[mlir][tensor] Add support for tensor.unpack static shapes inference. #81702
Conversation
This file contains 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
The revision does not refactor the inferStaticShape for pack and unpack ops because they can diverge quickly. Because there are more dimensions can be infered (i.e., with inner_tile_sizes) if the pack op does not have padding value.
|
@llvm/pr-subscribers-mlir-tensor @llvm/pr-subscribers-mlir Author: Han-Chung Wang (hanhanW) ChangesThe revision does not refactor the inferStaticShape for pack and unpack ops because they can diverge quickly. Because there are more dimensions can be inferred (i.e., with inner_tile_sizes) if the pack op does not have padding value. This is a follow-up of #80848 Full diff: https://github.com/llvm/llvm-project/pull/81702.diff 2 Files Affected:
diff --git a/mlir/lib/Dialect/Tensor/IR/TensorOps.cpp b/mlir/lib/Dialect/Tensor/IR/TensorOps.cpp
index bb72cba96ad935..1df15c0372e6e2 100644
--- a/mlir/lib/Dialect/Tensor/IR/TensorOps.cpp
+++ b/mlir/lib/Dialect/Tensor/IR/TensorOps.cpp
@@ -4212,6 +4212,40 @@ UnPackOp UnPackOp::createTransposedClone(OpBuilder &b, Location loc,
metadata.outerDimsPerm);
}
+/// Returns true if the `srcShape` or `destShape` is different from the one in
+/// `op` and populates each with the inferred static shape.
+static bool inferStaticShape(UnPackOp op, SmallVectorImpl<int64_t> &srcShape,
+ SmallVectorImpl<int64_t> &destShape) {
+ bool changeNeeded = false;
+ srcShape.assign(op.getSourceType().getShape().begin(),
+ op.getSourceType().getShape().end());
+ destShape.assign(op.getDestType().getShape().begin(),
+ op.getDestType().getShape().end());
+ llvm::SmallSetVector<int64_t, 4> innerDims;
+ innerDims.insert(op.getInnerDimsPos().begin(), op.getInnerDimsPos().end());
+ auto outerDimsPerm = op.getOuterDimsPerm();
+ int destRank = op.getDestRank();
+ for (auto i : llvm::seq<int64_t>(0, destRank)) {
+ if (innerDims.contains(i))
+ continue;
+ int64_t srcPos = i;
+ int64_t destPos = i;
+ if (!outerDimsPerm.empty())
+ srcPos = outerDimsPerm[destPos];
+ if (ShapedType::isDynamic(srcShape[srcPos]) ==
+ ShapedType::isDynamic(destShape[destPos])) {
+ continue;
+ }
+ int64_t size = srcShape[srcPos];
+ if (ShapedType::isDynamic(size))
+ size = destShape[destPos];
+ srcShape[srcPos] = size;
+ destShape[destPos] = size;
+ changeNeeded = true;
+ }
+ return changeNeeded;
+}
+
LogicalResult UnPackOp::canonicalize(UnPackOp unPackOp,
PatternRewriter &rewriter) {
/// pack(unpack(x)) -> x
@@ -4234,6 +4268,31 @@ LogicalResult UnPackOp::canonicalize(UnPackOp unPackOp,
[&]() { unPackOp.setDpsInitOperand(0, newDest); });
return success();
}
+
+ // Insert tensor.cast ops if static shape inference is available..
+ SmallVector<int64_t> srcShape, destShape;
+ if (inferStaticShape(unPackOp, srcShape, destShape)) {
+ Location loc = unPackOp.getLoc();
+ Value source = unPackOp.getSource();
+ if (srcShape != unPackOp.getSourceType().getShape()) {
+ auto newSrcType = unPackOp.getSourceType().clone(srcShape);
+ source = rewriter.create<tensor::CastOp>(loc, newSrcType,
+ unPackOp.getSource());
+ }
+ Value dest = unPackOp.getDest();
+ if (destShape != unPackOp.getDestType().getShape()) {
+ auto newDestType = unPackOp.getDestType().clone(destShape);
+ dest =
+ rewriter.create<tensor::CastOp>(loc, newDestType, unPackOp.getDest());
+ }
+ Value newOp = rewriter.create<tensor::UnPackOp>(
+ loc, source, dest, unPackOp.getInnerDimsPos(), unPackOp.getMixedTiles(),
+ unPackOp.getOuterDimsPerm());
+ rewriter.replaceOpWithNewOp<tensor::CastOp>(
+ unPackOp, unPackOp.getResult().getType(), newOp);
+ return success();
+ }
+
return failure();
}
diff --git a/mlir/test/Dialect/Tensor/canonicalize.mlir b/mlir/test/Dialect/Tensor/canonicalize.mlir
index 3b6cd799a6f348..35619d098f008a 100644
--- a/mlir/test/Dialect/Tensor/canonicalize.mlir
+++ b/mlir/test/Dialect/Tensor/canonicalize.mlir
@@ -909,6 +909,41 @@ func.func @fold_unpack_constant_splat(%dest : tensor<128x256xf32>) -> tensor<128
// -----
+func.func @infer_dest_shape_unpack(%src: tensor<10x20x30x40x16xf32>, %dest: tensor<?x?x?x?xf32>) -> tensor<?x?x?x?xf32> {
+ %unpack = tensor.unpack %src
+ outer_dims_perm = [2, 1, 3, 0]
+ inner_dims_pos = [2]
+ inner_tiles = [16]
+ into %dest : tensor<10x20x30x40x16xf32> -> tensor<?x?x?x?xf32>
+ return %unpack : tensor<?x?x?x?xf32>
+}
+// CHECK-LABEL: func.func @infer_dest_shape_unpack
+// CHECK-SAME: %[[SRC:[0-9a-zA-Z]+]]
+// CHECK-SAME: %[[DEST:[0-9a-zA-Z]+]]
+// CHECK: %[[CAST_DEST:.+]] = tensor.cast %[[DEST]] : tensor<?x?x?x?xf32> to tensor<30x20x?x10xf32>
+// CHECK: %[[UNPACK:.+]] = tensor.unpack %[[SRC]] {{.+}} into %[[CAST_DEST]]
+// CHECK: %[[CAST_UNPACK:.+]] = tensor.cast %[[UNPACK]] : tensor<30x20x?x10xf32> to tensor<?x?x?x?xf32>
+// CHECK: return %[[CAST_UNPACK]]
+
+// -----
+
+func.func @infer_src_shape_unpack(%src: tensor<?x?x?x?x16xf32>, %dest: tensor<30x20x?x10xf32>) -> tensor<30x20x?x10xf32> {
+ %unpack = tensor.unpack %src
+ outer_dims_perm = [2, 1, 3, 0]
+ inner_dims_pos = [2]
+ inner_tiles = [16]
+ into %dest : tensor<?x?x?x?x16xf32> -> tensor<30x20x?x10xf32>
+ return %unpack : tensor<30x20x?x10xf32>
+}
+// CHECK-LABEL: func.func @infer_src_shape_unpack
+// CHECK-SAME: %[[SRC:[0-9a-zA-Z]+]]
+// CHECK-SAME: %[[DEST:[0-9a-zA-Z]+]]
+// CHECK: %[[CAST_SRC:.+]] = tensor.cast %[[SRC]] : tensor<?x?x?x?x16xf32> to tensor<10x20x30x?x16xf32>
+// CHECK: %[[UNPACK:.+]] = tensor.unpack %[[CAST_SRC]]
+// CHECK: return %[[UNPACK]]
+
+// -----
+
// CHECK-LABEL: func @fold_overlapping_insert
// CHECK-SAME: %[[INPUT:.+]]: tensor<?x?x?xf32>, %{{.+}}: tensor<4x?x8xf32>, %[[SLICE2:.+]]: tensor<4x?x8xf32>
func.func @fold_overlapping_insert(%input : tensor<?x?x?xf32>, %slice1: tensor<4x?x8xf32>, %slice2: tensor<4x?x8xf32>, %i: index, %size: index) -> (tensor<?x?x?xf32>) {
@@ -2176,3 +2211,18 @@ func.func @generate_negative_size_verifies() -> tensor<?x8xi32> {
} : tensor<?x8xi32>
return %tensor : tensor<?x8xi32>
}
+
+// -----
+
+func.func @infer_and_fold_pack_unpack_same_tiles(%t: tensor<10x20x?x?xf32>,
+ %dim1: index, %dim2: index, %dim3: index, %dim4: index, %tile1: index,
+ %tile2: index) -> tensor<10x20x?x?xf32> {
+ %tensor_empty = tensor.empty(%dim1, %dim2) : tensor<?x?xf32>
+ %unpacked = tensor.unpack %t inner_dims_pos = [0, 1] inner_tiles = [%tile1, %tile2] into %tensor_empty : tensor<10x20x?x?xf32> -> tensor<?x?xf32>
+ %tensor_empty1 = tensor.empty(%dim3, %dim4) : tensor<10x20x?x?xf32>
+ %packed = tensor.pack %unpacked inner_dims_pos = [0, 1] inner_tiles = [%tile1, %tile2] into %tensor_empty1 : tensor<?x?xf32> -> tensor<10x20x?x?xf32>
+ return %packed : tensor<10x20x?x?xf32>
+}
+// CHECK-LABEL: func.func @infer_and_fold_pack_unpack_same_tiles
+// CHECK-SAME: %[[SRC:[0-9a-zA-Z]+]]
+// CHECK: return %[[SRC]]
|
qedawkins
approved these changes
Feb 14, 2024
| %tensor_empty = tensor.empty(%dim1, %dim2) : tensor<?x?xf32> | ||
| %unpacked = tensor.unpack %t inner_dims_pos = [0, 1] inner_tiles = [%tile1, %tile2] into %tensor_empty : tensor<10x20x?x?xf32> -> tensor<?x?xf32> | ||
| %tensor_empty1 = tensor.empty(%dim3, %dim4) : tensor<10x20x?x?xf32> | ||
| %packed = tensor.pack %unpacked inner_dims_pos = [0, 1] inner_tiles = [%tile1, %tile2] into %tensor_empty1 : tensor<?x?xf32> -> tensor<10x20x?x?xf32> |
There was a problem hiding this comment.
Does the static shape inference pattern kick in for this case? The tile sizes are dynamic here, so from what I can tell it doesn't apply. Can you add an additional test for something like this
%tensor_empty = tensor.empty(%dim1, %dim2) : tensor<?x?xf32>
%unpacked = tensor.unpack %t inner_dims_pos = [0, 1] inner_tiles = [4, 4] into %tensor_empty : tensor<10x20x4x4xf32> -> tensor<?x?xf32>
%cast = tensor.cast %unpacked : tensor<?x?xf32> to tensor<40x80xf32>
%tensor_empty1 = tensor.empty() : tensor<10x20x4x4xf32>
%packed = tensor.pack %cast inner_dims_pos = [0, 1] inner_tiles = [4, 4] into %tensor_empty1 : tensor<40x80xf32> -> tensor<10x20x4x4xf32>(might have written this wrong, but I'm wondering if we still get it to fold in this case).
There was a problem hiding this comment.
ah, good point! I verified that my test case does not work with -debug; yours is better. I'll update the test.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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.
The revision does not refactor the inferStaticShape for pack and unpack ops because they can diverge quickly. Because there are more dimensions can be inferred (i.e., with inner_tile_sizes) if the pack op does not have padding value.
This is a follow-up of #80848