This is an archive of the discontinued LLVM Phabricator instance.

Differential D110919

[mlir] Enable loop peeling for "reduction" dimensions of tiled_loop.
ClosedPublic

Authored by pifon2a on Oct 1 2021, 4:03 AM.

Details

Reviewers
tpopp
springerm
nicolasvasilache
Commits
rG693c61b2e050: [mlir] Enable loop peeling for "reduction" dimensions of tiled_loop.

Diff Detail

Event Timeline

pifon2a created this revision.Oct 1 2021, 4:03 AM
Herald added subscribers: wenzhicui, wrengr, Chia-hungDuan and 20 others. · View Herald TranscriptOct 1 2021, 4:03 AM
pifon2a requested review of this revision.Oct 1 2021, 4:03 AM
Herald added a reviewer: nicolasvasilache. · View Herald TranscriptOct 1 2021, 4:03 AM
Herald added a project: Restricted Project. · View Herald Transcript
Herald added subscribers: stephenneuendorffer, nicolasvasilache. · View Herald Transcript
Harbormaster completed remote builds in B126700: Diff 376473.Oct 1 2021, 4:03 AM
tpopp accepted this revision.Oct 1 2021, 4:07 AM
This revision is now accepted and ready to land.Oct 1 2021, 4:07 AM
Closed by commit rG693c61b2e050: [mlir] Enable loop peeling for "reduction" dimensions of tiled_loop. (authored by pifon2a). · Explain WhyOct 1 2021, 4:08 AM
This revision was automatically updated to reflect the committed changes.
pifon2a added a commit: rG693c61b2e050: [mlir] Enable loop peeling for "reduction" dimensions of tiled_loop..

Revision Contents

PathSize
mlir/
lib/
Dialect/
Linalg/
Transforms/
3 lines

Diff 376473

mlir/lib/Dialect/Linalg/Transforms/Loops.cpp

Show First 20 Lines • Show All 719 Lines • ▼ Show 20 Lines
LogicalResult mlir::linalg::peelAndCanonicalizeTiledLoop(RewriterBase &rewriter, LogicalResult mlir::linalg::peelAndCanonicalizeTiledLoop(RewriterBase &rewriter,
TiledLoopOp loopOp, TiledLoopOp loopOp,
int64_t idx, int64_t idx,
TiledLoopOp &result) { TiledLoopOp &result) {
int64_t numLoops = loopOp.iterator_types().size(); int64_t numLoops = loopOp.iterator_types().size();
if (idx < 0 || numLoops <= idx) if (idx < 0 || numLoops <= idx)
return failure(); return failure();
// Only parallel iterator supported.
if (!isParallelIterator(loopOp.iterator_types()[idx]))
return failure();
Value ub = loopOp.upperBound()[idx]; Value ub = loopOp.upperBound()[idx];
TiledLoopOp remainderLoop; TiledLoopOp remainderLoop;
Value splitBound; Value splitBound;
if (failed(peelTiledLoop(rewriter, loopOp, idx, remainderLoop, splitBound))) if (failed(peelTiledLoop(rewriter, loopOp, idx, remainderLoop, splitBound)))
return failure(); return failure();
// Rewrite affine.min and affine.max ops. // Rewrite affine.min and affine.max ops.
▲ Show 20 Lines • Show All 54 Lines • Show Last 20 Lines