[mlir][Linalg] Fix non-matmul linalg structured ops #116412
Merged
+79
−65
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
Groverkss
requested review from
dcaballe,
nicolasvasilache and
rengolin
as code owners
|
I need to add a test for these changes. Would the batch matmul test in the description be enough? |
MaheshRavishankar
approved these changes
Nov 15, 2024
|
The test failures look like parsing/printing issue. |
I fixed them, but they are coming because linalg.matmul indexing_maps attributes are being passed to matvec. This is fine for now, as long as they are printed as discardable attributes (which they are). |
|
@llvm/pr-subscribers-mlir-linalg @llvm/pr-subscribers-mlir Author: Kunwar Grover (Groverkss) Changes3ad0148 broke linalg structured ops other than MatmulOp. The patch:
These changes combined, hide the problem until you print the operation in it's generic form. Reproducer: func.func public @<!-- -->bug(%arg0 : tensor<5x10x20xf32>, %arg1 : tensor<5x20x40xf32>, %arg3 : tensor<5x10x40xf32>) -> tensor<5x10x40xf32> {
%out = linalg.batch_matmul ins(%arg0, %arg1 : tensor<5x10x20xf32>, tensor<5x20x40xf32>)
outs(%arg3 : tensor<5x10x40xf32>) -> tensor<5x10x40xf32>
func.return %out : tensor<5x10x40xf32>
}Prints fine, with The batch_matmul operation's builder now always inserts a indexing_map which is unrelated to the operation itself. This was caught when a transformation from one LinalgStructuredOp to another, tried to pass it's attributes to the other ops builder and there were multiple indexing_map attributes in the result. This patch fixes this by specializing the builders for MatmulOp with indexing map information. Full diff: https://github.com/llvm/llvm-project/pull/116412.diff 4 Files Affected:
diff --git a/mlir/include/mlir/Dialect/Linalg/IR/LinalgStructuredOps.td b/mlir/include/mlir/Dialect/Linalg/IR/LinalgStructuredOps.td
index e578f4b956ef5e..a90777c82bf63a 100644
--- a/mlir/include/mlir/Dialect/Linalg/IR/LinalgStructuredOps.td
+++ b/mlir/include/mlir/Dialect/Linalg/IR/LinalgStructuredOps.td
@@ -621,7 +621,7 @@ def MatmulOp : LinalgStructuredBase_Op<"matmul", [
(ins "ValueRange":$inputs, "ValueRange":$outputs,
CArg<"ArrayRef<NamedAttribute>", "{}">:$attributes),
[{
- buildStructuredOp($_builder, $_state, std::nullopt, inputs, outputs,
+ buildMatmulOp($_builder, $_state, std::nullopt, inputs, outputs,
attributes, MatmulOp::getRegionBuilder());
}]>,
OpBuilder<
@@ -629,7 +629,7 @@ def MatmulOp : LinalgStructuredBase_Op<"matmul", [
"ValueRange":$outputs,
CArg<"ArrayRef<NamedAttribute>", "{}">:$attributes),
[{
- buildStructuredOp($_builder, $_state, resultTensorTypes,
+ buildMatmulOp($_builder, $_state, resultTensorTypes,
inputs, outputs, attributes, MatmulOp::getRegionBuilder());
}]>,
OpBuilder<
@@ -647,7 +647,7 @@ def MatmulOp : LinalgStructuredBase_Op<"matmul", [
"Attribute":$cast, CArg<"ArrayRef<NamedAttribute>", "{}">:$attributes),
[{
$_state.addAttribute("cast", cast);
- buildStructuredOp($_builder, $_state, resultTensorTypes, inputs, outputs,
+ buildMatmulOp($_builder, $_state, resultTensorTypes, inputs, outputs,
attributes, MatmulOp::getRegionBuilder());
}]>
diff --git a/mlir/lib/Dialect/Linalg/IR/LinalgOps.cpp b/mlir/lib/Dialect/Linalg/IR/LinalgOps.cpp
index c909d13e4314b4..dee8a4e27e6b26 100644
--- a/mlir/lib/Dialect/Linalg/IR/LinalgOps.cpp
+++ b/mlir/lib/Dialect/Linalg/IR/LinalgOps.cpp
@@ -169,7 +169,8 @@ getDefaultIndexingMapsForMatmul(MLIRContext *context) {
}
/// Wrapper to return the typical indexing map array attribute for MatmulOp.
-static SmallVector<Attribute> getDefaultIndexingMapAttr(MLIRContext *context) {
+static SmallVector<Attribute>
+getDefaultMatmulIndexingMapAttr(MLIRContext *context) {
return llvm::map_to_vector(
getDefaultIndexingMapsForMatmul(context),
[](AffineMap map) -> Attribute { return AffineMapAttr::get(map); });
@@ -179,12 +180,11 @@ static SmallVector<Attribute> getDefaultIndexingMapAttr(MLIRContext *context) {
/// The result types are derived automatically if `resultTensorTypes` is none.
/// The body of the operation is filled using `regionBuilder`. All ods-gen
/// created structured operations use the method to implement their builders.
-static void buildStructuredOp(
- OpBuilder &b, OperationState &state,
- std::optional<TypeRange> resultTensorTypes, ValueRange inputs,
- ValueRange outputs, ArrayRef<NamedAttribute> attributes,
- RegionBuilderFn regionBuilder,
- std::optional<ArrayRef<AffineMap>> indexingMaps = std::nullopt) {
+static void buildStructuredOp(OpBuilder &b, OperationState &state,
+ std::optional<TypeRange> resultTensorTypes,
+ ValueRange inputs, ValueRange outputs,
+ ArrayRef<NamedAttribute> attributes,
+ RegionBuilderFn regionBuilder) {
// Derive the result types if needed.
SmallVector<Type> derivedResultTypes =
resultTensorTypes.value_or(TypeRange());
@@ -196,6 +196,24 @@ static void buildStructuredOp(
state.addOperands(outputs);
state.addTypes(derivedResultTypes);
+ state.addAttributes(attributes);
+ state.addAttribute(
+ "operandSegmentSizes",
+ b.getDenseI32ArrayAttr({static_cast<int32_t>(inputs.size()),
+ static_cast<int32_t>(outputs.size())}));
+
+ // Create and fill the region of the structured operation.
+ Region ®ion = *state.addRegion();
+ fillStructuredOpRegion(b, region, TypeRange(inputs), TypeRange(outputs),
+ state.attributes.getAttrs(), regionBuilder);
+}
+
+static void
+buildMatmulOp(OpBuilder &b, OperationState &state,
+ std::optional<TypeRange> resultTensorTypes, ValueRange inputs,
+ ValueRange outputs, ArrayRef<NamedAttribute> attributes,
+ RegionBuilderFn regionBuilder,
+ std::optional<ArrayRef<AffineMap>> indexingMaps = std::nullopt) {
// Initialize indexingMaps, for MatmulOp.
SmallVector<Attribute, 3> indexingMapsAttrVal;
if (indexingMaps.has_value()) {
@@ -205,20 +223,11 @@ static void buildStructuredOp(
}
state.addAttribute("indexing_maps", b.getArrayAttr(indexingMapsAttrVal));
} else {
- indexingMapsAttrVal = getDefaultIndexingMapAttr(b.getContext());
+ indexingMapsAttrVal = getDefaultMatmulIndexingMapAttr(b.getContext());
state.addAttribute("indexing_maps", b.getArrayAttr(indexingMapsAttrVal));
}
-
- state.addAttributes(attributes);
- state.addAttribute(
- "operandSegmentSizes",
- b.getDenseI32ArrayAttr({static_cast<int32_t>(inputs.size()),
- static_cast<int32_t>(outputs.size())}));
-
- // Create and fill the region of the structured operation.
- Region ®ion = *state.addRegion();
- fillStructuredOpRegion(b, region, TypeRange(inputs), TypeRange(outputs),
- state.attributes.getAttrs(), regionBuilder);
+ return buildStructuredOp(b, state, resultTensorTypes, inputs, outputs,
+ attributes, regionBuilder);
}
/// Common parsing used for both named structured ops created by ods-gen and by
@@ -340,39 +349,6 @@ static ParseResult parseNamedStructuredOp(OpAsmParser &parser,
OperationState &result,
unsigned numRegionArgs,
RegionBuilderFn regionBuilder) {
-
- SmallVector<Attribute, 3> indexingMapsAttr;
- Attribute mapAttr;
- if (succeeded(parser.parseOptionalKeyword("indexing_maps"))) {
- if (parser.parseEqual())
- return failure();
-
- if (parser.parseLSquare())
- return failure();
-
- do {
- if (parser.parseAttribute(mapAttr))
- return failure();
- if (!isa<AffineMapAttr>(mapAttr)) {
- return parser.emitError(parser.getCurrentLocation(),
- "expected affine map attribute");
- }
- indexingMapsAttr.push_back(mapAttr);
-
- if (parser.parseOptionalComma())
- break;
- } while (true);
-
- if (parser.parseRSquare())
- return failure();
- }
- // Initialize indexingMaps, if not supplied explicitly.
- if (indexingMapsAttr.empty()) {
- indexingMapsAttr = getDefaultIndexingMapAttr(result.getContext());
- }
- result.addAttribute("indexing_maps",
- parser.getBuilder().getArrayAttr(indexingMapsAttr));
-
// TODO: Enable when ods-gen supports captures.
SmallVector<Type, 1> inputTypes, outputTypes;
if (parseCommonStructuredOpParts(parser, result, inputTypes, outputTypes))
@@ -3503,9 +3479,11 @@ static LogicalResult verifyExtendedMatmulSemantic(MatmulOp matmulOp,
namespace mlir {
namespace linalg {
+
//===----------------------------------------------------------------------===//
// MatMulOp
//===----------------------------------------------------------------------===//
+
SmallVector<utils::IteratorType> MatmulOp::getIteratorTypesArray() {
return SmallVector<utils::IteratorType>{utils::IteratorType::parallel,
utils::IteratorType::parallel,
@@ -3520,8 +3498,8 @@ std::string MatmulOp::getLibraryCallName() {
bool MatmulOp::hasDynamicIndexingMaps() { return true; }
-/// Check if the op has broadcast and/or transpose semantic. Returns true if the
-/// user defined indexing maps are not equal to default map.
+/// Check if the op has broadcast and/or transpose semantic. Returns true if
+/// the user defined indexing maps are not equal to default map.
bool MatmulOp::hasUserDefinedMaps() {
SmallVector<AffineMap, 3> defaultMaps = getDefaultIndexingMaps();
SmallVector<AffineMap, 3> explicitMaps = getIndexingMapsArray();
@@ -3557,7 +3535,8 @@ void MatmulOp::regionBuilder(ImplicitLocOpBuilder &b, Block &block,
helper.yieldOutputs(yields);
}
-/// Returns a list of AffineMap with the typical matmul indexing charactristic.
+/// Returns a list of AffineMap with the typical matmul indexing
+/// charactristic.
SmallVector<AffineMap> MatmulOp::getDefaultIndexingMaps() {
MLIRContext *context = this->getContext();
return getDefaultIndexingMapsForMatmul(context);
@@ -3572,6 +3551,38 @@ bool MatmulOp::isValidLhsRhsBroadcastMap(AffineMap bcastMap) {
}
ParseResult MatmulOp::parse(OpAsmParser &parser, OperationState &result) {
+ SmallVector<Attribute, 3> indexingMapsAttr;
+ Attribute mapAttr;
+ if (succeeded(parser.parseOptionalKeyword("indexing_maps"))) {
+ if (parser.parseEqual())
+ return failure();
+
+ if (parser.parseLSquare())
+ return failure();
+
+ do {
+ if (parser.parseAttribute(mapAttr))
+ return failure();
+ if (!isa<AffineMapAttr>(mapAttr)) {
+ return parser.emitError(parser.getCurrentLocation(),
+ "expected affine map attribute");
+ }
+ indexingMapsAttr.push_back(mapAttr);
+
+ if (parser.parseOptionalComma())
+ break;
+ } while (true);
+
+ if (parser.parseRSquare())
+ return failure();
+ }
+ // Initialize indexingMaps, if not supplied explicitly.
+ if (indexingMapsAttr.empty()) {
+ indexingMapsAttr = getDefaultMatmulIndexingMapAttr(result.getContext());
+ }
+ result.addAttribute("indexing_maps",
+ parser.getBuilder().getArrayAttr(indexingMapsAttr));
+
return parseNamedStructuredOp(parser, result, MatmulOp::getNumRegionArgs(),
MatmulOp::getRegionBuilder());
}
@@ -3582,7 +3593,7 @@ void MatmulOp::print(OpAsmPrinter &p) {
elidedAttrs);
SmallVector<Attribute, 3> indexingMaps =
- getDefaultIndexingMapAttr(getContext());
+ getDefaultMatmulIndexingMapAttr(getContext());
if (!llvm::equal(getIndexingMaps(), indexingMaps)) {
p << " indexing_maps = [";
llvm::interleaveComma(getIndexingMaps(), p,
diff --git a/mlir/test/Dialect/Linalg/rank-reduce-contraction-ops.mlir b/mlir/test/Dialect/Linalg/rank-reduce-contraction-ops.mlir
index c086d0fd7e6332..ebdbe70ff46eb7 100644
--- a/mlir/test/Dialect/Linalg/rank-reduce-contraction-ops.mlir
+++ b/mlir/test/Dialect/Linalg/rank-reduce-contraction-ops.mlir
@@ -43,7 +43,8 @@ func.func @singleton_batch_matvec(%arg0 : tensor<1x128x512xf32>, %arg1 : tensor<
// CHECK-NEXT: %[[COLLAPSED_LHS:.*]] = tensor.collapse_shape %[[LHS]] {{\[}}[0, 1], [2]]
// CHECK-NEXT: %[[COLLAPSED_RHS:.*]] = tensor.collapse_shape %[[RHS]] {{\[}}[0, 1]]
// CHECK-NEXT: %[[COLLAPSED_INIT:.*]] = tensor.collapse_shape %[[INIT]] {{\[}}[0, 1]]
- // CHECK-NEXT: %[[MATMUL:.+]] = linalg.matvec ins(%[[COLLAPSED_LHS]], %[[COLLAPSED_RHS]] : tensor<128x512xf32>, tensor<512xf32>) outs(%[[COLLAPSED_INIT]] : tensor<128xf32>)
+ // CHECK-NEXT: %[[MATMUL:.+]] = linalg.matvec
+ // CHECK-SAME: ins(%[[COLLAPSED_LHS]], %[[COLLAPSED_RHS]] : tensor<128x512xf32>, tensor<512xf32>) outs(%[[COLLAPSED_INIT]] : tensor<128xf32>)
// CHECK-NEXT: %[[RES:.*]] = tensor.expand_shape %[[MATMUL]] {{\[}}[0, 1]] output_shape [1, 128]
// CHECK-NEXT: return %[[RES]]
%1 = linalg.batch_matvec ins(%arg0, %arg1 : tensor<1x128x512xf32>, tensor<1x512xf32>)
@@ -62,7 +63,8 @@ func.func @singleton_batch_vecmat(%arg0 : tensor<1x?xf32>, %arg1 : tensor<1x?x?x
// CHECK-NEXT: %[[COLLAPSED_LHS:.*]] = tensor.collapse_shape %[[LHS]] {{\[}}[0, 1]]
// CHECK-NEXT: %[[COLLAPSED_RHS:.*]] = tensor.collapse_shape %[[RHS]] {{\[}}[0, 1], [2]]
// CHECK-NEXT: %[[COLLAPSED_INIT:.*]] = tensor.collapse_shape %[[INIT]] {{\[}}[0, 1]]
- // CHECK-NEXT: %[[MATMUL:.+]] = linalg.vecmat ins(%[[COLLAPSED_LHS]], %[[COLLAPSED_RHS]] : tensor<?xf32>, tensor<?x?xf32>) outs(%[[COLLAPSED_INIT]] : tensor<?xf32>)
+ // CHECK-NEXT: %[[MATMUL:.+]] = linalg.vecmat
+ // CHECK-SAME: ins(%[[COLLAPSED_LHS]], %[[COLLAPSED_RHS]] : tensor<?xf32>, tensor<?x?xf32>) outs(%[[COLLAPSED_INIT]] : tensor<?xf32>)
// CHECK-NEXT: %[[DIM1:.*]] = tensor.dim %[[INIT]], %[[C1]]
// CHECK-NEXT: %[[RES:.*]] = tensor.expand_shape %[[MATMUL]] {{\[}}[0, 1]] output_shape [1, %[[DIM1]]]
// CHECK-NEXT: return %[[RES]]
@@ -113,7 +115,8 @@ func.func @matmul_to_matvec_tensor(%arg0: tensor<?x?xf32>, %arg1: tensor<?x1xf32
// CHECK-DAG: %[[C0:.*]] = arith.constant 0
// CHECK-NEXT: %[[COLLAPSED_RHS:.*]] = tensor.collapse_shape %[[RHS]] {{\[}}[0, 1]]
// CHECK-NEXT: %[[COLLAPSED_INIT:.*]] = tensor.collapse_shape %[[INIT]] {{\[}}[0, 1]]
- // CHECK-NEXT: %[[MATMUL:.+]] = linalg.matvec ins(%[[LHS]], %[[COLLAPSED_RHS]] : tensor<?x?xf32>, tensor<?xf32>) outs(%[[COLLAPSED_INIT]] : tensor<?xf32>)
+ // CHECK-NEXT: %[[MATMUL:.+]] = linalg.matvec
+ // CHECK-SAME: ins(%[[LHS]], %[[COLLAPSED_RHS]] : tensor<?x?xf32>, tensor<?xf32>) outs(%[[COLLAPSED_INIT]] : tensor<?xf32>)
// CHECK-NEXT: %[[DIM0:.*]] = tensor.dim %[[INIT]], %[[C0]]
// CHECK-NEXT: %[[RES:.*]] = tensor.expand_shape %[[MATMUL]] {{\[}}[0, 1]] output_shape [%[[DIM0]], 1]
// CHECK-NEXT: return %[[RES]]
@@ -140,7 +143,8 @@ func.func @matmul_to_vecmat_tensor(%arg0: tensor<1x?xf32>, %arg1: tensor<?x?xf32
// CHECK-DAG: %[[C1:.*]] = arith.constant 1
// CHECK-NEXT: %[[COLLAPSED_LHS:.*]] = tensor.collapse_shape %[[LHS]] {{\[}}[0, 1]]
// CHECK-NEXT: %[[COLLAPSED_INIT:.*]] = tensor.collapse_shape %[[INIT]] {{\[}}[0, 1]]
- // CHECK-NEXT: %[[RESULT:.*]] = linalg.vecmat ins(%[[COLLAPSED_LHS]], %[[RHS]] : tensor<?xf32>, tensor<?x?xf32>) outs(%[[COLLAPSED_INIT]] : tensor<?xf32>)
+ // CHECK-NEXT: %[[RESULT:.*]] = linalg.vecmat
+ // CHECK-SAME: ins(%[[COLLAPSED_LHS]], %[[RHS]] : tensor<?xf32>, tensor<?x?xf32>) outs(%[[COLLAPSED_INIT]] : tensor<?xf32>)
// CHECK-NEXT: %[[DIM1:.*]] = tensor.dim %[[INIT]], %[[C1]]
// CHECK-NEXT: %[[RES:.*]] = tensor.expand_shape %[[RESULT]] {{\[}}[0, 1]] output_shape [1, %[[DIM1]]]
// CHECK-NEXT: return %[[RES]]
diff --git a/mlir/tools/mlir-linalg-ods-gen/mlir-linalg-ods-yaml-gen.cpp b/mlir/tools/mlir-linalg-ods-gen/mlir-linalg-ods-yaml-gen.cpp
index 6be7d4320c6562..80d979864921d6 100644
--- a/mlir/tools/mlir-linalg-ods-gen/mlir-linalg-ods-yaml-gen.cpp
+++ b/mlir/tools/mlir-linalg-ods-gen/mlir-linalg-ods-yaml-gen.cpp
@@ -679,8 +679,7 @@ ParseResult {0}::parse(OpAsmParser &parser, OperationState &result) {{
}
void {0}::print(OpAsmPrinter &p) {{
SmallVector<StringRef, 3> elidedAttrs = {{"operandSegmentSizes",
- "linalg.memoized_indexing_maps",
- "indexing_maps"};
+ "linalg.memoized_indexing_maps"};
::printNamedStructuredOp(p, getOperation(), getInputs(), getOutputs(),
elidedAttrs);
}
|
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.
3ad0148 broke linalg structured ops other than MatmulOp.
The patch:
These changes combined, hide the problem until you print the operation in it's generic form.
Reproducer:
Prints fine, with
mlir-opt <file>, but if you domlir-opt --mlir-print-op-generic <file>:The batch_matmul operation's builder now always inserts a indexing_map which is unrelated to the operation itself. This was caught when a transformation from one LinalgStructuredOp to another, tried to pass it's attributes to the other ops builder and there were multiple indexing_map attributes in the result.
This patch fixes this by specializing the builders for MatmulOp with indexing map information.