-
Notifications
You must be signed in to change notification settings - Fork 14.4k
[mlir][Affine] Extend linearize/delinearize cancelation to partial tails #116872
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Conversation
xisting patterns would cancel out the linearize_index / delinearize_index pairs that had the exact same basis, like %0 = affine.linearize_index [%w, %x, %y, %z] by (X, Y, Z) : index %1:4 = affine.delinearize_index %0 into (W, X, Y, Z) : index, ... This commit extends the canonicalization to handle instances where the entire basis doesn't match, as in %0 = affine.linearize_index [%w, %x, %y, %z] by (X, Y, Z) : index %1:3 = affine.delinearize_index %0 into (XY, Y, Z) : index, ... where we can replace the last two results of the delinearize_index operation with the last two inputs of the linearize_index, creating a more canonical (fewer total computations to perform) result.
@llvm/pr-subscribers-mlir-affine @llvm/pr-subscribers-mlir Author: Krzysztof Drewniak (krzysz00) Changesxisting patterns would cancel out the linearize_index / delinearize_index pairs that had the exact same basis, like
This commit extends the canonicalization to handle instances where the entire basis doesn't match, as in
where we can replace the last two results of the delinearize_index operation with the last two inputs of the linearize_index, creating a more canonical (fewer total computations to perform) result. Full diff: https://github.com/llvm/llvm-project/pull/116872.diff 2 Files Affected:
diff --git a/mlir/lib/Dialect/Affine/IR/AffineOps.cpp b/mlir/lib/Dialect/Affine/IR/AffineOps.cpp
index 4cf07bc167eab9..67d7da622a3550 100644
--- a/mlir/lib/Dialect/Affine/IR/AffineOps.cpp
+++ b/mlir/lib/Dialect/Affine/IR/AffineOps.cpp
@@ -4666,14 +4666,16 @@ struct DropUnitExtentBasis
};
/// If a `affine.delinearize_index`'s input is a `affine.linearize_index
-/// disjoint` and the two operations have the same basis, replace the
-/// delinearizeation results with the inputs of the `affine.linearize_index`
-/// since they are exact inverses of each other.
+/// disjoint` and the two operations end with the same basis elements,
+/// cancel those parts of the operations out because they are inverses
+/// of each other.
+///
+/// If the operations have the same basis, cancel them entirely.
///
/// The `disjoint` flag is needed on the `affine.linearize_index` because
/// otherwise, there is no guarantee that the inputs to the linearization are
/// in-bounds the way the outputs of the delinearization would be.
-struct CancelDelinearizeOfLinearizeDisjointExact
+struct CancelDelinearizeOfLinearizeDisjointExactTail
: public OpRewritePattern<affine::AffineDelinearizeIndexOp> {
using OpRewritePattern::OpRewritePattern;
@@ -4685,12 +4687,45 @@ struct CancelDelinearizeOfLinearizeDisjointExact
return rewriter.notifyMatchFailure(delinearizeOp,
"index doesn't come from linearize");
- if (!linearizeOp.getDisjoint() ||
- linearizeOp.getEffectiveBasis() != delinearizeOp.getEffectiveBasis())
+ if (!linearizeOp.getDisjoint())
+ return rewriter.notifyMatchFailure(linearizeOp, "not disjoint");
+
+ ValueRange linearizeIns = linearizeOp.getMultiIndex();
+ // Note: we use the full basis so we don't lose outer bounds later.
+ SmallVector<OpFoldResult> linearizeBasis = linearizeOp.getMixedBasis();
+ SmallVector<OpFoldResult> delinearizeBasis = delinearizeOp.getMixedBasis();
+ size_t numMatches = 0;
+ for (auto [linSize, delinSize] : llvm::zip(
+ llvm::reverse(linearizeBasis), llvm::reverse(delinearizeBasis))) {
+ if (linSize != delinSize)
+ break;
+ ++numMatches;
+ }
+
+ if (numMatches == 0)
return rewriter.notifyMatchFailure(
- linearizeOp, "not disjoint or basis doesn't match delinearize");
+ delinearizeOp, "final basis element doesn't match linearize");
+
+ // The easy case: everything lines up and the basis match sup completely.
+ if (numMatches == linearizeBasis.size() &&
+ numMatches == delinearizeBasis.size() &&
+ linearizeIns.size() == delinearizeOp.getNumResults()) {
+ rewriter.replaceOp(delinearizeOp, linearizeOp.getMultiIndex());
+ return success();
+ }
- rewriter.replaceOp(delinearizeOp, linearizeOp.getMultiIndex());
+ Value newLinearize = rewriter.create<affine::AffineLinearizeIndexOp>(
+ linearizeOp.getLoc(), linearizeIns.drop_back(numMatches),
+ ArrayRef<OpFoldResult>{linearizeBasis}.drop_back(numMatches),
+ linearizeOp.getDisjoint());
+ auto newDelinearize = rewriter.create<affine::AffineDelinearizeIndexOp>(
+ delinearizeOp.getLoc(), newLinearize,
+ ArrayRef<OpFoldResult>{delinearizeBasis}.drop_back(numMatches),
+ delinearizeOp.hasOuterBound());
+ SmallVector<Value> mergedResults(newDelinearize.getResults());
+ mergedResults.append(linearizeIns.take_back(numMatches).begin(),
+ linearizeIns.take_back(numMatches).end());
+ rewriter.replaceOp(delinearizeOp, mergedResults);
return success();
}
};
@@ -4698,9 +4733,8 @@ struct CancelDelinearizeOfLinearizeDisjointExact
void affine::AffineDelinearizeIndexOp::getCanonicalizationPatterns(
RewritePatternSet &patterns, MLIRContext *context) {
- patterns
- .insert<CancelDelinearizeOfLinearizeDisjointExact, DropUnitExtentBasis>(
- context);
+ patterns.insert<CancelDelinearizeOfLinearizeDisjointExactTail,
+ DropUnitExtentBasis>(context);
}
//===----------------------------------------------------------------------===//
diff --git a/mlir/test/Dialect/Affine/canonicalize.mlir b/mlir/test/Dialect/Affine/canonicalize.mlir
index b54a13cffe7771..5384977151b47f 100644
--- a/mlir/test/Dialect/Affine/canonicalize.mlir
+++ b/mlir/test/Dialect/Affine/canonicalize.mlir
@@ -1739,6 +1739,24 @@ func.func @cancel_delinearize_linearize_disjoint_delinearize_extra_bound(%arg0:
// -----
+// CHECK-LABEL: func @cancel_delinearize_linearize_disjoint_partial(
+// CHECK-SAME: %[[ARG0:[a-zA-Z0-9]+]]: index,
+// CHECK-SAME: %[[ARG1:[a-zA-Z0-9]+]]: index,
+// CHECK-SAME: %[[ARG2:[a-zA-Z0-9]+]]: index,
+// CHECK-SAME: %[[ARG3:[a-zA-Z0-9]+]]: index,
+// CHECK-SAME: %[[ARG4:[a-zA-Z0-9]+]]: index)
+// CHECK: %[[LIN:.+]] = affine.linearize_index disjoint [%[[ARG0]], %[[ARG1]]] by (%[[ARG3]], 4) : index
+// CHECK: %[[DELIN:.+]]:2 = affine.delinearize_index %[[LIN]] into (8) : index, index
+// CHECK: return %[[DELIN]]#0, %[[DELIN]]#1, %[[ARG2]]
+func.func @cancel_delinearize_linearize_disjoint_partial(%arg0: index, %arg1: index, %arg2: index, %arg3: index, %arg4: index) -> (index, index, index) {
+ %0 = affine.linearize_index disjoint [%arg0, %arg1, %arg2] by (%arg3, 4, %arg4) : index
+ %1:3 = affine.delinearize_index %0 into (8, %arg4)
+ : index, index, index
+ return %1#0, %1#1, %1#2 : index, index, index
+}
+
+// -----
+
// Without `disjoint`, the cancelation isn't guaranteed to be the identity.
// CHECK-LABEL: func @no_cancel_delinearize_linearize_exact(
// CHECK-SAME: %[[ARG0:[a-zA-Z0-9]+]]: index,
|
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Cool! THanks! I think this better demonstrates the power of these operations....
xisting patterns would cancel out the linearize_index / delinearize_index pairs that had the exact same basis, like
This commit extends the canonicalization to handle instances where the entire basis doesn't match, as in
where we can replace the last two results of the delinearize_index operation with the last two inputs of the linearize_index, creating a more canonical (fewer total computations to perform) result.