Fix for TOSA-to-linalg lowering of tosa.transpose op #72698
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
|
@llvm/pr-subscribers-mlir-linalg @llvm/pr-subscribers-mlir Author: None (rafaelubalmw) ChangesThe TOSA-to-linalg conversion of Consider the following TOSA code using only static dimensions. The code transposes a tensor of shape 10x11x12 into 12x10x11 by permuting dimensions [2, 0, 1] into [0, 1, 2]. The code is correctly lowered to: Now let's make all dimensions dynamic in the TOSA code: The The output tensor shape is dynamically computed as 11x12x10 instead of 12x10x11. Since the total size of the output tensor is still the same, the code does not segfault after bufferization. However, index computations are invalid and lead to SWAs. Full diff: https://github.com/llvm/llvm-project/pull/72698.diff 2 Files Affected:
diff --git a/mlir/lib/Conversion/TosaToLinalg/TosaToLinalg.cpp b/mlir/lib/Conversion/TosaToLinalg/TosaToLinalg.cpp
index 3bf7bf12b5e96ff..ca37bd2b6643860 100644
--- a/mlir/lib/Conversion/TosaToLinalg/TosaToLinalg.cpp
+++ b/mlir/lib/Conversion/TosaToLinalg/TosaToLinalg.cpp
@@ -1072,12 +1072,11 @@ class TransposeConverter : public OpRewritePattern<tosa::TransposeOp> {
SmallVector<AffineExpr, 2> inputExprs;
inputExprs.resize(resultTy.getRank());
- auto operandTy = cast<ShapedType>(input.getType());
for (const auto &permutation : llvm::enumerate(perms.getValues<APInt>())) {
auto index = permutation.index();
auto value = permutation.value().getZExtValue();
- if (!operandTy.hasRank() || operandTy.isDynamicDim(index)) {
- dynDims[value] = rewriter.create<tensor::DimOp>(loc, input, index);
+ if (!resultTy.hasRank() || resultTy.isDynamicDim(index)) {
+ dynDims[index] = rewriter.create<tensor::DimOp>(loc, input, value);
}
inputExprs[value] = rewriter.getAffineDimExpr(index);
}
diff --git a/mlir/test/Conversion/TosaToLinalg/tosa-to-linalg.mlir b/mlir/test/Conversion/TosaToLinalg/tosa-to-linalg.mlir
index aa53b366f6da684..e0e041139fe4dc2 100644
--- a/mlir/test/Conversion/TosaToLinalg/tosa-to-linalg.mlir
+++ b/mlir/test/Conversion/TosaToLinalg/tosa-to-linalg.mlir
@@ -877,14 +877,14 @@ func.func @test_transpose_dyn(%arg0: tensor<1x?x3x4xi32>) -> () {
// CHECK: #[[$MAP0:.*]] = affine_map<(d0, d1) -> (d1, d0)>
// CHECK: #[[$MAP1:.*]] = affine_map<(d0, d1) -> (d0, d1)>
-// CHECK-LABEL: @test_transpose_dyn
+// CHECK-LABEL: @test_transpose_dyn_multiple_2d
// CHECK-SAME: (%[[ARG0:.+]]: tensor<?x?xf32>)
-func.func @test_transpose_dyn_multiple(%arg0: tensor<?x?xf32>) -> () {
+func.func @test_transpose_dyn_multiple_2d(%arg0: tensor<?x?xf32>) -> () {
%0 = arith.constant dense<[1, 0]> : tensor<2xi32>
- // CHECK: %[[C0:.+]] = arith.constant 0
- // CHECK: %[[DIM0:.+]] = tensor.dim %[[ARG0]], %[[C0]]
- // CHECK: %[[C1:.+]] = arith.constant 1
- // CHECK: %[[DIM1:.+]] = tensor.dim %[[ARG0]], %[[C1]]
+ // CHECK-DAG: %[[C0:.+]] = arith.constant 0
+ // CHECK-DAG: %[[DIM0:.+]] = tensor.dim %[[ARG0]], %[[C0]]
+ // CHECK-DAG: %[[C1:.+]] = arith.constant 1
+ // CHECK-DAG: %[[DIM1:.+]] = tensor.dim %[[ARG0]], %[[C1]]
// CHECK: %[[INIT:.+]] = tensor.empty(%[[DIM1]], %[[DIM0]])
// CHECK: %[[GENERIC:.+]] = linalg.generic {indexing_maps = [#[[$MAP0]], #[[$MAP1]]], iterator_types = ["parallel", "parallel"]} ins(%[[ARG0]] : tensor<?x?xf32>) outs([[OUT:%.+]] : tensor<?x?xf32>)
// CHECK: ^bb0([[ARG1:%.+]]: f32, [[ARG2:%.+]]: f32)
@@ -896,6 +896,29 @@ func.func @test_transpose_dyn_multiple(%arg0: tensor<?x?xf32>) -> () {
// -----
+// CHECK: #[[$MAP0:.+]] = affine_map<(d0, d1, d2) -> (d1, d2, d0)>
+// CHECK: #[[$MAP1:.+]] = affine_map<(d0, d1, d2) -> (d0, d1, d2)>
+
+// CHECK-LABEL: @test_transpose_dyn_multiple_3d
+// CHECK-SAME: (%[[ARG0:.+]]: tensor<?x?x?xf32>)
+func.func @test_transpose_dyn_multiple_3d(%arg0: tensor<?x?x?xf32>) {
+ %0 = arith.constant dense<[2, 0, 1]> : tensor<3xi32>
+ // CHECK-DAG: %[[C0:.*]] = arith.constant 0 : index
+ // CHECK-DAG: %[[DIM0:.*]] = tensor.dim %[[ARG0]], %[[C0]] : tensor<?x?x?xf32>
+ // CHECK-DAG: %[[C1:.*]] = arith.constant 1 : index
+ // CHECK-DAG: %[[DIM1:.*]] = tensor.dim %[[ARG0]], %[[C1]] : tensor<?x?x?xf32>
+ // CHECK-DAG: %[[C2:.*]] = arith.constant 2 : index
+ // CHECK-DAG: %[[DIM2:.*]] = tensor.dim %[[ARG0]], %[[C2]] : tensor<?x?x?xf32>
+ // CHECK: %[[INIT:.*]] = tensor.empty(%[[DIM2]], %[[DIM0]], %[[DIM1]]) : tensor<?x?x?xf32>
+ // CHECK: %[[GENERIC:.*]] = linalg.generic {indexing_maps = [#[[$MAP0]], #[[$MAP1]]], iterator_types = ["parallel", "parallel", "parallel"]} ins(%[[ARG0]] : tensor<?x?x?xf32>) outs(%[[INIT]] : tensor<?x?x?xf32>) {
+ // CHECK: ^bb0(%[[IN0:.*]]: f32, %[[OUT0:.*]]: f32):
+ // CHECK: linalg.yield %[[IN0]] : f32
+ // CHECK: } -> tensor<?x?x?xf32>
+ %1 = "tosa.transpose"(%arg0, %0) : (tensor<?x?x?xf32>, tensor<3xi32>) -> tensor<?x?x?xf32>
+ return
+}
+
+// -----
// CHECK-LABEL: @reduce_float
// CHECK-SAME: [[ARG0:%.+]]: tensor<5x4xf32>
|
matthias-springer
approved these changes
Nov 20, 2023
|
@matthias-springer Thank you for the review. If no additional approvals are expected, would you mind landing this change for me? |
|
@rafaelubalmw while I was trying to merge you patch I noticed that your email address is marked as private hence a noreply variant is set. Is this intentional? |
|
@GeorgeARM Thanks for pointing that out. I just updated my email pricacy settings. Please let me know if that addresses the issue. |
|
Thank you for addressing this @rafaelubalmw |
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 TOSA-to-linalg conversion of
tosa.transposecontains a bug in the computation of the result tensor shape when using dynamic dimensions. This bug may have widespread implications in projects such as Tensorflow, wheretosa.transposeis frequently generated.Consider the following TOSA code using only static dimensions. The code transposes a tensor of shape 10x11x12 into 12x10x11 by permuting dimensions [2, 0, 1] into [0, 1, 2].
The code is correctly lowered to:
Now let's make all dimensions dynamic in the TOSA code:
The
tensor.empty()op now needs additional information about the size of the output tensor, which is computed dynamically with a set oftensor.dimops. The comments below assume an input tensor of size 10x11x12, as before. The code is lowered as:The output tensor shape is dynamically computed as 11x12x10 instead of 12x10x11. Since the total size of the output tensor is still the same, the code does not segfault after bufferization. However, index computations are invalid and lead to SWAs.