[mlir][linalg] Add support for vectorizing dynamic elementwise named ops #71454
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
We are able to vectorize them in linalg.generic form. We just need to relax the condition, so it can also vectorize named ops.
hanhanW
requested review from
MaheshRavishankar,
nicolasvasilache and
dcaballe
|
@llvm/pr-subscribers-mlir Author: Han-Chung Wang (hanhanW) ChangesWe are able to vectorize them in linalg.generic form. We just need to relax the condition, so it can also vectorize named ops. Full diff: https://github.com/llvm/llvm-project/pull/71454.diff 2 Files Affected:
diff --git a/mlir/lib/Dialect/Linalg/Transforms/Vectorization.cpp b/mlir/lib/Dialect/Linalg/Transforms/Vectorization.cpp
index b427af33e3c4400..8b8eb5131669e13 100644
--- a/mlir/lib/Dialect/Linalg/Transforms/Vectorization.cpp
+++ b/mlir/lib/Dialect/Linalg/Transforms/Vectorization.cpp
@@ -1466,7 +1466,8 @@ static LogicalResult reductionPreconditions(LinalgOp op) {
static LogicalResult vectorizeDynamicLinalgOpPrecondition(linalg::LinalgOp op) {
// TODO: Masking only supports dynamic generic ops for now.
- if (!isa<linalg::GenericOp, linalg::FillOp, linalg::CopyOp,
+ if (!isElementwise(op) &&
+ !isa<linalg::GenericOp, linalg::FillOp, linalg::CopyOp,
linalg::ContractionOpInterface>(op.getOperation()))
return failure();
diff --git a/mlir/test/Dialect/Linalg/vectorization.mlir b/mlir/test/Dialect/Linalg/vectorization.mlir
index 7f4af344886f498..610339405d1c2c9 100644
--- a/mlir/test/Dialect/Linalg/vectorization.mlir
+++ b/mlir/test/Dialect/Linalg/vectorization.mlir
@@ -368,6 +368,34 @@ module attributes {transform.with_named_sequence} {
// -----
+// CHECK: #[[MAP:.*]] = affine_map<(d0, d1) -> (d1, d0)>
+// CHECK: func @test_masked_vectorize_linalg_transpose
+func.func @test_masked_vectorize_linalg_transpose(%arg0: tensor<?x?xf32>, %arg1: tensor<?x?xf32>) -> tensor<?x?xf32> {
+ // CHECK: %[[C0:.*]] = arith.constant 0 : index
+ // CHECK: %[[D0:.*]] = tensor.dim %arg0, %[[C0]]
+ // CHECK: %[[C1:.*]] = arith.constant 1 : index
+ // CHECK: %[[D1:.*]] = tensor.dim %arg0, %[[C1]]
+ // CHECK: %[[MASK0:.*]] = vector.create_mask %[[D0]], %[[D1]]
+ // CHECK: %[[LOAD:.*]] = vector.mask %[[MASK0]] { vector.transfer_read %arg0{{.+}} }
+ // CHECK-SAME: vector<2x4xi1> -> vector<2x4xf32>
+ // CHECK: %[[MASK1:.*]] = vector.create_mask %[[D1]], %[[D0]]
+ // CHECK: %[[WRITE:.*]] = vector.mask %[[MASK1]] { vector.transfer_write %[[LOAD]], %arg1{{.+}} permutation_map = #[[MAP]]{{.+}} }
+ // CHECK-SAME: vector<4x2xi1> -> tensor<?x?xf32>
+ // CHECK: return %[[WRITE]]
+ %0 = linalg.transpose ins(%arg0 : tensor<?x?xf32>) outs(%arg1 : tensor<?x?xf32>) permutation = [1, 0]
+ return %0 : tensor<?x?xf32>
+}
+
+module attributes {transform.with_named_sequence} {
+ transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {
+ %0 = transform.structured.match ops{["linalg.transpose"]} in %arg1 : (!transform.any_op) -> !transform.any_op
+ transform.structured.vectorize %0 vector_sizes [2, 4] : !transform.any_op
+ transform.yield
+ }
+}
+
+// -----
+
// CHECK-LABEL: func @test_masked_vectorize_linalg_copy
func.func @test_masked_vectorize_linalg_copy(%A : memref<?x?xf32>, %B : memref<?x?xf32>) {
// CHECK: %[[c0:.*]] = arith.constant 0 : index
|
|
@llvm/pr-subscribers-mlir-linalg Author: Han-Chung Wang (hanhanW) ChangesWe are able to vectorize them in linalg.generic form. We just need to relax the condition, so it can also vectorize named ops. Full diff: https://github.com/llvm/llvm-project/pull/71454.diff 2 Files Affected:
diff --git a/mlir/lib/Dialect/Linalg/Transforms/Vectorization.cpp b/mlir/lib/Dialect/Linalg/Transforms/Vectorization.cpp
index b427af33e3c4400..8b8eb5131669e13 100644
--- a/mlir/lib/Dialect/Linalg/Transforms/Vectorization.cpp
+++ b/mlir/lib/Dialect/Linalg/Transforms/Vectorization.cpp
@@ -1466,7 +1466,8 @@ static LogicalResult reductionPreconditions(LinalgOp op) {
static LogicalResult vectorizeDynamicLinalgOpPrecondition(linalg::LinalgOp op) {
// TODO: Masking only supports dynamic generic ops for now.
- if (!isa<linalg::GenericOp, linalg::FillOp, linalg::CopyOp,
+ if (!isElementwise(op) &&
+ !isa<linalg::GenericOp, linalg::FillOp, linalg::CopyOp,
linalg::ContractionOpInterface>(op.getOperation()))
return failure();
diff --git a/mlir/test/Dialect/Linalg/vectorization.mlir b/mlir/test/Dialect/Linalg/vectorization.mlir
index 7f4af344886f498..610339405d1c2c9 100644
--- a/mlir/test/Dialect/Linalg/vectorization.mlir
+++ b/mlir/test/Dialect/Linalg/vectorization.mlir
@@ -368,6 +368,34 @@ module attributes {transform.with_named_sequence} {
// -----
+// CHECK: #[[MAP:.*]] = affine_map<(d0, d1) -> (d1, d0)>
+// CHECK: func @test_masked_vectorize_linalg_transpose
+func.func @test_masked_vectorize_linalg_transpose(%arg0: tensor<?x?xf32>, %arg1: tensor<?x?xf32>) -> tensor<?x?xf32> {
+ // CHECK: %[[C0:.*]] = arith.constant 0 : index
+ // CHECK: %[[D0:.*]] = tensor.dim %arg0, %[[C0]]
+ // CHECK: %[[C1:.*]] = arith.constant 1 : index
+ // CHECK: %[[D1:.*]] = tensor.dim %arg0, %[[C1]]
+ // CHECK: %[[MASK0:.*]] = vector.create_mask %[[D0]], %[[D1]]
+ // CHECK: %[[LOAD:.*]] = vector.mask %[[MASK0]] { vector.transfer_read %arg0{{.+}} }
+ // CHECK-SAME: vector<2x4xi1> -> vector<2x4xf32>
+ // CHECK: %[[MASK1:.*]] = vector.create_mask %[[D1]], %[[D0]]
+ // CHECK: %[[WRITE:.*]] = vector.mask %[[MASK1]] { vector.transfer_write %[[LOAD]], %arg1{{.+}} permutation_map = #[[MAP]]{{.+}} }
+ // CHECK-SAME: vector<4x2xi1> -> tensor<?x?xf32>
+ // CHECK: return %[[WRITE]]
+ %0 = linalg.transpose ins(%arg0 : tensor<?x?xf32>) outs(%arg1 : tensor<?x?xf32>) permutation = [1, 0]
+ return %0 : tensor<?x?xf32>
+}
+
+module attributes {transform.with_named_sequence} {
+ transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {
+ %0 = transform.structured.match ops{["linalg.transpose"]} in %arg1 : (!transform.any_op) -> !transform.any_op
+ transform.structured.vectorize %0 vector_sizes [2, 4] : !transform.any_op
+ transform.yield
+ }
+}
+
+// -----
+
// CHECK-LABEL: func @test_masked_vectorize_linalg_copy
func.func @test_masked_vectorize_linalg_copy(%A : memref<?x?xf32>, %B : memref<?x?xf32>) {
// CHECK: %[[c0:.*]] = arith.constant 0 : index
|
dcaballe
approved these changes
Nov 6, 2023
| @@ -1466,7 +1466,8 @@ static LogicalResult reductionPreconditions(LinalgOp op) { | |||
|
|
|||
| static LogicalResult vectorizeDynamicLinalgOpPrecondition(linalg::LinalgOp op) { | |||
| // TODO: Masking only supports dynamic generic ops for now. | |||
There was a problem hiding this comment.
We could probably remove this TODO and add a comment on that these are the ops for which dynamic shapes vectorization is enabled.
There was a problem hiding this comment.
Good point, thanks! I also remove linalg.fill op from the list because it is an element-wise op.
|
thanks @hanhanW ! |
This was referenced Nov 8, 2023
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.
We are able to vectorize them in linalg.generic form. We just need to relax the condition, so it can also vectorize named ops.