[llvm] 78148eb - [Matrix] Fix crash during dot product lowering.
Florian Hahn via llvm-commits
llvm-commits at lists.llvm.org
Wed Apr 12 07:09:07 PDT 2023
Author: Florian Hahn
Date: 2023-04-12T15:08:39+01:00
New Revision: 78148eba491142ddda3b8de8199b6d80fb34ed1a
URL: https://github.com/llvm/llvm-project/commit/78148eba491142ddda3b8de8199b6d80fb34ed1a
DIFF: https://github.com/llvm/llvm-project/commit/78148eba491142ddda3b8de8199b6d80fb34ed1a.diff
LOG: [Matrix] Fix crash during dot product lowering.
Perform dot-product lowering before instruction fusion to avoid crash in
newly added test. Also update lowerDotProduct to properly mark optimized
matmul as fused.
Added:
Modified:
llvm/lib/Transforms/Scalar/LowerMatrixIntrinsics.cpp
llvm/test/Transforms/LowerMatrixIntrinsics/dot-product-int.ll
Removed:
################################################################################
diff --git a/llvm/lib/Transforms/Scalar/LowerMatrixIntrinsics.cpp b/llvm/lib/Transforms/Scalar/LowerMatrixIntrinsics.cpp
index 1147f018d5655..82e717fedb704 100644
--- a/llvm/lib/Transforms/Scalar/LowerMatrixIntrinsics.cpp
+++ b/llvm/lib/Transforms/Scalar/LowerMatrixIntrinsics.cpp
@@ -978,17 +978,15 @@ class LowerMatrixIntrinsics {
MatrixInsts.push_back(&I);
}
- // Second, try to fuse candidates.
+ // Second, try to lower any dot products
SmallPtrSet<Instruction *, 16> FusedInsts;
+ for (CallInst *CI : MaybeFusableInsts)
+ lowerDotProduct(CI, FusedInsts, getFastMathFlags(CI));
+
+ // Third, try to fuse candidates.
for (CallInst *CI : MaybeFusableInsts)
LowerMatrixMultiplyFused(CI, FusedInsts);
- // Third, try to lower any dot products
- for (CallInst *CI : MaybeFusableInsts) {
- if (FusedInsts.contains(CI)) // skip if already fused
- continue;
- lowerDotProduct(CI, FusedInsts, getFastMathFlags(CI));
- }
Changed = !FusedInsts.empty();
// Fourth, lower remaining instructions with shape information.
@@ -1324,7 +1322,8 @@ class LowerMatrixIntrinsics {
void lowerDotProduct(CallInst *MatMul,
SmallPtrSet<Instruction *, 16> &FusedInsts,
FastMathFlags FMF) {
- if (MatrixLayout != MatrixLayoutTy::ColumnMajor)
+ if (FusedInsts.contains(MatMul) ||
+ MatrixLayout != MatrixLayoutTy::ColumnMajor)
return;
ShapeInfo LShape(MatMul->getArgOperand(2), MatMul->getArgOperand(3));
ShapeInfo RShape(MatMul->getArgOperand(3), MatMul->getArgOperand(4));
@@ -1410,7 +1409,8 @@ class LowerMatrixIntrinsics {
Result = Builder.CreateInsertElement(PoisonValue::get(MatMul->getType()),
Result, uint64_t(0));
MatMul->replaceAllUsesWith(Result);
- MatMul->eraseFromParent();
+ FusedInsts.insert(MatMul);
+ ToRemove.push_back(MatMul);
}
/// Compute \p Result += \p A * \p B for input matrices with left-associating
diff --git a/llvm/test/Transforms/LowerMatrixIntrinsics/dot-product-int.ll b/llvm/test/Transforms/LowerMatrixIntrinsics/dot-product-int.ll
index 0129dff311216..dd5c4cf3f4b4f 100644
--- a/llvm/test/Transforms/LowerMatrixIntrinsics/dot-product-int.ll
+++ b/llvm/test/Transforms/LowerMatrixIntrinsics/dot-product-int.ll
@@ -624,3 +624,41 @@ entry:
}
declare <1 x i16> @llvm.matrix.multiply.v1i16.v6i16.v6i16(<6 x i16>, <6 x i16>, i32, i32, i32)
+
+define void @transposed_multiply_feeding_dot_produc_v4i32() {
+; CHECK-LABEL: @transposed_multiply_feeding_dot_produc_v4i32(
+; CHECK-NEXT: entry:
+; CHECK-NEXT: [[SCALAR_SPLAT_SPLAT_I_I_I_I:%.*]] = shufflevector <4 x i32> zeroinitializer, <4 x i32> zeroinitializer, <4 x i32> zeroinitializer
+; CHECK-NEXT: [[SPLIT:%.*]] = shufflevector <4 x i32> [[SCALAR_SPLAT_SPLAT_I_I_I_I]], <4 x i32> poison, <2 x i32> <i32 0, i32 1>
+; CHECK-NEXT: [[SPLIT1:%.*]] = shufflevector <4 x i32> [[SCALAR_SPLAT_SPLAT_I_I_I_I]], <4 x i32> poison, <2 x i32> <i32 2, i32 3>
+; CHECK-NEXT: [[BLOCK:%.*]] = shufflevector <2 x i32> [[SPLIT]], <2 x i32> poison, <2 x i32> <i32 0, i32 1>
+; CHECK-NEXT: [[TMP0:%.*]] = mul <2 x i32> [[BLOCK]], zeroinitializer
+; CHECK-NEXT: [[BLOCK2:%.*]] = shufflevector <2 x i32> [[SPLIT1]], <2 x i32> poison, <2 x i32> <i32 0, i32 1>
+; CHECK-NEXT: [[TMP1:%.*]] = mul <2 x i32> [[BLOCK2]], zeroinitializer
+; CHECK-NEXT: [[TMP2:%.*]] = add <2 x i32> [[TMP0]], [[TMP1]]
+; CHECK-NEXT: [[TMP3:%.*]] = shufflevector <2 x i32> [[TMP2]], <2 x i32> poison, <2 x i32> <i32 0, i32 1>
+; CHECK-NEXT: [[TMP4:%.*]] = shufflevector <2 x i32> undef, <2 x i32> [[TMP3]], <2 x i32> <i32 2, i32 3>
+; CHECK-NEXT: [[BLOCK3:%.*]] = shufflevector <2 x i32> [[SPLIT]], <2 x i32> poison, <2 x i32> <i32 0, i32 1>
+; CHECK-NEXT: [[TMP5:%.*]] = mul <2 x i32> [[BLOCK3]], zeroinitializer
+; CHECK-NEXT: [[BLOCK4:%.*]] = shufflevector <2 x i32> [[SPLIT1]], <2 x i32> poison, <2 x i32> <i32 0, i32 1>
+; CHECK-NEXT: [[TMP6:%.*]] = mul <2 x i32> [[BLOCK4]], zeroinitializer
+; CHECK-NEXT: [[TMP7:%.*]] = add <2 x i32> [[TMP5]], [[TMP6]]
+; CHECK-NEXT: [[TMP8:%.*]] = shufflevector <2 x i32> [[TMP7]], <2 x i32> poison, <2 x i32> <i32 0, i32 1>
+; CHECK-NEXT: [[TMP9:%.*]] = shufflevector <2 x i32> undef, <2 x i32> [[TMP8]], <2 x i32> <i32 2, i32 3>
+; CHECK-NEXT: [[TMP10:%.*]] = shufflevector <2 x i32> [[TMP4]], <2 x i32> [[TMP9]], <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+; CHECK-NEXT: [[TMP11:%.*]] = mul <4 x i32> [[TMP10]], zeroinitializer
+; CHECK-NEXT: [[TMP12:%.*]] = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> [[TMP11]])
+; CHECK-NEXT: [[TMP13:%.*]] = insertelement <1 x i32> poison, i32 [[TMP12]], i64 0
+; CHECK-NEXT: ret void
+;
+entry:
+ %scalar.splat.splat.i.i.i.i = shufflevector <4 x i32> zeroinitializer, <4 x i32> zeroinitializer, <4 x i32> zeroinitializer
+ %0 = call <4 x i32> @llvm.matrix.multiply.v4i32.v4i32.v4i32(<4 x i32> zeroinitializer, <4 x i32> %scalar.splat.splat.i.i.i.i, i32 2, i32 2, i32 2)
+ %1 = call <4 x i32> @llvm.matrix.transpose.v4i32(<4 x i32> %0, i32 4, i32 1)
+ %2 = call <1 x i32> @llvm.matrix.multiply.v1i32.v4i32.v4i32(<4 x i32> %1, <4 x i32> zeroinitializer, i32 1, i32 4, i32 1)
+ ret void
+}
+
+declare <4 x i32> @llvm.matrix.transpose.v4i32(<4 x i32>, i32 immarg, i32 immarg)
+
+declare <4 x i32> @llvm.matrix.multiply.v4i32.v4i32.v4i32(<4 x i32>, <4 x i32>, i32 immarg, i32 immarg, i32 immarg)
More information about the llvm-commits
mailing list