[Mlir-commits] [mlir] [mlir][linalg] Enable scalable vectorization of linalg.unpack (PR #149293)
Andrzej Warzyński
llvmlistbot at llvm.org
Mon Jul 28 01:02:05 PDT 2025
https://github.com/banach-space updated https://github.com/llvm/llvm-project/pull/149293
>From 6ba0d1791ab07bd0c505ab9684dff6a1f4732375 Mon Sep 17 00:00:00 2001
From: Andrzej Warzynski <andrzej.warzynski at arm.com>
Date: Wed, 16 Jul 2025 17:08:55 +0000
Subject: [PATCH 1/7] [mlir][linalg] Enable scalable vectorization of
linalg.unpack (WIP)
This patch updates `vectorizeAsTensorUnpackOp` to support scalable
vectorization by requiring user-specified vector sizes for both the
_read_ and _write_ operations involved in `linalg.unpack`. Detailed
rationale and an example are provided below.
Conceptually, `linalg.unpack` consists of the following high-level steps:
1. _Read_ from the source tensor.
2. Transpose the value read in step (1).
3. _Write_ the value from step (2) into the destination tensor.
Currently, when vectorizing with user-provided vector sizes, only the
sizes for the _write_ operation (step 3) are required. Sizes for the
_read_ operation (step 1) are inferred from static shapes and inner tile
sizes.
This logic breaks when the input shapes or tile sizes are dynamic
(indeed, `vectorizeUnPackOpPrecondition` rejects such cases ATM and the
vectorization fails). This patch addresses the issue by requiring
explicit vector sizes for both the read and write sides, enabling
scalable vectorization in such cases.
Example:
```mlir
func.func @unpack(%in: tensor<1x1x8x?xf32>, %out: tensor<8x?xf32>) -> tensor<8x?xf32> {
%vs = vector.vscale
%c8 = arith.constant 8 : index
%tile_size = arith.muli %vs, %c8 : index
%unpack = linalg.unpack %in
inner_dims_pos = [0, 1]
inner_tiles = [8, %tile_size]
into %out : tensor<1x1x8x?xf32> -> tensor<8x?xf32>
return %unpack : tensor<8x?xf32>
}
module attributes {transform.with_named_sequence} {
transform.named_sequence @__transform_main(%arg0: !transform.any_op {transform.readonly}) {
%0 = transform.structured.match ops{["linalg.unpack"]} in %arg0 : (!transform.any_op) -> !transform.any_op
transform.structured.vectorize %0 vector_sizes [1, 1, 8, [8], 8, [8]] : !transform.any_op
// \ / \ /
// read-sizes write-sizes
transform.yield
}
}
```
Finally, this patch also extends `createReadOrMaskedRead` and
`createWriteOrMaskedWrite` to take scalable flags.
---
.../mlir/Dialect/Vector/Utils/VectorUtils.h | 2 +-
.../Linalg/Transforms/Vectorization.cpp | 130 ++++++++++++------
mlir/lib/Dialect/Vector/Utils/VectorUtils.cpp | 22 +--
.../Linalg/vectorization/linalg-ops.mlir | 98 +++++++++++--
4 files changed, 190 insertions(+), 62 deletions(-)
diff --git a/mlir/include/mlir/Dialect/Vector/Utils/VectorUtils.h b/mlir/include/mlir/Dialect/Vector/Utils/VectorUtils.h
index 7cd70e42d363c..8bd54cf31b893 100644
--- a/mlir/include/mlir/Dialect/Vector/Utils/VectorUtils.h
+++ b/mlir/include/mlir/Dialect/Vector/Utils/VectorUtils.h
@@ -228,7 +228,7 @@ bool isLinearizableVector(VectorType type);
Value createReadOrMaskedRead(OpBuilder &builder, Location loc, Value source,
ArrayRef<int64_t> inputVectorSizes, Value padValue,
bool useInBoundsInsteadOfMasking = false,
- ArrayRef<bool> scalableDims = {});
+ ArrayRef<bool> inputScalableVecDims = {});
/// Returns success if `inputVectorSizes` is a valid masking configuraion for
/// given `shape`, i.e., it meets:
diff --git a/mlir/lib/Dialect/Linalg/Transforms/Vectorization.cpp b/mlir/lib/Dialect/Linalg/Transforms/Vectorization.cpp
index 793eec732aa03..cea72ef34e214 100644
--- a/mlir/lib/Dialect/Linalg/Transforms/Vectorization.cpp
+++ b/mlir/lib/Dialect/Linalg/Transforms/Vectorization.cpp
@@ -1805,7 +1805,8 @@ vectorizeAsTensorPackOp(RewriterBase &rewriter, linalg::PackOp packOp,
inputShape[innerDimsPos[idx]] *= size;
auto maskedRead = vector::createReadOrMaskedRead(
rewriter, loc, packOp.getSource(), inputShape, padValue,
- useInBoundsInsteadOfMasking);
+ useInBoundsInsteadOfMasking,
+ /*inputScalableVecSizes=*/{});
// Create ShapeCastOp.
SmallVector<int64_t> destShape(inputVectorSizes);
@@ -1831,18 +1832,23 @@ vectorizeAsTensorPackOp(RewriterBase &rewriter, linalg::PackOp packOp,
return success();
}
-/// Vectorize a `linalg::UnPackOp` to these 4 Ops:
-/// Vector::TransferReadOp - Reads a vector from the source tensor
-/// vector::TransposeOp - Transpose the Source tensor
-/// ShapeCastOp - Reshape the data based on the target.
-/// vector::TransferWriteOp. - Write the result vector back to the destination
-/// tensor.
-/// If the vector sizes are not provided:
+/// Vectorize `linalg.unpack %src into %dest` as:
+/// // Reads a vector from the source tensor
+/// %read = vector.transfer_read %src
+/// // Transpose %read as specified in `outer_dims_perm` attribute
+/// %tr = vector.transpose %read
+/// // Reshape the data based on the target
+/// %sc = vector.shape_cast %tr
+/// // Write the result vector to the destination tensor.
+/// vector.transfer_write %sc into %dest
+///
+/// If the vector sizes are not provided:
/// * the vector sizes are determined by the input operand and attributes,
/// * update the inBounds attribute instead of masking.
static LogicalResult
vectorizeAsTensorUnpackOp(RewriterBase &rewriter, linalg::UnPackOp unpackOp,
ArrayRef<int64_t> inputVectorSizes,
+ ArrayRef<bool> inputScalableVecDims,
SmallVectorImpl<Value> &newResults) {
// TODO: Introduce a parent class that will handle the insertion point update.
@@ -1859,25 +1865,54 @@ vectorizeAsTensorUnpackOp(RewriterBase &rewriter, linalg::UnPackOp unpackOp,
auto destSize = unpackOp.getDestRank();
- if (!inputVectorSizes.empty())
- assert(inputVectorSizes.size() == destSize &&
+ if (!inputVectorSizes.empty()) {
+ assert(inputVectorSizes.size() == destSize + sourceShape.size() &&
"Incorrect number of input vector sizes");
+ }
+
+ SmallVector<bool> readScalableVectorFlags;
+ SmallVector<bool> writeScalableVectorFlags;
+ SmallVector<int64_t> readVectorSizes;
+ SmallVector<int64_t> writeVectorSizes;
- // vectorSizes is the shape of the vector that will be used to do final
+ // Split input-vector-sizes into vector sizes for the read and write
+ // operations.
+ if (!inputVectorSizes.empty()) {
+ readVectorSizes.append(inputVectorSizes.begin(),
+ inputVectorSizes.begin() + sourceShape.size());
+ writeVectorSizes.append(inputVectorSizes.begin() + sourceShape.size(),
+ inputVectorSizes.end());
+ }
+ if (!inputScalableVecDims.empty()) {
+ readScalableVectorFlags.append(inputScalableVecDims.begin(),
+ inputScalableVecDims.begin() +
+ sourceShape.size());
+ writeScalableVectorFlags.append(inputScalableVecDims.begin() +
+ sourceShape.size(),
+ inputScalableVecDims.end());
+ } else {
+ readScalableVectorFlags = SmallVector<bool>(sourceShape.size(), false);
+ writeScalableVectorFlags = SmallVector<bool>(destSize, false);
+ }
+
+ // writeVectorSizes is the shape of the vector that will be used to do final
// write on the destination tensor. It is set like this: Let's say the
// source tensor is rank 'M' and the dest tensor rank 'N', where N <= M.
// Thus:
- // 1. vectorSizes = sourceShape.take_front(N)
- // 2. if outer_dims_perms is present: do that permutation on vectorSizes.
+ // 1. writeVectorSizes = sourceShape.take_front(N)
+ // 2. if outer_dims_perms is present: do that permutation on writeVectorSizes.
// 3. multiply all the locations in vectorSize pointed by innerDimPos by the
// innerTiles attribute value.
- SmallVector<int64_t> vectorSizes(inputVectorSizes);
- if (vectorSizes.empty()) {
- llvm::append_range(vectorSizes, sourceShape.take_front(destSize));
+ // SmallVector<int64_t> writeVectorSizes(inputVectorSizes);
+ if (writeVectorSizes.empty()) {
+ if (ShapedType::isDynamicShape(sourceShape))
+ return failure();
+
+ llvm::append_range(writeVectorSizes, sourceShape.take_front(destSize));
if (!outerDimsPerm.empty())
- applyPermutationToVector(vectorSizes, outerDimsPerm);
+ applyPermutationToVector(writeVectorSizes, outerDimsPerm);
for (auto [i, pos] : llvm::enumerate(innerDimPos))
- vectorSizes[pos] *= innerTiles[i];
+ writeVectorSizes[pos] *= innerTiles[i];
useInBoundsInsteadOfMasking = true;
}
@@ -1901,17 +1936,20 @@ vectorizeAsTensorUnpackOp(RewriterBase &rewriter, linalg::UnPackOp unpackOp,
// After applying outer_dims_perm: [8, 16]
// After appending the rest of the sourceShape: [8, 16, 32, 16]
- SmallVector<int64_t> readVectorSizes(vectorSizes.begin(), vectorSizes.end());
-
- for (auto [index, size] : enumerate(innerTiles)) {
- readVectorSizes[innerDimPos[index]] =
- llvm::divideCeil(readVectorSizes[innerDimPos[index]], size);
- }
- if (!outerDimsPerm.empty()) {
- applyPermutationToVector(readVectorSizes, outerDimsPerm);
+ if (readVectorSizes.empty()) {
+ // Compute read-vector-sizes based on the write-vector-sizes and inner tile
+ // sizes. Note, this will only work when all sizes are static.
+ readVectorSizes = writeVectorSizes;
+ for (auto [index, size] : enumerate(innerTiles)) {
+ readVectorSizes[innerDimPos[index]] =
+ llvm::divideCeil(readVectorSizes[innerDimPos[index]], size);
+ }
+ if (!outerDimsPerm.empty()) {
+ applyPermutationToVector(readVectorSizes, outerDimsPerm);
+ }
+ readVectorSizes.append(sourceShape.begin() + writeVectorSizes.size(),
+ sourceShape.end());
}
- readVectorSizes.append(sourceShape.begin() + vectorSizes.size(),
- sourceShape.end());
Location loc = unpackOp->getLoc();
@@ -1923,7 +1961,7 @@ vectorizeAsTensorUnpackOp(RewriterBase &rewriter, linalg::UnPackOp unpackOp,
// to shape of source, then a mask is necessary.
Value readResult = vector::createReadOrMaskedRead(
rewriter, loc, unpackOp.getSource(), readVectorSizes, padValue,
- /*useInBoundsInsteadOfMasking=*/false);
+ /*useInBoundsInsteadOfMasking=*/false, readScalableVectorFlags);
PackingMetadata packMetadata;
SmallVector<int64_t> lastDimToInsertPosPerm =
@@ -1942,15 +1980,17 @@ vectorizeAsTensorUnpackOp(RewriterBase &rewriter, linalg::UnPackOp unpackOp,
RankedTensorType collapsedType = tensor::CollapseShapeOp::inferCollapsedType(
stripMineTensorType, packMetadata.reassociations);
mlir::VectorType vecCollapsedType =
- VectorType::get(collapsedType.getShape(), collapsedType.getElementType());
+ VectorType::get(collapsedType.getShape(), collapsedType.getElementType(),
+ writeScalableVectorFlags);
vector::ShapeCastOp shapeCastOp = vector::ShapeCastOp::create(
rewriter, loc, vecCollapsedType, transposeOp->getResult(0));
- // writeVectorSizes had to match the shapecast shape for dynamic sizes,
+ // writeVectorSizesFinal had to match the shapecast shape for dynamic sizes,
// otherwise the validator complains that the mask size is invalid.
- SmallVector<int64_t> writeVectorSizes(
+ // FIXME: We should not override write-vector-sizes like this.
+ SmallVector<int64_t> writeVectorSizesFinal(
unpackOp.getDestType().hasStaticShape()
- ? vectorSizes
+ ? writeVectorSizes
: shapeCastOp.getResultVectorType().getShape());
Operation *write = createWriteOrMaskedWrite(
rewriter, loc, shapeCastOp.getResult(), unpackOp.getDest(),
@@ -1981,7 +2021,7 @@ vectorizeAsTensorPadOp(RewriterBase &rewriter, tensor::PadOp padOp,
assert(succeeded(status) && "failed to reify result shapes");
auto maskedRead = vector::createReadOrMaskedRead(
rewriter, loc, padOp.getSource(), inputVectorSizes, padValue,
- /*useInBoundsInsteadOfMasking=*/false);
+ /*useInBoundsInsteadOfMasking=*/false, /*inputScalableVecSizes=*/{});
// Create Xfer write Op
Value dest = tensor::EmptyOp::create(rewriter, loc, reifiedReturnShapes[0],
@@ -2065,6 +2105,9 @@ static LogicalResult
vectorizeUnPackOpPrecondition(linalg::UnPackOp unpackOp,
ArrayRef<int64_t> inputVectorSizes) {
+ // FIXME!!!
+ return success();
+
if (llvm::any_of(unpackOp.getInnerTiles(), [](OpFoldResult res) {
return !getConstantIntValue(res).has_value();
})) {
@@ -2401,6 +2444,7 @@ vectorizePackOpPrecondition(linalg::PackOp packOp,
LDBG() << "pad value is not constant: " << packOp;
return failure();
}
+
ArrayRef<int64_t> resultTensorShape = packOp.getDestType().getShape();
bool satisfyEmptyCond = true;
if (inputVectorSizes.empty()) {
@@ -2479,12 +2523,14 @@ vectorizeScalableVectorPrecondition(Operation *op,
if (numOfScalableDims == 0)
return success();
+ // TODO: Check the following!
auto linalgOp = dyn_cast<LinalgOp>(op);
- // Cond 1: There's been no need for scalable vectorisation of
- // non-linalg Ops so far
- if (!linalgOp)
- return failure();
+ // Cond 1: Reject Ops that don't implement the LinalgOp interface, with the
+ // exception of UnpackOp for which there is a dedicated hook.
+ if (!linalgOp) {
+ return isa<linalg::UnPackOp>(op) ? success() : failure();
+ }
// Cond 2: There's been no need for more than 2 scalable dims so far
if (numOfScalableDims > 2)
@@ -2582,7 +2628,7 @@ vectorizeScalableVectorPrecondition(Operation *op,
isa<linalg::MatmulTransposeAOp>(op) ||
isa<linalg::DepthwiseConv1DNwcWcOp>(op) ||
isa<linalg::MatvecOp>(op) || isa<linalg::Mmt4DOp>(op) ||
- hasReductionIterator(linalgOp));
+ isa<linalg::UnPackOp>(op) || hasReductionIterator(linalgOp));
}
LogicalResult mlir::linalg::vectorizeOpPrecondition(
@@ -2715,7 +2761,8 @@ FailureOr<VectorizationResult> mlir::linalg::vectorize(
})
.Case<linalg::UnPackOp>([&](auto unpackOp) {
return vectorizeAsTensorUnpackOp(rewriter, unpackOp,
- inputVectorSizes, results);
+ inputVectorSizes,
+ inputScalableVecDims, results);
})
.Case<tensor::InsertSliceOp>([&](auto sliceOp) {
return vectorizeAsInsertSliceOp(rewriter, sliceOp, inputVectorSizes,
@@ -3107,7 +3154,8 @@ vectorizeAsInsertSliceOp(RewriterBase &rewriter, tensor::InsertSliceOp sliceOp,
vecType.getRank(), arith::ConstantIndexOp::create(rewriter, loc, 0));
Value read = mlir::vector::createReadOrMaskedRead(
rewriter, loc, source, vecType.getShape(), padValue,
- /*useInBoundsInsteadOfMasking=*/inputVectorSizes.empty());
+ /*useInBoundsInsteadOfMasking=*/inputVectorSizes.empty(),
+ /*inputScalableVecSizes=*/{});
// Create write
auto writeIndices =
diff --git a/mlir/lib/Dialect/Vector/Utils/VectorUtils.cpp b/mlir/lib/Dialect/Vector/Utils/VectorUtils.cpp
index 10ed2bcfb35a3..34b1bdbd9e010 100644
--- a/mlir/lib/Dialect/Vector/Utils/VectorUtils.cpp
+++ b/mlir/lib/Dialect/Vector/Utils/VectorUtils.cpp
@@ -279,14 +279,16 @@ vector::createUnrollIterator(VectorType vType, int64_t targetRank) {
// Attempt to unroll until targetRank or the first scalable dimension (which
// cannot be unrolled).
auto shapeToUnroll = vType.getShape().drop_back(targetRank);
- auto scalableDimsToUnroll = vType.getScalableDims().drop_back(targetRank);
- auto it = llvm::find(scalableDimsToUnroll, true);
- auto firstScalableDim = it - scalableDimsToUnroll.begin();
+ auto inputScalableVecDimsToUnroll =
+ vType.getScalableDims().drop_back(targetRank);
+ auto it = llvm::find(inputScalableVecDimsToUnroll, true);
+ auto firstScalableDim = it - inputScalableVecDimsToUnroll.begin();
if (firstScalableDim == 0)
return {};
// All scalable dimensions should be removed now.
- scalableDimsToUnroll = scalableDimsToUnroll.slice(0, firstScalableDim);
- assert(!llvm::is_contained(scalableDimsToUnroll, true) &&
+ inputScalableVecDimsToUnroll =
+ inputScalableVecDimsToUnroll.slice(0, firstScalableDim);
+ assert(!llvm::is_contained(inputScalableVecDimsToUnroll, true) &&
"unexpected leading scalable dimension");
// Create an unroll iterator for leading dimensions.
shapeToUnroll = shapeToUnroll.slice(0, firstScalableDim);
@@ -319,15 +321,15 @@ Value vector::createReadOrMaskedRead(OpBuilder &builder, Location loc,
ArrayRef<int64_t> inputVectorSizes,
Value padValue,
bool useInBoundsInsteadOfMasking,
- ArrayRef<bool> scalableDims) {
+ ArrayRef<bool> inputScalableVecDims) {
assert(!llvm::is_contained(inputVectorSizes, ShapedType::kDynamic) &&
"invalid input vector sizes");
auto sourceShapedType = cast<ShapedType>(source.getType());
auto sourceShape = sourceShapedType.getShape();
assert(sourceShape.size() == inputVectorSizes.size() &&
"expected same ranks.");
- auto vectorType =
- VectorType::get(inputVectorSizes, padValue.getType(), scalableDims);
+ auto vectorType = VectorType::get(inputVectorSizes, padValue.getType(),
+ inputScalableVecDims);
assert(padValue.getType() == sourceShapedType.getElementType() &&
"expected same pad element type to match source element type");
int64_t readRank = inputVectorSizes.size();
@@ -356,8 +358,8 @@ Value vector::createReadOrMaskedRead(OpBuilder &builder, Location loc,
? memref::getMixedSizes(builder, loc, source)
: tensor::getMixedSizes(builder, loc, source);
- auto maskType =
- VectorType::get(inputVectorSizes, builder.getI1Type(), scalableDims);
+ auto maskType = VectorType::get(inputVectorSizes, builder.getI1Type(),
+ inputScalableVecDims);
Value mask =
vector::CreateMaskOp::create(builder, loc, maskType, mixedSourceDims);
return mlir::vector::maskOperation(builder, transferReadOp, mask)
diff --git a/mlir/test/Dialect/Linalg/vectorization/linalg-ops.mlir b/mlir/test/Dialect/Linalg/vectorization/linalg-ops.mlir
index d41d86117793b..ec227b46b409e 100644
--- a/mlir/test/Dialect/Linalg/vectorization/linalg-ops.mlir
+++ b/mlir/test/Dialect/Linalg/vectorization/linalg-ops.mlir
@@ -940,9 +940,9 @@ module attributes {transform.with_named_sequence} {
///----------------------------------------------------------------------------------------
// CHECK-LABEL: func @test_vectorize_dynamic_shapes_unpack
-// CHECK-SAME: %[[ARG_0:.*]]: tensor<?x?xf32>,
-// CHECK-SAME: %[[ARG_1:.*]]: tensor<?x?x16x2xf32>
-func.func @test_vectorize_dynamic_shapes_unpack(%arg0: tensor<?x?xf32>, %arg1: tensor<?x?x16x2xf32>) -> tensor<?x?xf32> {
+// CHECK-SAME: %[[DEST:.*]]: tensor<?x?xf32>,
+// CHECK-SAME: %[[SRC:.*]]: tensor<?x?x16x2xf32>
+func.func @test_vectorize_dynamic_shapes_unpack(%dest: tensor<?x?xf32>, %src: tensor<?x?x16x2xf32>) -> tensor<?x?xf32> {
// CHECK: %[[C0:.*]] = arith.constant 0
// CHECK: %[[C01:.*]] = arith.constant 0
// CHECK: %[[C02:.*]] = arith.constant 0
@@ -956,15 +956,93 @@ func.func @test_vectorize_dynamic_shapes_unpack(%arg0: tensor<?x?xf32>, %arg1: t
// CHECK: %[[trans0:.*]] = vector.transpose %[[read0]], [0, 3, 1, 2] : vector<2x1x16x2xf32> to vector<2x2x1x16xf32>
// CHECK: %[[sc0:.*]] = vector.shape_cast %[[trans0]] : vector<2x2x1x16xf32> to vector<4x16xf32>
// CHECK: %[[writeMsk0:.*]] = vector.create_mask {{.*}} : vector<4x16xi1>
-// CHECK: %[[write0:.*]] = vector.mask %[[writeMsk0:.*]] {{.*}} vector.transfer_write %[[sc0]], %[[ARG_0]]
+// CHECK: %[[write0:.*]] = vector.mask %[[writeMsk0:.*]] {{.*}} vector.transfer_write %[[sc0]], %[[SRC]]
// CHECK: return %[[write0]]
- %ret = linalg.unpack %arg1 inner_dims_pos = [1, 0] inner_tiles = [16, 2] into %arg0 : tensor<?x?x16x2xf32> -> tensor<?x?xf32>
+ %ret = linalg.unpack %src inner_dims_pos = [1, 0] inner_tiles = [16, 2] into %dest : tensor<?x?x16x2xf32> -> tensor<?x?xf32>
return %ret : tensor<?x?xf32>
}
module attributes {transform.with_named_sequence} {
transform.named_sequence @__transform_main(%arg0: !transform.any_op {transform.readonly}) {
%0 = transform.structured.match ops{["linalg.unpack"]} in %arg0 : (!transform.any_op) -> !transform.any_op
- transform.structured.vectorize %0 vector_sizes [4, 16] : !transform.any_op
+ transform.structured.vectorize %0 vector_sizes [2, 1, 16, 2, 4, 16] : !transform.any_op
+ transform.yield
+ }
+}
+
+// -----
+
+// CHECK-LABEL: func @test_vectorize_dynamic_shapes_unpack_scalable_vec
+// CHECK-SAME: %[[DEST:.*]]: tensor<?x?xf32>,
+// CHECK-SAME: %[[SRC:.*]]: tensor<?x?x16x2xf32>
+func.func @test_vectorize_dynamic_shapes_unpack_scalable_vec(%dest: tensor<?x?xf32>, %src: tensor<?x?x16x2xf32>) -> tensor<?x?xf32> {
+ // CHECK: %[[C0:.*]] = arith.constant 0
+ // CHECK: %[[DIM:.*]] = tensor.dim %[[DEST]], %[[C0]] : tensor<?x?xf32>
+ // CHECK: %[[C1:.*]] = arith.constant 1 : index
+ // CHECK: %[[DIM0:.*]] = tensor.dim %[[DEST]], %[[C1]] : tensor<?x?xf32>
+ // CHECK: %[[CST:.*]] = arith.constant 0.000000e+00
+ // CHECK: %[[C01:.*]] = arith.constant 0
+ // CHECK: %[[C02:.*]] = arith.constant 0
+ // CHECK: %[[DIM4:.*]] = tensor.dim %[[SRC]], %[[C02]] : tensor<?x?x16x2xf32>
+ // CHECK: %[[CNST14:.*]] = arith.constant 1
+ // CHECK: %[[DIM6:.*]] = tensor.dim %[[SRC]], %[[CNST14]] : tensor<?x?x16x2xf32>
+ // CHECK: %[[CNST16:.*]] = arith.constant 16 : index
+ // CHECK: %[[CNST2:.*]] = arith.constant 2 : index
+ // CHECK: %[[MASK_READ:.*]] = vector.create_mask %[[DIM4]], %[[DIM6]], %[[CNST16]], %[[CNST2]] : vector<2x1x[16]x2xi1>
+ // CHECK: %[[READ:.*]] = vector.mask %[[MASK_READ]] {{.*}} vector.transfer_read %{{.*}} : tensor<?x?x16x2xf32>, vector<2x1x[16]x2xf32> } : vector<2x1x[16]x2xi1> -> vector<2x1x[16]x2xf32>
+ // CHECK: %[[TR:.*]] = vector.transpose %[[READ]], [0, 3, 1, 2] : vector<2x1x[16]x2xf32> to vector<2x2x1x[16]xf32>
+ // CHECK: %[[SC:.*]] = vector.shape_cast %[[TR]] : vector<2x2x1x[16]xf32> to vector<4x[16]xf32>
+ // CHECK: %[[MASK_WRITE:.*]] = vector.create_mask {{.*}} : vector<4x[16]xi1>
+ // CHECK: %[[WRITE:.*]] = vector.mask %[[MASK_WRITE:.*]] {{.*}} vector.transfer_write %[[SC]], %[[DEST]]
+ // CHECK: return %[[WRITE]]
+ %ret = linalg.unpack %src inner_dims_pos = [1, 0] inner_tiles = [16, 2] into %dest : tensor<?x?x16x2xf32> -> tensor<?x?xf32>
+ return %ret : tensor<?x?xf32>
+}
+module attributes {transform.with_named_sequence} {
+ transform.named_sequence @__transform_main(%arg0: !transform.any_op {transform.readonly}) {
+ %0 = transform.structured.match ops{["linalg.unpack"]} in %arg0 : (!transform.any_op) -> !transform.any_op
+ transform.structured.vectorize %0 vector_sizes [2, 1, [16], 2, 4, [16]] : !transform.any_op
+ transform.yield
+ }
+}
+
+// -----
+
+// CHECK-LABEL: func @test_vectorize_dynamic_shapes_unpack_scalable_vec_and_tile_size
+// CHECK-SAME: %[[DEST:.*]]: tensor<?x?xf32>,
+// CHECK-SAME: %[[SRC:.*]]: tensor<?x?x?x2xf32>
+func.func @test_vectorize_dynamic_shapes_unpack_scalable_vec_and_tile_size(%dest: tensor<?x?xf32>, %src: tensor<?x?x?x2xf32>) -> tensor<?x?xf32> {
+ // CHECK: %[[C0:.*]] = arith.constant 0
+ // CHECK: %[[DIM:.*]] = tensor.dim %[[DEST]], %[[C0]] : tensor<?x?xf32>
+ // CHECK: %[[C1:.*]] = arith.constant 1 : index
+ // CHECK: %[[DIM0:.*]] = tensor.dim %[[DEST]], %[[C1]] : tensor<?x?xf32>
+ // CHECK: %[[CST:.*]] = arith.constant 0.000000e+00
+ // CHECK: %[[C01:.*]] = arith.constant 0
+ // CHECK: %[[C02:.*]] = arith.constant 0
+ // CHECK: %[[DIM4:.*]] = tensor.dim %[[SRC]], %[[C02]] : tensor<?x?x?x2xf32>
+ // CHECK: %[[C1_2:.*]] = arith.constant 1
+ // CHECK: %[[DIM6:.*]] = tensor.dim %[[SRC]], %[[C1_2]] : tensor<?x?x?x2xf32>
+ // CHECK: %[[C2:.*]] = arith.constant 2 : index
+ // CHECK: %[[DIM_2:.*]] = tensor.dim %[[SRC]], %[[C2]] : tensor<?x?x?x2xf32>
+ // CHECK: %[[C2_1:.*]] = arith.constant 2 : index
+ // CHECK: %[[MASK_READ:.*]] = vector.create_mask %[[DIM4]], %[[DIM6]], %[[DIM_2]], %[[C2_1]] : vector<2x1x[16]x2xi1>
+ // CHECK: %[[READ:.*]] = vector.mask %[[MASK_READ]] {{.*}} vector.transfer_read %{{.*}} : tensor<?x?x?x2xf32>, vector<2x1x[16]x2xf32> } : vector<2x1x[16]x2xi1> -> vector<2x1x[16]x2xf32>
+ // CHECK: %[[TR:.*]] = vector.transpose %[[READ]], [0, 3, 1, 2] : vector<2x1x[16]x2xf32> to vector<2x2x1x[16]xf32>
+ // CHECK: %[[SC:.*]] = vector.shape_cast %[[TR]] : vector<2x2x1x[16]xf32> to vector<4x[16]xf32>
+ // CHECK: %[[MASK_WRITE:.*]] = vector.create_mask {{.*}} : vector<4x[16]xi1>
+ // CHECK: %[[WRITE:.*]] = vector.mask %[[MASK_WRITE:.*]] {{.*}} vector.transfer_write %[[SC]], %[[DEST]]
+ // CHECK: return %[[WRITE]]
+
+ %vs = vector.vscale
+ %c16 = arith.constant 16 : index
+ %tile_size = arith.muli %vs, %c16 : index
+
+ %ret = linalg.unpack %src inner_dims_pos = [1, 0] inner_tiles = [%tile_size, 2] into %dest : tensor<?x?x?x2xf32> -> tensor<?x?xf32>
+ return %ret : tensor<?x?xf32>
+}
+module attributes {transform.with_named_sequence} {
+ transform.named_sequence @__transform_main(%arg0: !transform.any_op {transform.readonly}) {
+ %0 = transform.structured.match ops{["linalg.unpack"]} in %arg0 : (!transform.any_op) -> !transform.any_op
+ transform.structured.vectorize %0 vector_sizes [2, 1, [16], 2, 4, [16]] : !transform.any_op
transform.yield
}
}
@@ -997,7 +1075,7 @@ func.func @test_vectorize_unpack(%source: tensor<8x8x32x16xf32>, %dest: tensor<2
module attributes {transform.with_named_sequence} {
transform.named_sequence @__transform_main(%arg0: !transform.any_op {transform.readonly}) {
%0 = transform.structured.match ops{["linalg.unpack"]} in %arg0 : (!transform.any_op) -> !transform.any_op
- transform.structured.vectorize %0 vector_sizes [512, 128] : !transform.any_op
+ transform.structured.vectorize %0 vector_sizes [16, 8, 32, 16, 512, 128] : !transform.any_op
transform.yield
}
}
@@ -1022,7 +1100,7 @@ func.func @test_vectorize_unpack_no_masks(%source: tensor<8x8x32x16xf32>, %dest:
module attributes {transform.with_named_sequence} {
transform.named_sequence @__transform_main(%arg0: !transform.any_op {transform.readonly}) {
%0 = transform.structured.match ops{["linalg.unpack"]} in %arg0 : (!transform.any_op) -> !transform.any_op
- transform.structured.vectorize %0 vector_sizes [256, 128] : !transform.any_op
+ transform.structured.vectorize %0 vector_sizes [8, 8, 32, 16, 256, 128] : !transform.any_op
transform.yield
}
}
@@ -1047,7 +1125,7 @@ func.func @test_vectorize_unpack_no_masks(%source: tensor<8x8x32x16xf32>, %dest:
module attributes {transform.with_named_sequence} {
transform.named_sequence @__transform_main(%arg0: !transform.any_op {transform.readonly}) {
%0 = transform.structured.match ops{["linalg.unpack"]} in %arg0 : (!transform.any_op) -> !transform.any_op
- transform.structured.vectorize %0 vector_sizes [256, 128] : !transform.any_op
+ transform.structured.vectorize %0 vector_sizes [8, 8, 32, 16, 256, 128] : !transform.any_op
transform.yield
}
}
@@ -1170,7 +1248,7 @@ module attributes {transform.with_named_sequence} {
func.func @test_vectorize_padded_pack(%arg0: tensor<32x7x15xf32>, %arg1: tensor<32x4x1x16x2xf32>) -> tensor<32x4x1x16x2xf32> {
%pad = arith.constant 0.000000e+00 : f32
- %pack = linalg.pack %arg0 padding_value(%pad : f32) inner_dims_pos = [2, 1] inner_tiles = [16, 2] into %arg1 : tensor<32x7x15xf32> -> tensor<32x4x1x16x2xf32>
+ %pack = linalg.pack %arg0 padding_value(%pad : f32) inner_dims_pos = [2, 1] inner_tiles = [16, [2]] into %arg1 : tensor<32x7x15xf32> -> tensor<32x4x1x16x2xf32>
return %pack : tensor<32x4x1x16x2xf32>
}
// CHECK-DAG: %[[cst:.*]] = arith.constant 0.000000e+00 : f32
>From 9dd68bb73d528ca5f5ce3a0627a43d5b9a237158 Mon Sep 17 00:00:00 2001
From: Andrzej Warzynski <andrzej.warzynski at arm.com>
Date: Thu, 24 Jul 2025 20:52:12 +0000
Subject: [PATCH 2/7] fixup! [mlir][linalg] Enable scalable vectorization of
linalg.unpack (WIP)
Remove leftover code + comments
---
mlir/lib/Dialect/Linalg/Transforms/Vectorization.cpp | 4 ----
1 file changed, 4 deletions(-)
diff --git a/mlir/lib/Dialect/Linalg/Transforms/Vectorization.cpp b/mlir/lib/Dialect/Linalg/Transforms/Vectorization.cpp
index cea72ef34e214..dfd340d6681dc 100644
--- a/mlir/lib/Dialect/Linalg/Transforms/Vectorization.cpp
+++ b/mlir/lib/Dialect/Linalg/Transforms/Vectorization.cpp
@@ -1903,7 +1903,6 @@ vectorizeAsTensorUnpackOp(RewriterBase &rewriter, linalg::UnPackOp unpackOp,
// 2. if outer_dims_perms is present: do that permutation on writeVectorSizes.
// 3. multiply all the locations in vectorSize pointed by innerDimPos by the
// innerTiles attribute value.
- // SmallVector<int64_t> writeVectorSizes(inputVectorSizes);
if (writeVectorSizes.empty()) {
if (ShapedType::isDynamicShape(sourceShape))
return failure();
@@ -2105,9 +2104,6 @@ static LogicalResult
vectorizeUnPackOpPrecondition(linalg::UnPackOp unpackOp,
ArrayRef<int64_t> inputVectorSizes) {
- // FIXME!!!
- return success();
-
if (llvm::any_of(unpackOp.getInnerTiles(), [](OpFoldResult res) {
return !getConstantIntValue(res).has_value();
})) {
>From c67523cfdbf4a8e11063ecdb1ea459a770a4be29 Mon Sep 17 00:00:00 2001
From: Andrzej Warzynski <andrzej.warzynski at arm.com>
Date: Fri, 25 Jul 2025 09:24:12 +0000
Subject: [PATCH 3/7] fixup! fixup! [mlir][linalg] Enable scalable
vectorization of linalg.unpack (WIP)
Fix pre-condition calculation
---
.../Linalg/Transforms/Vectorization.cpp | 43 ++++++++++++++-----
1 file changed, 32 insertions(+), 11 deletions(-)
diff --git a/mlir/lib/Dialect/Linalg/Transforms/Vectorization.cpp b/mlir/lib/Dialect/Linalg/Transforms/Vectorization.cpp
index dfd340d6681dc..39fe614a27b8b 100644
--- a/mlir/lib/Dialect/Linalg/Transforms/Vectorization.cpp
+++ b/mlir/lib/Dialect/Linalg/Transforms/Vectorization.cpp
@@ -2099,24 +2099,45 @@ vectorizeDynamicLinalgOpPrecondition(linalg::LinalgOp op,
return success();
}
-/// Need to check if the inner-tiles are static/constant.
+//// This hook considers two cases:
+/// (1) If the input-vector-sizes are empty, then the vector sizes will be
+/// infered. This is only possible when all shapes are static.
+/// (2) If the input-vector-sizes are non-empty (i.e. user provided), then
+/// carry out basic sanity-checking.
static LogicalResult
vectorizeUnPackOpPrecondition(linalg::UnPackOp unpackOp,
ArrayRef<int64_t> inputVectorSizes) {
+ // If there are no input vector sizes and all shapes are static, there is
+ // nothing left to check.
+ if (inputVectorSizes.empty() && unpackOp.getDestType().hasStaticShape() &&
+ unpackOp.getSourceType().hasStaticShape())
+ return success();
- if (llvm::any_of(unpackOp.getInnerTiles(), [](OpFoldResult res) {
- return !getConstantIntValue(res).has_value();
- })) {
- LDBG() << "Inner-tiles must be constant: " << unpackOp;
+ // The input vector sizes must be equal to:
+ // * read-vector-rank + write-vector-rank
+ if (!inputVectorSizes.empty()) {
+ if (inputVectorSizes.size() !=
+ unpackOp.getDestRank() + unpackOp.getSourceRank()) {
+ LDBG("Incorrect number of input vector sizes");
+ return failure();
+ }
+ }
+
+ // Check the vector sizes for the write operation.
+ if (failed(vector::isValidMaskedInputVector(
+ unpackOp.getDestType().getShape(),
+ inputVectorSizes.take_back(unpackOp.getDestRank())))) {
+ LDBG("Incorrect number of input vector sizes");
return failure();
}
- ArrayRef<int64_t> resultShape = unpackOp.getDestType().getShape();
- bool satisfyEmptyCond = inputVectorSizes.empty() &&
- unpackOp.getDestType().hasStaticShape() &&
- unpackOp.getSourceType().hasStaticShape();
- if (!satisfyEmptyCond &&
- failed(vector::isValidMaskedInputVector(resultShape, inputVectorSizes)))
+
+ // Check the vector sizes for the read operation.
+ if (failed(vector::isValidMaskedInputVector(
+ unpackOp.getSourceType().getShape(),
+ inputVectorSizes.take_front(unpackOp.getSourceRank())))) {
+ LDBG("Incorrect number of input vector sizes");
return failure();
+ }
return success();
}
>From 7229a6d6f89ddf64c832a4e93611c4c924c921f4 Mon Sep 17 00:00:00 2001
From: Andrzej Warzynski <andrzej.warzynski at arm.com>
Date: Fri, 25 Jul 2025 10:20:12 +0000
Subject: [PATCH 4/7] fixup! fixup! [mlir][linalg] Enable scalable
vectorization of linalg.unpack (WIP)
Improve documentation + fix test after rebasing on top of
* https://github.com/llvm/llvm-project/pull/150602
---
.../Linalg/Transforms/Vectorization.cpp | 79 +++++++++----------
.../Linalg/vectorization/linalg-ops.mlir | 41 ++++------
2 files changed, 52 insertions(+), 68 deletions(-)
diff --git a/mlir/lib/Dialect/Linalg/Transforms/Vectorization.cpp b/mlir/lib/Dialect/Linalg/Transforms/Vectorization.cpp
index 39fe614a27b8b..0b0e71e8951a6 100644
--- a/mlir/lib/Dialect/Linalg/Transforms/Vectorization.cpp
+++ b/mlir/lib/Dialect/Linalg/Transforms/Vectorization.cpp
@@ -1850,6 +1850,13 @@ vectorizeAsTensorUnpackOp(RewriterBase &rewriter, linalg::UnPackOp unpackOp,
ArrayRef<int64_t> inputVectorSizes,
ArrayRef<bool> inputScalableVecDims,
SmallVectorImpl<Value> &newResults) {
+ if (!inputVectorSizes.empty()) {
+ assert(inputVectorSizes.size() ==
+ unpackOp.getDestRank() + unpackOp.getSourceRank() &&
+ "Invalid number of input vector sizes!");
+ assert(inputVectorSizes.size() == inputScalableVecDims.size() &&
+ "Incompatible number of vector sizes and vector scalable flags!");
+ }
// TODO: Introduce a parent class that will handle the insertion point update.
OpBuilder::InsertionGuard g(rewriter);
@@ -1865,44 +1872,41 @@ vectorizeAsTensorUnpackOp(RewriterBase &rewriter, linalg::UnPackOp unpackOp,
auto destSize = unpackOp.getDestRank();
- if (!inputVectorSizes.empty()) {
- assert(inputVectorSizes.size() == destSize + sourceShape.size() &&
- "Incorrect number of input vector sizes");
- }
-
- SmallVector<bool> readScalableVectorFlags;
- SmallVector<bool> writeScalableVectorFlags;
+ // 1. Obtain vector sizes for the read and write operation.s
SmallVector<int64_t> readVectorSizes;
SmallVector<int64_t> writeVectorSizes;
+ SmallVector<bool> readScalableVectorFlags;
+ SmallVector<bool> writeScalableVectorFlags;
- // Split input-vector-sizes into vector sizes for the read and write
- // operations.
+ // CASE 1: Vector sizes are user-specified.
+ // 1.0 This is the trivial case, simply split the input vector sizes.
if (!inputVectorSizes.empty()) {
readVectorSizes.append(inputVectorSizes.begin(),
inputVectorSizes.begin() + sourceShape.size());
writeVectorSizes.append(inputVectorSizes.begin() + sourceShape.size(),
inputVectorSizes.end());
- }
- if (!inputScalableVecDims.empty()) {
readScalableVectorFlags.append(inputScalableVecDims.begin(),
inputScalableVecDims.begin() +
sourceShape.size());
writeScalableVectorFlags.append(inputScalableVecDims.begin() +
sourceShape.size(),
inputScalableVecDims.end());
- } else {
- readScalableVectorFlags = SmallVector<bool>(sourceShape.size(), false);
- writeScalableVectorFlags = SmallVector<bool>(destSize, false);
}
- // writeVectorSizes is the shape of the vector that will be used to do final
- // write on the destination tensor. It is set like this: Let's say the
- // source tensor is rank 'M' and the dest tensor rank 'N', where N <= M.
- // Thus:
- // 1. writeVectorSizes = sourceShape.take_front(N)
- // 2. if outer_dims_perms is present: do that permutation on writeVectorSizes.
- // 3. multiply all the locations in vectorSize pointed by innerDimPos by the
- // innerTiles attribute value.
+ // CASE 2: Vector sizes have to be inferred.
+ //
+ // 1.1 Infer vector sizes for the write operation.
+ //
+ // Let:
+ // * rank(source tensor) = 'M'
+ // * rank(dest tensor) = 'N',
+ // and N <= M. The steps are:
+ // 1. writeVectorSizes = sourceShape.take_front(N)
+ // 2. Multiply all the locations in writeVectorSize pointed by inner_dims_pos
+ // by the corresponding values from the `inner_tiles` attribute value.
+ // 3. If outer_dims_perms is present, permutate writeVectorSizes accordingly.
+ //
+ // Note, this will only work when all sizes are static!
if (writeVectorSizes.empty()) {
if (ShapedType::isDynamicShape(sourceShape))
return failure();
@@ -1916,28 +1920,17 @@ vectorizeAsTensorUnpackOp(RewriterBase &rewriter, linalg::UnPackOp unpackOp,
useInBoundsInsteadOfMasking = true;
}
- // readVectorSizes is the size of tensor used to read and apply mask. It is
- // set like this: Let's say the vectorSize (VS) array is size 'N' and
- // the sourceShape(SS) is 'M' where M >= N and InnerTileSizes (IT) of
- // size M-N
- // Thus:
- // - initially: readVectorSizes = vectorInputSizes
- // - Divide all the readMaskShape locations pointed by innerDimPos
- // by the innerTileSize attribute value.
- // - if outer_dims_perms is present: do that permutation on readVectorSizes.
- // - Append the remaining shape from SS
- // E.g. let's say let's say unpackTensorType.getShape() = <8x8x32x16>
- // inner Dim Pos = [0, 1] and Inner Tiles = [32, 16], vector_sizes are [512,
- // 128] and outer_dims_perm is [1, 0] then read shape is:
- // ReadVectorSizes(initial): [512, 128]
- // Final Value(after innerDim Adjustment): [512/32, 128/16]
- // = [16, 8]
- // After applying outer_dims_perm: [8, 16]
- // After appending the rest of the sourceShape: [8, 16, 32, 16]
-
+ // 1.2 Infer vector sizes for the read operation.
+ //
+ // The steps are:
+ // 1. readVectorSizes = vectorInputSizes
+ // 2. Take readVectorSizes from 1. and divide all locations pointed by
+ // the inner_dims_pos attribyte by the `inner_tiles` attribute value.
+ // 3. If outer_dims_perms is present, permutate readVectorSizes accordingly.
+ // 4. Append the remaining sizes from the source tensor.
+ //
+ // Note, this will only work when all sizes are static!
if (readVectorSizes.empty()) {
- // Compute read-vector-sizes based on the write-vector-sizes and inner tile
- // sizes. Note, this will only work when all sizes are static.
readVectorSizes = writeVectorSizes;
for (auto [index, size] : enumerate(innerTiles)) {
readVectorSizes[innerDimPos[index]] =
diff --git a/mlir/test/Dialect/Linalg/vectorization/linalg-ops.mlir b/mlir/test/Dialect/Linalg/vectorization/linalg-ops.mlir
index ec227b46b409e..fcb8b02d3faa3 100644
--- a/mlir/test/Dialect/Linalg/vectorization/linalg-ops.mlir
+++ b/mlir/test/Dialect/Linalg/vectorization/linalg-ops.mlir
@@ -943,23 +943,22 @@ module attributes {transform.with_named_sequence} {
// CHECK-SAME: %[[DEST:.*]]: tensor<?x?xf32>,
// CHECK-SAME: %[[SRC:.*]]: tensor<?x?x16x2xf32>
func.func @test_vectorize_dynamic_shapes_unpack(%dest: tensor<?x?xf32>, %src: tensor<?x?x16x2xf32>) -> tensor<?x?xf32> {
-// CHECK: %[[C0:.*]] = arith.constant 0
-// CHECK: %[[C01:.*]] = arith.constant 0
-// CHECK: %[[C02:.*]] = arith.constant 0
-// CHECK: %[[DIM_0:.*]] = tensor.dim %[[ARG_1]], %[[C02]] : tensor<?x?x16x2xf32>
-// CHECK: %[[C1:.*]] = arith.constant 1
-// CHECK: %[[DIM6:.*]] = tensor.dim %[[ARG_1]], %[[C1]] : tensor<?x?x16x2xf32>
-// CHECK: %[[CNST16:.*]] = arith.constant 16 : index
-// CHECK: %[[CNST2:.*]] = arith.constant 2 : index
-// CHECK: %[[readMsk0:.*]] = vector.create_mask %[[DIM_0]], %[[DIM6]], %[[CNST16]], %[[CNST2]] : vector<2x1x16x2xi1>
-// CHECK: %[[read0:.*]] = vector.mask %[[readMsk0]] {{.*}} vector.transfer_read %{{.*}} : tensor<?x?x16x2xf32>, vector<2x1x16x2xf32> } : vector<2x1x16x2xi1> -> vector<2x1x16x2xf32>
-// CHECK: %[[trans0:.*]] = vector.transpose %[[read0]], [0, 3, 1, 2] : vector<2x1x16x2xf32> to vector<2x2x1x16xf32>
-// CHECK: %[[sc0:.*]] = vector.shape_cast %[[trans0]] : vector<2x2x1x16xf32> to vector<4x16xf32>
-// CHECK: %[[writeMsk0:.*]] = vector.create_mask {{.*}} : vector<4x16xi1>
-// CHECK: %[[write0:.*]] = vector.mask %[[writeMsk0:.*]] {{.*}} vector.transfer_write %[[sc0]], %[[SRC]]
-// CHECK: return %[[write0]]
- %ret = linalg.unpack %src inner_dims_pos = [1, 0] inner_tiles = [16, 2] into %dest : tensor<?x?x16x2xf32> -> tensor<?x?xf32>
- return %ret : tensor<?x?xf32>
+ // CHECK: %[[C0:.*]] = arith.constant 0 : index
+ // CHECK: %[[C0_1:.*]] = arith.constant 0 : index
+ // CHECK: %[[DIM_0:.*]] = tensor.dim %[[SRC]], %[[C0_1]] : tensor<?x?x16x2xf32>
+ // CHECK: %[[C1:.*]] = arith.constant 1
+ // CHECK: %[[DIM6:.*]] = tensor.dim %[[SRC]], %[[C1]] : tensor<?x?x16x2xf32>
+ // CHECK: %[[CNST16:.*]] = arith.constant 16 : index
+ // CHECK: %[[CNST2:.*]] = arith.constant 2 : index
+ // CHECK: %[[MASK_READ:.*]] = vector.create_mask %[[DIM_0]], %[[DIM6]], %[[CNST16]], %[[CNST2]] : vector<2x1x16x2xi1>
+ // CHECK: %[[READ:.*]] = vector.mask %[[MASK_READ]] {{.*}} vector.transfer_read %{{.*}} : tensor<?x?x16x2xf32>, vector<2x1x16x2xf32> } : vector<2x1x16x2xi1> -> vector<2x1x16x2xf32>
+ // CHECK: %[[TR:.*]] = vector.transpose %[[READ]], [0, 3, 1, 2] : vector<2x1x16x2xf32> to vector<2x2x1x16xf32>
+ // CHECK: %[[SC:.*]] = vector.shape_cast %[[TR]] : vector<2x2x1x16xf32> to vector<4x16xf32>
+ // CHECK: %[[MASK_WRITE:.*]] = vector.create_mask {{.*}} : vector<4x16xi1>
+ // CHECK: %[[WRITE:.*]] = vector.mask %[[MASK_WRITE:.*]] {{.*}} vector.transfer_write %[[SC]], %[[DEST]]
+ // CHECK: return %[[WRITE]]
+ %ret = linalg.unpack %src inner_dims_pos = [1, 0] inner_tiles = [16, 2] into %dest : tensor<?x?x16x2xf32> -> tensor<?x?xf32>
+ return %ret : tensor<?x?xf32>
}
module attributes {transform.with_named_sequence} {
transform.named_sequence @__transform_main(%arg0: !transform.any_op {transform.readonly}) {
@@ -975,10 +974,6 @@ module attributes {transform.with_named_sequence} {
// CHECK-SAME: %[[DEST:.*]]: tensor<?x?xf32>,
// CHECK-SAME: %[[SRC:.*]]: tensor<?x?x16x2xf32>
func.func @test_vectorize_dynamic_shapes_unpack_scalable_vec(%dest: tensor<?x?xf32>, %src: tensor<?x?x16x2xf32>) -> tensor<?x?xf32> {
- // CHECK: %[[C0:.*]] = arith.constant 0
- // CHECK: %[[DIM:.*]] = tensor.dim %[[DEST]], %[[C0]] : tensor<?x?xf32>
- // CHECK: %[[C1:.*]] = arith.constant 1 : index
- // CHECK: %[[DIM0:.*]] = tensor.dim %[[DEST]], %[[C1]] : tensor<?x?xf32>
// CHECK: %[[CST:.*]] = arith.constant 0.000000e+00
// CHECK: %[[C01:.*]] = arith.constant 0
// CHECK: %[[C02:.*]] = arith.constant 0
@@ -1011,10 +1006,6 @@ module attributes {transform.with_named_sequence} {
// CHECK-SAME: %[[DEST:.*]]: tensor<?x?xf32>,
// CHECK-SAME: %[[SRC:.*]]: tensor<?x?x?x2xf32>
func.func @test_vectorize_dynamic_shapes_unpack_scalable_vec_and_tile_size(%dest: tensor<?x?xf32>, %src: tensor<?x?x?x2xf32>) -> tensor<?x?xf32> {
- // CHECK: %[[C0:.*]] = arith.constant 0
- // CHECK: %[[DIM:.*]] = tensor.dim %[[DEST]], %[[C0]] : tensor<?x?xf32>
- // CHECK: %[[C1:.*]] = arith.constant 1 : index
- // CHECK: %[[DIM0:.*]] = tensor.dim %[[DEST]], %[[C1]] : tensor<?x?xf32>
// CHECK: %[[CST:.*]] = arith.constant 0.000000e+00
// CHECK: %[[C01:.*]] = arith.constant 0
// CHECK: %[[C02:.*]] = arith.constant 0
>From 78b335b06f64ccb809d02971dced1580c9aa2dad Mon Sep 17 00:00:00 2001
From: Andrzej Warzynski <andrzej.warzynski at arm.com>
Date: Fri, 25 Jul 2025 10:49:19 +0000
Subject: [PATCH 5/7] fixup! fixup! fixup! [mlir][linalg] Enable scalable
vectorization of linalg.unpack (WIP)
Remove unintended test change
---
mlir/test/Dialect/Linalg/vectorization/linalg-ops.mlir | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/mlir/test/Dialect/Linalg/vectorization/linalg-ops.mlir b/mlir/test/Dialect/Linalg/vectorization/linalg-ops.mlir
index fcb8b02d3faa3..9c9ddb54d1d5f 100644
--- a/mlir/test/Dialect/Linalg/vectorization/linalg-ops.mlir
+++ b/mlir/test/Dialect/Linalg/vectorization/linalg-ops.mlir
@@ -1239,7 +1239,7 @@ module attributes {transform.with_named_sequence} {
func.func @test_vectorize_padded_pack(%arg0: tensor<32x7x15xf32>, %arg1: tensor<32x4x1x16x2xf32>) -> tensor<32x4x1x16x2xf32> {
%pad = arith.constant 0.000000e+00 : f32
- %pack = linalg.pack %arg0 padding_value(%pad : f32) inner_dims_pos = [2, 1] inner_tiles = [16, [2]] into %arg1 : tensor<32x7x15xf32> -> tensor<32x4x1x16x2xf32>
+ %pack = linalg.pack %arg0 padding_value(%pad : f32) inner_dims_pos = [2, 1] inner_tiles = [16, 2] into %arg1 : tensor<32x7x15xf32> -> tensor<32x4x1x16x2xf32>
return %pack : tensor<32x4x1x16x2xf32>
}
// CHECK-DAG: %[[cst:.*]] = arith.constant 0.000000e+00 : f32
>From 724cc9a5b6a0fc5cc9a9aa81938e05370f0c5d4e Mon Sep 17 00:00:00 2001
From: Andrzej Warzynski <andrzej.warzynski at arm.com>
Date: Fri, 25 Jul 2025 10:56:00 +0000
Subject: [PATCH 6/7] fixup! fixup! fixup! fixup! [mlir][linalg] Enable
scalable vectorization of linalg.unpack (WIP)
Remove TODO
---
mlir/lib/Dialect/Linalg/Transforms/Vectorization.cpp | 1 -
1 file changed, 1 deletion(-)
diff --git a/mlir/lib/Dialect/Linalg/Transforms/Vectorization.cpp b/mlir/lib/Dialect/Linalg/Transforms/Vectorization.cpp
index 0b0e71e8951a6..8c3a7c3545667 100644
--- a/mlir/lib/Dialect/Linalg/Transforms/Vectorization.cpp
+++ b/mlir/lib/Dialect/Linalg/Transforms/Vectorization.cpp
@@ -2533,7 +2533,6 @@ vectorizeScalableVectorPrecondition(Operation *op,
if (numOfScalableDims == 0)
return success();
- // TODO: Check the following!
auto linalgOp = dyn_cast<LinalgOp>(op);
// Cond 1: Reject Ops that don't implement the LinalgOp interface, with the
>From 1d9f73778a74103bdfddb9112d2fcb7af8025981 Mon Sep 17 00:00:00 2001
From: Andrzej Warzynski <andrzej.warzynski at arm.com>
Date: Fri, 25 Jul 2025 10:57:56 +0000
Subject: [PATCH 7/7] fixup! fixup! fixup! fixup! fixup! [mlir][linalg] Enable
scalable vectorization of linalg.unpack (WIP)
Fix comment
---
mlir/lib/Dialect/Linalg/Transforms/Vectorization.cpp | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/mlir/lib/Dialect/Linalg/Transforms/Vectorization.cpp b/mlir/lib/Dialect/Linalg/Transforms/Vectorization.cpp
index 8c3a7c3545667..a8cc678396f57 100644
--- a/mlir/lib/Dialect/Linalg/Transforms/Vectorization.cpp
+++ b/mlir/lib/Dialect/Linalg/Transforms/Vectorization.cpp
@@ -1923,7 +1923,7 @@ vectorizeAsTensorUnpackOp(RewriterBase &rewriter, linalg::UnPackOp unpackOp,
// 1.2 Infer vector sizes for the read operation.
//
// The steps are:
- // 1. readVectorSizes = vectorInputSizes
+ // 1. readVectorSizes = writeVectorSizes
// 2. Take readVectorSizes from 1. and divide all locations pointed by
// the inner_dims_pos attribyte by the `inner_tiles` attribute value.
// 3. If outer_dims_perms is present, permutate readVectorSizes accordingly.
More information about the Mlir-commits
mailing list