[Mlir-commits] [mlir] [mlir][vector] Add mask elimination transform (PR #99314)
llvmlistbot at llvm.org
llvmlistbot at llvm.org
Wed Jul 17 05:38:34 PDT 2024
llvmbot wrote:
<!--LLVM PR SUMMARY COMMENT-->
@llvm/pr-subscribers-mlir
@llvm/pr-subscribers-mlir-vector
Author: Benjamin Maxwell (MacDue)
<details>
<summary>Changes</summary>
This adds a new transform `eliminateVectorMasks()` which aims at removing scalable `vector.create_masks` that will be all-true at runtime. It attempts to do this by simply pattern-matching the mask operands (similar to some canonicalizations), if that does not lead to an answer (is all-true? yes/no), then value bounds analysis will be used to find the lower bound of the unknown operands. If the lower bound is >= to the corresponding mask vector type dim, then that dimension of the mask is all true.
Note: Eliminating create_masks here means replacing them with all-true constants (which will then lead to the masks folding away).
---
Full diff: https://github.com/llvm/llvm-project/pull/99314.diff
5 Files Affected:
- (modified) mlir/include/mlir/Dialect/Vector/Transforms/VectorTransforms.h (+17)
- (modified) mlir/lib/Dialect/Vector/Transforms/CMakeLists.txt (+1)
- (added) mlir/lib/Dialect/Vector/Transforms/VectorMaskElimination.cpp (+117)
- (added) mlir/test/Dialect/Vector/eliminate-masks.mlir (+138)
- (modified) mlir/test/lib/Dialect/Vector/TestVectorTransforms.cpp (+34)
``````````diff
diff --git a/mlir/include/mlir/Dialect/Vector/Transforms/VectorTransforms.h b/mlir/include/mlir/Dialect/Vector/Transforms/VectorTransforms.h
index 1f7d6411cd5a4..847f333d6a931 100644
--- a/mlir/include/mlir/Dialect/Vector/Transforms/VectorTransforms.h
+++ b/mlir/include/mlir/Dialect/Vector/Transforms/VectorTransforms.h
@@ -11,6 +11,7 @@
#include "mlir/Dialect/Vector/Transforms/VectorRewritePatterns.h"
#include "mlir/Dialect/Vector/Utils/VectorUtils.h"
+#include "mlir/Interfaces/FunctionInterfaces.h"
namespace mlir {
class MLIRContext;
@@ -115,6 +116,22 @@ castAwayContractionLeadingOneDim(vector::ContractionOp contractOp,
MaskingOpInterface maskingOp,
RewriterBase &rewriter);
+/// Structure to hold the range [vscaleMin, vscaleMax] `vector.vscale` can take.
+struct VscaleRange {
+ unsigned vscaleMin;
+ unsigned vscaleMax;
+};
+
+/// Attempts to eliminate redundant vector masks by replacing them with all-true
+/// constants at the top of the function (which results in the masks folding
+/// away). Note: Currently, this only runs for vector.create_mask ops and
+/// requires `vscaleRange`. If `vscaleRange` is not provided this transform does
+/// nothing. This is because these redundant masks are much more likely for
+/// scalable code which requires memref/tensor dynamic sizes, whereas fixed-size
+/// code has static sizes, so simpler folds remove the masks.
+void eliminateVectorMasks(IRRewriter &rewriter, FunctionOpInterface function,
+ std::optional<VscaleRange> vscaleRange = {});
+
} // namespace vector
} // namespace mlir
diff --git a/mlir/lib/Dialect/Vector/Transforms/CMakeLists.txt b/mlir/lib/Dialect/Vector/Transforms/CMakeLists.txt
index 723b2f62d65d4..2639a67e1c8b3 100644
--- a/mlir/lib/Dialect/Vector/Transforms/CMakeLists.txt
+++ b/mlir/lib/Dialect/Vector/Transforms/CMakeLists.txt
@@ -22,6 +22,7 @@ add_mlir_dialect_library(MLIRVectorTransforms
VectorTransferSplitRewritePatterns.cpp
VectorTransforms.cpp
VectorUnroll.cpp
+ VectorMaskElimination.cpp
ADDITIONAL_HEADER_DIRS
${MLIR_MAIN_INCLUDE_DIR}/mlir/Dialect/Vector/Transforms
diff --git a/mlir/lib/Dialect/Vector/Transforms/VectorMaskElimination.cpp b/mlir/lib/Dialect/Vector/Transforms/VectorMaskElimination.cpp
new file mode 100644
index 0000000000000..abec8c75b8fc9
--- /dev/null
+++ b/mlir/lib/Dialect/Vector/Transforms/VectorMaskElimination.cpp
@@ -0,0 +1,117 @@
+#include "mlir/Dialect/Arith/IR/Arith.h"
+#include "mlir/Dialect/Utils/StaticValueUtils.h"
+#include "mlir/Dialect/Vector/IR/ScalableValueBoundsConstraintSet.h"
+#include "mlir/Dialect/Vector/Transforms/VectorRewritePatterns.h"
+#include "mlir/Dialect/Vector/Transforms/VectorTransforms.h"
+#include "mlir/Interfaces/FunctionInterfaces.h"
+
+using namespace mlir;
+using namespace mlir::vector;
+namespace {
+
+/// If `value` is a constant multiple of `vector.vscale` return the multiplier.
+std::optional<int64_t> getConstantVscaleMultiplier(Value value) {
+ if (value.getDefiningOp<vector::VectorScaleOp>())
+ return 1;
+ auto mul = value.getDefiningOp<arith::MulIOp>();
+ if (!mul)
+ return {};
+ auto lhs = mul.getLhs();
+ auto rhs = mul.getRhs();
+ if (lhs.getDefiningOp<vector::VectorScaleOp>())
+ return getConstantIntValue(rhs);
+ if (rhs.getDefiningOp<vector::VectorScaleOp>())
+ return getConstantIntValue(lhs);
+ return {};
+}
+
+/// Attempts to resolve a (scalable) CreateMaskOp to an all-true constant mask.
+/// All-true masks can then be eliminated by simple folds.
+LogicalResult resolveAllTrueCreateMaskOp(IRRewriter &rewriter,
+ vector::CreateMaskOp createMaskOp,
+ VscaleRange vscaleRange) {
+ auto maskType = createMaskOp.getVectorType();
+ auto maskTypeDimScalableFlags = maskType.getScalableDims();
+ auto maskTypeDimSizes = maskType.getShape();
+
+ struct UnknownMaskDim {
+ size_t position;
+ Value dimSize;
+ };
+
+ // Check for any dims that could be (partially) false before doing the more
+ // expensive value bounds computations.
+ SmallVector<UnknownMaskDim> unknownDims;
+ for (auto [i, dimSize] : llvm::enumerate(createMaskOp.getOperands())) {
+ if (auto intSize = getConstantIntValue(dimSize)) {
+ // Mask not all-true for this dim.
+ if (maskTypeDimScalableFlags[i] || intSize < maskTypeDimSizes[i])
+ return failure();
+ } else if (auto vscaleMultiplier = getConstantVscaleMultiplier(dimSize)) {
+ // Mask not all-true for this dim.
+ if (vscaleMultiplier < maskTypeDimSizes[i])
+ return failure();
+ } else {
+ // Unknown (without further analysis).
+ unknownDims.push_back(UnknownMaskDim{i, dimSize});
+ }
+ }
+
+ for (auto [i, dimSize] : unknownDims) {
+ // Compute the lower bound for the unknown dimension (i.e. the smallest
+ // value it could be).
+ auto lowerBound =
+ vector::ScalableValueBoundsConstraintSet::computeScalableBound(
+ dimSize, {}, vscaleRange.vscaleMin, vscaleRange.vscaleMax,
+ presburger::BoundType::LB);
+ if (failed(lowerBound))
+ return failure();
+ auto boundSize = lowerBound->getSize();
+ if (failed(boundSize))
+ return failure();
+ if (boundSize->scalable) {
+ // If the lower bound is scalable and >= to the mask dim size then this
+ // dim is all-true.
+ if (boundSize->baseSize < maskTypeDimSizes[i])
+ return failure();
+ } else {
+ // If the lower bound is a constant and >= to the _fixed-size_ mask dim
+ // size then this dim is all-true.
+ if (maskTypeDimScalableFlags[i])
+ return failure();
+ if (boundSize->baseSize < maskTypeDimSizes[i])
+ return failure();
+ }
+ }
+
+ // Replace createMaskOp with an all-true constant. This should result in the
+ // mask being removed in most cases (as xfer ops + vector.mask have folds to
+ // remove all-true masks).
+ auto allTrue = rewriter.create<arith::ConstantOp>(
+ createMaskOp.getLoc(), maskType, DenseElementsAttr::get(maskType, true));
+ rewriter.replaceAllUsesWith(createMaskOp, allTrue);
+ return success();
+}
+
+} // namespace
+
+namespace mlir::vector {
+
+void eliminateVectorMasks(IRRewriter &rewriter, FunctionOpInterface function,
+ std::optional<VscaleRange> vscaleRange) {
+ // TODO: Support fixed-size case. This is less likely to be useful as for
+ // fixed-size code dimensions are all static so masks tend to fold away.
+ if (!vscaleRange)
+ return;
+
+ OpBuilder::InsertionGuard g(rewriter);
+ SmallVector<vector::CreateMaskOp> worklist;
+ function.walk([&](vector::CreateMaskOp createMaskOp) {
+ worklist.push_back(createMaskOp);
+ });
+ rewriter.setInsertionPointToStart(&function.front());
+ for (auto mask : worklist)
+ (void)resolveAllTrueCreateMaskOp(rewriter, mask, *vscaleRange);
+}
+
+} // namespace mlir::vector
diff --git a/mlir/test/Dialect/Vector/eliminate-masks.mlir b/mlir/test/Dialect/Vector/eliminate-masks.mlir
new file mode 100644
index 0000000000000..99c9a60a09fac
--- /dev/null
+++ b/mlir/test/Dialect/Vector/eliminate-masks.mlir
@@ -0,0 +1,138 @@
+// RUN: mlir-opt %s -split-input-file -test-eliminate-vector-masks | FileCheck %s
+
+// This tests a general pattern the vectorizer tends to emit.
+
+// CHECK-LABEL: @eliminate_redundant_masks_through_insert_and_extracts
+// CHECK: %[[ALL_TRUE_MASK:.*]] = arith.constant dense<true> : vector<[4]xi1>
+// CHECK: vector.transfer_read {{.*}} %[[ALL_TRUE_MASK]]
+// CHECK: vector.transfer_write {{.*}} %[[ALL_TRUE_MASK]]
+func.func @eliminate_redundant_masks_through_insert_and_extracts(%tensor: tensor<1x1000xf32>) {
+ %c0 = arith.constant 0 : index
+ %c4 = arith.constant 4 : index
+ %c1000 = arith.constant 1000 : index
+ %c0_f32 = arith.constant 0.0 : f32
+ %vscale = vector.vscale
+ %c4_vscale = arith.muli %vscale, %c4 : index
+ %extracted_slice_0 = tensor.extract_slice %tensor[0, 0] [1, %c4_vscale] [1, 1] : tensor<1x1000xf32> to tensor<1x?xf32>
+ %output_tensor = scf.for %i = %c0 to %c1000 step %c4_vscale iter_args(%arg = %extracted_slice_0) -> tensor<1x?xf32> {
+ // 1. Extract a slice.
+ %extracted_slice_1 = tensor.extract_slice %arg[0, 0] [1, %c4_vscale] [1, 1] : tensor<1x?xf32> to tensor<?xf32>
+
+ // 2. Create a mask for the slice.
+ %dim_1 = tensor.dim %extracted_slice_1, %c0 : tensor<?xf32>
+ %mask = vector.create_mask %dim_1 : vector<[4]xi1>
+
+ // 3. Read the slice and do some computation.
+ %vec = vector.transfer_read %extracted_slice_1[%c0], %c0_f32, %mask {in_bounds = [true]} : tensor<?xf32>, vector<[4]xf32>
+ %new_vec = "test.some_computation"(%vec) : (vector<[4]xf32>) -> (vector<[4]xf32>)
+
+ // 4. Write the new value.
+ %write = vector.transfer_write %new_vec, %extracted_slice_1[%c0], %mask {in_bounds = [true]} : vector<[4]xf32>, tensor<?xf32>
+
+ // 5. Insert and yield the new tensor value.
+ %result = tensor.insert_slice %write into %arg[0, 0] [1, %c4_vscale] [1, 1] : tensor<?xf32> into tensor<1x?xf32>
+ scf.yield %result : tensor<1x?xf32>
+ }
+ "test.some_use"(%output_tensor) : (tensor<1x?xf32>) -> ()
+ return
+}
+
+// -----
+
+// CHECK-LABEL: @negative_extract_slice_size_shrink
+// CHECK-NOT: arith.constant dense<true> : vector<[4]xi1>
+// CHECK: %[[MASK:.*]] = vector.create_mask
+// CHECK: "test.some_use"(%[[MASK]]) : (vector<[4]xi1>) -> ()
+func.func @negative_extract_slice_size_shrink(%tensor: tensor<1000xf32>) {
+ %c0 = arith.constant 0 : index
+ %c4 = arith.constant 4 : index
+ %c1000 = arith.constant 1000 : index
+ %vscale = vector.vscale
+ %c4_vscale = arith.muli %vscale, %c4 : index
+ %extracted_slice = tensor.extract_slice %tensor[0] [%c4_vscale] [1] : tensor<1000xf32> to tensor<?xf32>
+ %slice = scf.for %i = %c0 to %c1000 step %c4_vscale iter_args(%arg = %extracted_slice) -> tensor<?xf32> {
+ // This mask cannot be eliminated even though looking at the above operations
+ // it appears `tensor.dim` will always be c4_vscale (so the mask all-true).
+ %dim = tensor.dim %arg, %c0 : tensor<?xf32>
+ %mask = vector.create_mask %dim : vector<[4]xi1>
+ "test.some_use"(%mask) : (vector<[4]xi1>) -> ()
+ // !!! Here the size of the mask could shrink in the next iteration.
+ %next_num_els = affine.min affine_map<(d0)[s0] -> (-d0 + 1000, s0)>(%i)[%c4_vscale]
+ %new_extracted_slice = tensor.extract_slice %tensor[%c4_vscale] [%next_num_els] [1] : tensor<1000xf32> to tensor<?xf32>
+ scf.yield %new_extracted_slice : tensor<?xf32>
+ }
+ "test.some_use"(%slice) : (tensor<?xf32>) -> ()
+ return
+}
+
+// -----
+
+// CHECK-LABEL: @negative_constant_dim_not_all_true
+// CHECK-NOT: arith.constant dense<true> : vector<2x[4]xi1>
+// CHECK: %[[MASK:.*]] = vector.create_mask
+// CHECK: "test.some_use"(%[[MASK]]) : (vector<2x[4]xi1>) -> ()
+func.func @negative_constant_dim_not_all_true()
+{
+ %c1 = arith.constant 1 : index
+ %c4 = arith.constant 4 : index
+ %vscale = vector.vscale
+ %c4_vscale = arith.muli %vscale, %c4 : index
+ %mask = vector.create_mask %c1, %c4_vscale : vector<2x[4]xi1>
+ "test.some_use"(%mask) : (vector<2x[4]xi1>) -> ()
+ return
+}
+
+// -----
+
+// CHECK-LABEL: @negative_constant_vscale_multiple_not_all_true
+// CHECK-NOT: arith.constant dense<true> : vector<2x[4]xi1>
+// CHECK: %[[MASK:.*]] = vector.create_mask
+// CHECK: "test.some_use"(%[[MASK]]) : (vector<2x[4]xi1>) -> ()
+func.func @negative_constant_vscale_multiple_not_all_true() {
+ %c2 = arith.constant 2 : index
+ %c3 = arith.constant 3 : index
+ %vscale = vector.vscale
+ %c3_vscale = arith.muli %vscale, %c3 : index
+ %mask = vector.create_mask %c2, %c3_vscale : vector<2x[4]xi1>
+ "test.some_use"(%mask) : (vector<2x[4]xi1>) -> ()
+ return
+}
+
+// -----
+
+// CHECK-LABEL: @negative_value_bounds_fixed_dim_not_all_true
+// CHECK-NOT: arith.constant dense<true> : vector<3x[4]xi1>
+// CHECK: %[[MASK:.*]] = vector.create_mask
+// CHECK: "test.some_use"(%[[MASK]]) : (vector<3x[4]xi1>) -> ()
+func.func @negative_value_bounds_fixed_dim_not_all_true(%tensor: tensor<2x?xf32>)
+{
+ %c0 = arith.constant 0 : index
+ %c4 = arith.constant 4 : index
+ %vscale = vector.vscale
+ %c4_vscale = arith.muli %vscale, %c4 : index
+ // This is _very_ simple but since addi is not a constant value bounds will
+ // be used to resolve it.
+ %dim = tensor.dim %tensor, %c0 : tensor<2x?xf32>
+ %mask = vector.create_mask %dim, %c4_vscale : vector<3x[4]xi1>
+ "test.some_use"(%mask) : (vector<3x[4]xi1>) -> ()
+ return
+}
+
+// -----
+
+// CHECK-LABEL: @negative_value_bounds_scalable_dim_not_all_true
+// CHECK-NOT: arith.constant dense<true> : vector<3x[4]xi1>
+// CHECK: %[[MASK:.*]] = vector.create_mask
+// CHECK: "test.some_use"(%[[MASK]]) : (vector<3x[4]xi1>) -> ()
+func.func @negative_value_bounds_scalable_dim_not_all_true(%tensor: tensor<2x100xf32>) {
+ %c1 = arith.constant 1 : index
+ %c3 = arith.constant 3 : index
+ %vscale = vector.vscale
+ %c3_vscale = arith.muli %vscale, %c3 : index
+ %slice = tensor.extract_slice %tensor[0, 0] [2, %c3_vscale] [1, 1] : tensor<2x100xf32> to tensor<2x?xf32>
+ // Another simple example, but value bounds will be used to resolve the tensor.dim.
+ %dim = tensor.dim %slice, %c1 : tensor<2x?xf32>
+ %mask = vector.create_mask %c3, %dim : vector<3x[4]xi1>
+ "test.some_use"(%mask) : (vector<3x[4]xi1>) -> ()
+ return
+}
diff --git a/mlir/test/lib/Dialect/Vector/TestVectorTransforms.cpp b/mlir/test/lib/Dialect/Vector/TestVectorTransforms.cpp
index c978699e179fc..f74ff2725f815 100644
--- a/mlir/test/lib/Dialect/Vector/TestVectorTransforms.cpp
+++ b/mlir/test/lib/Dialect/Vector/TestVectorTransforms.cpp
@@ -874,6 +874,38 @@ struct TestVectorLinearize final
return signalPassFailure();
}
};
+
+struct TestEliminateVectorMasks
+ : public PassWrapper<TestEliminateVectorMasks,
+ OperationPass<func::FuncOp>> {
+ MLIR_DEFINE_EXPLICIT_INTERNAL_INLINE_TYPE_ID(TestEliminateVectorMasks)
+
+ TestEliminateVectorMasks() = default;
+ TestEliminateVectorMasks(const TestEliminateVectorMasks &pass)
+ : PassWrapper(pass) {}
+
+ Option<unsigned> vscaleMin{
+ *this, "vscale-min",
+ llvm::cl::desc(
+ "Minimum value `vector.vscale` can possibly be at runtime."),
+ llvm::cl::init(1)};
+
+ Option<unsigned> vscaleMax{
+ *this, "vscale-max",
+ llvm::cl::desc(
+ "Maximum value `vector.vscale` can possibly be at runtime."),
+ llvm::cl::init(16)};
+
+ StringRef getArgument() const final { return "test-eliminate-vector-masks"; }
+ StringRef getDescription() const final {
+ return "Test eliminating vector masks";
+ }
+ void runOnOperation() override {
+ IRRewriter rewriter(&getContext());
+ eliminateVectorMasks(rewriter, getOperation(),
+ VscaleRange{vscaleMin, vscaleMax});
+ }
+};
} // namespace
namespace mlir {
@@ -920,6 +952,8 @@ void registerTestVectorLowerings() {
PassRegistration<TestVectorEmulateMaskedLoadStore>();
PassRegistration<TestVectorLinearize>();
+
+ PassRegistration<TestEliminateVectorMasks>();
}
} // namespace test
} // namespace mlir
``````````
</details>
https://github.com/llvm/llvm-project/pull/99314
More information about the Mlir-commits
mailing list