Adaptive profile floating-point computations extension #2078
Conversation
idavis
force-pushed
the
iadavis/adaptive-float
branch
from
71011e3 to
bcfe36d
Compare
idavis
requested review from
billti,
sezna,
swernli and
minestarks
as code owners
pip/src/interpreter.rs
Outdated
| /// | ||
| /// This profile includes all of the required Adaptive Profile and Adaptive_RI | ||
| /// capabilities, as well as the optional floating point computation and qubit | ||
| /// reset capabilities, as defined by the QIR specification. |
There was a problem hiding this comment.
As prior comment, including Adaptive_RI already includes reset capabilities.
There was a problem hiding this comment.
Reworded them. The adaptive profile requires qubit reset now, so I updated the adaptive RI docs as well.
| !8 = !{i32 1, !"classical_fixed_points", i1 false} | ||
| !9 = !{i32 1, !"user_functions", i1 false} | ||
| !10 = !{i32 1, !"multiple_target_branching", i1 false} | ||
| !4 = !{i32 1, !"int_computations", !"i64"} |
There was a problem hiding this comment.
Will this change in module flags metadata name from classical_ints to int_computations be a breaking change anywhere in our existing flow (including any service validation)? Hopefully Quantinuum can already handle either, as they helped write the spec, but maybe worth a check before changing.
| !8 = !{i32 1, !"classical_fixed_points", i1 false} | ||
| !9 = !{i32 1, !"user_functions", i1 false} | ||
| !10 = !{i32 1, !"multiple_target_branching", i1 false} | ||
| !4 = !{i32 1, !"int_computations", !"i64"} |
There was a problem hiding this comment.
Should we have a test case in this folder that includes an .ll file using the new capability? (Which maybe means a new sample)
Adds dynamic float support implementing the floating-point computations for the AdaptiveRIF profile. This also implements
fcmpmissing from the spec which was filed as a spec bug.The
fcmpinstructions chosen were the ordered set. The spec says operations like dividing by0are undefined behavior and diving by0usually gives aNANvalue. For anyfcmpcalls we need to choose ordered or unordered. I'm not sure that it matters which we decide.This PR does not change the default profiles for the defined hardware targets.