[Mlir-commits] [mlir] [mlir] Fix bug in pack op canonicalization for folding dynamic dims (PR #82539)
llvmlistbot at llvm.org
llvmlistbot at llvm.org
Fri Feb 23 07:57:50 PST 2024
https://github.com/Max191 updated https://github.com/llvm/llvm-project/pull/82539
>From 86cec8cec4adb35edde36834111ad5740e14d091 Mon Sep 17 00:00:00 2001
From: Max Dawkins <max.dawkins at gmail.com>
Date: Wed, 21 Feb 2024 16:02:20 -0500
Subject: [PATCH 1/4] [mlir] Fix bug in pack op canonicalization for folding
dynamic dims
---
mlir/lib/Dialect/Tensor/IR/TensorOps.cpp | 8 +++++---
1 file changed, 5 insertions(+), 3 deletions(-)
diff --git a/mlir/lib/Dialect/Tensor/IR/TensorOps.cpp b/mlir/lib/Dialect/Tensor/IR/TensorOps.cpp
index e6efec14e31a60..b687bc8768056b 100644
--- a/mlir/lib/Dialect/Tensor/IR/TensorOps.cpp
+++ b/mlir/lib/Dialect/Tensor/IR/TensorOps.cpp
@@ -4012,15 +4012,17 @@ static bool inferStaticShape(PackOp packOp, SmallVectorImpl<int64_t> &srcShape,
llvm::SmallSetVector<int64_t, 4> innerDims;
innerDims.insert(packOp.getInnerDimsPos().begin(),
packOp.getInnerDimsPos().end());
- auto outerDimsPerm = packOp.getOuterDimsPerm();
+ SmallVector<int64_t> inverseOuterDimsPerm;
+ if (!packOp.getOuterDimsPerm().empty())
+ inverseOuterDimsPerm = invertPermutationVector(packOp.getOuterDimsPerm());
int srcRank = packOp.getSourceRank();
for (auto i : llvm::seq<int64_t>(0, srcRank)) {
if (innerDims.contains(i))
continue;
int64_t srcPos = i;
int64_t destPos = i;
- if (!outerDimsPerm.empty())
- destPos = outerDimsPerm[srcPos];
+ if (!inverseOuterDimsPerm.empty())
+ destPos = inverseOuterDimsPerm[srcPos];
if (ShapedType::isDynamic(srcShape[srcPos]) ==
ShapedType::isDynamic(destShape[destPos])) {
continue;
>From 3a5a83cb89ac931d3a9c55e9a5761e45b080a9ca Mon Sep 17 00:00:00 2001
From: Max Dawkins <max.dawkins at gmail.com>
Date: Wed, 21 Feb 2024 16:42:00 -0500
Subject: [PATCH 2/4] fix tests
---
mlir/test/Dialect/Tensor/canonicalize.mlir | 6 +++---
1 file changed, 3 insertions(+), 3 deletions(-)
diff --git a/mlir/test/Dialect/Tensor/canonicalize.mlir b/mlir/test/Dialect/Tensor/canonicalize.mlir
index e123c77aabd57c..5a754dd0d61cf5 100644
--- a/mlir/test/Dialect/Tensor/canonicalize.mlir
+++ b/mlir/test/Dialect/Tensor/canonicalize.mlir
@@ -822,7 +822,7 @@ func.func @infer_src_shape_pack(%src: tensor<?x?x?x?xf32>, %dest: tensor<10x20x3
// CHECK-LABEL: func.func @infer_src_shape_pack
// CHECK-SAME: %[[SRC:[0-9a-zA-Z]+]]
// CHECK-SAME: %[[DEST:[0-9a-zA-Z]+]]
-// CHECK: %[[CAST_SRC:.+]] = tensor.cast %[[SRC]] : tensor<?x?x?x?xf32> to tensor<30x20x?x10xf32>
+// CHECK: %[[CAST_SRC:.+]] = tensor.cast %[[SRC]] : tensor<?x?x?x?xf32> to tensor<40x20x?x30xf32>
// CHECK: %[[PACK:.+]] = tensor.pack %[[CAST_SRC]] {{.+}} into %[[DEST]]
// CHECK: return %[[PACK]]
@@ -841,9 +841,9 @@ func.func @infer_dest_shape_pack(%src: tensor<30x20x?x10xf32>, %dest: tensor<?x?
// CHECK-LABEL: func.func @infer_dest_shape_pack
// CHECK-SAME: %[[SRC:[0-9a-zA-Z]+]]
// CHECK-SAME: %[[DEST:[0-9a-zA-Z]+]]
-// CHECK: %[[CAST_DEST:.+]] = tensor.cast %[[DEST]] : tensor<?x?x?x?x16xf32> to tensor<10x20x30x?x16xf32>
+// CHECK: %[[CAST_DEST:.+]] = tensor.cast %[[DEST]] : tensor<?x?x?x?x16xf32> to tensor<?x20x10x30x16xf32>
// CHECK: %[[PACK:.+]] = tensor.pack %[[SRC]] {{.+}} into %[[CAST_DEST]]
-// CHECK: %[[CAST_PACK:.+]] = tensor.cast %[[PACK]] : tensor<10x20x30x?x16xf32> to tensor<?x?x?x?x16xf32>
+// CHECK: %[[CAST_PACK:.+]] = tensor.cast %[[PACK]] : tensor<?x20x10x30x16xf32> to tensor<?x?x?x?x16xf32>
// CHECK: return %[[CAST_PACK]]
// -----
>From 914a4cbc0a41dd3d0938ba4563bdd5a5defaa16a Mon Sep 17 00:00:00 2001
From: Max Dawkins <max.dawkins at gmail.com>
Date: Wed, 21 Feb 2024 16:59:49 -0500
Subject: [PATCH 3/4] fix unpack
---
mlir/lib/Dialect/Tensor/IR/TensorOps.cpp | 8 +++++---
mlir/test/Dialect/Tensor/canonicalize.mlir | 6 +++---
2 files changed, 8 insertions(+), 6 deletions(-)
diff --git a/mlir/lib/Dialect/Tensor/IR/TensorOps.cpp b/mlir/lib/Dialect/Tensor/IR/TensorOps.cpp
index b687bc8768056b..fe2f250e6b9290 100644
--- a/mlir/lib/Dialect/Tensor/IR/TensorOps.cpp
+++ b/mlir/lib/Dialect/Tensor/IR/TensorOps.cpp
@@ -4242,15 +4242,17 @@ static bool inferStaticShape(UnPackOp op, SmallVectorImpl<int64_t> &srcShape,
op.getDestType().getShape().end());
llvm::SmallSetVector<int64_t, 4> innerDims;
innerDims.insert(op.getInnerDimsPos().begin(), op.getInnerDimsPos().end());
- auto outerDimsPerm = op.getOuterDimsPerm();
+ SmallVector<int64_t> inverseOuterDimsPerm;
+ if (!op.getOuterDimsPerm().empty())
+ inverseOuterDimsPerm = invertPermutationVector(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 (!inverseOuterDimsPerm.empty())
+ srcPos = inverseOuterDimsPerm[destPos];
if (ShapedType::isDynamic(srcShape[srcPos]) ==
ShapedType::isDynamic(destShape[destPos])) {
continue;
diff --git a/mlir/test/Dialect/Tensor/canonicalize.mlir b/mlir/test/Dialect/Tensor/canonicalize.mlir
index 5a754dd0d61cf5..9e239f1874159a 100644
--- a/mlir/test/Dialect/Tensor/canonicalize.mlir
+++ b/mlir/test/Dialect/Tensor/canonicalize.mlir
@@ -920,9 +920,9 @@ func.func @infer_dest_shape_unpack(%src: tensor<10x20x30x40x16xf32>, %dest: tens
// 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: %[[CAST_DEST:.+]] = tensor.cast %[[DEST]] : tensor<?x?x?x?xf32> to tensor<40x20x?x30xf32>
// CHECK: %[[UNPACK:.+]] = tensor.unpack %[[SRC]] {{.+}} into %[[CAST_DEST]]
-// CHECK: %[[CAST_UNPACK:.+]] = tensor.cast %[[UNPACK]] : tensor<30x20x?x10xf32> to tensor<?x?x?x?xf32>
+// CHECK: %[[CAST_UNPACK:.+]] = tensor.cast %[[UNPACK]] : tensor<40x20x?x30xf32> to tensor<?x?x?x?xf32>
// CHECK: return %[[CAST_UNPACK]]
// -----
@@ -938,7 +938,7 @@ func.func @infer_src_shape_unpack(%src: tensor<?x?x?x?x16xf32>, %dest: tensor<30
// 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: %[[CAST_SRC:.+]] = tensor.cast %[[SRC]] : tensor<?x?x?x?x16xf32> to tensor<?x20x10x30x16xf32>
// CHECK: %[[UNPACK:.+]] = tensor.unpack %[[CAST_SRC]]
// CHECK: return %[[UNPACK]]
>From d454ff4c88af570ee4b8f6020e7cace02d83d861 Mon Sep 17 00:00:00 2001
From: Max Dawkins <max.dawkins at gmail.com>
Date: Wed, 21 Feb 2024 18:36:05 -0500
Subject: [PATCH 4/4] add new test cases
---
mlir/test/Dialect/Tensor/canonicalize.mlir | 23 ++++++++++++++++++++++
1 file changed, 23 insertions(+)
diff --git a/mlir/test/Dialect/Tensor/canonicalize.mlir b/mlir/test/Dialect/Tensor/canonicalize.mlir
index 9e239f1874159a..d17c23adfb14d8 100644
--- a/mlir/test/Dialect/Tensor/canonicalize.mlir
+++ b/mlir/test/Dialect/Tensor/canonicalize.mlir
@@ -848,6 +848,17 @@ func.func @infer_dest_shape_pack(%src: tensor<30x20x?x10xf32>, %dest: tensor<?x?
// -----
+func.func @no_infer_pack_shape(%arg0: tensor<?x32x100xf32>, %arg1: index) -> tensor<32x7x?x16x1xf32> {
+ %cst = arith.constant 0.000000e+00 : f32
+ %0 = tensor.empty(%arg1) : tensor<32x7x?x16x1xf32>
+ %pack = tensor.pack %arg0 padding_value(%cst : f32) outer_dims_perm = [1, 2, 0] inner_dims_pos = [2, 0] inner_tiles = [16, 1] into %0 : tensor<?x32x100xf32> -> tensor<32x7x?x16x1xf32>
+ return %pack : tensor<32x7x?x16x1xf32>
+}
+// CHECK-LABEL: func.func @no_infer_pack_shape
+// CHECK-NOT: tensor.cast
+
+// -----
+
func.func @fold_padding_value_pack_negative1(%arg0: tensor<1200x499999xf32>) -> tensor<31250x1200x16x1xf32> {
%cst = arith.constant 0.000000e+00 : f32
%0 = tensor.empty() : tensor<31250x1200x16x1xf32>
@@ -944,6 +955,18 @@ func.func @infer_src_shape_unpack(%src: tensor<?x?x?x?x16xf32>, %dest: tensor<30
// -----
+func.func @no_infer_unpack_shape(%arg1: tensor<32x7x?x16x1xf32>, %arg2: index) -> tensor<?x32x100xf32> {
+ %cst = arith.constant 0.000000e+00 : f32
+ %0 = tensor.empty(%arg2) : tensor<?x32x100xf32>
+ %unpack = tensor.unpack %arg1 outer_dims_perm = [1, 2, 0] inner_dims_pos = [2, 0] inner_tiles = [16, 1] into %0 : tensor<32x7x?x16x1xf32> -> tensor<?x32x100xf32>
+ return %unpack : tensor<?x32x100xf32>
+}
+// CHECK-LABEL: func.func @no_infer_unpack_shape
+// CHECK-NOT: tensor.cast
+
+// -----
+
+
// 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>) {
More information about the Mlir-commits
mailing list