[llvm] [MatrixIntrinsics] Add row-major support for llvm.matrix.transpose

[aartbik]

Summary

Only column-major was supported so far. This adds row-major support as well.
Note that we probably also want very efficient SIMD implementations for the
various target platforms.

Bug:
https://bugs.llvm.org/show_bug.cgi?id=46085

[fhahn]

Thank you very much for the patch! A few suggestions inline.

llvm/lib/Transforms/Scalar/LowerMatrixIntrinsics.cpp

1300

I think it might be possible to share most of the transpose implementation for both row/column major transpose, as we are iterating over the indices of the second dimension while extracting from the vectors of the first dimension in both cases.

Instead of using ArgShape.NumRows, we could just use the second dimension (InputMatrix.isColumnMajor() : ArgShape.NumRows : ArgShape.NumColumns and instead of iterating over columns/rows we could just iterate over the vectors of the leading dimension (InputMatrix.vectors()) and instead of ArgShape.NumColumns/NumRows, we could use ArgShape.getNumVectors().

llvm/test/Transforms/LowerMatrixIntrinsics/pr46085-row.ll

1 ↗

(On Diff #266691)

for multiply, we just have row/column major test variants in multiply-i32-row-major.ll/multiply-i32.ll and so on. For consistency, it might be slightly preferable to extend the existing transpose-*.ll tests (if the cases in the new test are not covered) and call the row-major equivalent transpose-*-row-major.ll.

[aartbik]

addressed comments, thanks!

[aartbik]

llvm/lib/Transforms/Scalar/LowerMatrixIntrinsics.cpp

1300

That is a good idea. Fits nicely with the provided utilities. Thanks!

llvm/test/Transforms/LowerMatrixIntrinsics/pr46085-row.ll

1 ↗

(On Diff #266691)

Sounds good. I removed the pr* tests and used the existing transpose tests for RM.

[fhahn]

LGTM, thanks!

llvm/lib/Transforms/Scalar/LowerMatrixIntrinsics.cpp

public:
   /// Lowers llvm.matrix.transpose.
   void LowerTranspose(CallInst *Inst) {
     MatrixTy Result;
     IRBuilder<> Builder(Inst);
     Value *InputVal = Inst->getArgOperand(0);
     VectorType *VectorTy = cast<VectorType>(InputVal->getType());
     ShapeInfo ArgShape(Inst->getArgOperand(1), Inst->getArgOperand(2));
     MatrixTy InputMatrix = getMatrix(InputVal, ArgShape, Builder);
-    assert(InputMatrix.isColumnMajor() &&
-           "Row-major code-gen not supported yet!");
 
-    for (unsigned Row = 0; Row < ArgShape.NumRows; ++Row) {
-      // Build a single column vector for this row. First initialize it.
-      Value *ResultColumn = UndefValue::get(
-          VectorType::get(VectorTy->getElementType(), ArgShape.NumColumns));
-
-      // Go through the elements of this row and insert it into the resulting
-      // column vector.
-      for (auto C : enumerate(InputMatrix.columns())) {
-        Value *Elt = Builder.CreateExtractElement(C.value(), Row);
-        // We insert at index Column since that is the row index after the
-        // transpose.
-        ResultColumn =
-            Builder.CreateInsertElement(ResultColumn, Elt, C.index());
+    const unsigned NewNumVecs =
+        InputMatrix.isColumnMajor() ? ArgShape.NumRows : ArgShape.NumColumns;
+    const unsigned NewNumElts =
+        InputMatrix.isColumnMajor() ? ArgShape.NumColumns : ArgShape.NumRows;
+
+    for (unsigned I = 0; I < NewNumVecs; ++I) {
+      // Build a single result vector. First initialize it.
+      Value *ResultVector = UndefValue::get(
+          VectorType::get(VectorTy->getElementType(), NewNumElts));
+      // Go through the old elements and insert it into the resulting vector.
+      for (auto J : enumerate(InputMatrix.vectors())) {
+        Value *Elt = Builder.CreateExtractElement(J.value(), I);
+        // Row and column indices are transposed.
+        ResultVector =
+            Builder.CreateInsertElement(ResultVector, Elt, J.index());
       }
-      Result.addVector(ResultColumn);
+      Result.addVector(ResultVector);
     }
 
     // TODO: Improve estimate of operations needed for transposes. Currently we
     // just count the insertelement/extractelement instructions, but do not
     // account for later simplifications/combines.
     finalizeLowering(
         Inst,
         Result.addNumComputeOps(2 * ArgShape.NumRows * ArgShape.NumColumns),

llvm/test/Transforms/LowerMatrixIntrinsics/transpose-double-row-major.ll

+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt -lower-matrix-intrinsics -matrix-default-layout=row-major -S < %s | FileCheck --check-prefix=RM %s
+
+define <8 x double> @transpose(<8 x double> %a) {
+; RM-LABEL: @transpose(
+; RM-NEXT:  entry:
+; RM-NEXT:    [[SPLIT:%.*]] = shufflevector <8 x double> [[A:%.*]], <8 x double> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+; RM-NEXT:    [[SPLIT1:%.*]] = shufflevector <8 x double> [[A]], <8 x double> undef, <4 x i32> <i32 4, i32 5, i32 6, i32 7>
+; RM-NEXT:    [[TMP0:%.*]] = extractelement <4 x double> [[SPLIT]], i64 0
+; RM-NEXT:    [[TMP1:%.*]] = insertelement <2 x double> undef, double [[TMP0]], i64 0
+; RM-NEXT:    [[TMP2:%.*]] = extractelement <4 x double> [[SPLIT1]], i64 0
+; RM-NEXT:    [[TMP3:%.*]] = insertelement <2 x double> [[TMP1]], double [[TMP2]], i64 1
+; RM-NEXT:    [[TMP4:%.*]] = extractelement <4 x double> [[SPLIT]], i64 1
+; RM-NEXT:    [[TMP5:%.*]] = insertelement <2 x double> undef, double [[TMP4]], i64 0
+; RM-NEXT:    [[TMP6:%.*]] = extractelement <4 x double> [[SPLIT1]], i64 1
+; RM-NEXT:    [[TMP7:%.*]] = insertelement <2 x double> [[TMP5]], double [[TMP6]], i64 1
+; RM-NEXT:    [[TMP8:%.*]] = extractelement <4 x double> [[SPLIT]], i64 2
+; RM-NEXT:    [[TMP9:%.*]] = insertelement <2 x double> undef, double [[TMP8]], i64 0
+; RM-NEXT:    [[TMP10:%.*]] = extractelement <4 x double> [[SPLIT1]], i64 2
+; RM-NEXT:    [[TMP11:%.*]] = insertelement <2 x double> [[TMP9]], double [[TMP10]], i64 1
+; RM-NEXT:    [[TMP12:%.*]] = extractelement <4 x double> [[SPLIT]], i64 3
+; RM-NEXT:    [[TMP13:%.*]] = insertelement <2 x double> undef, double [[TMP12]], i64 0
+; RM-NEXT:    [[TMP14:%.*]] = extractelement <4 x double> [[SPLIT1]], i64 3
+; RM-NEXT:    [[TMP15:%.*]] = insertelement <2 x double> [[TMP13]], double [[TMP14]], i64 1
+; RM-NEXT:    [[TMP16:%.*]] = shufflevector <2 x double> [[TMP3]], <2 x double> [[TMP7]], <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+; RM-NEXT:    [[TMP17:%.*]] = shufflevector <2 x double> [[TMP11]], <2 x double> [[TMP15]], <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+; RM-NEXT:    [[TMP18:%.*]] = shufflevector <4 x double> [[TMP16]], <4 x double> [[TMP17]], <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
+; RM-NEXT:    ret <8 x double> [[TMP18]]
+;
+entry:
+  %c  = call <8 x double> @llvm.matrix.transpose(<8 x double> %a, i32 2, i32 4)
+  ret <8 x double> %c
+}
+
+declare <8 x double> @llvm.matrix.transpose(<8 x double>, i32, i32)
+
+define <8 x double> @transpose_single_column(<8 x double> %a) {
+; RM-LABEL: @transpose_single_column(
+; RM-NEXT:  entry:
+; RM-NEXT:    [[SPLIT:%.*]] = shufflevector <8 x double> [[A:%.*]], <8 x double> undef, <1 x i32> zeroinitializer
+; RM-NEXT:    [[SPLIT1:%.*]] = shufflevector <8 x double> [[A]], <8 x double> undef, <1 x i32> <i32 1>
+; RM-NEXT:    [[SPLIT2:%.*]] = shufflevector <8 x double> [[A]], <8 x double> undef, <1 x i32> <i32 2>
+; RM-NEXT:    [[SPLIT3:%.*]] = shufflevector <8 x double> [[A]], <8 x double> undef, <1 x i32> <i32 3>
+; RM-NEXT:    [[SPLIT4:%.*]] = shufflevector <8 x double> [[A]], <8 x double> undef, <1 x i32> <i32 4>
+; RM-NEXT:    [[SPLIT5:%.*]] = shufflevector <8 x double> [[A]], <8 x double> undef, <1 x i32> <i32 5>
+; RM-NEXT:    [[SPLIT6:%.*]] = shufflevector <8 x double> [[A]], <8 x double> undef, <1 x i32> <i32 6>
+; RM-NEXT:    [[SPLIT7:%.*]] = shufflevector <8 x double> [[A]], <8 x double> undef, <1 x i32> <i32 7>
+; RM-NEXT:    [[TMP0:%.*]] = extractelement <1 x double> [[SPLIT]], i64 0
+; RM-NEXT:    [[TMP1:%.*]] = insertelement <8 x double> undef, double [[TMP0]], i64 0
+; RM-NEXT:    [[TMP2:%.*]] = extractelement <1 x double> [[SPLIT1]], i64 0
+; RM-NEXT:    [[TMP3:%.*]] = insertelement <8 x double> [[TMP1]], double [[TMP2]], i64 1
+; RM-NEXT:    [[TMP4:%.*]] = extractelement <1 x double> [[SPLIT2]], i64 0
+; RM-NEXT:    [[TMP5:%.*]] = insertelement <8 x double> [[TMP3]], double [[TMP4]], i64 2
+; RM-NEXT:    [[TMP6:%.*]] = extractelement <1 x double> [[SPLIT3]], i64 0
+; RM-NEXT:    [[TMP7:%.*]] = insertelement <8 x double> [[TMP5]], double [[TMP6]], i64 3
+; RM-NEXT:    [[TMP8:%.*]] = extractelement <1 x double> [[SPLIT4]], i64 0
+; RM-NEXT:    [[TMP9:%.*]] = insertelement <8 x double> [[TMP7]], double [[TMP8]], i64 4
+; RM-NEXT:    [[TMP10:%.*]] = extractelement <1 x double> [[SPLIT5]], i64 0
+; RM-NEXT:    [[TMP11:%.*]] = insertelement <8 x double> [[TMP9]], double [[TMP10]], i64 5
+; RM-NEXT:    [[TMP12:%.*]] = extractelement <1 x double> [[SPLIT6]], i64 0
+; RM-NEXT:    [[TMP13:%.*]] = insertelement <8 x double> [[TMP11]], double [[TMP12]], i64 6
+; RM-NEXT:    [[TMP14:%.*]] = extractelement <1 x double> [[SPLIT7]], i64 0
+; RM-NEXT:    [[TMP15:%.*]] = insertelement <8 x double> [[TMP13]], double [[TMP14]], i64 7
+; RM-NEXT:    ret <8 x double> [[TMP15]]
+;
+entry:
+  %c  = call <8 x double> @llvm.matrix.transpose(<8 x double> %a, i32 8, i32 1)
+  ret <8 x double> %c
+}
+
+declare <12 x double> @llvm.matrix.transpose.v12f64(<12 x double>, i32, i32)
+
+define <12 x double> @transpose_double_3x4(<12 x double> %a) {
+; RM-LABEL: @transpose_double_3x4(
+; RM-NEXT:  entry:
+; RM-NEXT:    [[SPLIT:%.*]] = shufflevector <12 x double> [[A:%.*]], <12 x double> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+; RM-NEXT:    [[SPLIT1:%.*]] = shufflevector <12 x double> [[A]], <12 x double> undef, <4 x i32> <i32 4, i32 5, i32 6, i32 7>
+; RM-NEXT:    [[SPLIT2:%.*]] = shufflevector <12 x double> [[A]], <12 x double> undef, <4 x i32> <i32 8, i32 9, i32 10, i32 11>
+; RM-NEXT:    [[TMP0:%.*]] = extractelement <4 x double> [[SPLIT]], i64 0
+; RM-NEXT:    [[TMP1:%.*]] = insertelement <3 x double> undef, double [[TMP0]], i64 0
+; RM-NEXT:    [[TMP2:%.*]] = extractelement <4 x double> [[SPLIT1]], i64 0
+; RM-NEXT:    [[TMP3:%.*]] = insertelement <3 x double> [[TMP1]], double [[TMP2]], i64 1
+; RM-NEXT:    [[TMP4:%.*]] = extractelement <4 x double> [[SPLIT2]], i64 0
+; RM-NEXT:    [[TMP5:%.*]] = insertelement <3 x double> [[TMP3]], double [[TMP4]], i64 2
+; RM-NEXT:    [[TMP6:%.*]] = extractelement <4 x double> [[SPLIT]], i64 1
+; RM-NEXT:    [[TMP7:%.*]] = insertelement <3 x double> undef, double [[TMP6]], i64 0
+; RM-NEXT:    [[TMP8:%.*]] = extractelement <4 x double> [[SPLIT1]], i64 1
+; RM-NEXT:    [[TMP9:%.*]] = insertelement <3 x double> [[TMP7]], double [[TMP8]], i64 1
+; RM-NEXT:    [[TMP10:%.*]] = extractelement <4 x double> [[SPLIT2]], i64 1
+; RM-NEXT:    [[TMP11:%.*]] = insertelement <3 x double> [[TMP9]], double [[TMP10]], i64 2
+; RM-NEXT:    [[TMP12:%.*]] = extractelement <4 x double> [[SPLIT]], i64 2
+; RM-NEXT:    [[TMP13:%.*]] = insertelement <3 x double> undef, double [[TMP12]], i64 0
+; RM-NEXT:    [[TMP14:%.*]] = extractelement <4 x double> [[SPLIT1]], i64 2
+; RM-NEXT:    [[TMP15:%.*]] = insertelement <3 x double> [[TMP13]], double [[TMP14]], i64 1
+; RM-NEXT:    [[TMP16:%.*]] = extractelement <4 x double> [[SPLIT2]], i64 2
+; RM-NEXT:    [[TMP17:%.*]] = insertelement <3 x double> [[TMP15]], double [[TMP16]], i64 2
+; RM-NEXT:    [[TMP18:%.*]] = extractelement <4 x double> [[SPLIT]], i64 3
+; RM-NEXT:    [[TMP19:%.*]] = insertelement <3 x double> undef, double [[TMP18]], i64 0
+; RM-NEXT:    [[TMP20:%.*]] = extractelement <4 x double> [[SPLIT1]], i64 3
+; RM-NEXT:    [[TMP21:%.*]] = insertelement <3 x double> [[TMP19]], double [[TMP20]], i64 1
+; RM-NEXT:    [[TMP22:%.*]] = extractelement <4 x double> [[SPLIT2]], i64 3
+; RM-NEXT:    [[TMP23:%.*]] = insertelement <3 x double> [[TMP21]], double [[TMP22]], i64 2
+; RM-NEXT:    [[TMP24:%.*]] = shufflevector <3 x double> [[TMP5]], <3 x double> [[TMP11]], <6 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5>
+; RM-NEXT:    [[TMP25:%.*]] = shufflevector <3 x double> [[TMP17]], <3 x double> [[TMP23]], <6 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5>
+; RM-NEXT:    [[TMP26:%.*]] = shufflevector <6 x double> [[TMP24]], <6 x double> [[TMP25]], <12 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11>
+; RM-NEXT:    ret <12 x double> [[TMP26]]
+;
+entry:
+  %c  = call <12 x double> @llvm.matrix.transpose.v12f64(<12 x double> %a, i32 3, i32 4)
+  ret <12 x double> %c
+}

llvm/test/Transforms/LowerMatrixIntrinsics/transpose-float-row-major.ll

+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt -lower-matrix-intrinsics -matrix-default-layout=row-major -S < %s | FileCheck --check-prefix=RM %s
+
+define <8 x float> @transpose(<8 x float> %a) {
+; RM-LABEL: @transpose(
+; RM-NEXT:  entry:
+; RM-NEXT:    [[SPLIT:%.*]] = shufflevector <8 x float> [[A:%.*]], <8 x float> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+; RM-NEXT:    [[SPLIT1:%.*]] = shufflevector <8 x float> [[A]], <8 x float> undef, <4 x i32> <i32 4, i32 5, i32 6, i32 7>
+; RM-NEXT:    [[TMP0:%.*]] = extractelement <4 x float> [[SPLIT]], i64 0
+; RM-NEXT:    [[TMP1:%.*]] = insertelement <2 x float> undef, float [[TMP0]], i64 0
+; RM-NEXT:    [[TMP2:%.*]] = extractelement <4 x float> [[SPLIT1]], i64 0
+; RM-NEXT:    [[TMP3:%.*]] = insertelement <2 x float> [[TMP1]], float [[TMP2]], i64 1
+; RM-NEXT:    [[TMP4:%.*]] = extractelement <4 x float> [[SPLIT]], i64 1
+; RM-NEXT:    [[TMP5:%.*]] = insertelement <2 x float> undef, float [[TMP4]], i64 0
+; RM-NEXT:    [[TMP6:%.*]] = extractelement <4 x float> [[SPLIT1]], i64 1
+; RM-NEXT:    [[TMP7:%.*]] = insertelement <2 x float> [[TMP5]], float [[TMP6]], i64 1
+; RM-NEXT:    [[TMP8:%.*]] = extractelement <4 x float> [[SPLIT]], i64 2
+; RM-NEXT:    [[TMP9:%.*]] = insertelement <2 x float> undef, float [[TMP8]], i64 0
+; RM-NEXT:    [[TMP10:%.*]] = extractelement <4 x float> [[SPLIT1]], i64 2
+; RM-NEXT:    [[TMP11:%.*]] = insertelement <2 x float> [[TMP9]], float [[TMP10]], i64 1
+; RM-NEXT:    [[TMP12:%.*]] = extractelement <4 x float> [[SPLIT]], i64 3
+; RM-NEXT:    [[TMP13:%.*]] = insertelement <2 x float> undef, float [[TMP12]], i64 0
+; RM-NEXT:    [[TMP14:%.*]] = extractelement <4 x float> [[SPLIT1]], i64 3
+; RM-NEXT:    [[TMP15:%.*]] = insertelement <2 x float> [[TMP13]], float [[TMP14]], i64 1
+; RM-NEXT:    [[TMP16:%.*]] = shufflevector <2 x float> [[TMP3]], <2 x float> [[TMP7]], <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+; RM-NEXT:    [[TMP17:%.*]] = shufflevector <2 x float> [[TMP11]], <2 x float> [[TMP15]], <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+; RM-NEXT:    [[TMP18:%.*]] = shufflevector <4 x float> [[TMP16]], <4 x float> [[TMP17]], <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
+; RM-NEXT:    ret <8 x float> [[TMP18]]
+;
+entry:
+  %c  = call <8 x float> @llvm.matrix.transpose(<8 x float> %a, i32 2, i32 4)
+  ret <8 x float> %c
+}
+
+declare <8 x float> @llvm.matrix.transpose(<8 x float>, i32, i32)
+
+define <8 x float> @transpose_single_column(<8 x float> %a) {
+; RM-LABEL: @transpose_single_column(
+; RM-NEXT:  entry:
+; RM-NEXT:    [[SPLIT:%.*]] = shufflevector <8 x float> [[A:%.*]], <8 x float> undef, <1 x i32> zeroinitializer
+; RM-NEXT:    [[SPLIT1:%.*]] = shufflevector <8 x float> [[A]], <8 x float> undef, <1 x i32> <i32 1>
+; RM-NEXT:    [[SPLIT2:%.*]] = shufflevector <8 x float> [[A]], <8 x float> undef, <1 x i32> <i32 2>
+; RM-NEXT:    [[SPLIT3:%.*]] = shufflevector <8 x float> [[A]], <8 x float> undef, <1 x i32> <i32 3>
+; RM-NEXT:    [[SPLIT4:%.*]] = shufflevector <8 x float> [[A]], <8 x float> undef, <1 x i32> <i32 4>
+; RM-NEXT:    [[SPLIT5:%.*]] = shufflevector <8 x float> [[A]], <8 x float> undef, <1 x i32> <i32 5>
+; RM-NEXT:    [[SPLIT6:%.*]] = shufflevector <8 x float> [[A]], <8 x float> undef, <1 x i32> <i32 6>
+; RM-NEXT:    [[SPLIT7:%.*]] = shufflevector <8 x float> [[A]], <8 x float> undef, <1 x i32> <i32 7>
+; RM-NEXT:    [[TMP0:%.*]] = extractelement <1 x float> [[SPLIT]], i64 0
+; RM-NEXT:    [[TMP1:%.*]] = insertelement <8 x float> undef, float [[TMP0]], i64 0
+; RM-NEXT:    [[TMP2:%.*]] = extractelement <1 x float> [[SPLIT1]], i64 0
+; RM-NEXT:    [[TMP3:%.*]] = insertelement <8 x float> [[TMP1]], float [[TMP2]], i64 1
+; RM-NEXT:    [[TMP4:%.*]] = extractelement <1 x float> [[SPLIT2]], i64 0
+; RM-NEXT:    [[TMP5:%.*]] = insertelement <8 x float> [[TMP3]], float [[TMP4]], i64 2
+; RM-NEXT:    [[TMP6:%.*]] = extractelement <1 x float> [[SPLIT3]], i64 0
+; RM-NEXT:    [[TMP7:%.*]] = insertelement <8 x float> [[TMP5]], float [[TMP6]], i64 3
+; RM-NEXT:    [[TMP8:%.*]] = extractelement <1 x float> [[SPLIT4]], i64 0
+; RM-NEXT:    [[TMP9:%.*]] = insertelement <8 x float> [[TMP7]], float [[TMP8]], i64 4
+; RM-NEXT:    [[TMP10:%.*]] = extractelement <1 x float> [[SPLIT5]], i64 0
+; RM-NEXT:    [[TMP11:%.*]] = insertelement <8 x float> [[TMP9]], float [[TMP10]], i64 5
+; RM-NEXT:    [[TMP12:%.*]] = extractelement <1 x float> [[SPLIT6]], i64 0
+; RM-NEXT:    [[TMP13:%.*]] = insertelement <8 x float> [[TMP11]], float [[TMP12]], i64 6
+; RM-NEXT:    [[TMP14:%.*]] = extractelement <1 x float> [[SPLIT7]], i64 0
+; RM-NEXT:    [[TMP15:%.*]] = insertelement <8 x float> [[TMP13]], float [[TMP14]], i64 7
+; RM-NEXT:    ret <8 x float> [[TMP15]]
+;
+entry:
+  %c  = call <8 x float> @llvm.matrix.transpose(<8 x float> %a, i32 8, i32 1)
+  ret <8 x float> %c
+}
+
+declare <12 x float> @llvm.matrix.transpose.v12f32(<12 x float>, i32, i32)
+
+define <12 x float> @transpose_float_3x4(<12 x float> %a) {
+; RM-LABEL: @transpose_float_3x4(
+; RM-NEXT:  entry:
+; RM-NEXT:    [[SPLIT:%.*]] = shufflevector <12 x float> [[A:%.*]], <12 x float> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+; RM-NEXT:    [[SPLIT1:%.*]] = shufflevector <12 x float> [[A]], <12 x float> undef, <4 x i32> <i32 4, i32 5, i32 6, i32 7>
+; RM-NEXT:    [[SPLIT2:%.*]] = shufflevector <12 x float> [[A]], <12 x float> undef, <4 x i32> <i32 8, i32 9, i32 10, i32 11>
+; RM-NEXT:    [[TMP0:%.*]] = extractelement <4 x float> [[SPLIT]], i64 0
+; RM-NEXT:    [[TMP1:%.*]] = insertelement <3 x float> undef, float [[TMP0]], i64 0
+; RM-NEXT:    [[TMP2:%.*]] = extractelement <4 x float> [[SPLIT1]], i64 0
+; RM-NEXT:    [[TMP3:%.*]] = insertelement <3 x float> [[TMP1]], float [[TMP2]], i64 1
+; RM-NEXT:    [[TMP4:%.*]] = extractelement <4 x float> [[SPLIT2]], i64 0
+; RM-NEXT:    [[TMP5:%.*]] = insertelement <3 x float> [[TMP3]], float [[TMP4]], i64 2
+; RM-NEXT:    [[TMP6:%.*]] = extractelement <4 x float> [[SPLIT]], i64 1
+; RM-NEXT:    [[TMP7:%.*]] = insertelement <3 x float> undef, float [[TMP6]], i64 0
+; RM-NEXT:    [[TMP8:%.*]] = extractelement <4 x float> [[SPLIT1]], i64 1
+; RM-NEXT:    [[TMP9:%.*]] = insertelement <3 x float> [[TMP7]], float [[TMP8]], i64 1
+; RM-NEXT:    [[TMP10:%.*]] = extractelement <4 x float> [[SPLIT2]], i64 1
+; RM-NEXT:    [[TMP11:%.*]] = insertelement <3 x float> [[TMP9]], float [[TMP10]], i64 2
+; RM-NEXT:    [[TMP12:%.*]] = extractelement <4 x float> [[SPLIT]], i64 2
+; RM-NEXT:    [[TMP13:%.*]] = insertelement <3 x float> undef, float [[TMP12]], i64 0
+; RM-NEXT:    [[TMP14:%.*]] = extractelement <4 x float> [[SPLIT1]], i64 2
+; RM-NEXT:    [[TMP15:%.*]] = insertelement <3 x float> [[TMP13]], float [[TMP14]], i64 1
+; RM-NEXT:    [[TMP16:%.*]] = extractelement <4 x float> [[SPLIT2]], i64 2
+; RM-NEXT:    [[TMP17:%.*]] = insertelement <3 x float> [[TMP15]], float [[TMP16]], i64 2
+; RM-NEXT:    [[TMP18:%.*]] = extractelement <4 x float> [[SPLIT]], i64 3
+; RM-NEXT:    [[TMP19:%.*]] = insertelement <3 x float> undef, float [[TMP18]], i64 0
+; RM-NEXT:    [[TMP20:%.*]] = extractelement <4 x float> [[SPLIT1]], i64 3
+; RM-NEXT:    [[TMP21:%.*]] = insertelement <3 x float> [[TMP19]], float [[TMP20]], i64 1
+; RM-NEXT:    [[TMP22:%.*]] = extractelement <4 x float> [[SPLIT2]], i64 3
+; RM-NEXT:    [[TMP23:%.*]] = insertelement <3 x float> [[TMP21]], float [[TMP22]], i64 2
+; RM-NEXT:    [[TMP24:%.*]] = shufflevector <3 x float> [[TMP5]], <3 x float> [[TMP11]], <6 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5>
+; RM-NEXT:    [[TMP25:%.*]] = shufflevector <3 x float> [[TMP17]], <3 x float> [[TMP23]], <6 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5>
+; RM-NEXT:    [[TMP26:%.*]] = shufflevector <6 x float> [[TMP24]], <6 x float> [[TMP25]], <12 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11>
+; RM-NEXT:    ret <12 x float> [[TMP26]]
+;
+entry:
+  %c  = call <12 x float> @llvm.matrix.transpose.v12f32(<12 x float> %a, i32 3, i32 4)
+  ret <12 x float> %c
+}

llvm/test/Transforms/LowerMatrixIntrinsics/transpose-i32-row-major.ll

+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt -lower-matrix-intrinsics -matrix-default-layout=row-major -S < %s | FileCheck --check-prefix=RM %s
+
+define <8 x i32> @transpose(<8 x i32> %a) {
+; RM-LABEL: @transpose(
+; RM-NEXT:  entry:
+; RM-NEXT:    [[SPLIT:%.*]] = shufflevector <8 x i32> [[A:%.*]], <8 x i32> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+; RM-NEXT:    [[SPLIT1:%.*]] = shufflevector <8 x i32> [[A]], <8 x i32> undef, <4 x i32> <i32 4, i32 5, i32 6, i32 7>
+; RM-NEXT:    [[TMP0:%.*]] = extractelement <4 x i32> [[SPLIT]], i64 0
+; RM-NEXT:    [[TMP1:%.*]] = insertelement <2 x i32> undef, i32 [[TMP0]], i64 0
+; RM-NEXT:    [[TMP2:%.*]] = extractelement <4 x i32> [[SPLIT1]], i64 0
+; RM-NEXT:    [[TMP3:%.*]] = insertelement <2 x i32> [[TMP1]], i32 [[TMP2]], i64 1
+; RM-NEXT:    [[TMP4:%.*]] = extractelement <4 x i32> [[SPLIT]], i64 1
+; RM-NEXT:    [[TMP5:%.*]] = insertelement <2 x i32> undef, i32 [[TMP4]], i64 0
+; RM-NEXT:    [[TMP6:%.*]] = extractelement <4 x i32> [[SPLIT1]], i64 1
+; RM-NEXT:    [[TMP7:%.*]] = insertelement <2 x i32> [[TMP5]], i32 [[TMP6]], i64 1
+; RM-NEXT:    [[TMP8:%.*]] = extractelement <4 x i32> [[SPLIT]], i64 2
+; RM-NEXT:    [[TMP9:%.*]] = insertelement <2 x i32> undef, i32 [[TMP8]], i64 0
+; RM-NEXT:    [[TMP10:%.*]] = extractelement <4 x i32> [[SPLIT1]], i64 2
+; RM-NEXT:    [[TMP11:%.*]] = insertelement <2 x i32> [[TMP9]], i32 [[TMP10]], i64 1
+; RM-NEXT:    [[TMP12:%.*]] = extractelement <4 x i32> [[SPLIT]], i64 3
+; RM-NEXT:    [[TMP13:%.*]] = insertelement <2 x i32> undef, i32 [[TMP12]], i64 0
+; RM-NEXT:    [[TMP14:%.*]] = extractelement <4 x i32> [[SPLIT1]], i64 3
+; RM-NEXT:    [[TMP15:%.*]] = insertelement <2 x i32> [[TMP13]], i32 [[TMP14]], i64 1
+; RM-NEXT:    [[TMP16:%.*]] = shufflevector <2 x i32> [[TMP3]], <2 x i32> [[TMP7]], <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+; RM-NEXT:    [[TMP17:%.*]] = shufflevector <2 x i32> [[TMP11]], <2 x i32> [[TMP15]], <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+; RM-NEXT:    [[TMP18:%.*]] = shufflevector <4 x i32> [[TMP16]], <4 x i32> [[TMP17]], <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
+; RM-NEXT:    ret <8 x i32> [[TMP18]]
+;
+entry:
+  %c  = call <8 x i32> @llvm.matrix.transpose(<8 x i32> %a, i32 2, i32 4)
+  ret <8 x i32> %c
+}
+
+declare <8 x i32> @llvm.matrix.transpose(<8 x i32>, i32, i32)
+
+define <8 x i32> @transpose_single_column(<8 x i32> %a) {
+; RM-LABEL: @transpose_single_column(
+; RM-NEXT:  entry:
+; RM-NEXT:    [[SPLIT:%.*]] = shufflevector <8 x i32> [[A:%.*]], <8 x i32> undef, <1 x i32> zeroinitializer
+; RM-NEXT:    [[SPLIT1:%.*]] = shufflevector <8 x i32> [[A]], <8 x i32> undef, <1 x i32> <i32 1>
+; RM-NEXT:    [[SPLIT2:%.*]] = shufflevector <8 x i32> [[A]], <8 x i32> undef, <1 x i32> <i32 2>
+; RM-NEXT:    [[SPLIT3:%.*]] = shufflevector <8 x i32> [[A]], <8 x i32> undef, <1 x i32> <i32 3>
+; RM-NEXT:    [[SPLIT4:%.*]] = shufflevector <8 x i32> [[A]], <8 x i32> undef, <1 x i32> <i32 4>
+; RM-NEXT:    [[SPLIT5:%.*]] = shufflevector <8 x i32> [[A]], <8 x i32> undef, <1 x i32> <i32 5>
+; RM-NEXT:    [[SPLIT6:%.*]] = shufflevector <8 x i32> [[A]], <8 x i32> undef, <1 x i32> <i32 6>
+; RM-NEXT:    [[SPLIT7:%.*]] = shufflevector <8 x i32> [[A]], <8 x i32> undef, <1 x i32> <i32 7>
+; RM-NEXT:    [[TMP0:%.*]] = extractelement <1 x i32> [[SPLIT]], i64 0
+; RM-NEXT:    [[TMP1:%.*]] = insertelement <8 x i32> undef, i32 [[TMP0]], i64 0
+; RM-NEXT:    [[TMP2:%.*]] = extractelement <1 x i32> [[SPLIT1]], i64 0
+; RM-NEXT:    [[TMP3:%.*]] = insertelement <8 x i32> [[TMP1]], i32 [[TMP2]], i64 1
+; RM-NEXT:    [[TMP4:%.*]] = extractelement <1 x i32> [[SPLIT2]], i64 0
+; RM-NEXT:    [[TMP5:%.*]] = insertelement <8 x i32> [[TMP3]], i32 [[TMP4]], i64 2
+; RM-NEXT:    [[TMP6:%.*]] = extractelement <1 x i32> [[SPLIT3]], i64 0
+; RM-NEXT:    [[TMP7:%.*]] = insertelement <8 x i32> [[TMP5]], i32 [[TMP6]], i64 3
+; RM-NEXT:    [[TMP8:%.*]] = extractelement <1 x i32> [[SPLIT4]], i64 0
+; RM-NEXT:    [[TMP9:%.*]] = insertelement <8 x i32> [[TMP7]], i32 [[TMP8]], i64 4
+; RM-NEXT:    [[TMP10:%.*]] = extractelement <1 x i32> [[SPLIT5]], i64 0
+; RM-NEXT:    [[TMP11:%.*]] = insertelement <8 x i32> [[TMP9]], i32 [[TMP10]], i64 5
+; RM-NEXT:    [[TMP12:%.*]] = extractelement <1 x i32> [[SPLIT6]], i64 0
+; RM-NEXT:    [[TMP13:%.*]] = insertelement <8 x i32> [[TMP11]], i32 [[TMP12]], i64 6
+; RM-NEXT:    [[TMP14:%.*]] = extractelement <1 x i32> [[SPLIT7]], i64 0
+; RM-NEXT:    [[TMP15:%.*]] = insertelement <8 x i32> [[TMP13]], i32 [[TMP14]], i64 7
+; RM-NEXT:    ret <8 x i32> [[TMP15]]
+;
+entry:
+  %c  = call <8 x i32> @llvm.matrix.transpose(<8 x i32> %a, i32 8, i32 1)
+  ret <8 x i32> %c
+}
+
+declare <12 x i32> @llvm.matrix.transpose.v12i32(<12 x i32>, i32, i32)
+
+define <12 x i32> @transpose_i32_3x4(<12 x i32> %a) {
+; RM-LABEL: @transpose_i32_3x4(
+; RM-NEXT:  entry:
+; RM-NEXT:    [[SPLIT:%.*]] = shufflevector <12 x i32> [[A:%.*]], <12 x i32> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+; RM-NEXT:    [[SPLIT1:%.*]] = shufflevector <12 x i32> [[A]], <12 x i32> undef, <4 x i32> <i32 4, i32 5, i32 6, i32 7>
+; RM-NEXT:    [[SPLIT2:%.*]] = shufflevector <12 x i32> [[A]], <12 x i32> undef, <4 x i32> <i32 8, i32 9, i32 10, i32 11>
+; RM-NEXT:    [[TMP0:%.*]] = extractelement <4 x i32> [[SPLIT]], i64 0
+; RM-NEXT:    [[TMP1:%.*]] = insertelement <3 x i32> undef, i32 [[TMP0]], i64 0
+; RM-NEXT:    [[TMP2:%.*]] = extractelement <4 x i32> [[SPLIT1]], i64 0
+; RM-NEXT:    [[TMP3:%.*]] = insertelement <3 x i32> [[TMP1]], i32 [[TMP2]], i64 1
+; RM-NEXT:    [[TMP4:%.*]] = extractelement <4 x i32> [[SPLIT2]], i64 0
+; RM-NEXT:    [[TMP5:%.*]] = insertelement <3 x i32> [[TMP3]], i32 [[TMP4]], i64 2
+; RM-NEXT:    [[TMP6:%.*]] = extractelement <4 x i32> [[SPLIT]], i64 1
+; RM-NEXT:    [[TMP7:%.*]] = insertelement <3 x i32> undef, i32 [[TMP6]], i64 0
+; RM-NEXT:    [[TMP8:%.*]] = extractelement <4 x i32> [[SPLIT1]], i64 1
+; RM-NEXT:    [[TMP9:%.*]] = insertelement <3 x i32> [[TMP7]], i32 [[TMP8]], i64 1
+; RM-NEXT:    [[TMP10:%.*]] = extractelement <4 x i32> [[SPLIT2]], i64 1
+; RM-NEXT:    [[TMP11:%.*]] = insertelement <3 x i32> [[TMP9]], i32 [[TMP10]], i64 2
+; RM-NEXT:    [[TMP12:%.*]] = extractelement <4 x i32> [[SPLIT]], i64 2
+; RM-NEXT:    [[TMP13:%.*]] = insertelement <3 x i32> undef, i32 [[TMP12]], i64 0
+; RM-NEXT:    [[TMP14:%.*]] = extractelement <4 x i32> [[SPLIT1]], i64 2
+; RM-NEXT:    [[TMP15:%.*]] = insertelement <3 x i32> [[TMP13]], i32 [[TMP14]], i64 1
+; RM-NEXT:    [[TMP16:%.*]] = extractelement <4 x i32> [[SPLIT2]], i64 2
+; RM-NEXT:    [[TMP17:%.*]] = insertelement <3 x i32> [[TMP15]], i32 [[TMP16]], i64 2
+; RM-NEXT:    [[TMP18:%.*]] = extractelement <4 x i32> [[SPLIT]], i64 3
+; RM-NEXT:    [[TMP19:%.*]] = insertelement <3 x i32> undef, i32 [[TMP18]], i64 0
+; RM-NEXT:    [[TMP20:%.*]] = extractelement <4 x i32> [[SPLIT1]], i64 3
+; RM-NEXT:    [[TMP21:%.*]] = insertelement <3 x i32> [[TMP19]], i32 [[TMP20]], i64 1
+; RM-NEXT:    [[TMP22:%.*]] = extractelement <4 x i32> [[SPLIT2]], i64 3
+; RM-NEXT:    [[TMP23:%.*]] = insertelement <3 x i32> [[TMP21]], i32 [[TMP22]], i64 2
+; RM-NEXT:    [[TMP24:%.*]] = shufflevector <3 x i32> [[TMP5]], <3 x i32> [[TMP11]], <6 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5>
+; RM-NEXT:    [[TMP25:%.*]] = shufflevector <3 x i32> [[TMP17]], <3 x i32> [[TMP23]], <6 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5>
+; RM-NEXT:    [[TMP26:%.*]] = shufflevector <6 x i32> [[TMP24]], <6 x i32> [[TMP25]], <12 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11>
+; RM-NEXT:    ret <12 x i32> [[TMP26]]
+;
+entry:
+  %c  = call <12 x i32> @llvm.matrix.transpose.v12i32(<12 x i32> %a, i32 3, i32 4)
+  ret <12 x i32> %c
+}