[py] Fix loop invariant attribute. #2767
Conversation
boschmitt
force-pushed
the
fix_py_for_loop
branch
from
56a2a4a to
1c0d1cb
Compare
|
CUDA Quantum Docs Bot: A preview of the documentation can be found here. |
| bodyBuilder, | ||
| startVal=startVal, | ||
| stepVal=stepVal, | ||
| isDecrementing=isDecrementing) |
There was a problem hiding this comment.
Something like this must be normalized. Is the cc-loop-normalize pass being run?
There was a problem hiding this comment.
Hard to know. The pass does get run in various places but I'm not sure if it cover all paths used to execute Python kernels. Is this pass a requirement in all cases or only when doing other loop transformations ?
There was a problem hiding this comment.
Normalization transforms the iteration space of a loop to a counted or invariant form.
More precisely, if we have a loop like
for (i = start_expr; comparison(i, stop_expr); i = step_expr(i)) ....this loop in normalized form is more like
for (i' = 0; i' < new_uppser_bound_expr; i' = i' + 1) {
i = some_expr(i');
....
} Once in this form, the loop is much more amenable to loop transformations such as unrolling.
There was a problem hiding this comment.
So it is not a requirement for loops that are already normalized, but if Python is creating a loop that descends or uses a step other than 1, the loop must be normalized.
Normalization ought to be part of the new "classical optimizations" pipeline by default.
boschmitt
force-pushed
the
fix_py_for_loop
branch
from
1c0d1cb to
1de30f8
Compare
Python's for-loops are not guaranteed to be invariant. The AST bridge would need to do further analysis to guarantee the property. This PR does not add such an analysis, but fixes the places where the bridge might wrongly label a for-loop invariant when, in reality, it might not be. Signed-off-by: boschmitt <[email protected]>
boschmitt
force-pushed
the
fix_py_for_loop
branch
from
1de30f8 to
5cd7492
Compare
|
CUDA Quantum Docs Bot: A preview of the documentation can be found here. |
Description
Python's for-loops are not guaranteed to be invariant. The AST bridge would need to do further analysis to guarantee the property. This PR does not add such an analysis, but fixes the places where the bridge might wrongly label a for-loop invariant when, in reality, it might not be.