[Polly] Data flow reduction detection to cover more cases #84901
Conversation
The base concept is same as existing reduction algorithm where we get
the list of candidate pairs <store,load>. But the existing algorithm
works only if there is single binary operation between the load and
store.
Example sum += a[i];
This algorithm extends to work with more than single binary operation
as well. It is implemented using data flow reduction detection on
basic block level. We propagate the loads, the number of times the
load is used(flows into instruction) and binary operation performed
until we reach a store.
Example sum += a[i] + b[i];
sum(Ld) a[i](Ld)
\ + /
tmp b[i](Ld)
\+/
sum(St)
In the above case the candidate pairs are formed by associating sum
with all of its load inputs which are sum, a[i] and b[i]. Then check
functions are used to filter a valid reduction pair ie {sum,sum}.
There was a problem hiding this comment.
Thanks for the patch. I really appreciate the good code comments.
Could you add a test that checks two independent reductions in the same loop, e.g.:
for (int i = 0; i < n; ++i) {
*sum += A[i];
*prod += B[i];
}
ScopBuilder::buildEqivClassBlockStmts should create two ScopStmts for the same BasicBlock here.
The base concept is same as existing reduction algorithm where we get
the list of candidate pairs <store,load>. But the existing algorithm
works only if there is single binary operation between the load and
store.
Example sum += a[i];
This algorithm extends to work with more than single binary operation
as well. It is implemented using data flow reduction detection on
basic block level. We propagate the loads, the number of times the
load is used(flows into instruction) and binary operation performed
until we reach a store.
Example sum += a[i] + b[i];
sum(Ld) a[i](Ld)
\ + /
tmp b[i](Ld)
\+/
sum(St)
In the above case the candidate pairs are formed by associating sum
with all of its load inputs which are sum, a[i] and b[i]. Then check
functions are used to filter a valid reduction pair ie {sum,sum}.
|
Ping @Meinersbur @efriedma-quic |
|
Ping @Meinersbur |
1 similar comment
|
Ping @Meinersbur |
| @@ -0,0 +1,39 @@ | |||
| ; RUN: opt %loadPolly -basic-aa -polly-print-dependences -polly-allow-nonaffine -disable-output < %s | FileCheck %s | |||
There was a problem hiding this comment.
-passes=print<polly-dependences> (see #92918). I think legacy-pm -polly-print-scops wasn't ported yet, so it should be okay to continue using it for now.
Otherwise, patch looks fine.
There was a problem hiding this comment.
Please don't forget to update this.
There was a problem hiding this comment.
Thanks for pointing out Eli. Updated.
polly/lib/Analysis/ScopBuilder.cpp
Outdated
| @@ -2557,7 +2568,6 @@ bool checkCandidatePairAccesses(MemoryAccess *LoadMA, MemoryAccess *StoreMA, | |||
| .intersect_domain(isl::manage(Domain.copy())); | |||
| isl::set RS = R.range(); | |||
| isl::set WS = W.range(); | |||
|
|
|||
There was a problem hiding this comment.
Please try to avoid unrelated changes.
Co-authored-by: Michael Kruse <[email protected]>
|
✅ With the latest revision this PR passed the C/C++ code formatter. |
|
Thanks @Meinersbur and @efriedma-quic for your time to review and approvals. |
The base concept is same as existing reduction algorithm where we get the list of candidate pairs <store,load>. But the existing algorithm works only if there is single binary operation between the load and store.
Example sum += a[i];
This algorithm extends to work with more than single binary operation as well. It is implemented using data flow reduction detection on basic block level. We propagate the loads, the number of times the load is used(flows into instruction) and binary operation performed until we reach a store.
Example sum += a[i] + b[i];
In the above case the candidate pairs are formed by associating sum with all of its load inputs which are sum, a[i] and b[i]. Then check functions are used to filter a valid reduction pair ie {sum,sum}.