[Mlir-commits] [mlir] [MLIR] Specify new padOp's output type in DropPadUnitDims (PR #150706)
Daniel Garvey
llvmlistbot at llvm.org
Tue Jul 29 15:00:09 PDT 2025
https://github.com/dan-garvey updated https://github.com/llvm/llvm-project/pull/150706
>From 617b335dec04d6927943b45b551bfbd5e0eaeca1 Mon Sep 17 00:00:00 2001
From: dan <danimal197 at gmail.com>
Date: Fri, 25 Jul 2025 16:15:46 -0500
Subject: [PATCH 1/2] [MLIR] Specify outputType in pad op unitDim drop
Previously when dropping unit dim from a pad with mixed
dynamic/static input/output shapes, the resulting shape
would take on the Type of the input, resulting in invalid IR
Signed-off-by: dan <danimal197 at gmail.com>
---
.../Linalg/Transforms/DropUnitDims.cpp | 14 +++--
.../Dialect/Linalg/drop-unit-extent-dims.mlir | 55 +++++++++++++------
2 files changed, 48 insertions(+), 21 deletions(-)
diff --git a/mlir/lib/Dialect/Linalg/Transforms/DropUnitDims.cpp b/mlir/lib/Dialect/Linalg/Transforms/DropUnitDims.cpp
index e0062d15e61ca..7194662145bcc 100644
--- a/mlir/lib/Dialect/Linalg/Transforms/DropUnitDims.cpp
+++ b/mlir/lib/Dialect/Linalg/Transforms/DropUnitDims.cpp
@@ -603,6 +603,7 @@ struct DropPadUnitDims : public OpRewritePattern<tensor::PadOp> {
}
ArrayRef<int64_t> sourceShape = padOp.getSourceType().getShape();
+ ArrayRef<int64_t> resultShape = padOp.getResultType().getShape();
int64_t padRank = sourceShape.size();
auto isStaticZero = [](OpFoldResult f) {
@@ -613,16 +614,18 @@ struct DropPadUnitDims : public OpRewritePattern<tensor::PadOp> {
allowedUnitDims.end());
llvm::SmallDenseSet<unsigned> unitDims;
SmallVector<int64_t> newShape;
+ SmallVector<int64_t> newResultShape;
SmallVector<OpFoldResult> newLowPad;
SmallVector<OpFoldResult> newHighPad;
- for (const auto [dim, size, low, high] :
- zip_equal(llvm::seq(static_cast<int64_t>(0), padRank), sourceShape,
- padOp.getMixedLowPad(), padOp.getMixedHighPad())) {
+ for (const auto [dim, size, outSize, low, high] : zip_equal(
+ llvm::seq(static_cast<int64_t>(0), padRank), sourceShape,
+ resultShape, padOp.getMixedLowPad(), padOp.getMixedHighPad())) {
if (unitDimsFilter.contains(dim) && size == 1 && isStaticZero(low) &&
isStaticZero(high)) {
unitDims.insert(dim);
} else {
newShape.push_back(size);
+ newResultShape.push_back(outSize);
newLowPad.push_back(low);
newHighPad.push_back(high);
}
@@ -652,8 +655,11 @@ struct DropPadUnitDims : public OpRewritePattern<tensor::PadOp> {
collapseValue(rewriter, padOp.getLoc(), padOp.getSource(), newShape,
reassociationMap, options.rankReductionStrategy);
+ auto resultType = RankedTensorType::get(
+ newResultShape, padOp.getResultType().getElementType()); //,
+ // padOp.getResultType().getEncoding());
auto newPadOp = rewriter.create<tensor::PadOp>(
- padOp.getLoc(), /*result=*/Type(), collapsedSource, newLowPad,
+ padOp.getLoc(), /*result=*/resultType, collapsedSource, newLowPad,
newHighPad, paddingVal, padOp.getNofold());
Value dest = padOp.getResult();
diff --git a/mlir/test/Dialect/Linalg/drop-unit-extent-dims.mlir b/mlir/test/Dialect/Linalg/drop-unit-extent-dims.mlir
index a00c798197e5a..5df113b038469 100644
--- a/mlir/test/Dialect/Linalg/drop-unit-extent-dims.mlir
+++ b/mlir/test/Dialect/Linalg/drop-unit-extent-dims.mlir
@@ -1076,6 +1076,44 @@ func.func @drop_known_unit_constant_low_high(%arg0: tensor<1x383x128xf32>) -> te
// -----
+func.func @drop_unit_dim_mixed_static_dynamic(%arg0: tensor<1x?xf32>) -> tensor<1x16xf32> {
+ %c0 = arith.constant 0 : index
+ %c1 = arith.constant 1 : index
+ %cst = arith.constant 0.000000e+00 : f32
+ %padded = tensor.pad %arg0 low[%c0, %c1] high[%c0, %c0] {
+ ^bb0(%arg1: index, %arg2: index):
+ tensor.yield %cst : f32
+ } : tensor<1x?xf32> to tensor<1x16xf32>
+ return %padded : tensor<1x16xf32>
+}
+// CHECK-LABEL: func @drop_unit_dim_mixed_static_dynamic
+// CHECK: %[[CST:.*]] = arith.constant 0.000000e+00 : f32
+// CHECK: %[[COLLAPSE:.+]] = tensor.collapse_shape
+// CHECK: %[[PADDED:.*]] = tensor.pad %[[COLLAPSE]] low[1] high[0] {
+// CHECK: ^bb0(%[[IDX:.*]]: index):
+// CHECK: tensor.yield %[[CST]] : f32
+// CHECK: } : tensor<?xf32> to tensor<16xf32>
+// CHECK: %[[EXPANDED:.*]] = tensor.expand_shape %[[PADDED]]
+// CHECK: return %[[EXPANDED]] : tensor<1x16xf32>
+
+// -----
+
+#map = affine_map<(d0, d1, d2, d3, d4, d5, d6) -> (d0, d1 + d4, d2 + d5, d6)>
+#map1 = affine_map<(d0, d1, d2, d3, d4, d5, d6) -> (d4, d5, d6, d3)>
+#map2 = affine_map<(d0, d1, d2, d3, d4, d5, d6) -> (d0, d1, d2, d3)>
+module {
+ func.func @drop_unit_dim_corresponding_to_dynamic_dim(%arg0: tensor<1x?x?x1xf32>, %arg1: index) -> tensor<?x1x61x1xf32> {
+ %cst = arith.constant dense<1.000000e+00> : tensor<1x1x1x1xf32>
+ %0 = tensor.empty(%arg1) : tensor<?x1x61x1xf32>
+ %1 = linalg.generic {indexing_maps = [#map, #map1, #map2], iterator_types = ["parallel", "parallel", "parallel", "parallel", "reduction", "reduction", "reduction"]} ins(%arg0, %cst : tensor<1x?x?x1xf32>, tensor<1x1x1x1xf32>) outs(%0 : tensor<?x1x61x1xf32>) {
+ ^bb0(%in: f32, %in_0: f32, %out: f32):
+ %2 = arith.mulf %in, %in_0 : f32
+ %3 = arith.addf %out, %2 : f32
+ linalg.yield %3 : f32
+ } -> tensor<?x1x61x1xf32>
+ return %1 : tensor<?x1x61x1xf32>
+ }
+}
// CHECK: #[[$MAP1:.+]] = affine_map<(d0) -> (0, d0)>
// CHECK: #[[$MAP2:.+]] = affine_map<(d0) -> ()>
@@ -1097,23 +1135,6 @@ func.func @drop_known_unit_constant_low_high(%arg0: tensor<1x383x128xf32>) -> te
// CHECK: return %[[VAL_14]] : tensor<?x1x61x1xf32>
// CHECK: }
-#map = affine_map<(d0, d1, d2, d3, d4, d5, d6) -> (d0, d1 + d4, d2 + d5, d6)>
-#map1 = affine_map<(d0, d1, d2, d3, d4, d5, d6) -> (d4, d5, d6, d3)>
-#map2 = affine_map<(d0, d1, d2, d3, d4, d5, d6) -> (d0, d1, d2, d3)>
-module {
- func.func @drop_unit_dim_corresponding_to_dynamic_dim(%arg0: tensor<1x?x?x1xf32>, %arg1: index) -> tensor<?x1x61x1xf32> {
- %cst = arith.constant dense<1.000000e+00> : tensor<1x1x1x1xf32>
- %0 = tensor.empty(%arg1) : tensor<?x1x61x1xf32>
- %1 = linalg.generic {indexing_maps = [#map, #map1, #map2], iterator_types = ["parallel", "parallel", "parallel", "parallel", "reduction", "reduction", "reduction"]} ins(%arg0, %cst : tensor<1x?x?x1xf32>, tensor<1x1x1x1xf32>) outs(%0 : tensor<?x1x61x1xf32>) {
- ^bb0(%in: f32, %in_0: f32, %out: f32):
- %2 = arith.mulf %in, %in_0 : f32
- %3 = arith.addf %out, %2 : f32
- linalg.yield %3 : f32
- } -> tensor<?x1x61x1xf32>
- return %1 : tensor<?x1x61x1xf32>
- }
-}
-
// -----
func.func @no_fold_empty_tensor_dim_out_of_bounds(%arg0: tensor<1x?x10xf32>) -> tensor<1x?xf32> {
>From a7b2f0de5c25d70a869fb9ea83227dba692830ea Mon Sep 17 00:00:00 2001
From: dan <danimal197 at gmail.com>
Date: Tue, 29 Jul 2025 16:58:32 -0500
Subject: [PATCH 2/2] address comments
---
mlir/lib/Dialect/Linalg/Transforms/DropUnitDims.cpp | 4 ++--
mlir/test/Dialect/Linalg/drop-unit-extent-dims.mlir | 4 ++--
2 files changed, 4 insertions(+), 4 deletions(-)
diff --git a/mlir/lib/Dialect/Linalg/Transforms/DropUnitDims.cpp b/mlir/lib/Dialect/Linalg/Transforms/DropUnitDims.cpp
index a66b704d9a0f4..bf66ed01ef158 100644
--- a/mlir/lib/Dialect/Linalg/Transforms/DropUnitDims.cpp
+++ b/mlir/lib/Dialect/Linalg/Transforms/DropUnitDims.cpp
@@ -689,10 +689,10 @@ struct DropPadUnitDims : public OpRewritePattern<tensor::PadOp> {
collapseValue(rewriter, padOp.getLoc(), padOp.getSource(), newShape,
reassociationMap, options.rankReductionStrategy);
- auto resultType = RankedTensorType::get(
+ auto newResultType = RankedTensorType::get(
newResultShape, padOp.getResultType().getElementType());
auto newPadOp = rewriter.create<tensor::PadOp>(
- padOp.getLoc(), /*result=*/resultType, collapsedSource, newLowPad,
+ padOp.getLoc(), /*result=*/newResultType, collapsedSource, newLowPad,
newHighPad, paddingVal, padOp.getNofold());
Value dest = padOp.getResult();
diff --git a/mlir/test/Dialect/Linalg/drop-unit-extent-dims.mlir b/mlir/test/Dialect/Linalg/drop-unit-extent-dims.mlir
index 5df113b038469..5f42938244db6 100644
--- a/mlir/test/Dialect/Linalg/drop-unit-extent-dims.mlir
+++ b/mlir/test/Dialect/Linalg/drop-unit-extent-dims.mlir
@@ -1088,12 +1088,12 @@ func.func @drop_unit_dim_mixed_static_dynamic(%arg0: tensor<1x?xf32>) -> tensor<
}
// CHECK-LABEL: func @drop_unit_dim_mixed_static_dynamic
// CHECK: %[[CST:.*]] = arith.constant 0.000000e+00 : f32
-// CHECK: %[[COLLAPSE:.+]] = tensor.collapse_shape
+// CHECK: %[[COLLAPSE:.+]] = tensor.collapse_shape %[[ARGS:.*]] : tensor<1x?xf32> into tensor<?xf32>
// CHECK: %[[PADDED:.*]] = tensor.pad %[[COLLAPSE]] low[1] high[0] {
// CHECK: ^bb0(%[[IDX:.*]]: index):
// CHECK: tensor.yield %[[CST]] : f32
// CHECK: } : tensor<?xf32> to tensor<16xf32>
-// CHECK: %[[EXPANDED:.*]] = tensor.expand_shape %[[PADDED]]
+// CHECK: %[[EXPANDED:.*]] = tensor.expand_shape %[[PADDED]] {{\[\[}}0, 1]] output_shape [1, 16] : tensor<16xf32> into tensor<1x16xf32>
// CHECK: return %[[EXPANDED]] : tensor<1x16xf32>
// -----
More information about the Mlir-commits
mailing list