[mlir][vector] Remove MatrixMultiplyOp and FlatTransposeOp from Vector dialect

This patch deletes `vector.matrix_multiply` and `vector.flat_transpose`,
which are thin wrappers around the corresponding LLVM intrinsics:
  - `llvm.intr.matrix.multiply`
  - `llvm.intr.matrix.transpose`

These Vector dialect ops did not provide additional semantics or
abstraction beyond the LLVM intrinsics. Their removal simplifies the
lowering pipeline without losing any functionality.

The lowering chains:
  - `vector.contract` → `vector.matrix_multiply` → `llvm.intr.matrix.multiply`
  - `vector.transpose` → `vector.flat_transpose` → `llvm.intr.matrix.transpose`

are now replaced with:
  - `vector.contract` → `llvm.intr.matrix.multiply`
  - `vector.transpose` → `llvm.intr.matrix.transpose`

This was accomplished by directly replacing:
  - `vector::MatrixMultiplyOp` with `LLVM::MatrixMultiplyOp`
  - `vector::FlatTransposeOp` with `LLVM::MatrixTransposeOp`

Note: This change introduces a build-time dependency from `Vector` to
`LLVM`. Ideally, such dependencies should be confined to dialect
conversion (`ConvertVectorToLLVM`). However, moving the lowering code
there would introduce notable churn, so this patch leaves the new
dependency in place for now.

Author: banach-space

mlir/include/mlir/Conversion/VectorToLLVM/ConvertVectorToLLVM.h

@@ -13,12 +13,6 @@
 namespace mlir {
 class LLVMTypeConverter;
 
-/// Collect a set of patterns to convert from Vector contractions to LLVM Matrix
-/// Intrinsics. To lower to assembly, the LLVM flag -lower-matrix-intrinsics
-/// will be needed when invoking LLVM.
-void populateVectorToLLVMMatrixConversionPatterns(
-    const LLVMTypeConverter &converter, RewritePatternSet &patterns);
-
 /// Collect a set of patterns to convert from the Vector dialect to LLVM.
 void populateVectorToLLVMConversionPatterns(
     const LLVMTypeConverter &converter, RewritePatternSet &patterns,

mlir/include/mlir/Dialect/Vector/IR/VectorOps.td

@@ -2788,124 +2788,6 @@ def Vector_PrintOp :
     }];
 }
 
-//===----------------------------------------------------------------------===//
-// Ops used for supporting progressive lowering and conversion type changes.
-// The Ops are typically not used directly by higher level dialects, but are
-// used by intra-dialect rewriting rules to bring vector operations closer
-// to the hardware ISA.
-//===----------------------------------------------------------------------===//
-
-/// Vector dialect matrix multiplication op that operates on flattened 1-D
-/// MLIR vectors. This is the counterpart of llvm.matrix.multiply in MLIR.
-/// This may seem redundant with vector.contract but it serves the purposes of
-/// more progressive lowering and localized type conversion on the path:
-///   `vector<...x...xf32> -> vector<...xf32> -> !llvm<... x float>`.
-def Vector_MatmulOp : Vector_Op<"matrix_multiply", [Pure,
-        PredOpTrait<"lhs operand and result have same element type",
-                    TCresVTEtIsSameAsOpBase<0, 0>>,
-        PredOpTrait<"rhs operand and result have same element type",
-                    TCresVTEtIsSameAsOpBase<0, 1>>]>,
-      Arguments<(
-        // TODO: tighten vector element types that make sense.
-        ins FixedVectorOfRankAndType<[1],
-              [AnySignlessInteger, AnySignedInteger, Index, AnyFloat]>:$lhs,
-            FixedVectorOfRankAndType<[1],
-              [AnySignlessInteger, AnySignedInteger, Index, AnyFloat]>:$rhs,
-            I32Attr:$lhs_rows, I32Attr:$lhs_columns, I32Attr:$rhs_columns)>,
-      Results<(
-        outs FixedVectorOfRankAndType<[1],
-               [AnySignlessInteger, AnySignedInteger, Index, AnyFloat]>:$res)>
-{
-  let summary = "Vector matrix multiplication op that operates on flattened 1-D"
-    " MLIR vectors";
-  let description = [{
-    This is the counterpart of llvm.matrix.multiply in MLIR. It serves the
-    purposes of more progressive lowering and localized type conversion.
-    Higher levels typically lower matrix multiplications into 'vector.contract'
-    operations. Subsequent rewriting rule progressively lower these operations
-    into 'vector.matrix_multiply' operations to bring the operations closer
-    to the hardware ISA.
-
-    The ‘vector.matrix_multiply’ op treats `lhs` as matrix with <lhs_rows> rows
-    and <lhs_columns> columns, `rhs` as matrix with <lhs_columns> rows and
-    <rhs_columns> and multiplies them. The result matrix is returned embedded in
-    the result vector.
-
-    Note, the corresponding LLVM intrinsic, `@llvm.matrix.multiply.*`, does not
-    support scalable vectors. Hence, this Op is only available for fixed-width
-    vectors. Also see:
-
-    http://llvm.org/docs/LangRef.html#llvm-matrix-multiply-intrinsic
-
-    Example:
-
-    ```mlir
-    %C = vector.matrix_multiply %A, %B
-      { lhs_rows = 4: i32, lhs_columns = 16: i32 , rhs_columns = 3: i32 } :
-      (vector<64xf64>, vector<48xf64>) -> vector<12xf64>
-    ```
-  }];
-  let builders = [
-   OpBuilder<(ins "Value":$lhs, "Value":$rhs, "unsigned":$lhsRows,
-     "unsigned":$lhsColumns, "unsigned":$rhsColumns),
-   [{
-     $_state.addOperands({lhs, rhs});
-     $_state.addAttribute("lhs_rows",$_builder.getI32IntegerAttr(lhsRows));
-     $_state.addAttribute("lhs_columns",$_builder.getI32IntegerAttr(lhsColumns));
-     $_state.addAttribute("rhs_columns",$_builder.getI32IntegerAttr(rhsColumns));
-     $_state.addTypes(VectorType::get(lhsRows * rhsColumns,
-       ::llvm::cast<VectorType>(lhs.getType()).getElementType()));
-   }]>,
-  ];
-  let assemblyFormat = "$lhs `,` $rhs attr-dict "
-    "`:` `(` type($lhs) `,` type($rhs) `)` `->` type($res)";
-}
-
-/// Vector dialect matrix transposition op that operates on flattened 1-D
-/// MLIR vectors. This is the counterpart of llvm.matrix.transpose in MLIR.
-/// This may seem redundant with vector.transpose but it serves the purposes of
-/// more progressive lowering and localized type conversion on the path:
-///   `vector<...x...xf32> -> vector<...xf32> -> !llvm<... x float>`.
-def Vector_FlatTransposeOp : Vector_Op<"flat_transpose", [Pure,
-  PredOpTrait<"source operand and result have same element type",
-                 TCresVTEtIsSameAsOpBase<0, 0>>]>,
-    Arguments<(
-      // TODO: tighten vector element types that make sense.
-      ins FixedVectorOfRankAndType<[1],
-            [AnySignlessInteger, AnySignedInteger, Index, AnyFloat]>:$matrix,
-          I32Attr:$rows, I32Attr:$columns)>,
-    Results<(
-      outs FixedVectorOfRankAndType<[1],
-             [AnySignlessInteger, AnySignedInteger, Index, AnyFloat]>:$res)> {
-  let summary = "Vector matrix transposition on flattened 1-D MLIR vectors";
-  let description = [{
-    This is the counterpart of llvm.matrix.transpose in MLIR. It serves
-    the purposes of more progressive lowering and localized type conversion.
-    Higher levels typically lower matrix transpositions into 'vector.transpose'
-    operations. Subsequent rewriting rule progressively lower these operations
-    into 'vector.flat_transpose' operations to bring the operations closer
-    to the hardware ISA.
-
-    The `vector.flat_transpose` op treats the 1-D input `matrix` as
-    a 2-D matrix with <rows> rows and <columns> columns, and returns the
-    transposed matrix in flattened form in 'res'.
-
-    Note, the corresponding LLVM intrinsic, `@llvm.matrix.transpose.*`, does not
-    support scalable vectors. Hence, this Op is only available for fixed-width
-    vectors. Also see:
-
-    http://llvm.org/docs/LangRef.html#llvm-matrix-transpose-intrinsic
-
-    Example:
-
-    ```mlir
-    %1 = vector.flat_transpose %0 {columns = 4 : i32, rows = 4 : i32}
-       : vector<16xf32> -> vector<16xf32>
-    ```
-  }];
-  let assemblyFormat = "$matrix attr-dict `:` type($matrix) `->` type($res)";
-}
-
 //===----------------------------------------------------------------------===//
 // SplatOp
 //===----------------------------------------------------------------------===//

mlir/lib/Conversion/VectorToLLVM/ConvertVectorToLLVM.cpp

@@ -184,41 +184,6 @@ class VectorBitCastOpConversion
   }
 };
 
-/// Conversion pattern for a vector.matrix_multiply.
-/// This is lowered directly to the proper llvm.intr.matrix.multiply.
-class VectorMatmulOpConversion
-    : public ConvertOpToLLVMPattern<vector::MatmulOp> {
-public:
-  using ConvertOpToLLVMPattern<vector::MatmulOp>::ConvertOpToLLVMPattern;
-
-  LogicalResult
-  matchAndRewrite(vector::MatmulOp matmulOp, OpAdaptor adaptor,
-                  ConversionPatternRewriter &rewriter) const override {
-    rewriter.replaceOpWithNewOp<LLVM::MatrixMultiplyOp>(
-        matmulOp, typeConverter->convertType(matmulOp.getRes().getType()),
-        adaptor.getLhs(), adaptor.getRhs(), matmulOp.getLhsRows(),
-        matmulOp.getLhsColumns(), matmulOp.getRhsColumns());
-    return success();
-  }
-};
-
-/// Conversion pattern for a vector.flat_transpose.
-/// This is lowered directly to the proper llvm.intr.matrix.transpose.
-class VectorFlatTransposeOpConversion
-    : public ConvertOpToLLVMPattern<vector::FlatTransposeOp> {
-public:
-  using ConvertOpToLLVMPattern<vector::FlatTransposeOp>::ConvertOpToLLVMPattern;
-
-  LogicalResult
-  matchAndRewrite(vector::FlatTransposeOp transOp, OpAdaptor adaptor,
-                  ConversionPatternRewriter &rewriter) const override {
-    rewriter.replaceOpWithNewOp<LLVM::MatrixTransposeOp>(
-        transOp, typeConverter->convertType(transOp.getRes().getType()),
-        adaptor.getMatrix(), transOp.getRows(), transOp.getColumns());
-    return success();
-  }
-};
-
 /// Overloaded utility that replaces a vector.load, vector.store,
 /// vector.maskedload and vector.maskedstore with their respective LLVM
 /// couterparts.
@@ -2071,12 +2036,6 @@ void mlir::populateVectorToLLVMConversionPatterns(
       converter);
 }
 
-void mlir::populateVectorToLLVMMatrixConversionPatterns(
-    const LLVMTypeConverter &converter, RewritePatternSet &patterns) {
-  patterns.add<VectorMatmulOpConversion>(converter);
-  patterns.add<VectorFlatTransposeOpConversion>(converter);
-}
-
 namespace {
 struct VectorToLLVMDialectInterface : public ConvertToLLVMPatternInterface {
   using ConvertToLLVMPatternInterface::ConvertToLLVMPatternInterface;

mlir/lib/Conversion/VectorToLLVM/ConvertVectorToLLVMPass.cpp

@@ -96,11 +96,9 @@ void ConvertVectorToLLVMPass::runOnOperation() {
   LLVMTypeConverter converter(&getContext(), options);
   RewritePatternSet patterns(&getContext());
   populateVectorTransferLoweringPatterns(patterns);
-  populateVectorToLLVMMatrixConversionPatterns(converter, patterns);
   populateVectorToLLVMConversionPatterns(
       converter, patterns, reassociateFPReductions, force32BitVectorIndices,
       useVectorAlignment);
-  populateVectorToLLVMMatrixConversionPatterns(converter, patterns);
 
   // Architecture specific augmentations.
   LLVMConversionTarget target(getContext());

mlir/lib/Dialect/Arith/Transforms/EmulateUnsupportedFloats.cpp

@@ -118,9 +118,9 @@ void mlir::arith::populateEmulateUnsupportedFloatsLegality(
         return converter.isLegal(op);
       });
   // Manually mark arithmetic-performing vector instructions.
-  target.addDynamicallyLegalOp<
-      vector::ContractionOp, vector::ReductionOp, vector::MultiDimReductionOp,
-      vector::FMAOp, vector::OuterProductOp, vector::MatmulOp, vector::ScanOp>(
+  target.addDynamicallyLegalOp<vector::ContractionOp, vector::ReductionOp,
+                               vector::MultiDimReductionOp, vector::FMAOp,
+                               vector::OuterProductOp, vector::ScanOp>(
       [&](Operation *op) { return converter.isLegal(op); });
   target.addLegalOp<arith::BitcastOp, arith::ExtFOp, arith::TruncFOp,
                     arith::ConstantOp, vector::SplatOp>();

mlir/lib/Dialect/Vector/Transforms/CMakeLists.txt

@@ -50,6 +50,7 @@ add_mlir_dialect_library(MLIRVectorTransforms
   MLIRTensorDialect
   MLIRTransforms
   MLIRVectorDialect
+  MLIRLLVMDialect
   MLIRVectorInterfaces
   MLIRVectorUtils
   )

mlir/lib/Dialect/Vector/Transforms/LowerVectorContract.cpp

@@ -1271,12 +1271,11 @@ class OuterProductOpLowering : public OpRewritePattern<vector::OuterProductOp> {
 ///    %mtb = maybe_transpose
 ///    %flattened_a = vector.shape_cast %mta
 ///    %flattened_b = vector.shape_cast %mtb
-///    %flattened_d = vector.matrix_multiply %flattened_a, %flattened_b
+///    %flattened_d = llvm.intr.matrix.multiply %flattened_a, %flattened_b
 ///    %mtd = vector.shape_cast %flattened_d
 ///    %d = maybe_untranspose %mtd
 ///    %e = add %c, %d
 /// ```
-/// `vector.matrix_multiply` later lowers to `llvm.matrix.multiply`.
 //
 /// This only kicks in when vectorContractLowering is set to `Matmul`.
 /// vector.transpose operations are inserted if the vector.contract op is not a
@@ -1353,8 +1352,12 @@ FailureOr<Value> ContractionOpToMatmulOpLowering::matchAndRewriteMaskableOp(
       VectorType::get(rhsType.getNumElements(), rhsType.getElementType());
   rhs = rew.create<vector::ShapeCastOp>(loc, flattenedRHSType, rhs);
 
-  Value mul = rew.create<vector::MatmulOp>(loc, lhs, rhs, lhsRows, lhsColumns,
-                                           rhsColumns);
+  Value mul = rew.create<LLVM::MatrixMultiplyOp>(
+      loc,
+      VectorType::get(lhsRows * rhsColumns,
+                      cast<VectorType>(lhs.getType()).getElementType()),
+      lhs, rhs, lhsRows, lhsColumns, rhsColumns);
+
   mul = rew.create<vector::ShapeCastOp>(
       loc,
       VectorType::get({lhsRows, rhsColumns},

mlir/lib/Dialect/Vector/Transforms/LowerVectorTranspose.cpp

@@ -337,7 +337,7 @@ class TransposeOpLowering : public OpRewritePattern<vector::TransposeOp> {
           rewriter.create<vector::ShapeCastOp>(loc, flattenedType, input);
       auto rows = rewriter.getI32IntegerAttr(resType.getShape()[0]);
       auto columns = rewriter.getI32IntegerAttr(resType.getShape()[1]);
-      Value trans = rewriter.create<vector::FlatTransposeOp>(
+      Value trans = rewriter.create<LLVM::MatrixTransposeOp>(
           loc, flattenedType, matrix, rows, columns);
       rewriter.replaceOpWithNewOp<vector::ShapeCastOp>(op, resType, trans);
       return success();

mlir/test/Conversion/VectorToLLVM/vector-to-llvm.mlir

@@ -1424,36 +1424,6 @@ func.func @fma_scalable(%vec_1d: vector<[8]xf32>, %vec_2d: vector<2x[4]xf32>, %v
 
   return %0, %1, %2: vector<[8]xf32>, vector<2x[4]xf32>, vector<1x1x[1]xf32>
 }
-// -----
-
-//===----------------------------------------------------------------------===//
-// vector.matrix_multiply
-//===----------------------------------------------------------------------===//
-
-//                          4x16                16x3               4x3
-func.func @matrix_ops(%A: vector<64xf64>, %B: vector<48xf64>) -> vector<12xf64> {
-  %C = vector.matrix_multiply %A, %B
-    { lhs_rows = 4: i32, lhs_columns = 16: i32 , rhs_columns = 3: i32 } :
-    (vector<64xf64>, vector<48xf64>) -> vector<12xf64>
-  return %C: vector<12xf64>
-}
-// CHECK-LABEL: @matrix_ops
-//       CHECK:   llvm.intr.matrix.multiply %{{.*}}, %{{.*}} {
-//  CHECK-SAME: lhs_columns = 16 : i32, lhs_rows = 4 : i32, rhs_columns = 3 : i32
-//  CHECK-SAME: } : (vector<64xf64>, vector<48xf64>) -> vector<12xf64>
-
-// -----
-
-func.func @matrix_ops_index(%A: vector<64xindex>, %B: vector<48xindex>) -> vector<12xindex> {
-  %C = vector.matrix_multiply %A, %B
-    { lhs_rows = 4: i32, lhs_columns = 16: i32 , rhs_columns = 3: i32 } :
-    (vector<64xindex>, vector<48xindex>) -> vector<12xindex>
-  return %C: vector<12xindex>
-}
-// CHECK-LABEL: @matrix_ops_index
-//       CHECK:   llvm.intr.matrix.multiply %{{.*}}, %{{.*}} {
-//  CHECK-SAME: lhs_columns = 16 : i32, lhs_rows = 4 : i32, rhs_columns = 3 : i32
-//  CHECK-SAME: } : (vector<64xi64>, vector<48xi64>) -> vector<12xi64>
 
 // -----
 
@@ -1602,56 +1572,6 @@ func.func @create_mask_1d_scalable(%num_elems : index) -> vector<[4]xi1> {
 
 // -----
 
-//===----------------------------------------------------------------------===//
-// vector.flat_transpose
-//===----------------------------------------------------------------------===//
-
-func.func @flat_transpose(%arg0: vector<16xf32>) -> vector<16xf32> {
-  %0 = vector.flat_transpose %arg0 { rows = 4: i32, columns = 4: i32 }
-     : vector<16xf32> -> vector<16xf32>
-  return %0 : vector<16xf32>
-}
-
-// CHECK-LABEL: func @flat_transpose
-// CHECK-SAME:  %[[A:.*]]: vector<16xf32>
-// CHECK:       %[[T:.*]] = llvm.intr.matrix.transpose %[[A]]
-// CHECK-SAME:      {columns = 4 : i32, rows = 4 : i32} :
-// CHECK-SAME:      vector<16xf32> into vector<16xf32>
-// CHECK:       return %[[T]] : vector<16xf32>
-
-// -----
-
-func.func @flat_transpose_index(%arg0: vector<16xindex>) -> vector<16xindex> {
-  %0 = vector.flat_transpose %arg0 { rows = 4: i32, columns = 4: i32 }
-     : vector<16xindex> -> vector<16xindex>
-  return %0 : vector<16xindex>
-}
-// CHECK-LABEL: func @flat_transpose_index
-// CHECK-SAME:  %[[A:.*]]: vector<16xindex>
-// CHECK:       %[[T0:.*]] = builtin.unrealized_conversion_cast %[[A]] : vector<16xindex> to vector<16xi64>
-// CHECK:       %[[T1:.*]] = llvm.intr.matrix.transpose %[[T0]]
-// CHECK-SAME:      {columns = 4 : i32, rows = 4 : i32} :
-// CHECK-SAME:      vector<16xi64> into vector<16xi64>
-// CHECK:       %[[T2:.*]] = builtin.unrealized_conversion_cast %[[T1]] : vector<16xi64> to vector<16xindex>
-// CHECK:       return %[[T2]] : vector<16xindex>
-
-// -----
-
-func.func @flat_transpose(%arg0: vector<16xf32>) -> vector<16xf32> {
-  %0 = vector.flat_transpose %arg0 { rows = 4: i32, columns = 4: i32 }
-     : vector<16xf32> -> vector<16xf32>
-  return %0 : vector<16xf32>
-}
-
-// CHECK-LABEL: func @flat_transpose
-// CHECK-SAME:  %[[A:.*]]: vector<16xf32>
-// CHECK:       %[[T:.*]] = llvm.intr.matrix.transpose %[[A]]
-// CHECK-SAME:      {columns = 4 : i32, rows = 4 : i32} :
-// CHECK-SAME:      vector<16xf32> into vector<16xf32>
-// CHECK:       return %[[T]] : vector<16xf32>
-
-// -----
-
 //===----------------------------------------------------------------------===//
 // vector.gather
 //

mlir/test/Dialect/Vector/invalid.mlir

@@ -1321,13 +1321,6 @@ func.func @transpose_dim_size_mismatch(%arg0: vector<11x7x3x2xi32>) {
 
 // -----
 
-func.func @flat_transpose_type_mismatch(%arg0: vector<16xf32>) {
-  // expected-error@+1 {{'vector.flat_transpose' op failed to verify that source operand and result have same element type}}
-  %0 = vector.flat_transpose %arg0 { rows = 4: i32, columns = 4: i32 } : vector<16xf32> -> vector<16xf64>
-}
-
-// -----
-
 func.func @type_cast_layout(%arg0: memref<4x3xf32, affine_map<(d0, d1)[s0, s1, s2] -> (d0 * s0 + d1 * s1 + s2)>>) {
   // expected-error@+1 {{expects operand to be a memref with identity layout}}
   %0 = vector.type_cast %arg0: memref<4x3xf32, affine_map<(d0, d1)[s0, s1, s2] -> (d0 * s0 + d1 * s1 + s2)>> to memref<vector<4x3xf32>>
@@ -1939,26 +1932,6 @@ func.func @invalid_step_2d() {
 
 // -----
 
-func.func @matrix_multiply_scalable(%a: vector<[4]xf64>, %b: vector<4xf64>) {
-  // expected-error @+1 {{'vector.matrix_multiply' op operand #0 must be fixed-length vector of signless integer or signed integer or index or floating-point values of ranks 1, but got 'vector<[4]xf64>'}}
-  %c = vector.matrix_multiply %a, %b {
-    lhs_rows = 2: i32,
-    lhs_columns = 2: i32 ,
-    rhs_columns = 2: i32 }
-  : (vector<[4]xf64>, vector<4xf64>) -> vector<4xf64>
-
-  return
-}
-
-// -----
-
-func.func @flat_transpose_scalable(%arg0: vector<[16]xf32>) -> vector<[16]xf32> {
-  // expected-error @+1 {{'vector.flat_transpose' op operand #0 must be fixed-length vector of signless integer or signed integer or index or floating-point values of ranks 1, but got 'vector<[16]xf32>'}}
-  %0 = vector.flat_transpose %arg0 { rows = 4: i32, columns = 4: i32 }
-     : vector<[16]xf32> -> vector<[16]xf32>
-  return %0 : vector<[16]xf32>
-}
-
 //===----------------------------------------------------------------------===//
 // vector.splat
 //===----------------------------------------------------------------------===//