[mlir][vector] Lower vector.fma to llvm.intr.fma #143089
Open
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
Now it lowers to llvm.intr.fmuladd.
jacquesguan
requested review from
banach-space,
dcaballe and
nicolasvasilache
as code owners
|
@llvm/pr-subscribers-mlir-llvm @llvm/pr-subscribers-mlir Author: Jianjian Guan (jacquesguan) ChangesNow it lowers to llvm.intr.fmuladd. Full diff: https://github.com/llvm/llvm-project/pull/143089.diff 3 Files Affected:
diff --git a/mlir/lib/Conversion/VectorToLLVM/ConvertVectorToLLVM.cpp b/mlir/lib/Conversion/VectorToLLVM/ConvertVectorToLLVM.cpp
index f725993635672..26a2de43e2836 100644
--- a/mlir/lib/Conversion/VectorToLLVM/ConvertVectorToLLVM.cpp
+++ b/mlir/lib/Conversion/VectorToLLVM/ConvertVectorToLLVM.cpp
@@ -1208,7 +1208,7 @@ class VectorExtractOpConversion
};
/// Conversion pattern that turns a vector.fma on a 1-D vector
-/// into an llvm.intr.fmuladd. This is a trivial 1-1 conversion.
+/// into an llvm.intr.fma. This is a trivial 1-1 conversion.
/// This does not match vectors of n >= 2 rank.
///
/// Example:
@@ -1217,7 +1217,7 @@ class VectorExtractOpConversion
/// ```
/// is converted to:
/// ```
-/// llvm.intr.fmuladd %va, %va, %va:
+/// llvm.intr.fma %va, %va, %va:
/// (!llvm."<8 x f32>">, !llvm<"<8 x f32>">, !llvm<"<8 x f32>">)
/// -> !llvm."<8 x f32>">
/// ```
@@ -1232,7 +1232,7 @@ class VectorFMAOp1DConversion : public ConvertOpToLLVMPattern<vector::FMAOp> {
if (vType.getRank() > 1)
return failure();
- rewriter.replaceOpWithNewOp<LLVM::FMulAddOp>(
+ rewriter.replaceOpWithNewOp<LLVM::FMAOp>(
fmaOp, adaptor.getLhs(), adaptor.getRhs(), adaptor.getAcc());
return success();
}
diff --git a/mlir/test/Conversion/VectorToLLVM/vector-to-llvm.mlir b/mlir/test/Conversion/VectorToLLVM/vector-to-llvm.mlir
index 64e51f5554628..9d48a3a93e1bb 100644
--- a/mlir/test/Conversion/VectorToLLVM/vector-to-llvm.mlir
+++ b/mlir/test/Conversion/VectorToLLVM/vector-to-llvm.mlir
@@ -660,14 +660,14 @@ func.func @outerproduct_add(%arg0: vector<2xf32>, %arg1: vector<3xf32>, %arg2: v
// CHECK: %[[T6Insert:.*]] = llvm.insertelement %[[T5]]
// CHECK: %[[T6:.*]] = llvm.shufflevector %[[T6Insert]]
// CHECK: %[[T8:.*]] = llvm.extractvalue %[[T7]][0] : !llvm.array<2 x vector<3xf32>>
-// CHECK: %[[T9:.*]] = llvm.intr.fmuladd(%[[T6]], %[[B]], %[[T8]]) : (vector<3xf32>, vector<3xf32>, vector<3xf32>) -> vector<3xf32>
+// CHECK: %[[T9:.*]] = llvm.intr.fma(%[[T6]], %[[B]], %[[T8]]) : (vector<3xf32>, vector<3xf32>, vector<3xf32>) -> vector<3xf32>
// CHECK: %[[T11:.*]] = llvm.insertvalue %[[T9]], %[[T10]][0] : !llvm.array<2 x vector<3xf32>>
// CHECK: %[[T12:.*]] = llvm.mlir.constant(1 : i64) : i64
// CHECK: %[[T13:.*]] = llvm.extractelement %[[A]]{{\[}}%[[T12]] : i64] : vector<2xf32>
// CHECK: %[[T14Insert:.*]] = llvm.insertelement %[[T13]]
// CHECK: %[[T14:.*]] = llvm.shufflevector %[[T14Insert]]
// CHECK: %[[T16:.*]] = llvm.extractvalue %[[T7]][1] : !llvm.array<2 x vector<3xf32>>
-// CHECK: %[[T17:.*]] = llvm.intr.fmuladd(%[[T14]], %[[B]], %[[T16]]) : (vector<3xf32>, vector<3xf32>, vector<3xf32>) -> vector<3xf32>
+// CHECK: %[[T17:.*]] = llvm.intr.fma(%[[T14]], %[[B]], %[[T16]]) : (vector<3xf32>, vector<3xf32>, vector<3xf32>) -> vector<3xf32>
// CHECK: %[[T18:.*]] = llvm.insertvalue %[[T17]], %[[T11]][1] : !llvm.array<2 x vector<3xf32>>
// CHECK: %[[T19:.*]] = builtin.unrealized_conversion_cast %[[T18]] : !llvm.array<2 x vector<3xf32>> to vector<2x3xf32>
// CHECK: return %[[T19]] : vector<2x3xf32>
@@ -690,14 +690,14 @@ func.func @outerproduct_add_scalable(%arg0: vector<2xf32>, %arg1: vector<[3]xf32
// CHECK: %[[T6Insert:.*]] = llvm.insertelement %[[T5]]
// CHECK: %[[T6:.*]] = llvm.shufflevector %[[T6Insert]]
// CHECK: %[[T8:.*]] = llvm.extractvalue %[[T7]][0] : !llvm.array<2 x vector<[3]xf32>>
-// CHECK: %[[T9:.*]] = llvm.intr.fmuladd(%[[T6]], %[[B]], %[[T8]]) : (vector<[3]xf32>, vector<[3]xf32>, vector<[3]xf32>) -> vector<[3]xf32>
+// CHECK: %[[T9:.*]] = llvm.intr.fma(%[[T6]], %[[B]], %[[T8]]) : (vector<[3]xf32>, vector<[3]xf32>, vector<[3]xf32>) -> vector<[3]xf32>
// CHECK: %[[T11:.*]] = llvm.insertvalue %[[T9]], %[[T10]][0] : !llvm.array<2 x vector<[3]xf32>>
// CHECK: %[[T12:.*]] = llvm.mlir.constant(1 : i64) : i64
// CHECK: %[[T13:.*]] = llvm.extractelement %[[A]]{{\[}}%[[T12]] : i64] : vector<2xf32>
// CHECK: %[[T14Insert:.*]] = llvm.insertelement %[[T13]]
// CHECK: %[[T14:.*]] = llvm.shufflevector %[[T14Insert]]
// CHECK: %[[T16:.*]] = llvm.extractvalue %[[T7]][1] : !llvm.array<2 x vector<[3]xf32>>
-// CHECK: %[[T17:.*]] = llvm.intr.fmuladd(%[[T14]], %[[B]], %[[T16]]) : (vector<[3]xf32>, vector<[3]xf32>, vector<[3]xf32>) -> vector<[3]xf32>
+// CHECK: %[[T17:.*]] = llvm.intr.fma(%[[T14]], %[[B]], %[[T16]]) : (vector<[3]xf32>, vector<[3]xf32>, vector<[3]xf32>) -> vector<[3]xf32>
// CHECK: %[[T18:.*]] = llvm.insertvalue %[[T17]], %[[T11]][1] : !llvm.array<2 x vector<[3]xf32>>
// CHECK: %[[T19:.*]] = builtin.unrealized_conversion_cast %[[T18]] : !llvm.array<2 x vector<[3]xf32>> to vector<2x[3]xf32>
// CHECK: return %[[T19]] : vector<2x[3]xf32>
@@ -715,7 +715,7 @@ func.func @masked_float_add_outerprod(%arg0: vector<2xf32>, %arg1: f32, %arg2: v
// CHECK-LABEL: func.func @masked_float_add_outerprod(
// CHECK-SAME: %[[VAL_0:.*]]: vector<2xf32>, %[[VAL_1:.*]]: f32, %[[VAL_2:.*]]: vector<2xf32>, %[[VAL_3:.*]]: vector<2xi1>) -> vector<2xf32> {
-// CHECK: %[[VAL_8:.*]] = llvm.intr.fmuladd(%[[VAL_0]], %{{.*}}, %[[VAL_2]]) : (vector<2xf32>, vector<2xf32>, vector<2xf32>) -> vector<2xf32>
+// CHECK: %[[VAL_8:.*]] = llvm.intr.fma(%[[VAL_0]], %{{.*}}, %[[VAL_2]]) : (vector<2xf32>, vector<2xf32>, vector<2xf32>) -> vector<2xf32>
// CHECK: %[[VAL_9:.*]] = arith.select %[[VAL_3]], %[[VAL_8]], %[[VAL_2]] : vector<2xi1>, vector<2xf32>
// -----
@@ -727,7 +727,7 @@ func.func @masked_float_add_outerprod_scalable(%arg0: vector<[2]xf32>, %arg1: f3
// CHECK-LABEL: func.func @masked_float_add_outerprod_scalable(
// CHECK-SAME: %[[VAL_0:.*]]: vector<[2]xf32>, %[[VAL_1:.*]]: f32, %[[VAL_2:.*]]: vector<[2]xf32>, %[[VAL_3:.*]]: vector<[2]xi1>) -> vector<[2]xf32> {
-// CHECK: %[[VAL_8:.*]] = llvm.intr.fmuladd(%[[VAL_0]], %{{.*}}, %[[VAL_2]]) : (vector<[2]xf32>, vector<[2]xf32>, vector<[2]xf32>) -> vector<[2]xf32>
+// CHECK: %[[VAL_8:.*]] = llvm.intr.fma(%[[VAL_0]], %{{.*}}, %[[VAL_2]]) : (vector<[2]xf32>, vector<[2]xf32>, vector<[2]xf32>) -> vector<[2]xf32>
// CHECK: %[[VAL_9:.*]] = arith.select %[[VAL_3]], %[[VAL_8]], %[[VAL_2]] : vector<[2]xi1>, vector<[2]xf32>
// -----
@@ -1363,29 +1363,29 @@ func.func @fma(%vec_1d: vector<8xf32>, %vec_2d: vector<2x4xf32>, %vec_3d: vector
// CHECK-SAME: %[[VEC_2D:.*]]: vector<2x4xf32>
// CHECK-SAME: %[[VEC_3D:.*]]: vector<1x1x1xf32>
// CHECK: %[[VEC_2D_CAST:.*]] = builtin.unrealized_conversion_cast %[[VEC_2D]] : vector<2x4xf32> to !llvm.array<2 x vector<4xf32>>
- // CHECK: llvm.intr.fmuladd
+ // CHECK: llvm.intr.fma
// CHECK-SAME: (vector<8xf32>, vector<8xf32>, vector<8xf32>) -> vector<8xf32>
%0 = vector.fma %vec_1d, %vec_1d, %vec_1d : vector<8xf32>
// CHECK: %[[VEC_2D_00:.*]] = llvm.extractvalue %[[VEC_2D_CAST]][0] : !llvm.array<2 x vector<4xf32>>
// CHECK: %[[VEC_2D_01:.*]] = llvm.extractvalue %[[VEC_2D_CAST]][0] : !llvm.array<2 x vector<4xf32>>
// CHECK: %[[VEC_2D_02:.*]] = llvm.extractvalue %[[VEC_2D_CAST]][0] : !llvm.array<2 x vector<4xf32>>
- // CHECK: %[[VEC_2D_ADD_1:.*]] = llvm.intr.fmuladd(%[[VEC_2D_00]], %[[VEC_2D_01]], %[[VEC_2D_02]]) :
+ // CHECK: %[[VEC_2D_ADD_1:.*]] = llvm.intr.fma(%[[VEC_2D_00]], %[[VEC_2D_01]], %[[VEC_2D_02]]) :
// CHECK-SAME: (vector<4xf32>, vector<4xf32>, vector<4xf32>) -> vector<4xf32>
// CHECK: llvm.insertvalue %[[VEC_2D_ADD_1]], {{.*}}[0] : !llvm.array<2 x vector<4xf32>>
// CHECK: %[[VEC_2D_10:.*]] = llvm.extractvalue %[[VEC_2D_CAST]][1] : !llvm.array<2 x vector<4xf32>>
// CHECK: %[[VEC_2D_11:.*]] = llvm.extractvalue %[[VEC_2D_CAST]][1] : !llvm.array<2 x vector<4xf32>>
// CHECK: %[[VEC_2D_12:.*]] = llvm.extractvalue %[[VEC_2D_CAST]][1] : !llvm.array<2 x vector<4xf32>>
- // CHECK: %[[VEC_2D_ADD_2:.*]] = llvm.intr.fmuladd(%[[VEC_2D_10]], %[[VEC_2D_11]], %[[VEC_2D_12]]) :
+ // CHECK: %[[VEC_2D_ADD_2:.*]] = llvm.intr.fma(%[[VEC_2D_10]], %[[VEC_2D_11]], %[[VEC_2D_12]]) :
// CHECK-SAME: (vector<4xf32>, vector<4xf32>, vector<4xf32>) -> vector<4xf32>
// CHECK: llvm.insertvalue %[[VEC_2D_ADD_2]], {{.*}}[1] : !llvm.array<2 x vector<4xf32>>
%1 = vector.fma %vec_2d, %vec_2d, %vec_2d : vector<2x4xf32>
- // CHECK: %[[C0:.*]] = llvm.intr.fmuladd
+ // CHECK: %[[C0:.*]] = llvm.intr.fma
// CHECK-SAME: (vector<1xf32>, vector<1xf32>, vector<1xf32>) -> vector<1xf32>
%2 = vector.fma %vec_3d, %vec_3d, %vec_3d : vector<1x1x1xf32>
- // CHECK: %[[D0:.*]] = llvm.intr.fmuladd
+ // CHECK: %[[D0:.*]] = llvm.intr.fma
// CHECK-SAME: (vector<1xf32>, vector<1xf32>, vector<1xf32>) -> vector<1xf32>
%3 = vector.fma %vec_0d, %vec_0d, %vec_0d : vector<f32>
@@ -1400,25 +1400,25 @@ func.func @fma_scalable(%vec_1d: vector<[8]xf32>, %vec_2d: vector<2x[4]xf32>, %v
// CHECK-SAME: %[[VEC_2D:.*]]: vector<2x[4]xf32>
// CHECK-SAME: %[[VEC_3D:.*]]: vector<1x1x[1]xf32>
// CHECK: %[[VEC_2D_CAST:.*]] = builtin.unrealized_conversion_cast %[[VEC_2D]] : vector<2x[4]xf32> to !llvm.array<2 x vector<[4]xf32>>
- // CHECK: llvm.intr.fmuladd
+ // CHECK: llvm.intr.fma
// CHECK-SAME: (vector<[8]xf32>, vector<[8]xf32>, vector<[8]xf32>) -> vector<[8]xf32>
%0 = vector.fma %vec_1d, %vec_1d, %vec_1d : vector<[8]xf32>
// CHECK: %[[VEC_2D_00:.*]] = llvm.extractvalue %[[VEC_2D_CAST]][0] : !llvm.array<2 x vector<[4]xf32>>
// CHECK: %[[VEC_2D_01:.*]] = llvm.extractvalue %[[VEC_2D_CAST]][0] : !llvm.array<2 x vector<[4]xf32>>
// CHECK: %[[VEC_2D_02:.*]] = llvm.extractvalue %[[VEC_2D_CAST]][0] : !llvm.array<2 x vector<[4]xf32>>
- // CHECK: %[[VEC_2D_ADD_1:.*]] = llvm.intr.fmuladd(%[[VEC_2D_00]], %[[VEC_2D_01]], %[[VEC_2D_02]]) :
+ // CHECK: %[[VEC_2D_ADD_1:.*]] = llvm.intr.fma(%[[VEC_2D_00]], %[[VEC_2D_01]], %[[VEC_2D_02]]) :
// CHECK-SAME: (vector<[4]xf32>, vector<[4]xf32>, vector<[4]xf32>) -> vector<[4]xf32>
// CHECK: llvm.insertvalue %[[VEC_2D_ADD_1]], {{.*}}[0] : !llvm.array<2 x vector<[4]xf32>>
// CHECK: %[[VEC_2D_10:.*]] = llvm.extractvalue %[[VEC_2D_CAST]][1] : !llvm.array<2 x vector<[4]xf32>>
// CHECK: %[[VEC_2D_11:.*]] = llvm.extractvalue %[[VEC_2D_CAST]][1] : !llvm.array<2 x vector<[4]xf32>>
// CHECK: %[[VEC_2D_12:.*]] = llvm.extractvalue %[[VEC_2D_CAST]][1] : !llvm.array<2 x vector<[4]xf32>>
- // CHECK: %[[VEC_2D_ADD_2:.*]] = llvm.intr.fmuladd(%[[VEC_2D_10]], %[[VEC_2D_11]], %[[VEC_2D_12]]) :
+ // CHECK: %[[VEC_2D_ADD_2:.*]] = llvm.intr.fma(%[[VEC_2D_10]], %[[VEC_2D_11]], %[[VEC_2D_12]]) :
// CHECK-SAME: (vector<[4]xf32>, vector<[4]xf32>, vector<[4]xf32>) -> vector<[4]xf32>
// CHECK: llvm.insertvalue %[[VEC_2D_ADD_2]], {{.*}}[1] : !llvm.array<2 x vector<[4]xf32>>
%1 = vector.fma %vec_2d, %vec_2d, %vec_2d : vector<2x[4]xf32>
- // CHECK: %[[C0:.*]] = llvm.intr.fmuladd
+ // CHECK: %[[C0:.*]] = llvm.intr.fma
// CHECK-SAME: (vector<[1]xf32>, vector<[1]xf32>, vector<[1]xf32>) -> vector<[1]xf32>
%2 = vector.fma %vec_3d, %vec_3d, %vec_3d : vector<1x1x[1]xf32>
diff --git a/mlir/test/Dialect/LLVM/transform-e2e.mlir b/mlir/test/Dialect/LLVM/transform-e2e.mlir
index c00b47fb936e9..fc87d66e9d2c8 100644
--- a/mlir/test/Dialect/LLVM/transform-e2e.mlir
+++ b/mlir/test/Dialect/LLVM/transform-e2e.mlir
@@ -5,7 +5,7 @@ func.func @matmul_tensors(
%arg0: tensor<2x4xf32>, %arg1: tensor<4x6xf32>, %arg2: tensor<2x6xf32>)
-> tensor<2x6xf32> {
// CHECK-NOT: linalg
-// CHECK: llvm.intr.fmuladd{{.*}}
+// CHECK: llvm.intr.fma{{.*}}
%0 = linalg.matmul ins(%arg0, %arg1: tensor<2x4xf32>, tensor<4x6xf32>)
outs(%arg2: tensor<2x6xf32>)
-> tensor<2x6xf32>
|
|
Hey @jacquesguan , thanks for sending this! This PR lacks justification for the change. Importantly, how does it relate to: ? |
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.
Now it lowers to llvm.intr.fmuladd.