[mlir][Linalg] implement bufferization for linalg.pack
#177982
Merged
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Add a BufferizableOpInterface implementation for linalg.pack now that pack supports memref semantics llvm@4b066c7. This completes the op’s bufferization path and avoids copy-before-write for destination operands. Signed-off-by: Ryutaro Okada <[email protected]>
sakupan102
requested review from
dcaballe,
nicolasvasilache and
rengolin
as code owners
Member
|
@llvm/pr-subscribers-mlir-linalg Author: Ryutaro Okada (sakupan102) ChangesAdd a BufferizableOpInterface implementation for linalg.pack now that pack supports memref semantics 4b066c7. This completes the op’s bufferization path and avoids copy-before-write for destination operands. Full diff: https://github.com/llvm/llvm-project/pull/177982.diff 2 Files Affected:
diff --git a/mlir/lib/Dialect/Linalg/Transforms/BufferizableOpInterfaceImpl.cpp b/mlir/lib/Dialect/Linalg/Transforms/BufferizableOpInterfaceImpl.cpp
index 3512ecd9d2eb2..60c685578682a 100644
--- a/mlir/lib/Dialect/Linalg/Transforms/BufferizableOpInterfaceImpl.cpp
+++ b/mlir/lib/Dialect/Linalg/Transforms/BufferizableOpInterfaceImpl.cpp
@@ -191,6 +191,47 @@ struct SoftmaxOpInterface
return success();
}
};
+
+struct PackOpInterface
+ : public DstBufferizableOpInterfaceExternalModel<PackOpInterface,
+ linalg::PackOp> {
+ bool bufferizesToMemoryRead(Operation *op, OpOperand &opOperand,
+ const AnalysisState &state) const {
+ auto packOp = cast<linalg::PackOp>(op);
+ return !packOp.isDpsInit(&opOperand);
+ }
+
+ LogicalResult bufferize(Operation *op, RewriterBase &rewriter,
+ const BufferizationOptions &options,
+ BufferizationState &state) const {
+ auto packOp = cast<linalg::PackOp>(op);
+ if (packOp.hasPureBufferSemantics())
+ return success();
+ if (!packOp.hasPureTensorSemantics())
+ return packOp.emitError() << "op does not have pure tensor semantics";
+
+ FailureOr<Value> sourceBuffer =
+ getBuffer(rewriter, packOp.getSource(), options, state);
+ if (failed(sourceBuffer))
+ return failure();
+ FailureOr<Value> destBuffer =
+ getBuffer(rewriter, packOp.getDest(), options, state);
+ if (failed(destBuffer))
+ return failure();
+
+ SmallVector<Value> operands;
+ operands.push_back(*sourceBuffer);
+ operands.push_back(*destBuffer);
+ if (auto val = packOp.getPaddingValue())
+ operands.push_back(val);
+ llvm::append_range(operands, packOp.getInnerTiles());
+
+ linalg::PackOp::create(rewriter, packOp.getLoc(), TypeRange{}, operands,
+ op->getAttrs());
+ replaceOpWithBufferizedValues(rewriter, op, *destBuffer);
+ return success();
+ }
+};
} // namespace
void mlir::linalg::registerBufferizableOpInterfaceExternalModels(
@@ -206,5 +247,6 @@ void mlir::linalg::registerBufferizableOpInterfaceExternalModels(
>::registerOpInterface(ctx);
SoftmaxOp::attachInterface<SoftmaxOpInterface>(*ctx);
+ PackOp::attachInterface<PackOpInterface>(*ctx);
});
}
diff --git a/mlir/test/Dialect/Linalg/bufferize.mlir b/mlir/test/Dialect/Linalg/bufferize.mlir
index 1c6cb88fa028b..2cb09c39b5776 100644
--- a/mlir/test/Dialect/Linalg/bufferize.mlir
+++ b/mlir/test/Dialect/Linalg/bufferize.mlir
@@ -206,3 +206,20 @@ func.func @bufferize_softmax(%arg0: tensor<2x16x32xf32>, %arg1: tensor<2x16x32xf
outs(%arg1: tensor<2x16x32xf32>) -> tensor<2x16x32xf32>
return %1 : tensor<2x16x32xf32>
}
+
+// -----
+
+// CHECK-LABEL: func @bufferize_pack(
+// CHECK-SAME: %[[SRC:.*]]: tensor<128x256xf32>, %[[DST:.*]]: tensor<16x8x8x32xf32>) -> tensor<16x8x8x32xf32> {
+// CHECK-DAG: %[[SRC_BUF:.*]] = bufferization.to_buffer %[[SRC]] : tensor<128x256xf32> to memref<128x256xf32>
+// CHECK-DAG: %[[DST_BUF:.*]] = memref.alloc() {{.*}} : memref<16x8x8x32xf32>
+// CHECK-NOT: memref.copy
+// CHECK: linalg.pack %[[SRC_BUF]] inner_dims_pos = [0, 1] inner_tiles = [8, 32] into %[[DST_BUF]] : memref<128x256xf32> -> memref<16x8x8x32xf32>
+// CHECK: %[[RESULT:.*]] = bufferization.to_tensor %[[DST_BUF]] : memref<16x8x8x32xf32> to tensor<16x8x8x32xf32>
+// CHECK: return %[[RESULT]] : tensor<16x8x8x32xf32>
+func.func @bufferize_pack(%source: tensor<128x256xf32>, %dest: tensor<16x8x8x32xf32>) -> tensor<16x8x8x32xf32> {
+ %0 = linalg.pack %source inner_dims_pos = [0, 1] inner_tiles = [8, 32]
+ into %dest : tensor<128x256xf32> -> tensor<16x8x8x32xf32>
+ return %0 : tensor<16x8x8x32xf32>
+}
+
|
Member
|
@llvm/pr-subscribers-mlir Author: Ryutaro Okada (sakupan102) ChangesAdd a BufferizableOpInterface implementation for linalg.pack now that pack supports memref semantics 4b066c7. This completes the op’s bufferization path and avoids copy-before-write for destination operands. Full diff: https://github.com/llvm/llvm-project/pull/177982.diff 2 Files Affected:
diff --git a/mlir/lib/Dialect/Linalg/Transforms/BufferizableOpInterfaceImpl.cpp b/mlir/lib/Dialect/Linalg/Transforms/BufferizableOpInterfaceImpl.cpp
index 3512ecd9d2eb2..60c685578682a 100644
--- a/mlir/lib/Dialect/Linalg/Transforms/BufferizableOpInterfaceImpl.cpp
+++ b/mlir/lib/Dialect/Linalg/Transforms/BufferizableOpInterfaceImpl.cpp
@@ -191,6 +191,47 @@ struct SoftmaxOpInterface
return success();
}
};
+
+struct PackOpInterface
+ : public DstBufferizableOpInterfaceExternalModel<PackOpInterface,
+ linalg::PackOp> {
+ bool bufferizesToMemoryRead(Operation *op, OpOperand &opOperand,
+ const AnalysisState &state) const {
+ auto packOp = cast<linalg::PackOp>(op);
+ return !packOp.isDpsInit(&opOperand);
+ }
+
+ LogicalResult bufferize(Operation *op, RewriterBase &rewriter,
+ const BufferizationOptions &options,
+ BufferizationState &state) const {
+ auto packOp = cast<linalg::PackOp>(op);
+ if (packOp.hasPureBufferSemantics())
+ return success();
+ if (!packOp.hasPureTensorSemantics())
+ return packOp.emitError() << "op does not have pure tensor semantics";
+
+ FailureOr<Value> sourceBuffer =
+ getBuffer(rewriter, packOp.getSource(), options, state);
+ if (failed(sourceBuffer))
+ return failure();
+ FailureOr<Value> destBuffer =
+ getBuffer(rewriter, packOp.getDest(), options, state);
+ if (failed(destBuffer))
+ return failure();
+
+ SmallVector<Value> operands;
+ operands.push_back(*sourceBuffer);
+ operands.push_back(*destBuffer);
+ if (auto val = packOp.getPaddingValue())
+ operands.push_back(val);
+ llvm::append_range(operands, packOp.getInnerTiles());
+
+ linalg::PackOp::create(rewriter, packOp.getLoc(), TypeRange{}, operands,
+ op->getAttrs());
+ replaceOpWithBufferizedValues(rewriter, op, *destBuffer);
+ return success();
+ }
+};
} // namespace
void mlir::linalg::registerBufferizableOpInterfaceExternalModels(
@@ -206,5 +247,6 @@ void mlir::linalg::registerBufferizableOpInterfaceExternalModels(
>::registerOpInterface(ctx);
SoftmaxOp::attachInterface<SoftmaxOpInterface>(*ctx);
+ PackOp::attachInterface<PackOpInterface>(*ctx);
});
}
diff --git a/mlir/test/Dialect/Linalg/bufferize.mlir b/mlir/test/Dialect/Linalg/bufferize.mlir
index 1c6cb88fa028b..2cb09c39b5776 100644
--- a/mlir/test/Dialect/Linalg/bufferize.mlir
+++ b/mlir/test/Dialect/Linalg/bufferize.mlir
@@ -206,3 +206,20 @@ func.func @bufferize_softmax(%arg0: tensor<2x16x32xf32>, %arg1: tensor<2x16x32xf
outs(%arg1: tensor<2x16x32xf32>) -> tensor<2x16x32xf32>
return %1 : tensor<2x16x32xf32>
}
+
+// -----
+
+// CHECK-LABEL: func @bufferize_pack(
+// CHECK-SAME: %[[SRC:.*]]: tensor<128x256xf32>, %[[DST:.*]]: tensor<16x8x8x32xf32>) -> tensor<16x8x8x32xf32> {
+// CHECK-DAG: %[[SRC_BUF:.*]] = bufferization.to_buffer %[[SRC]] : tensor<128x256xf32> to memref<128x256xf32>
+// CHECK-DAG: %[[DST_BUF:.*]] = memref.alloc() {{.*}} : memref<16x8x8x32xf32>
+// CHECK-NOT: memref.copy
+// CHECK: linalg.pack %[[SRC_BUF]] inner_dims_pos = [0, 1] inner_tiles = [8, 32] into %[[DST_BUF]] : memref<128x256xf32> -> memref<16x8x8x32xf32>
+// CHECK: %[[RESULT:.*]] = bufferization.to_tensor %[[DST_BUF]] : memref<16x8x8x32xf32> to tensor<16x8x8x32xf32>
+// CHECK: return %[[RESULT]] : tensor<16x8x8x32xf32>
+func.func @bufferize_pack(%source: tensor<128x256xf32>, %dest: tensor<16x8x8x32xf32>) -> tensor<16x8x8x32xf32> {
+ %0 = linalg.pack %source inner_dims_pos = [0, 1] inner_tiles = [8, 32]
+ into %dest : tensor<128x256xf32> -> tensor<16x8x8x32xf32>
+ return %0 : tensor<16x8x8x32xf32>
+}
+
|
7 tasks
adam-smnk
reviewed
Jan 26, 2026
adam-smnk
approved these changes
Jan 26, 2026
Contributor
Author
|
If everything looks good, could you merge it? |
Signed-off-by: Ryutaro Okada <[email protected]>
sakupan102
force-pushed
the
bufferize-pack
branch
from
340a435 to
f9ea1db
Compare
Member
Sure but let’s give others a chance to have a look too. |
matthias-springer
approved these changes
Jan 27, 2026
mlir/lib/Dialect/Linalg/Transforms/BufferizableOpInterfaceImpl.cpp
Outdated
Show resolved
Hide resolved
mlir/lib/Dialect/Linalg/Transforms/BufferizableOpInterfaceImpl.cpp
Outdated
Show resolved
Hide resolved
Signed-off-by: Ryutaro Okada <[email protected]>
stomfaig
pushed a commit
to stomfaig/llvm-project
that referenced
this pull request
Jan 28, 2026
Add a BufferizableOpInterface implementation for linalg.pack now that pack supports memref semantics llvm@4b066c7. This completes the op’s bufferization path and avoids copy-before-write for destination operands. --------- Signed-off-by: Ryutaro Okada <[email protected]>
IanWood1
added a commit
to iree-org/iree
that referenced
this pull request
Jan 28, 2026
Reverts: - Carries local revert of llvm/llvm-project#169614 due to #22649. - Adds revert of llvm/llvm-project#177982. `reifyResultShapes()` is unimplemented for pack ops on memrefs causing a crash in [getPackUnPackIterationDomain](https://github.com/iree-org/llvm-project/blob/b24bd7161ca7eb6e9652c34b92e200ac16af3628/mlir/lib/Dialect/Linalg/Transforms/TilingInterfaceImpl.cpp#L718) Fixes: - Passes `LLVM::DIFlags::Zero` to address the added argument for L`LVM::DIDerivedTypeAttr::get()` [llvm/llvm-project#177889](llvm/llvm-project#177889) - Removes he `firstIndex` parameter to address the API change to `visitNonControlFlowArguments()` [llvm/llvm-project#175210](llvm/llvm-project#175210) https://github.com/iree-org/llvm-project/tree/sm-iree-integrates/llvm-20260128 Signed-off-by: Ian Wood <[email protected]>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
4 participants
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Add a BufferizableOpInterface implementation for linalg.pack now that pack supports memref semantics 4b066c7. This completes the op’s bufferization path and avoids copy-before-write for destination operands.