[mlir][Vector] Move mask materialization patterns to greedy rewrite #119973
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
matthias-springer
requested review from
banach-space,
dcaballe and
nicolasvasilache
as code owners
|
@llvm/pr-subscribers-mlir Author: Matthias Springer (matthias-springer) ChangesThe mask materialization patterns during The current combination of rewrite patterns and conversion patterns triggered an edge case when merging the 1:1 and 1:N dialect conversions. Full diff: https://github.com/llvm/llvm-project/pull/119973.diff 4 Files Affected:
diff --git a/mlir/lib/Conversion/VectorToLLVM/ConvertVectorToLLVMPass.cpp b/mlir/lib/Conversion/VectorToLLVM/ConvertVectorToLLVMPass.cpp
index 4623b9667998cc..6e02cd2dad8149 100644
--- a/mlir/lib/Conversion/VectorToLLVM/ConvertVectorToLLVMPass.cpp
+++ b/mlir/lib/Conversion/VectorToLLVM/ConvertVectorToLLVMPass.cpp
@@ -61,8 +61,8 @@ struct ConvertVectorToLLVMPass
} // namespace
void ConvertVectorToLLVMPass::runOnOperation() {
- // Perform progressive lowering of operations on slices and
- // all contraction operations. Also applies folding and DCE.
+ // Perform progressive lowering of operations on slices and all contraction
+ // operations. Also materializes masks, applies folding and DCE.
{
RewritePatternSet patterns(&getContext());
populateVectorToVectorCanonicalizationPatterns(patterns);
@@ -76,6 +76,7 @@ void ConvertVectorToLLVMPass::runOnOperation() {
VectorTransformsOptions());
// Vector transfer ops with rank > 1 should be lowered with VectorToSCF.
populateVectorTransferLoweringPatterns(patterns, /*maxTransferRank=*/1);
+ populateVectorMaskMaterializationPatterns(patterns, force32BitVectorIndices);
(void)applyPatternsAndFoldGreedily(getOperation(), std::move(patterns));
}
@@ -83,7 +84,6 @@ void ConvertVectorToLLVMPass::runOnOperation() {
LowerToLLVMOptions options(&getContext());
LLVMTypeConverter converter(&getContext(), options);
RewritePatternSet patterns(&getContext());
- populateVectorMaskMaterializationPatterns(patterns, force32BitVectorIndices);
populateVectorTransferLoweringPatterns(patterns);
populateVectorToLLVMMatrixConversionPatterns(converter, patterns);
populateVectorToLLVMConversionPatterns(
diff --git a/mlir/test/Conversion/VectorToLLVM/vector-mask-to-llvm.mlir b/mlir/test/Conversion/VectorToLLVM/vector-mask-to-llvm.mlir
index 82351eb7c98a43..91e5358622b69d 100644
--- a/mlir/test/Conversion/VectorToLLVM/vector-mask-to-llvm.mlir
+++ b/mlir/test/Conversion/VectorToLLVM/vector-mask-to-llvm.mlir
@@ -7,7 +7,7 @@
// CMP32: %[[T1:.*]] = arith.index_cast %[[ARG]] : index to i32
// CMP32: %[[T2:.*]] = llvm.insertelement %[[T1]], %{{.*}}[%{{.*}} : i32] : vector<11xi32>
// CMP32: %[[T3:.*]] = llvm.shufflevector %[[T2]], %{{.*}} [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0] : vector<11xi32>
-// CMP32: %[[T4:.*]] = arith.cmpi slt, %[[T0]], %[[T3]] : vector<11xi32>
+// CMP32: %[[T4:.*]] = arith.cmpi sgt, %[[T3]], %[[T0]] : vector<11xi32>
// CMP32: return %[[T4]] : vector<11xi1>
// CMP64-LABEL: @genbool_var_1d(
@@ -16,7 +16,7 @@
// CMP64: %[[T1:.*]] = arith.index_cast %[[ARG]] : index to i64
// CMP64: %[[T2:.*]] = llvm.insertelement %[[T1]], %{{.*}}[%{{.*}} : i32] : vector<11xi64>
// CMP64: %[[T3:.*]] = llvm.shufflevector %[[T2]], %{{.*}} [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0] : vector<11xi64>
-// CMP64: %[[T4:.*]] = arith.cmpi slt, %[[T0]], %[[T3]] : vector<11xi64>
+// CMP64: %[[T4:.*]] = arith.cmpi sgt, %[[T3]], %[[T0]] : vector<11xi64>
// CMP64: return %[[T4]] : vector<11xi1>
func.func @genbool_var_1d(%arg0: index) -> vector<11xi1> {
diff --git a/mlir/test/Conversion/VectorToLLVM/vector-to-llvm.mlir b/mlir/test/Conversion/VectorToLLVM/vector-to-llvm.mlir
index 2473fe933ffcb2..ea88fece9e662d 100644
--- a/mlir/test/Conversion/VectorToLLVM/vector-to-llvm.mlir
+++ b/mlir/test/Conversion/VectorToLLVM/vector-to-llvm.mlir
@@ -3097,7 +3097,7 @@ func.func @create_mask_0d(%num_elems : index) -> vector<i1> {
// CHECK: %[[NUM_ELEMS_i32:.*]] = arith.index_cast %[[NUM_ELEMS]] : index to i32
// CHECK: %[[BOUNDS:.*]] = llvm.insertelement %[[NUM_ELEMS_i32]]
// CHECK: %[[BOUNDS_CAST:.*]] = builtin.unrealized_conversion_cast %[[BOUNDS]] : vector<1xi32> to vector<i32>
-// CHECK: %[[RESULT:.*]] = arith.cmpi slt, %[[INDICES]], %[[BOUNDS_CAST]] : vector<i32>
+// CHECK: %[[RESULT:.*]] = arith.cmpi sgt, %[[BOUNDS_CAST]], %[[INDICES]] : vector<i32>
// CHECK: return %[[RESULT]] : vector<i1>
// -----
@@ -3113,7 +3113,7 @@ func.func @create_mask_1d(%num_elems : index) -> vector<4xi1> {
// CHECK: %[[NUM_ELEMS_i32:.*]] = arith.index_cast %[[NUM_ELEMS]] : index to i32
// CHECK: %[[BOUNDS_INSERT:.*]] = llvm.insertelement %[[NUM_ELEMS_i32]]
// CHECK: %[[BOUNDS:.*]] = llvm.shufflevector %[[BOUNDS_INSERT]]
-// CHECK: %[[RESULT:.*]] = arith.cmpi slt, %[[INDICES]], %[[BOUNDS]] : vector<4xi32>
+// CHECK: %[[RESULT:.*]] = arith.cmpi sgt, %[[BOUNDS]], %[[INDICES]] : vector<4xi32>
// CHECK: return %[[RESULT]] : vector<4xi1>
// -----
diff --git a/mlir/test/Conversion/VectorToLLVM/vector-xfer-to-llvm.mlir b/mlir/test/Conversion/VectorToLLVM/vector-xfer-to-llvm.mlir
index 8f01cc2b8d44c3..d3f6d7eca90b41 100644
--- a/mlir/test/Conversion/VectorToLLVM/vector-xfer-to-llvm.mlir
+++ b/mlir/test/Conversion/VectorToLLVM/vector-xfer-to-llvm.mlir
@@ -14,30 +14,28 @@ func.func @transfer_read_write_1d(%A : memref<?xf32>, %base: index) -> vector<17
// CHECK-LABEL: func @transfer_read_write_1d
// CHECK-SAME: %[[MEM:.*]]: memref<?xf32>,
// CHECK-SAME: %[[BASE:.*]]: index) -> vector<17xf32>
-// CHECK: %[[C7:.*]] = arith.constant 7.0
-//
-// 1. Let dim be the memref dimension, compute the in-bound index (dim - offset)
-// CHECK: %[[C0:.*]] = arith.constant 0 : index
-// CHECK: %[[DIM:.*]] = memref.dim %[[MEM]], %[[C0]] : memref<?xf32>
-// CHECK: %[[BOUND:.*]] = arith.subi %[[DIM]], %[[BASE]] : index
+// 1. Create pass-through vector.
+// CHECK-DAG: %[[PASS_THROUGH:.*]] = arith.constant dense<7.000000e+00> : vector<17xf32>
//
// 2. Create a vector with linear indices [ 0 .. vector_length - 1 ].
-// CHECK: %[[linearIndex:.*]] = arith.constant dense
+// CHECK-DAG: %[[linearIndex:.*]] = arith.constant dense
// CHECK-SAME: <[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16]> : vector<17x[[$IDX_TYPE]]>
//
-// 3. Create bound vector to compute in-bound mask:
+// 3. Let dim be the memref dimension, compute the in-bound index (dim - offset)
+// CHECK-DAG: %[[C0:.*]] = arith.constant 0 : index
+// CHECK: %[[DIM:.*]] = memref.dim %[[MEM]], %[[C0]] : memref<?xf32>
+// CHECK: %[[BOUND:.*]] = arith.subi %[[DIM]], %[[BASE]] : index
+//
+// 4. Create bound vector to compute in-bound mask:
// [ 0 .. vector_length - 1 ] < [ dim - offset .. dim - offset ]
// CHECK: %[[btrunc:.*]] = arith.index_cast %[[BOUND]] :
// CMP32-SAME: index to i32
// CMP64-SAME: index to i64
// CHECK: %[[boundVecInsert:.*]] = llvm.insertelement %[[btrunc]]
// CHECK: %[[boundVect:.*]] = llvm.shufflevector %[[boundVecInsert]]
-// CHECK: %[[mask:.*]] = arith.cmpi slt, %[[linearIndex]], %[[boundVect]] : vector<17x[[$IDX_TYPE]]>
+// CHECK: %[[mask:.*]] = arith.cmpi sgt, %[[boundVect]], %[[linearIndex]] : vector<17x[[$IDX_TYPE]]>
// CMP64-SAME: : vector<17xi64>
//
-// 4. Create pass-through vector.
-// CHECK: %[[PASS_THROUGH:.*]] = arith.constant dense<7.{{.*}}> : vector<17xf32>
-//
// 5. Bitcast to vector form.
// CHECK: %[[gep:.*]] = llvm.getelementptr %{{.*}} :
// CHECK-SAME: (!llvm.ptr, i64) -> !llvm.ptr, f32
@@ -48,28 +46,23 @@ func.func @transfer_read_write_1d(%A : memref<?xf32>, %base: index) -> vector<17
// CHECK-SAME: -> vector<17xf32>
//
// 1. Let dim be the memref dimension, compute the in-bound index (dim - offset)
-// CHECK: %[[C0_b:.*]] = arith.constant 0 : index
-// CHECK: %[[DIM_b:.*]] = memref.dim %[[MEM]], %[[C0_b]] : memref<?xf32>
+// CHECK: %[[DIM_b:.*]] = memref.dim %[[MEM]], %[[C0]] : memref<?xf32>
// CHECK: %[[BOUND_b:.*]] = arith.subi %[[DIM_b]], %[[BASE]] : index
//
-// 2. Create a vector with linear indices [ 0 .. vector_length - 1 ].
-// CHECK: %[[linearIndex_b:.*]] = arith.constant dense
-// CHECK-SAME: <[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16]> : vector<17x[[$IDX_TYPE]]>
-//
-// 3. Create bound vector to compute in-bound mask:
+// 2. Create bound vector to compute in-bound mask:
// [ 0 .. vector_length - 1 ] < [ dim - offset .. dim - offset ]
// CHECK: %[[btrunc_b:.*]] = arith.index_cast %[[BOUND_b]]
// CMP32-SAME: index to i32
// CHECK: %[[boundVecInsert_b:.*]] = llvm.insertelement %[[btrunc_b]]
// CHECK: %[[boundVect_b:.*]] = llvm.shufflevector %[[boundVecInsert_b]]
-// CHECK: %[[mask_b:.*]] = arith.cmpi slt, %[[linearIndex_b]],
-// CHECK-SAME: %[[boundVect_b]] : vector<17x[[$IDX_TYPE]]>
+// CHECK: %[[mask_b:.*]] = arith.cmpi sgt, %[[boundVect_b]],
+// CHECK-SAME: %[[linearIndex]] : vector<17x[[$IDX_TYPE]]>
//
-// 4. Bitcast to vector form.
+// 3. Bitcast to vector form.
// CHECK: %[[gep_b:.*]] = llvm.getelementptr {{.*}} :
// CHECK-SAME: (!llvm.ptr, i64) -> !llvm.ptr, f32
//
-// 5. Rewrite as a masked write.
+// 4. Rewrite as a masked write.
// CHECK: llvm.intr.masked.store %[[loaded]], %[[gep_b]], %[[mask_b]]
// CHECK-SAME: {alignment = 4 : i32} :
// CHECK-SAME: vector<17xf32>, vector<17xi1> into !llvm.ptr
@@ -87,17 +80,18 @@ func.func @transfer_read_write_1d_scalable(%A : memref<?xf32>, %base: index) ->
// CHECK-LABEL: func @transfer_read_write_1d_scalable
// CHECK-SAME: %[[MEM:.*]]: memref<?xf32>,
// CHECK-SAME: %[[BASE:.*]]: index) -> vector<[17]xf32>
-// CHECK: %[[C7:.*]] = arith.constant 7.0
+// 1. Create pass-through vector.
+// CHECK-DAG: %[[PASS_THROUGH:.*]] = arith.constant dense<7.000000e+00> : vector<[17]xf32>
//
-// 1. Let dim be the memref dimension, compute the in-bound index (dim - offset)
-// CHECK: %[[C0:.*]] = arith.constant 0 : index
+// 2. Let dim be the memref dimension, compute the in-bound index (dim - offset)
+// CHECK-DAG: %[[C0:.*]] = arith.constant 0 : index
// CHECK: %[[DIM:.*]] = memref.dim %[[MEM]], %[[C0]] : memref<?xf32>
// CHECK: %[[BOUND:.*]] = arith.subi %[[DIM]], %[[BASE]] : index
//
-// 2. Create a vector with linear indices [ 0 .. vector_length - 1 ].
+// 3. Create a vector with linear indices [ 0 .. vector_length - 1 ].
// CHECK: %[[linearIndex:.*]] = llvm.intr.stepvector : vector<[17]x[[$IDX_TYPE]]>
//
-// 3. Create bound vector to compute in-bound mask:
+// 4. Create bound vector to compute in-bound mask:
// [ 0 .. vector_length - 1 ] < [ dim - offset .. dim - offset ]
// CHECK: %[[btrunc:.*]] = arith.index_cast %[[BOUND]] : index to [[$IDX_TYPE]]
// CHECK: %[[boundVecInsert:.*]] = llvm.insertelement %[[btrunc]]
@@ -105,9 +99,6 @@ func.func @transfer_read_write_1d_scalable(%A : memref<?xf32>, %base: index) ->
// CHECK: %[[mask:.*]] = arith.cmpi slt, %[[linearIndex]], %[[boundVect]]
// CHECK-SAME: : vector<[17]x[[$IDX_TYPE]]>
//
-// 4. Create pass-through vector.
-// CHECK: %[[PASS_THROUGH:.*]] = arith.constant dense<7.{{.*}}> : vector<[17]xf32>
-//
// 5. Bitcast to vector form.
// CHECK: %[[gep:.*]] = llvm.getelementptr %{{.*}} :
// CHECK-SAME: (!llvm.ptr, i64) -> !llvm.ptr, f32
@@ -118,8 +109,7 @@ func.func @transfer_read_write_1d_scalable(%A : memref<?xf32>, %base: index) ->
// CHECK-SAME: -> vector<[17]xf32>
//
// 1. Let dim be the memref dimension, compute the in-bound index (dim - offset)
-// CHECK: %[[C0_b:.*]] = arith.constant 0 : index
-// CHECK: %[[DIM_b:.*]] = memref.dim %[[MEM]], %[[C0_b]] : memref<?xf32>
+// CHECK: %[[DIM_b:.*]] = memref.dim %[[MEM]], %[[C0]] : memref<?xf32>
// CHECK: %[[BOUND_b:.*]] = arith.subi %[[DIM_b]], %[[BASE]] : index
//
// 2. Create a vector with linear indices [ 0 .. vector_length - 1 ].
@@ -197,23 +187,23 @@ func.func @transfer_read_2d_to_1d(%A : memref<?x?xf32>, %base0: index, %base1: i
}
// CHECK-LABEL: func @transfer_read_2d_to_1d
// CHECK-SAME: %[[BASE_0:[a-zA-Z0-9]*]]: index, %[[BASE_1:[a-zA-Z0-9]*]]: index) -> vector<17xf32>
-// CHECK: %[[c1:.*]] = arith.constant 1 : index
+//
+// Create a vector with linear indices [ 0 .. vector_length - 1 ].
+// CHECK-DAG: %[[linearIndex:.*]] = arith.constant dense<[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16]> :
+// CHECK-SAME: vector<17x[[$IDX_TYPE]]>
+//
+// CHECK-DAG: %[[c1:.*]] = arith.constant 1 : index
// CHECK: %[[DIM:.*]] = memref.dim %{{.*}}, %[[c1]] : memref<?x?xf32>
//
// Compute the in-bound index (dim - offset)
// CHECK: %[[BOUND:.*]] = arith.subi %[[DIM]], %[[BASE_1]] : index
//
-// Create a vector with linear indices [ 0 .. vector_length - 1 ].
-// CHECK: %[[linearIndex:.*]] = arith.constant dense
-// CHECK-SAME: <[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16]> :
-// CHECK-SAME: vector<17x[[$IDX_TYPE]]>
-//
// Create bound vector to compute in-bound mask:
// [ 0 .. vector_length - 1 ] < [ dim - offset .. dim - offset ]
// CHECK: %[[btrunc:.*]] = arith.index_cast %[[BOUND]] : index to [[$IDX_TYPE]]
// CHECK: %[[boundVecInsert:.*]] = llvm.insertelement %[[btrunc]]
// CHECK: %[[boundVect:.*]] = llvm.shufflevector %[[boundVecInsert]]
-// CHECK: %[[mask:.*]] = arith.cmpi slt, %[[linearIndex]], %[[boundVect]]
+// CHECK: %[[mask:.*]] = arith.cmpi sgt, %[[boundVect]], %[[linearIndex]]
func.func @transfer_read_2d_to_1d_scalable(%A : memref<?x?xf32>, %base0: index, %base1: index) -> vector<[17]xf32> {
%f7 = arith.constant 7.0: f32
@@ -255,12 +245,13 @@ func.func @transfer_read_write_1d_non_zero_addrspace(%A : memref<?xf32, 3>, %bas
// CHECK-LABEL: func @transfer_read_write_1d_non_zero_addrspace
// CHECK-SAME: %[[BASE:[a-zA-Z0-9]*]]: index) -> vector<17xf32>
//
+// CHECK: %[[c0:.*]] = arith.constant 0 : index
+//
// 1. Check address space for GEP is correct.
// CHECK: %[[gep:.*]] = llvm.getelementptr {{.*}} :
// CHECK-SAME: (!llvm.ptr<3>, i64) -> !llvm.ptr<3>, f32
//
// 2. Check address space of the memref is correct.
-// CHECK: %[[c0:.*]] = arith.constant 0 : index
// CHECK: %[[DIM:.*]] = memref.dim %{{.*}}, %[[c0]] : memref<?xf32, 3>
//
// 3. Check address space for GEP is correct.
@@ -280,12 +271,13 @@ func.func @transfer_read_write_1d_non_zero_addrspace_scalable(%A : memref<?xf32,
// CHECK-LABEL: func @transfer_read_write_1d_non_zero_addrspace_scalable
// CHECK-SAME: %[[BASE:[a-zA-Z0-9]*]]: index) -> vector<[17]xf32>
//
+// CHECK: %[[c0:.*]] = arith.constant 0 : index
+//
// 1. Check address space for GEP is correct.
// CHECK: %[[gep:.*]] = llvm.getelementptr {{.*}} :
// CHECK-SAME: (!llvm.ptr<3>, i64) -> !llvm.ptr<3>, f32
//
// 2. Check address space of the memref is correct.
-// CHECK: %[[c0:.*]] = arith.constant 0 : index
// CHECK: %[[DIM:.*]] = memref.dim %{{.*}}, %[[c0]] : memref<?xf32, 3>
//
// 3. Check address space for GEP is correct.
@@ -330,10 +322,10 @@ func.func @transfer_read_1d_inbounds_scalable(%A : memref<?xf32>, %base: index)
// CHECK-LABEL: func @transfer_read_write_1d_mask
// CHECK: %[[mask1:.*]] = arith.constant dense<[false, false, true, false, true]>
-// CHECK: %[[cmpi:.*]] = arith.cmpi slt
+// CHECK: %[[cmpi:.*]] = arith.cmpi sgt
// CHECK: %[[mask2:.*]] = arith.andi %[[cmpi]], %[[mask1]]
// CHECK: %[[r:.*]] = llvm.intr.masked.load %{{.*}}, %[[mask2]]
-// CHECK: %[[cmpi_1:.*]] = arith.cmpi slt
+// CHECK: %[[cmpi_1:.*]] = arith.cmpi sgt
// CHECK: %[[mask3:.*]] = arith.andi %[[cmpi_1]], %[[mask1]]
// CHECK: llvm.intr.masked.store %[[r]], %{{.*}}, %[[mask3]]
// CHECK: return %[[r]]
|
|
✅ With the latest revision this PR passed the C/C++ code formatter. |
The mask materialization patterns during `VectorToLLVM` are rewrite patterns. They should run as part of the greedy pattern rewrite and not the dialect conversion. (Rewrite patterns and conversion patterns are not generally compatible.) The current combination of rewrite patterns and conversion patterns triggered an edge case when merging the 1:1 and 1:N dialect conversions.
matthias-springer
force-pushed
the
users/matthias-springer/mask_mat
branch
from
e896497 to
bf0d135
Compare
banach-space
approved these changes
Dec 14, 2024
matthias-springer
added a commit
that referenced
this pull request
Dec 17, 2024
…9975) Clean up `populateVectorToLLVMConversionPatterns` so that it populates only conversion patterns. All rewrite patterns that do not lower to LLVM should be populated into a separate greedy pattern rewrite. The current combination of rewrite patterns and conversion patterns triggered an edge case when merging the 1:1 and 1:N dialect conversions. Depends on #119973.
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 mask materialization patterns during
VectorToLLVMare rewrite patterns. They should run as part of the greedy pattern rewrite and not the dialect conversion. (Rewrite patterns and conversion patterns are not generally compatible.)The current combination of rewrite patterns and conversion patterns triggered an edge case when merging the 1:1 and 1:N dialect conversions.