method that generates parsable code. This also helps debugging (unittest); i.e. the reference is always text, which is very robust.
There was a problem hiding this comment.
(note that I don't suggest implementing the code in this PR, since you need to start from defining AST...)
|
Thanks @nkanazawa1989 for the quick review. With the exception of About |
Summary
This PR continues the work on the new Pulse Compiler and IR (RFC) by introducing a skeleton for the IR representation.
Details and comments
The implementation follows the form first presented in #11234. Only basic functionalities were introduced at this point, because #11726 is still pending, and the work on the compiler (following #11743) will better reveal what functionality is missing.