[Mlir-commits] [mlir] [mlir][ArmSME] Support widening outer products (PR #78975)

Cullen Rhodes llvmlistbot at llvm.org
Wed Jan 24 05:22:49 PST 2024


https://github.com/c-rhodes updated https://github.com/llvm/llvm-project/pull/78975

>From f1fe6a22df1ef9ac6d54e4d26afcdb03c2faa173 Mon Sep 17 00:00:00 2001
From: Cullen Rhodes <cullen.rhodes at arm.com>
Date: Wed, 24 Jan 2024 10:15:14 +0000
Subject: [PATCH 1/2] [mlir][ArmSVE] add zip1 intrinsic

---
 mlir/include/mlir/Dialect/ArmSVE/IR/ArmSVE.td | 4 ++++
 mlir/test/Target/LLVMIR/arm-sve.mlir          | 7 +++++++
 2 files changed, 11 insertions(+)

diff --git a/mlir/include/mlir/Dialect/ArmSVE/IR/ArmSVE.td b/mlir/include/mlir/Dialect/ArmSVE/IR/ArmSVE.td
index e3f3d9e62e8fb3..754413a1ad491e 100644
--- a/mlir/include/mlir/Dialect/ArmSVE/IR/ArmSVE.td
+++ b/mlir/include/mlir/Dialect/ArmSVE/IR/ArmSVE.td
@@ -410,4 +410,8 @@ def ConvertToSvboolIntrOp :
     /*overloadedResults=*/[]>,
     Arguments<(ins SVEPredicate:$mask)>;
 
+def Zip1IntrOp :
+  ArmSVE_IntrBinaryOverloadedOp<"zip1">,
+  Arguments<(ins AnyScalableVector, AnyScalableVector)>;
+
 #endif // ARMSVE_OPS
diff --git a/mlir/test/Target/LLVMIR/arm-sve.mlir b/mlir/test/Target/LLVMIR/arm-sve.mlir
index b63d3f06515690..002b1f9d804a7c 100644
--- a/mlir/test/Target/LLVMIR/arm-sve.mlir
+++ b/mlir/test/Target/LLVMIR/arm-sve.mlir
@@ -314,3 +314,10 @@ llvm.func @arm_sve_convert_to_svbool(
     : (vector<[1]xi1>) -> vector<[16]xi1>
   llvm.return
 }
+
+// CHECK-LABEL: @arm_sve_zip1
+// CHECK-NEXT: call <vscale x 8 x half> @llvm.aarch64.sve.zip1.nxv8f16(<vscale x 8 x half> %{{.*}}, <vscale x 8 x half> {{.*}})
+llvm.func @arm_sve_zip1(%arg0 : vector<[8]xf16>) -> vector<[8]xf16> {
+  %0 = "arm_sve.intr.zip1"(%arg0, %arg0) : (vector<[8]xf16>, vector<[8]xf16>) -> vector<[8]xf16>
+  llvm.return %0 : vector<[8]xf16>
+}

>From bc9df962fd466c372f37404ae6d9b9732d7ba08f Mon Sep 17 00:00:00 2001
From: Cullen Rhodes <cullen.rhodes at arm.com>
Date: Tue, 12 Dec 2023 15:03:34 +0000
Subject: [PATCH 2/2] [mlir][ArmSME] Support 2-way widening outer products

This patch introduces support for 2-way widening outer products. This
enables the folding of 2 'arm_sme.outerproduct' operations that are
chained via the accumulator into single widened operations.

Changes:

- Add 'llvm.aarch64.sme.[us]mop[as].za32' intrinsics for 2-way variants.
  These map to instruction variants added in SME2 and use different
  intrinsics. Intrinsics are already implemented for widening variants
  from SME1.
- Adds the following operations:
  - fmopa_wide_2way, fmops_wide_2way
  - smopa_wide_2way, smops_wide_2way
  - umopa_wide_2way, umops_wide_2way
- Implements conversions for the above ops to intrinsics in ArmSMEToLLVM.
- Adds a pass 'arm-sme-outer-product' widening that folds
  'arm_sme.outerproduct' operations.

For a detailed description of these operations see the
'arm_sme.fmopa_wide_2way' description.

The reason for introducing many operations rather than one is the
signed/unsigned variants can't be distinguished with types (e.g., ui16,
si16) since 'arith.extui' and 'arith.extsi' only support signless
integers. A single operation would require this information and an
attribute (for example) for the sign doesn't feel right if
floating-point types are also supported where this wouldn't apply.
Furthermore, the SME FP8 extensions (FEAT_SME_F8F16, FEAT_SME_F8F32)
introduce FMOPA 2-way (FP8 to FP16) and 4-way (FP8 to FP32) variants but
no subtract variant. Whilst these are not supported in this patch, it
felt simpler to have separate ops for add/subtract given this.
---
 .../Dialect/ArmSME/IR/ArmSMEIntrinsicOps.td   |   4 +
 .../mlir/Dialect/ArmSME/IR/ArmSMEOps.td       | 294 ++++++++++++++++++
 .../mlir/Dialect/ArmSME/Transforms/Passes.h   |   3 +
 .../mlir/Dialect/ArmSME/Transforms/Passes.td  |  39 +++
 .../Dialect/ArmSME/Transforms/Transforms.h    |   4 +
 .../Conversion/ArmSMEToLLVM/ArmSMEToLLVM.cpp  |  66 +++-
 .../Dialect/ArmSME/Transforms/CMakeLists.txt  |   2 +
 .../Transforms/OuterProductWidening.cpp       | 239 ++++++++++++++
 .../ArmSMEToLLVM/arm-sme-to-llvm.mlir         |  96 ++++++
 mlir/test/Dialect/ArmSME/invalid.mlir         |  53 ++++
 .../ArmSME/outer-product-widening.mlir        | 192 ++++++++++++
 mlir/test/Dialect/ArmSME/roundtrip.mlir       | 112 +++++++
 .../ArmSME/test-outerproduct-f16f16f32.mlir   | 100 ++++++
 mlir/test/Target/LLVMIR/arm-sme.mlir          |  12 +
 14 files changed, 1214 insertions(+), 2 deletions(-)
 create mode 100644 mlir/lib/Dialect/ArmSME/Transforms/OuterProductWidening.cpp
 create mode 100644 mlir/test/Dialect/ArmSME/outer-product-widening.mlir
 create mode 100644 mlir/test/Integration/Dialect/Vector/CPU/ArmSME/test-outerproduct-f16f16f32.mlir

diff --git a/mlir/include/mlir/Dialect/ArmSME/IR/ArmSMEIntrinsicOps.td b/mlir/include/mlir/Dialect/ArmSME/IR/ArmSMEIntrinsicOps.td
index d85ef963ae5dc4..f051e03efbcda6 100644
--- a/mlir/include/mlir/Dialect/ArmSME/IR/ArmSMEIntrinsicOps.td
+++ b/mlir/include/mlir/Dialect/ArmSME/IR/ArmSMEIntrinsicOps.td
@@ -105,6 +105,10 @@ def LLVM_aarch64_sme_sumopa_wide : ArmSME_IntrMopOverloadedOp<"sumopa.wide">;
 def LLVM_aarch64_sme_sumops_wide : ArmSME_IntrMopOverloadedOp<"sumops.wide">;
 def LLVM_aarch64_sme_usmopa_wide : ArmSME_IntrMopOverloadedOp<"usmopa.wide">;
 def LLVM_aarch64_sme_usmops_wide : ArmSME_IntrMopOverloadedOp<"usmops.wide">;
+def LLVM_aarch64_sme_smopa_za32 : ArmSME_IntrMopOverloadedOp<"smopa.za32">;
+def LLVM_aarch64_sme_umopa_za32 : ArmSME_IntrMopOverloadedOp<"umopa.za32">;
+def LLVM_aarch64_sme_smops_za32 : ArmSME_IntrMopOverloadedOp<"smops.za32">;
+def LLVM_aarch64_sme_umops_za32 : ArmSME_IntrMopOverloadedOp<"umops.za32">;
 
 class ArmSME_IntrLoadStoreOp<string mnemonic>
     : ArmSME_IntrOp<mnemonic,
diff --git a/mlir/include/mlir/Dialect/ArmSME/IR/ArmSMEOps.td b/mlir/include/mlir/Dialect/ArmSME/IR/ArmSMEOps.td
index 8a34ad7e52012f..3544df494d33dd 100644
--- a/mlir/include/mlir/Dialect/ArmSME/IR/ArmSMEOps.td
+++ b/mlir/include/mlir/Dialect/ArmSME/IR/ArmSMEOps.td
@@ -814,6 +814,300 @@ let arguments = (ins
   }];
 }
 
+class OuterProductWideBase<string mnemonic,
+                           list<Type> allowedInputVectorTypes,
+                           list<Type> allowedResultVectorTypes,
+                           int numOuterProducts> :
+  ArmSME_Op<mnemonic, [
+    ArmSMETileOpInterface,
+    AttrSizedOperandSegments,
+    AllTypesMatch<["lhs", "rhs"]>,
+    HasMatchingMaskTypeConstraint<"lhs", "lhsMask">,
+    HasMatchingMaskTypeConstraint<"rhs", "rhsMask">,
+    PredOpTrait<
+      "both `lhsMask` and `rhsMask` should be provided or neither",
+      CPred<"bool(getLhsMask()) == bool(getRhsMask())">
+    >,
+    OptionalTypesMatchWith<"result and acc have the same type",
+                           "result", "acc", "::llvm::cast<Type>($_self)">,
+    // this trait ensures the input type match the correct output type for ops
+    // that takes multiple inputs and outputs (i.e., 4-way).
+    PredOpTrait<
+      "tile element size equals lhs element size * " # numOuterProducts,
+      CPred<"getTileType().getElementTypeBitWidth() == "
+            "(getLhsType().getElementTypeBitWidth() * " # numOuterProducts # ")">
+    >,
+  ]> {
+
+  let arguments = (ins
+    AnyTypeOf<allowedInputVectorTypes>:$lhs, AnyVector:$rhs,
+    Optional<AnyVector>:$lhsMask, Optional<AnyVector>:$rhsMask,
+    Optional<AnyVector>:$acc);
+  let results = (outs AnyTypeOf<allowedResultVectorTypes>:$result);
+
+  let assemblyFormat = [{
+    $lhs `,` $rhs
+    oilist(
+        `acc` `` `(` $acc `)`
+      | `masks` `` `(` $lhsMask `,` $rhsMask `)`
+    ) attr-dict `:` type($lhs) `,` type($rhs) `into` type($result)
+  }];
+
+  let extraClassDeclaration = [{
+    VectorType getLhsType() { return llvm::cast<VectorType>(getLhs().getType()); }
+    VectorType getRhsType() { return llvm::cast<VectorType>(getRhs().getType()); }
+    VectorType getResultType() { return llvm::cast<VectorType>(getResult().getType()); }
+    std::optional<arm_sme::ArmSMETileType> getAllocatedTileType() {
+      // The outerproduct op allocates a new tile if no accumulator is passed.
+      if (!getAcc())
+        return arm_sme::getSMETileType(getResultType());
+      return std::nullopt;
+    }
+    VectorType getTileType() {
+      return getResultType();
+    }
+  }];
+}
+
+class OuterProductWide2Way<string mnemonic,
+                           list<Type> allowedInputVectorTypes,
+                           list<Type> allowedResultVectorTypes>
+  : OuterProductWideBase<mnemonic, allowedInputVectorTypes,
+                         allowedResultVectorTypes, /*numOuterProducts=*/2>;
+
+def FMopaWide2WayOp
+  : OuterProductWide2Way<"fmopa_wide_2way",
+      [ScalableVectorOfRankAndLengthAndType<[1], [8], [F16, BF16]>],
+      [nxnxv4f32]> {
+  let summary = "Floating-point sum of 2 outer products and accumulate";
+
+  let description = [{
+    This operation represents a sum of 2 widened outer products. It takes 2 1-D
+    scalable vectors as input and a 2-D scalable vector (ZA tile) as output.
+
+    For example (fp16 to fp32):
+
+    ```mlir
+    %result = arm_sme.fmopa_wide_2way %lhs, %rhs :
+      vector<[8]xf16>, vector<[8]xf16> into vector<[4]x[4]xf32>
+    ```
+
+    The `lhs` encodes a matrix of shape SVLSx2 and the `rhs` a matrix of
+    2xSVLS, where SVLS (spec [1], section B2.1) is the number of 32-bit
+    elements in a vector of SVL bits. To illustrate, below is a breakdown of
+    this operation for SVL=128 (i.e., vscale=1):
+
+    ```
+                          LHS                          RHS
+               [A0 A1 A2 A3 A4 A5 A6 A7]    [B0 B1 B2 B3 B4 B5 B6 B7]
+
+    ----------------------------------------------------------------------------
+
+                                  implicit layout
+
+                              [A0 A1]    |
+                              [A2 A3]    |    [B0 B2 B4 B6]
+                              [A4 A5]    |    [B1 B3 B5 B7]
+                              [A6 A7]    |
+
+    ----------------------------------------------------------------------------
+
+                                  2 outer products
+
+                      Acol0 ⊗ Brow0      |           Acol1 ⊗ Brow1
+                      -------------      |           -------------
+                                         |
+                  [B0 B2 B4 B6]          |       [B1 B3 B5 B7]
+                                         |
+             [A0  [A0B0 A0B2 A0B4 A0B6]  |  [A1  [A1B1 A1B3 A1B5 A1B7]
+              A2  [A2B0 A2B2 A2B4 A2B6]  |   A3  [A3B1 A3B3 A3B5 A3B7]
+              A4  [A4B0 A4B2 A4B4 A4B6]  |   A5  [A5B1 A5B3 A5B5 A5B7]
+              A6] [A6B0 A6B2 A6B4 A6B6]  |   A7] [A7B1 A7B3 A7B5 A7B7]
+                                         |
+
+    ----------------------------------------------------------------------------
+
+                              sum of 2 outer products
+
+                           Acol0 ⊗ Brow0 + Acol1 ⊗ Brow1
+
+                 [A0B0 + A1B1 A0B2 + A1B3 A0B4 + A1B5 A0B6 + A1B7]
+                 [A2B0 + A3B1 A2B2 + A3B3 A2B4 + A3B5 A2B6 + A3B7]
+                 [A4B0 + A5B1 A4B2 + A5B3 A4B4 + A5B5 A4B6 + A5B7]
+                 [A6B0 + A7B1 A6B2 + A7B3 A6B4 + A7B5 A6B6 + A7B7]
+
+    ----------------------------------------------------------------------------
+    ```
+
+    This operation enables the folding of 2 outer products chained via the
+    accumulator into a single outer product.
+
+    For example:
+
+    ```mlir
+    %a0_ext = arith.extf %a0 : vector<[4]xf16> to vector<[4]xf32>
+    %b0_ext = arith.extf %b0 : vector<[4]xf16> to vector<[4]xf32>
+    %a1_ext = arith.extf %a1 : vector<[4]xf16> to vector<[4]xf32>
+    %b1_ext = arith.extf %b1 : vector<[4]xf16> to vector<[4]xf32>
+
+    %0 = arm_sme.outerproduct %a0_ext, %b0_ext : vector<[4]xf32>, vector<[4]xf32>
+    %1 = arm_sme.outerproduct %a1_ext, %b1_ext acc(%0) : vector<[4]xf32>, vector<[4]xf32>
+    ```
+
+    The 2 outer products in the example above can be fused into a single outer
+    product as follows:
+
+	```mlir
+    %undef = llvm.mlir.undef : vector<[8]xf16>
+    %a0_ins = vector.scalable.insert %a0, %undef[0] : vector<[4]xf16> into vector<[8]xf16>
+    %a1_ins = vector.scalable.insert %a1, %undef[0] : vector<[4]xf16> into vector<[8]xf16>
+    %a_packed = "arm_sve.intr.zip1"(%a0_ins, %a1_ins) : (vector<[8]xf16>, vector<[8]xf16>) -> vector<[8]xf16>
+    %b0_ins = vector.scalable.insert %b0, %undef[0] : vector<[4]xf16> into vector<[8]xf16>
+    %b1_ins = vector.scalable.insert %b1, %undef[0] : vector<[4]xf16> into vector<[8]xf16>
+    %b_packed = "arm_sve.intr.zip1"(%b0_ins, %b1_ins) : (vector<[8]xf16>, vector<[8]xf16>) -> vector<[8]xf16>
+    %0 = arm_sme.fmopa_wide_2way %a_packed, %b_packed : vector<[8]xf16>, vector<[8]xf16> into vector<[4]x[4]xf32>
+	```
+
+    This is implemented in the `-arm-sme-outer-product-widening` pass.
+
+    Example: FP16 to FP32
+    ```mlir
+    %result = arm_sme.fmopa_wide_2way $lhs, $rhs : vector<[8]xf16>, vector<[8]xf16> into vector<[4]x[4]xf32>
+    ```
+
+    Example: BF16 to FP32
+    ```mlir
+    %result = arm_sme.fmopa_wide_2way $lhs, $rhs : vector<[8]xbf16>, vector<[8]xbf16> into vector<[4]x[4]xf32>
+    ```
+
+    | Spec | Features |
+    | ---- | -------- |
+    | [FMOPA (widening, 2-way, FP16 to FP32)](https://developer.arm.com/documentation/ddi0602/2023-09/SME-Instructions/FMOPA--widening--2-way--FP16-to-FP32---Half-precision-floating-point-sum-of-outer-products-and-accumulate-) | +sme |
+    | [BFMOPA (widening, 2-way, BF16 to FP32)](https://developer.arm.com/documentation/ddi0602/2023-09/SME-Instructions/BFMOPA--widening---BFloat16-sum-of-outer-products-and-accumulate-) | +sme |
+
+    [1] https://developer.arm.com/documentation/ddi0616
+  }];
+}
+
+// TODO: support:
+// - FMOPA 2-way FP8 to FP16
+// - FMOPA 4-way FP16 to FP32
+// once intrinsic support lands in the backend.
+
+def FMopsWide2WayOp
+  : OuterProductWide2Way<"fmops_wide_2way",
+      [ScalableVectorOfRankAndLengthAndType<[1], [8], [F16, BF16]>],
+      [nxnxv4f32]> {
+  let summary = "Floating-point sum of 2 outer products and subtract";
+  let description = [{
+    Equivalent to `fmopa_wide_2way` but outer products are subtracted from
+    destination `result`.
+
+    Example: FP16 to FP32
+    ```mlir
+    %result = arm_sme.fmops_wide_2way $lhs, $rhs : vector<[8]xf16>, vector<[8]xf16> into vector<[4]x[4]xf32>
+    ```
+
+    Example: BF16 to FP32
+    ```mlir
+    %result = arm_sme.fmops_wide_2way $lhs, $rhs : vector<[8]xbf16>, vector<[8]xbf16> into vector<[4]x[4]xf32>
+
+    Refer to
+    [fmopa_wide_2way](#arm_smefmopa_wide_2way-arm_smefmopa_wide_2wayop) for a
+    detailed description of 2-way outer products.
+
+    | Spec | Features |
+    | ---- | -------- |
+    | [FMOPS (widening, 2-way, FP16 to FP32)](https://developer.arm.com/documentation/ddi0602/2023-09/SME-Instructions/FMOPS--widening---Half-precision-floating-point-sum-of-outer-products-and-subtract-) | +sme |
+    | [BFMOPS (widening, 2-way, BF16 to FP32)](https://developer.arm.com/documentation/ddi0602/2023-09/SME-Instructions/BMOPS--Bitwise-exclusive-NOR-population-count-outer-product-and-subtract-) | +sme |
+    ```
+  }];
+}
+
+def SMopaWide2WayOp
+  : OuterProductWide2Way<"smopa_wide_2way",
+      [ScalableVectorOfRankAndLengthAndType<[1], [8], [I16]>],
+      [nxnxv4i32]> {
+  let summary = "Signed integer sum of 2 outer products and accumulate";
+  let description = [{
+    Example:
+    ```mlir
+    %result = arm_sme.smopa_wide_2way $lhs, $rhs : vector<[8]xi16>, vector<[8]xi16> into vector<[4]x[4]xi32>
+
+    Refer to
+    [fmopa_wide_2way](#arm_smefmopa_wide_2way-arm_smefmopa_wide_2wayop) for a
+    detailed description of 2-way outer products.
+
+    | Spec | Features |
+    | ---- | -------- |
+    | [SMOPA (2-way)](https://developer.arm.com/documentation/ddi0602/2023-09/SME-Instructions/SMOPA--2-way---Signed-integer-sum-of-outer-products-and-accumulate-) | +sme2 |
+    ```
+  }];
+}
+
+def SMopsWide2WayOp
+  : OuterProductWide2Way<"smops_wide_2way",
+      [ScalableVectorOfRankAndLengthAndType<[1], [8], [I16]>],
+      [nxnxv4i32]> {
+  let summary = "Signed integer sum of 2 outer products and subtract";
+  let description = [{
+    Example:
+    ```mlir
+    %result = arm_sme.smops_wide_2way $lhs, $rhs : vector<[8]xi16>, vector<[8]xi16> into vector<[4]x[4]xi32>
+
+    Refer to
+    [fmopa_wide_2way](#arm_smefmopa_wide_2way-arm_smefmopa_wide_2wayop) for a
+    detailed description of 2-way outer products.
+
+    | Spec | Features |
+    | ---- | -------- |
+    | [SMOPS (2-way)](https://developer.arm.com/documentation/ddi0602/2023-09/SME-Instructions/SMOPS--2-way---Signed-integer-sum-of-outer-products-and-subtract-) | +sme2 |
+    ```
+  }];
+}
+
+def UMopaWide2WayOp
+  : OuterProductWide2Way<"umopa_wide_2way",
+      [ScalableVectorOfRankAndLengthAndType<[1], [8], [I16]>],
+      [nxnxv4i32]> {
+  let summary = "Unsiged integer sum of 2 outer products and accumulate";
+  let description = [{
+    Example:
+    ```mlir
+    %result = arm_sme.umopa_wide_2way $lhs, $rhs : vector<[8]xi16>, vector<[8]xi16> into vector<[4]x[4]xi32>
+
+    Refer to
+    [fmopa_wide_2way](#arm_smefmopa_wide_2way-arm_smefmopa_wide_2wayop) for a
+    detailed description of 2-way outer products.
+
+    | Spec | Features |
+    | ---- | -------- |
+    | [UMOPA (2-way)](https://developer.arm.com/documentation/ddi0602/2023-09/SME-Instructions/UMOPA--2-way---Unsigned-integer-sum-of-outer-products-and-accumulate-) | +sme2 |
+    ```
+  }];
+}
+
+def UMopsWide2WayOp
+  : OuterProductWide2Way<"umops_wide_2way",
+      [ScalableVectorOfRankAndLengthAndType<[1], [8], [I16]>],
+      [nxnxv4i32]> {
+  let summary = "Unsiged integer sum of 2 outer products and subtract";
+  let description = [{
+    Example:
+    ```mlir
+    %result = arm_sme.umops_wide_2way $lhs, $rhs : vector<[8]xi16>, vector<[8]xi16> into vector<[4]x[4]xi32>
+
+    Refer to
+    [fmopa_wide_2way](#arm_smefmopa_wide_2way-arm_smefmopa_wide_2wayop) for a
+    detailed description of 2-way outer products.
+
+    | Spec | Features |
+    | ---- | -------- |
+    | [UMOPS (2-way)](https://developer.arm.com/documentation/ddi0602/2023-09/SME-Instructions/UMOPS--2-way---Unsigned-integer-sum-of-outer-products-and-subtract-) | +sme2 |
+    ```
+  }];
+}
+
 def StreamingVLOp : ArmSME_Op<"streaming_vl", [Pure]>
 {
   let summary = "Query the streaming vector length";
diff --git a/mlir/include/mlir/Dialect/ArmSME/Transforms/Passes.h b/mlir/include/mlir/Dialect/ArmSME/Transforms/Passes.h
index aef2959265a7cd..d3e4fccd628489 100644
--- a/mlir/include/mlir/Dialect/ArmSME/Transforms/Passes.h
+++ b/mlir/include/mlir/Dialect/ArmSME/Transforms/Passes.h
@@ -32,6 +32,9 @@ std::unique_ptr<Pass> createEnableArmStreamingPass(
 /// Pass that allocates tile IDs to ArmSME operations.
 std::unique_ptr<Pass> createTileAllocationPass();
 
+/// Pass that folds 'arm_sme.outerproduct' ops into widening variants.
+std::unique_ptr<Pass> createOuterProductWideningPass();
+
 //===----------------------------------------------------------------------===//
 // Registration
 //===----------------------------------------------------------------------===//
diff --git a/mlir/include/mlir/Dialect/ArmSME/Transforms/Passes.td b/mlir/include/mlir/Dialect/ArmSME/Transforms/Passes.td
index 8d1ba6ed34e805..aa9ad9b2e3340b 100644
--- a/mlir/include/mlir/Dialect/ArmSME/Transforms/Passes.td
+++ b/mlir/include/mlir/Dialect/ArmSME/Transforms/Passes.td
@@ -122,4 +122,43 @@ def TileAllocation
   let dependentDialects = ["func::FuncDialect"];
 }
 
+def OuterProductWidening
+    : Pass<"arm-sme-outer-product-widening", "mlir::func::FuncOp"> {
+  let summary = "Fold 'arm_sme.outerproduct' operations into widening variants";
+  let description = [{
+    This pass folds 'arm_sme.outerproduct' operations that are chained via the
+    accumulator into 2-way or 4-way ArmSME outer product operations.
+
+    For example:
+    ```mlir
+    %a0_ext = arith.extf %a0 : vector<[4]xf16> to vector<[4]xf32>
+    %b0_ext = arith.extf %b0 : vector<[4]xf16> to vector<[4]xf32>
+    %a1_ext = arith.extf %a1 : vector<[4]xf16> to vector<[4]xf32>
+    %b1_ext = arith.extf %b1 : vector<[4]xf16> to vector<[4]xf32>
+
+    %0 = arm_sme.outerproduct %a0_ext, %b0_ext : vector<[4]xf32>, vector<[4]xf32>
+    %1 = arm_sme.outerproduct %a1_ext, %b1_ext acc(%0) : vector<[4]xf32>, vector<[4]xf32>
+    ```
+
+    Becomes:
+
+    ```mlir
+    %undef = llvm.mlir.undef : vector<[8]xf16>
+    %a0_ins = vector.scalable.insert %a0, %undef[0] : vector<[4]xf16> into vector<[8]xf16>
+    %a1_ins = vector.scalable.insert %a1, %undef[0] : vector<[4]xf16> into vector<[8]xf16>
+    %a_packed = "arm_sve.intr.zip1"(%a0_ins, %a1_ins) : (vector<[8]xf16>, vector<[8]xf16>) -> vector<[8]xf16>
+    %b0_ins = vector.scalable.insert %b0, %undef[0] : vector<[4]xf16> into vector<[8]xf16>
+    %b1_ins = vector.scalable.insert %b1, %undef[0] : vector<[4]xf16> into vector<[8]xf16>
+    %b_packed = "arm_sve.intr.zip1"(%b0_ins, %b1_ins) : (vector<[8]xf16>, vector<[8]xf16>) -> vector<[8]xf16>
+    %0 = arm_sme.fmopa_wide_2way %a_packed, %b_packed : vector<[8]xf16>, vector<[8]xf16> into vector<[4]x[4]xf32>
+    ```
+
+    For further information on the widening ops see:
+    https://mlir.llvm.org/docs/Dialects/ArmSME/#arm_smefmopa_wide_2way-arm_smefmopa_wide_2wayop
+    https://mlir.llvm.org/docs/Dialects/ArmSME/#arm_smesmopa_wide_4way-arm_smesmopa_wide_4wayop
+  }];
+  let constructor = "mlir::arm_sme::createOuterProductWideningPass()";
+  let dependentDialects = ["func::FuncDialect", "arm_sme::ArmSMEDialect", "arm_sve::ArmSVEDialect", "LLVM::LLVMDialect"];
+}
+
 #endif // MLIR_DIALECT_ARMSME_TRANSFORMS_PASSES_TD
diff --git a/mlir/include/mlir/Dialect/ArmSME/Transforms/Transforms.h b/mlir/include/mlir/Dialect/ArmSME/Transforms/Transforms.h
index f622bc0562e9e1..09e3b4375fa5f9 100644
--- a/mlir/include/mlir/Dialect/ArmSME/Transforms/Transforms.h
+++ b/mlir/include/mlir/Dialect/ArmSME/Transforms/Transforms.h
@@ -15,6 +15,10 @@ class LLVMConversionTarget;
 class LLVMTypeConverter;
 class RewritePatternSet;
 
+namespace arm_sme {
+void populateOuterProductWideningPatterns(RewritePatternSet &patterns);
+} // namespace arm_sme
+
 } // namespace mlir
 
 #endif // MLIR_DIALECT_ARMSME_TRANSFORMS_H
diff --git a/mlir/lib/Conversion/ArmSMEToLLVM/ArmSMEToLLVM.cpp b/mlir/lib/Conversion/ArmSMEToLLVM/ArmSMEToLLVM.cpp
index bbef3b996e40b8..0871658bc3653c 100644
--- a/mlir/lib/Conversion/ArmSMEToLLVM/ArmSMEToLLVM.cpp
+++ b/mlir/lib/Conversion/ArmSMEToLLVM/ArmSMEToLLVM.cpp
@@ -776,6 +776,49 @@ struct OuterProductOpConversion
   }
 };
 
+/// Lower 2-way and 4-way outer products to intrinsics.
+template <class OuterProductWideOp, class OuterProductWideIntrOp>
+struct OuterProductWideOpConversion
+    : public ConvertArmSMEOpToLLVMPattern<OuterProductWideOp> {
+  using ConvertArmSMEOpToLLVMPattern<
+      OuterProductWideOp>::ConvertArmSMEOpToLLVMPattern;
+
+  LogicalResult
+  matchAndRewrite(OuterProductWideOp op,
+                  typename OuterProductWideOp::Adaptor adaptor,
+                  ConversionPatternRewriter &rewriter) const override {
+    auto tileId = getTileIdOrError(op);
+    if (!tileId)
+      return failure();
+
+    Value acc = op.getAcc();
+    if (!acc)
+      // Initalize accumulator with zero.
+      acc = op.template createOpAndForwardTileId<arm_sme::ZeroOp>(
+          rewriter, op.getLoc(), op.getResultType());
+
+    Value lhsMask = op.getLhsMask();
+    Value rhsMask = op.getRhsMask();
+    if (!lhsMask || !rhsMask) {
+      auto predTy = op.getLhsType().cloneWith({}, rewriter.getI1Type());
+      Value allActiveMask = rewriter.create<arith::ConstantOp>(
+          op.getLoc(), DenseElementsAttr::get(predTy, true));
+      lhsMask = allActiveMask;
+      rhsMask = allActiveMask;
+    }
+
+    rewriter.create<OuterProductWideIntrOp>(op.getLoc(), tileId, lhsMask,
+                                            rhsMask, adaptor.getLhs(),
+                                            adaptor.getRhs());
+
+    // The outerproduct intrinsics have no result, replace
+    // 'arm_sme.outerproduct' with the input tile to preserve dataflow.
+    rewriter.replaceOp(op, acc);
+
+    return success();
+  }
+};
+
 /// Lower `arm_sme.streaming_vl` to SME CNTS intrinsics.
 ///
 /// Example:
@@ -854,6 +897,13 @@ void mlir::configureArmSMEToLLVMConversionLegality(ConversionTarget &target) {
       arm_sme::aarch64_sme_st1q_vert, arm_sme::aarch64_sme_read_horiz,
       arm_sme::aarch64_sme_read_vert, arm_sme::aarch64_sme_write_horiz,
       arm_sme::aarch64_sme_write_vert, arm_sme::aarch64_sme_mopa,
+      arm_sme::aarch64_sme_mopa_wide, arm_sme::aarch64_sme_mops_wide,
+      arm_sme::aarch64_sme_smopa_wide, arm_sme::aarch64_sme_smops_wide,
+      arm_sme::aarch64_sme_umopa_wide, arm_sme::aarch64_sme_umops_wide,
+      arm_sme::aarch64_sme_smopa_za32, arm_sme::aarch64_sme_smops_za32,
+      arm_sme::aarch64_sme_umopa_za32, arm_sme::aarch64_sme_umops_za32,
+      arm_sme::aarch64_sme_sumopa_wide, arm_sme::aarch64_sme_sumops_wide,
+      arm_sme::aarch64_sme_usmopa_wide, arm_sme::aarch64_sme_usmops_wide,
       arm_sme::aarch64_sme_cntsb, arm_sme::aarch64_sme_cntsh,
       arm_sme::aarch64_sme_cntsw, arm_sme::aarch64_sme_cntsd>();
   target.addLegalDialect<arith::ArithDialect,
@@ -876,8 +926,20 @@ void mlir::populateArmSMEToLLVMConversionPatterns(LLVMTypeConverter &converter,
   addArmSMEConversionPatterns<
       LoadTileSliceConversion, MoveTileSliceToVectorConversion,
       MoveVectorToTileSliceConversion, StoreTileSliceConversion,
-      OuterProductOpConversion, ZeroOpConversion, GetTileConversion,
-      StreamingVLOpConversion>(patterns, converter);
+      StreamingVLOpConversion, OuterProductOpConversion,
+      OuterProductWideOpConversion<arm_sme::FMopaWide2WayOp,
+                                   arm_sme::aarch64_sme_mopa_wide>,
+      OuterProductWideOpConversion<arm_sme::FMopsWide2WayOp,
+                                   arm_sme::aarch64_sme_mops_wide>,
+      OuterProductWideOpConversion<arm_sme::SMopaWide2WayOp,
+                                   arm_sme::aarch64_sme_smopa_za32>,
+      OuterProductWideOpConversion<arm_sme::SMopsWide2WayOp,
+                                   arm_sme::aarch64_sme_smops_za32>,
+      OuterProductWideOpConversion<arm_sme::UMopaWide2WayOp,
+                                   arm_sme::aarch64_sme_umopa_za32>,
+      OuterProductWideOpConversion<arm_sme::UMopsWide2WayOp,
+                                   arm_sme::aarch64_sme_umops_za32>,
+      ZeroOpConversion, GetTileConversion>(patterns, converter);
 }
 
 std::unique_ptr<Pass> mlir::createConvertArmSMEToLLVMPass() {
diff --git a/mlir/lib/Dialect/ArmSME/Transforms/CMakeLists.txt b/mlir/lib/Dialect/ArmSME/Transforms/CMakeLists.txt
index 96eb5844204384..24942b6f28d2cd 100644
--- a/mlir/lib/Dialect/ArmSME/Transforms/CMakeLists.txt
+++ b/mlir/lib/Dialect/ArmSME/Transforms/CMakeLists.txt
@@ -1,5 +1,6 @@
 add_mlir_dialect_library(MLIRArmSMETransforms
   EnableArmStreaming.cpp
+  OuterProductWidening.cpp
   TileAllocation.cpp
 
   ADDITIONAL_HEADER_DIRS
@@ -10,6 +11,7 @@ add_mlir_dialect_library(MLIRArmSMETransforms
 
   LINK_LIBS PUBLIC
   MLIRArmSMEDialect
+  MLIRArmSVEDialect
   MLIRFuncDialect
   MLIRLLVMCommonConversion
   MLIRVectorDialect
diff --git a/mlir/lib/Dialect/ArmSME/Transforms/OuterProductWidening.cpp b/mlir/lib/Dialect/ArmSME/Transforms/OuterProductWidening.cpp
new file mode 100644
index 00000000000000..b0a52b31121de6
--- /dev/null
+++ b/mlir/lib/Dialect/ArmSME/Transforms/OuterProductWidening.cpp
@@ -0,0 +1,239 @@
+//===- OuterProductWidening.cpp - Widen 'arm_sme.outerproduct' ops --------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+// This file implements rewrites that fold 'arm_sme.outerproduct' operations
+// into the 2-way or 4-way widening outerproduct operations.
+//
+//===----------------------------------------------------------------------===//
+
+#include "mlir/Dialect/ArmSME/IR/ArmSME.h"
+#include "mlir/Dialect/ArmSME/Transforms/Passes.h"
+#include "mlir/Dialect/ArmSME/Transforms/Transforms.h"
+#include "mlir/Dialect/ArmSVE/IR/ArmSVEDialect.h"
+#include "mlir/Dialect/Func/IR/FuncOps.h"
+#include "mlir/Dialect/LLVMIR/LLVMDialect.h"
+#include "mlir/IR/PatternMatch.h"
+#include "mlir/Transforms/GreedyPatternRewriteDriver.h"
+
+#define DEBUG_TYPE "arm-sme-outerproduct-widening"
+
+namespace mlir::arm_sme {
+#define GEN_PASS_DEF_OUTERPRODUCTWIDENING
+#include "mlir/Dialect/ArmSME/Transforms/Passes.h.inc"
+} // namespace mlir::arm_sme
+
+using namespace mlir;
+using namespace mlir::arm_sme;
+
+namespace {
+// Fold two 'arm_sme.outerproduct' operations that are chained via the
+// accumulator into 2-way outer product operation.
+//
+// For example:
+//
+//  %a0_ext = arith.extf %a0 : vector<[4]xf16> to vector<[4]xf32>
+//  %b0_ext = arith.extf %b0 : vector<[4]xf16> to vector<[4]xf32>
+//  %0 = arm_sme.outerproduct %a0_ext, %b0_ext : vector<[4]xf32>,
+//                                               vector<[4]xf32>
+//
+//  %a1_ext = arith.extf %a1 : vector<[4]xf16> to vector<[4]xf32>
+//  %b1_ext = arith.extf %b1 : vector<[4]xf16> to vector<[4]xf32>
+//  %1 = arm_sme.outerproduct %a1_ext, %b1_ext, %0 : vector<[4]xf32>,
+//                                                   vector<[4]xf32>
+//
+// Becomes:
+//
+//  %a_packed = arm_sve.zip %a0, %a1 : vector<[8]xf16> to vector<[8]xf16>
+//  %b_packed = arm_sve.zip %b0, %b1 : vector<[8]xf16> to vector<[8]xf16>
+//  %0 = arm_sme.fmopa_wide_2way %a_packed, %b_packed : vector<[8]xf16>,
+//                                                      vector<[4]xf32>
+class OuterProduct2WayWidening
+    : public OpRewritePattern<arm_sme::OuterProductOp> {
+public:
+  using OpRewritePattern::OpRewritePattern;
+
+  LogicalResult matchAndRewrite(arm_sme::OuterProductOp op,
+                                PatternRewriter &rewriter) const override {
+    Value acc = op.getAcc();
+    if (!acc)
+      return rewriter.notifyMatchFailure(op, "no accumulator operand");
+
+    arm_sme::OuterProductOp op1 = acc.getDefiningOp<arm_sme::OuterProductOp>();
+    arm_sme::OuterProductOp op2 = op;
+    if (!op1)
+      return rewriter.notifyMatchFailure(op,
+                                         "defining op of accumulator operand "
+                                         "must be an 'arm_sme.outerproduct'");
+
+    if (op1.getKind() != op2.getKind())
+      return rewriter.notifyMatchFailure(
+          op, "combining kind (add or sub) of outer products must match");
+
+    if (!llvm::hasSingleElement(op1->getUses())) {
+      // We could still widen, but if the first outer product has an
+      // accumulator it will be used as the root for tile allocation and since
+      // the widening outer product uses the same accumulator it will get
+      // assigned the same tile ID, resulting in 3 outer products and incorrect
+      // results. No accumulator would be ok, but it's simpler to prevent this
+      // altogether, since it has no benefit.
+      return rewriter.notifyMatchFailure(
+          op, "first outer product is not single use and cannot be removed, "
+              "no benefit to widening");
+    }
+
+    auto nxnxv4i32 =
+        VectorType::get({4, 4}, rewriter.getI32Type(), {true, true});
+    auto nxnxv4f32 =
+        VectorType::get({4, 4}, rewriter.getF32Type(), {true, true});
+    auto nxv4i16 = VectorType::get({4}, rewriter.getI16Type(), true);
+    auto nxv4f16 = VectorType::get({4}, rewriter.getF16Type(), true);
+    auto nxv4bf16 = VectorType::get({4}, rewriter.getBF16Type(), true);
+    if ((failed(
+             isWidenable<arith::ExtFOp>(rewriter, op1, nxnxv4f32, nxv4f16)) ||
+         failed(
+             isWidenable<arith::ExtFOp>(rewriter, op2, nxnxv4f32, nxv4f16))) &&
+        (failed(
+             isWidenable<arith::ExtFOp>(rewriter, op1, nxnxv4f32, nxv4bf16)) ||
+         failed(
+             isWidenable<arith::ExtFOp>(rewriter, op2, nxnxv4f32, nxv4bf16))) &&
+        (failed(
+             isWidenable<arith::ExtSIOp>(rewriter, op1, nxnxv4i32, nxv4i16)) ||
+         failed(
+             isWidenable<arith::ExtSIOp>(rewriter, op2, nxnxv4i32, nxv4i16))) &&
+        (failed(
+             isWidenable<arith::ExtUIOp>(rewriter, op1, nxnxv4i32, nxv4i16)) ||
+         failed(
+             isWidenable<arith::ExtUIOp>(rewriter, op2, nxnxv4i32, nxv4i16))))
+      return failure();
+
+    auto loc = op.getLoc();
+
+    // zip(lhs, rhs)
+    auto packInputs = [&](VectorType type, Value lhs, Value rhs) {
+      auto undef = rewriter.create<LLVM::UndefOp>(loc, type);
+      auto insertLHS =
+          rewriter.create<vector::ScalableInsertOp>(loc, lhs, undef, 0);
+      auto insertRHS =
+          rewriter.create<vector::ScalableInsertOp>(loc, rhs, undef, 0);
+      return rewriter.create<arm_sve::Zip1IntrOp>(loc, type, insertLHS,
+                                                  insertRHS);
+    };
+
+    auto extOp = op.getLhs().getDefiningOp();
+    VectorType extSourceVectorType =
+        cast<VectorType>(extOp->getOperand(0).getType());
+    VectorType widenedVectorType =
+        VectorType::Builder(extSourceVectorType)
+            .setDim(0, extSourceVectorType.getShape()[0] * 2);
+    auto lhs = packInputs(widenedVectorType,
+                          op1.getLhs().getDefiningOp()->getOperand(0),
+                          op2.getLhs().getDefiningOp()->getOperand(0));
+    auto rhs = packInputs(widenedVectorType,
+                          op1.getRhs().getDefiningOp()->getOperand(0),
+                          op2.getRhs().getDefiningOp()->getOperand(0));
+
+    Value lhsMask, rhsMask;
+    if (op1.getLhsMask() || op2.getLhsMask()) {
+      if (!(op1.getLhsMask() && op2.getLhsMask()))
+        return rewriter.notifyMatchFailure(
+            op, "unsupported masking, either both outerproducts are masked "
+                "or neither");
+
+      VectorType maskType = VectorType::Builder(widenedVectorType)
+                                .setElementType(rewriter.getI1Type());
+      lhsMask = packInputs(maskType, op1.getLhsMask(), op2.getLhsMask());
+      rhsMask = packInputs(maskType, op1.getRhsMask(), op2.getRhsMask());
+    }
+
+    arm_sme::CombiningKind kind = op.getKind();
+    assert((kind == arm_sme::CombiningKind::Add ||
+            kind == arm_sme::CombiningKind::Sub) &&
+           "unhandled arm_sme::CombiningKind!");
+
+    if (isa<arith::ExtFOp>(extOp)) {
+      if (kind == arm_sme::CombiningKind::Add)
+        rewriter.replaceOpWithNewOp<arm_sme::FMopaWide2WayOp>(
+            op2, op.getResultType(), lhs, rhs, lhsMask, rhsMask, op1.getAcc());
+      else
+        rewriter.replaceOpWithNewOp<arm_sme::FMopsWide2WayOp>(
+            op2, op.getResultType(), lhs, rhs, lhsMask, rhsMask, op1.getAcc());
+    } else if (isa<arith::ExtSIOp>(extOp)) {
+      if (kind == arm_sme::CombiningKind::Add)
+        rewriter.replaceOpWithNewOp<arm_sme::SMopaWide2WayOp>(
+            op2, op.getResultType(), lhs, rhs, lhsMask, rhsMask, op1.getAcc());
+      else
+        rewriter.replaceOpWithNewOp<arm_sme::SMopsWide2WayOp>(
+            op2, op.getResultType(), lhs, rhs, lhsMask, rhsMask, op1.getAcc());
+    } else if (isa<arith::ExtUIOp>(extOp)) {
+      if (kind == arm_sme::CombiningKind::Add)
+        rewriter.replaceOpWithNewOp<arm_sme::UMopaWide2WayOp>(
+            op2, op.getResultType(), lhs, rhs, lhsMask, rhsMask, op1.getAcc());
+      else
+        rewriter.replaceOpWithNewOp<arm_sme::UMopsWide2WayOp>(
+            op2, op.getResultType(), lhs, rhs, lhsMask, rhsMask, op1.getAcc());
+    } else
+      llvm_unreachable("unexpected extend op!");
+
+    op1.erase();
+
+    return success();
+  }
+
+private:
+  template <typename ExtOp>
+  LogicalResult isWidenable(PatternRewriter &rewriter,
+                            arm_sme::OuterProductOp op, VectorType resultType,
+                            VectorType inputType) const {
+    if (op.getResultType() != resultType)
+      return rewriter.notifyMatchFailure(
+          op, "unsupported result type, expected 'vector<[4]x[4]xi32>' or "
+              "'vector<[4]x[4]xf32>'");
+
+    auto lhsDefOp = op.getLhs().getDefiningOp<ExtOp>();
+    auto rhsDefOp = op.getRhs().getDefiningOp<ExtOp>();
+
+    if (!lhsDefOp || !rhsDefOp)
+      return rewriter.notifyMatchFailure(
+          op, "defining op of outerproduct operands must be 'arith.extf' or "
+              "'arith.extsi' or 'arith.extui'");
+
+    auto lhsInType = cast<VectorType>(lhsDefOp->getOperand(0).getType());
+    auto rhsInType = cast<VectorType>(rhsDefOp->getOperand(0).getType());
+
+    if (lhsInType != inputType || rhsInType != inputType)
+      return rewriter.notifyMatchFailure(
+          op, "unsupported input types, expected 'vector<[4]xi16>' or "
+              "'vector<[4]xf16>' or 'vector<[4]xbf16>'");
+    return success();
+  }
+};
+
+struct OuterProductWideningPass
+    : public arm_sme::impl::OuterProductWideningBase<OuterProductWideningPass> {
+
+  void runOnOperation() override {
+    RewritePatternSet patterns(&getContext());
+    populateOuterProductWideningPatterns(patterns);
+
+    if (failed(
+            applyPatternsAndFoldGreedily(getOperation(), std::move(patterns))))
+      signalPassFailure();
+  }
+};
+
+} // namespace
+
+void mlir::arm_sme::populateOuterProductWideningPatterns(
+    RewritePatternSet &patterns) {
+  patterns.add<OuterProduct2WayWidening>(
+      patterns.getContext());
+}
+
+std::unique_ptr<Pass> mlir::arm_sme::createOuterProductWideningPass() {
+  return std::make_unique<OuterProductWideningPass>();
+}
diff --git a/mlir/test/Conversion/ArmSMEToLLVM/arm-sme-to-llvm.mlir b/mlir/test/Conversion/ArmSMEToLLVM/arm-sme-to-llvm.mlir
index f9cf77ca15ffb9..e0e0d90d8e85f6 100644
--- a/mlir/test/Conversion/ArmSMEToLLVM/arm-sme-to-llvm.mlir
+++ b/mlir/test/Conversion/ArmSMEToLLVM/arm-sme-to-llvm.mlir
@@ -601,3 +601,99 @@ func.func @arm_sme_streaming_vl_double_words() -> index {
   %svl_d = arm_sme.streaming_vl <double>
   return %svl_d : index
 }
+
+//===----------------------------------------------------------------------===//
+// arm_sme.fmopa_wide_2way
+//===----------------------------------------------------------------------===//
+
+// -----
+
+// CHECK-LABEL: arm_sme_fmopa_wide_2way_f16f16_to_f32
+// CHECK: "arm_sme.intr.mopa.wide"({{.*}}) <{tile_id = 0 : i32}> : (vector<[8]xi1>, vector<[8]xi1>, vector<[8]xf16>, vector<[8]xf16>) -> ()
+func.func @arm_sme_fmopa_wide_2way_f16f16_to_f32(%vecA: vector<[8]xf16>, %vecB: vector<[8]xf16>) -> vector<[4]x[4]xf32> {
+  %result = arm_sme.fmopa_wide_2way %vecA, %vecB : vector<[8]xf16>, vector<[8]xf16> into vector<[4]x[4]xf32>
+  return %result : vector<[4]x[4]xf32>
+}
+
+// -----
+
+// CHECK-LABEL: arm_sme_fmopa_wide_2way_bf16bf16_to_f32
+// CHECK: "arm_sme.intr.mopa.wide"({{.*}}) <{tile_id = 0 : i32}> : (vector<[8]xi1>, vector<[8]xi1>, vector<[8]xbf16>, vector<[8]xbf16>) -> ()
+func.func @arm_sme_fmopa_wide_2way_bf16bf16_to_f32(%vecA: vector<[8]xbf16>, %vecB: vector<[8]xbf16>) -> vector<[4]x[4]xf32> {
+  %result = arm_sme.fmopa_wide_2way %vecA, %vecB : vector<[8]xbf16>, vector<[8]xbf16> into vector<[4]x[4]xf32>
+  return %result : vector<[4]x[4]xf32>
+}
+
+//===----------------------------------------------------------------------===//
+// arm_sme.fmops_wide_2way
+//===----------------------------------------------------------------------===//
+
+// -----
+
+// CHECK-LABEL: arm_sme_fmops_wide_2way_f16f16_to_f32
+// CHECK: "arm_sme.intr.mops.wide"({{.*}}) <{tile_id = 0 : i32}> : (vector<[8]xi1>, vector<[8]xi1>, vector<[8]xf16>, vector<[8]xf16>) -> ()
+func.func @arm_sme_fmops_wide_2way_f16f16_to_f32(%vecA: vector<[8]xf16>, %vecB: vector<[8]xf16>) -> vector<[4]x[4]xf32> {
+  %result = arm_sme.fmops_wide_2way %vecA, %vecB : vector<[8]xf16>, vector<[8]xf16> into vector<[4]x[4]xf32>
+  return %result : vector<[4]x[4]xf32>
+}
+
+// -----
+
+// CHECK-LABEL: arm_sme_fmops_wide_2way_bf16bf16_to_f32
+// CHECK: "arm_sme.intr.mops.wide"({{.*}}) <{tile_id = 0 : i32}> : (vector<[8]xi1>, vector<[8]xi1>, vector<[8]xbf16>, vector<[8]xbf16>) -> ()
+func.func @arm_sme_fmops_wide_2way_bf16bf16_to_f32(%vecA: vector<[8]xbf16>, %vecB: vector<[8]xbf16>) -> vector<[4]x[4]xf32> {
+  %result = arm_sme.fmops_wide_2way %vecA, %vecB : vector<[8]xbf16>, vector<[8]xbf16> into vector<[4]x[4]xf32>
+  return %result : vector<[4]x[4]xf32>
+}
+
+//===----------------------------------------------------------------------===//
+// arm_sme.smopa_wide_2way
+//===----------------------------------------------------------------------===//
+
+// -----
+
+// CHECK-LABEL: arm_sme_smopa_wide_2way_i16i16_to_i32
+// CHECK: "arm_sme.intr.smopa.za32"({{.*}}) <{tile_id = 0 : i32}> : (vector<[8]xi1>, vector<[8]xi1>, vector<[8]xi16>, vector<[8]xi16>) -> ()
+func.func @arm_sme_smopa_wide_2way_i16i16_to_i32(%vecA: vector<[8]xi16>, %vecB: vector<[8]xi16>) -> vector<[4]x[4]xi32> {
+  %result = arm_sme.smopa_wide_2way %vecA, %vecB : vector<[8]xi16>, vector<[8]xi16> into vector<[4]x[4]xi32>
+  return %result : vector<[4]x[4]xi32>
+}
+
+//===----------------------------------------------------------------------===//
+// arm_sme.smops_wide_2way
+//===----------------------------------------------------------------------===//
+
+// -----
+
+// CHECK-LABEL: arm_sme_smops_wide_2way_i16i16_to_i32
+// CHECK: "arm_sme.intr.smops.za32"({{.*}}) <{tile_id = 0 : i32}> : (vector<[8]xi1>, vector<[8]xi1>, vector<[8]xi16>, vector<[8]xi16>) -> ()
+func.func @arm_sme_smops_wide_2way_i16i16_to_i32(%vecA: vector<[8]xi16>, %vecB: vector<[8]xi16>) -> vector<[4]x[4]xi32> {
+  %result = arm_sme.smops_wide_2way %vecA, %vecB : vector<[8]xi16>, vector<[8]xi16> into vector<[4]x[4]xi32>
+  return %result : vector<[4]x[4]xi32>
+}
+
+//===----------------------------------------------------------------------===//
+// arm_sme.umopa_wide_2way
+//===----------------------------------------------------------------------===//
+
+// -----
+
+// CHECK-LABEL: arm_sme_umopa_wide_2way_i16i16_to_i32
+// CHECK: "arm_sme.intr.umopa.za32"({{.*}}) <{tile_id = 0 : i32}> : (vector<[8]xi1>, vector<[8]xi1>, vector<[8]xi16>, vector<[8]xi16>) -> ()
+func.func @arm_sme_umopa_wide_2way_i16i16_to_i32(%vecA: vector<[8]xi16>, %vecB: vector<[8]xi16>) -> vector<[4]x[4]xi32> {
+  %result = arm_sme.umopa_wide_2way %vecA, %vecB : vector<[8]xi16>, vector<[8]xi16> into vector<[4]x[4]xi32>
+  return %result : vector<[4]x[4]xi32>
+}
+
+//===----------------------------------------------------------------------===//
+// arm_sme.umops_wide_2way
+//===----------------------------------------------------------------------===//
+
+// -----
+
+// CHECK-LABEL: arm_sme_umops_wide_2way_i16i16_to_i32
+// CHECK: "arm_sme.intr.umops.za32"({{.*}}) <{tile_id = 0 : i32}> : (vector<[8]xi1>, vector<[8]xi1>, vector<[8]xi16>, vector<[8]xi16>) -> ()
+func.func @arm_sme_umops_wide_2way_i16i16_to_i32(%vecA: vector<[8]xi16>, %vecB: vector<[8]xi16>) -> vector<[4]x[4]xi32> {
+  %result = arm_sme.umops_wide_2way %vecA, %vecB : vector<[8]xi16>, vector<[8]xi16> into vector<[4]x[4]xi32>
+  return %result : vector<[4]x[4]xi32>
+}
diff --git a/mlir/test/Dialect/ArmSME/invalid.mlir b/mlir/test/Dialect/ArmSME/invalid.mlir
index 85b95a8b6cf12b..1f63de927ea006 100644
--- a/mlir/test/Dialect/ArmSME/invalid.mlir
+++ b/mlir/test/Dialect/ArmSME/invalid.mlir
@@ -173,3 +173,56 @@ func.func @arm_sme_outerproduct__bad_vector_type(%vecA: vector<[4]xf32>, %vecB:
   %0 = arm_sme.outerproduct %vecA, %vecB : vector<[4]xf32>, vector<[8]xf32>
   return %0 : vector<[4]x[4]xf32>
 }
+
+//===----------------------------------------------------------------------===//
+// arm_sme.fmopa_wide_2way
+//===----------------------------------------------------------------------===//
+
+// -----
+
+func.func @arm_sme_fmopa_wide_2way__bad_rhs_vector_type(%vecA: vector<[8]xf16>, %vecB: vector<[4]xf32>) -> vector<[4]x[4]xf32>
+{
+  // expected-error at +1 {{op failed to verify that all of {lhs, rhs} have same type}}
+  %0 = arm_sme.fmopa_wide_2way %vecA, %vecB : vector<[8]xf16>, vector<[4]xf32> into vector<[4]x[4]xf32>
+  return %0 : vector<[4]x[4]xf32>
+}
+
+// -----
+
+func.func @arm_sme_fmopa_wide_2way__bad_lhs_mask_type(%vecA: vector<[8]xf16>, %vecB: vector<[8]xf16>, %maskA : vector<[4]xi1>, %maskB : vector<[8]xi1>) -> vector<[4]x[4]xf32>
+{
+  // expected-note at -2 {{prior use here}}
+  // expected-error at +1 {{use of value '%maskA' expects different type than prior uses: 'vector<[8]xi1>' vs 'vector<[4]xi1>}}
+  %0 = arm_sme.fmopa_wide_2way %vecA, %vecB masks(%maskA, %maskB) : vector<[8]xf16>, vector<[8]xf16> into vector<[4]x[4]xf32>
+  return %0 : vector<[4]x[4]xf32>
+}
+
+// -----
+
+func.func @arm_sme_fmopa_wide_2way__bad_rhs_mask_type(%vecA: vector<[8]xf16>, %vecB: vector<[8]xf16>, %maskA : vector<[8]xi1>, %maskB : vector<[4]xi1>) -> vector<[4]x[4]xf32>
+{
+  // expected-note at -2 {{prior use here}}
+  // expected-error at +1 {{use of value '%maskB' expects different type than prior uses: 'vector<[8]xi1>' vs 'vector<[4]xi1>}}
+  %0 = arm_sme.fmopa_wide_2way %vecA, %vecB masks(%maskA, %maskB) : vector<[8]xf16>, vector<[8]xf16> into vector<[4]x[4]xf32>
+  return %0 : vector<[4]x[4]xf32>
+}
+
+// -----
+
+func.func @arm_sme_fmopa_wide_2way__no_rhs_mask(%vecA: vector<[8]xf16>, %vecB: vector<[8]xf16>, %maskA : vector<[8]xi1>) -> vector<[4]x[4]xf32>
+{
+  // expected-error at +1 {{op failed to verify that both `lhsMask` and `rhsMask` should be provided or neither}}
+  %0 = arm_sme.fmopa_wide_2way %vecA, %vecB masks(%maskA,) : vector<[8]xf16>, vector<[8]xf16> into vector<[4]x[4]xf32>
+  return %0 : vector<[4]x[4]xf32>
+}
+
+// -----
+
+func.func @arm_sme_fmopa_wide_2way__bad_acc_type(%vecA: vector<[8]xf16>, %vecB: vector<[8]xf16>) -> vector<[4]x[4]xf32>
+{
+  %acc = arm_sme.zero : vector<[2]x[2]xi64>
+  // expected-note at -1 {{prior use here}}
+  // expected-error at +1 {{use of value '%acc' expects different type than prior uses: 'vector<[4]x[4]xf32>' vs 'vector<[2]x[2]xi64>'}}
+  %0 = arm_sme.fmopa_wide_2way %vecA, %vecB masks(%maskA, %maskB) acc(%acc) : vector<[8]xf16>, vector<[8]xf16> into vector<[4]x[4]xf32>
+  return %0 : vector<[4]x[4]xf32>
+}
diff --git a/mlir/test/Dialect/ArmSME/outer-product-widening.mlir b/mlir/test/Dialect/ArmSME/outer-product-widening.mlir
new file mode 100644
index 00000000000000..0f3bba67146672
--- /dev/null
+++ b/mlir/test/Dialect/ArmSME/outer-product-widening.mlir
@@ -0,0 +1,192 @@
+// RUN: mlir-opt %s -arm-sme-outer-product-widening -cse -split-input-file | FileCheck %s
+
+// CHECK-LABEL: @outerproduct_add_widening_2way_f16f16f32
+// CHECK-SAME:    %[[A0:.*]]: vector<[4]xf16>, %[[B0:.*]]: vector<[4]xf16>, %[[A1:.*]]: vector<[4]xf16>, %[[B1:.*]]: vector<[4]xf16>,
+// CHECK-SAME:    %[[A0_MASK:.*]]: vector<[4]xi1>, %[[B0_MASK:.*]]: vector<[4]xi1>, %[[A1_MASK:.*]]: vector<[4]xi1>, %[[B1_MASK:.*]]: vector<[4]xi1>
+// CHECK-DAG: %[[ACC:.*]] = arith.constant dense<0.000000e+00> : vector<[4]x[4]xf32>
+// CHECK-DAG: %[[VEC_UNDEF:.*]] = llvm.mlir.undef : vector<[8]xf16>
+// CHECK-DAG: %[[A0_INSERT:.*]] = vector.scalable.insert %[[A0]], %[[VEC_UNDEF]][0] : vector<[4]xf16> into vector<[8]xf16>
+// CHECK-DAG: %[[B0_INSERT:.*]] = vector.scalable.insert %[[B0]], %[[VEC_UNDEF]][0] : vector<[4]xf16> into vector<[8]xf16>
+// CHECK-DAG: %[[A1_INSERT:.*]] = vector.scalable.insert %[[A1]], %[[VEC_UNDEF]][0] : vector<[4]xf16> into vector<[8]xf16>
+// CHECK-DAG: %[[B1_INSERT:.*]] = vector.scalable.insert %[[B1]], %[[VEC_UNDEF]][0] : vector<[4]xf16> into vector<[8]xf16>
+// CHECK-DAG: %[[LHS:.*]] = "arm_sve.intr.zip1"(%[[A0_INSERT]], %[[A1_INSERT]]) : (vector<[8]xf16>, vector<[8]xf16>) -> vector<[8]xf16>
+// CHECK-DAG: %[[RHS:.*]] = "arm_sve.intr.zip1"(%[[B0_INSERT]], %[[B1_INSERT]]) : (vector<[8]xf16>, vector<[8]xf16>) -> vector<[8]xf16>
+// CHECK-DAG: %[[MASK_UNDEF:.*]] = llvm.mlir.undef : vector<[8]xi1>
+// CHECK-DAG: %[[A0_MASK_INSERT:.*]] = vector.scalable.insert %[[A0_MASK]], %[[MASK_UNDEF]][0] : vector<[4]xi1> into vector<[8]xi1>
+// CHECK-DAG: %[[B0_MASK_INSERT:.*]] = vector.scalable.insert %[[B0_MASK]], %[[MASK_UNDEF]][0] : vector<[4]xi1> into vector<[8]xi1>
+// CHECK-DAG: %[[A1_MASK_INSERT:.*]] = vector.scalable.insert %[[A1_MASK]], %[[MASK_UNDEF]][0] : vector<[4]xi1> into vector<[8]xi1>
+// CHECK-DAG: %[[B1_MASK_INSERT:.*]] = vector.scalable.insert %[[B1_MASK]], %[[MASK_UNDEF]][0] : vector<[4]xi1> into vector<[8]xi1>
+// CHECK-DAG: %[[LHS_MASK:.*]] = "arm_sve.intr.zip1"(%[[A0_MASK_INSERT]], %[[A1_MASK_INSERT]]) : (vector<[8]xi1>, vector<[8]xi1>) -> vector<[8]xi1>
+// CHECK-DAG: %[[RHS_MASK:.*]] = "arm_sve.intr.zip1"(%[[B0_MASK_INSERT]], %[[B1_MASK_INSERT]]) : (vector<[8]xi1>, vector<[8]xi1>) -> vector<[8]xi1>
+// CHECK-DAG: arm_sme.fmopa_wide_2way %[[LHS]], %[[RHS]] acc(%[[ACC]]) masks(%[[LHS_MASK]], %[[RHS_MASK]]) : vector<[8]xf16>, vector<[8]xf16> into vector<[4]x[4]xf32>
+func.func @outerproduct_add_widening_2way_f16f16f32(
+    %a0 : vector<[4]xf16>, %b0 : vector<[4]xf16>,
+    %a1 : vector<[4]xf16>, %b1 : vector<[4]xf16>,
+    %a0_mask : vector<[4]xi1>, %b0_mask : vector<[4]xi1>,
+    %a1_mask : vector<[4]xi1>, %b1_mask : vector<[4]xi1>) -> vector<[4]x[4]xf32> {
+  %a0_ext = arith.extf %a0 : vector<[4]xf16> to vector<[4]xf32>
+  %b0_ext = arith.extf %b0 : vector<[4]xf16> to vector<[4]xf32>
+  %a1_ext = arith.extf %a1 : vector<[4]xf16> to vector<[4]xf32>
+  %b1_ext = arith.extf %b1 : vector<[4]xf16> to vector<[4]xf32>
+
+  %acc = arith.constant dense<0.0> : vector<[4]x[4]xf32>
+
+  %0 = arm_sme.outerproduct %a0_ext, %b0_ext acc(%acc) masks(%a0_mask, %b0_mask) : vector<[4]xf32>, vector<[4]xf32>
+  %1 = arm_sme.outerproduct %a1_ext, %b1_ext acc(%0) masks(%a1_mask, %b1_mask) : vector<[4]xf32>, vector<[4]xf32>
+
+  return %1 : vector<[4]x[4]xf32>
+}
+
+// -----
+
+// CHECK-LABEL: @outerproduct_sub_widening_2way_f16f16f32
+// CHECK: arm_sme.fmops_wide_2way %{{.*}}, %{{.*}} acc(%{{.*}}) masks(%{{.*}}, %{{.*}}) : vector<[8]xf16>, vector<[8]xf16> into vector<[4]x[4]xf32>
+func.func @outerproduct_sub_widening_2way_f16f16f32(
+    %a0 : vector<[4]xf16>, %b0 : vector<[4]xf16>,
+    %a1 : vector<[4]xf16>, %b1 : vector<[4]xf16>,
+    %a0_mask : vector<[4]xi1>, %b0_mask : vector<[4]xi1>,
+    %a1_mask : vector<[4]xi1>, %b1_mask : vector<[4]xi1>) -> vector<[4]x[4]xf32> {
+  %a0_ext = arith.extf %a0 : vector<[4]xf16> to vector<[4]xf32>
+  %b0_ext = arith.extf %b0 : vector<[4]xf16> to vector<[4]xf32>
+  %a1_ext = arith.extf %a1 : vector<[4]xf16> to vector<[4]xf32>
+  %b1_ext = arith.extf %b1 : vector<[4]xf16> to vector<[4]xf32>
+
+  %acc = arith.constant dense<0.0> : vector<[4]x[4]xf32>
+
+  %0 = arm_sme.outerproduct %a0_ext, %b0_ext kind<sub> acc(%acc) masks(%a0_mask, %b0_mask) : vector<[4]xf32>, vector<[4]xf32>
+  %1 = arm_sme.outerproduct %a1_ext, %b1_ext kind<sub> acc(%0) masks(%a1_mask, %b1_mask) : vector<[4]xf32>, vector<[4]xf32>
+
+  return %1 : vector<[4]x[4]xf32>
+}
+
+// -----
+
+// CHECK-LABEL: @outerproduct_add_widening_2way_bf16bf16f32
+// CHECK: arm_sme.fmopa_wide_2way %{{.*}}, %{{.*}} acc(%{{.*}}) masks(%{{.*}}, %{{.*}}) : vector<[8]xbf16>, vector<[8]xbf16> into vector<[4]x[4]xf32>
+func.func @outerproduct_add_widening_2way_bf16bf16f32(
+    %a0 : vector<[4]xbf16>, %b0 : vector<[4]xbf16>,
+    %a1 : vector<[4]xbf16>, %b1 : vector<[4]xbf16>,
+    %a0_mask : vector<[4]xi1>, %b0_mask : vector<[4]xi1>,
+    %a1_mask : vector<[4]xi1>, %b1_mask : vector<[4]xi1>) -> vector<[4]x[4]xf32> {
+  %a0_ext = arith.extf %a0 : vector<[4]xbf16> to vector<[4]xf32>
+  %b0_ext = arith.extf %b0 : vector<[4]xbf16> to vector<[4]xf32>
+  %a1_ext = arith.extf %a1 : vector<[4]xbf16> to vector<[4]xf32>
+  %b1_ext = arith.extf %b1 : vector<[4]xbf16> to vector<[4]xf32>
+
+  %acc = arith.constant dense<0.0> : vector<[4]x[4]xf32>
+
+  %0 = arm_sme.outerproduct %a0_ext, %b0_ext acc(%acc) masks(%a0_mask, %b0_mask) : vector<[4]xf32>, vector<[4]xf32>
+  %1 = arm_sme.outerproduct %a1_ext, %b1_ext acc(%0) masks(%a1_mask, %b1_mask) : vector<[4]xf32>, vector<[4]xf32>
+
+  return %1 : vector<[4]x[4]xf32>
+}
+
+// -----
+
+// CHECK-LABEL: @outerproduct_sub_widening_2way_bf16bf16f32
+// CHECK: arm_sme.fmops_wide_2way %{{.*}}, %{{.*}} acc(%{{.*}}) masks(%{{.*}}, %{{.*}}) : vector<[8]xbf16>, vector<[8]xbf16> into vector<[4]x[4]xf32>
+func.func @outerproduct_sub_widening_2way_bf16bf16f32(
+    %a0 : vector<[4]xbf16>, %b0 : vector<[4]xbf16>,
+    %a1 : vector<[4]xbf16>, %b1 : vector<[4]xbf16>,
+    %a0_mask : vector<[4]xi1>, %b0_mask : vector<[4]xi1>,
+    %a1_mask : vector<[4]xi1>, %b1_mask : vector<[4]xi1>) -> vector<[4]x[4]xf32> {
+  %a0_ext = arith.extf %a0 : vector<[4]xbf16> to vector<[4]xf32>
+  %b0_ext = arith.extf %b0 : vector<[4]xbf16> to vector<[4]xf32>
+  %a1_ext = arith.extf %a1 : vector<[4]xbf16> to vector<[4]xf32>
+  %b1_ext = arith.extf %b1 : vector<[4]xbf16> to vector<[4]xf32>
+
+  %acc = arith.constant dense<0.0> : vector<[4]x[4]xf32>
+
+  %0 = arm_sme.outerproduct %a0_ext, %b0_ext kind<sub> acc(%acc) masks(%a0_mask, %b0_mask) : vector<[4]xf32>, vector<[4]xf32>
+  %1 = arm_sme.outerproduct %a1_ext, %b1_ext kind<sub> acc(%0) masks(%a1_mask, %b1_mask) : vector<[4]xf32>, vector<[4]xf32>
+
+  return %1 : vector<[4]x[4]xf32>
+}
+
+// -----
+
+// CHECK-LABEL: @outerproduct_add_widening_2way_signed_i16i16i32
+// CHECK: arm_sme.smopa_wide_2way %{{.*}}, %{{.*}} acc(%{{.*}}) masks(%{{.*}}, %{{.*}}) : vector<[8]xi16>, vector<[8]xi16> into vector<[4]x[4]xi32>
+func.func @outerproduct_add_widening_2way_signed_i16i16i32(
+    %a0 : vector<[4]xi16>, %b0 : vector<[4]xi16>,
+    %a1 : vector<[4]xi16>, %b1 : vector<[4]xi16>,
+    %a0_mask : vector<[4]xi1>, %b0_mask : vector<[4]xi1>,
+    %a1_mask : vector<[4]xi1>, %b1_mask : vector<[4]xi1>) -> vector<[4]x[4]xi32> {
+  %a0_ext = arith.extsi %a0 : vector<[4]xi16> to vector<[4]xi32>
+  %b0_ext = arith.extsi %b0 : vector<[4]xi16> to vector<[4]xi32>
+  %a1_ext = arith.extsi %a1 : vector<[4]xi16> to vector<[4]xi32>
+  %b1_ext = arith.extsi %b1 : vector<[4]xi16> to vector<[4]xi32>
+
+  %acc = arith.constant dense<0> : vector<[4]x[4]xi32>
+
+  %0 = arm_sme.outerproduct %a0_ext, %b0_ext acc(%acc) masks(%a0_mask, %b0_mask) : vector<[4]xi32>, vector<[4]xi32>
+  %1 = arm_sme.outerproduct %a1_ext, %b1_ext acc(%0) masks(%a1_mask, %b1_mask) : vector<[4]xi32>, vector<[4]xi32>
+
+  return %1 : vector<[4]x[4]xi32>
+}
+
+// -----
+
+// CHECK-LABEL: @outerproduct_sub_widening_2way_signed_i16i16i32
+// CHECK: arm_sme.smops_wide_2way %{{.*}}, %{{.*}} acc(%{{.*}}) masks(%{{.*}}, %{{.*}}) : vector<[8]xi16>, vector<[8]xi16> into vector<[4]x[4]xi32>
+func.func @outerproduct_sub_widening_2way_signed_i16i16i32(
+    %a0 : vector<[4]xi16>, %b0 : vector<[4]xi16>,
+    %a1 : vector<[4]xi16>, %b1 : vector<[4]xi16>,
+    %a0_mask : vector<[4]xi1>, %b0_mask : vector<[4]xi1>,
+    %a1_mask : vector<[4]xi1>, %b1_mask : vector<[4]xi1>) -> vector<[4]x[4]xi32> {
+  %a0_ext = arith.extsi %a0 : vector<[4]xi16> to vector<[4]xi32>
+  %b0_ext = arith.extsi %b0 : vector<[4]xi16> to vector<[4]xi32>
+  %a1_ext = arith.extsi %a1 : vector<[4]xi16> to vector<[4]xi32>
+  %b1_ext = arith.extsi %b1 : vector<[4]xi16> to vector<[4]xi32>
+
+  %acc = arith.constant dense<0> : vector<[4]x[4]xi32>
+
+  %0 = arm_sme.outerproduct %a0_ext, %b0_ext kind<sub> acc(%acc) masks(%a0_mask, %b0_mask) : vector<[4]xi32>, vector<[4]xi32>
+  %1 = arm_sme.outerproduct %a1_ext, %b1_ext kind<sub> acc(%0) masks(%a1_mask, %b1_mask) : vector<[4]xi32>, vector<[4]xi32>
+
+  return %1 : vector<[4]x[4]xi32>
+}
+
+// -----
+
+// CHECK-LABEL: @outerproduct_add_widening_2way_unsigned_i16i16i32
+// CHECK: arm_sme.umopa_wide_2way %{{.*}}, %{{.*}} acc(%{{.*}}) masks(%{{.*}}, %{{.*}}) : vector<[8]xi16>, vector<[8]xi16> into vector<[4]x[4]xi32>
+func.func @outerproduct_add_widening_2way_unsigned_i16i16i32(
+    %a0 : vector<[4]xi16>, %b0 : vector<[4]xi16>,
+    %a1 : vector<[4]xi16>, %b1 : vector<[4]xi16>,
+    %a0_mask : vector<[4]xi1>, %b0_mask : vector<[4]xi1>,
+    %a1_mask : vector<[4]xi1>, %b1_mask : vector<[4]xi1>) -> vector<[4]x[4]xi32> {
+  %a0_ext = arith.extui %a0 : vector<[4]xi16> to vector<[4]xi32>
+  %b0_ext = arith.extui %b0 : vector<[4]xi16> to vector<[4]xi32>
+  %a1_ext = arith.extui %a1 : vector<[4]xi16> to vector<[4]xi32>
+  %b1_ext = arith.extui %b1 : vector<[4]xi16> to vector<[4]xi32>
+
+  %acc = arith.constant dense<0> : vector<[4]x[4]xi32>
+
+  %0 = arm_sme.outerproduct %a0_ext, %b0_ext acc(%acc) masks(%a0_mask, %b0_mask) : vector<[4]xi32>, vector<[4]xi32>
+  %1 = arm_sme.outerproduct %a1_ext, %b1_ext acc(%0) masks(%a1_mask, %b1_mask) : vector<[4]xi32>, vector<[4]xi32>
+
+  return %1 : vector<[4]x[4]xi32>
+}
+
+// -----
+
+// CHECK-LABEL: @outerproduct_sub_widening_2way_unsigned_i16i16i32
+// CHECK: arm_sme.umops_wide_2way %{{.*}}, %{{.*}} acc(%{{.*}}) masks(%{{.*}}, %{{.*}}) : vector<[8]xi16>, vector<[8]xi16> into vector<[4]x[4]xi32>
+func.func @outerproduct_sub_widening_2way_unsigned_i16i16i32(
+    %a0 : vector<[4]xi16>, %b0 : vector<[4]xi16>,
+    %a1 : vector<[4]xi16>, %b1 : vector<[4]xi16>,
+    %a0_mask : vector<[4]xi1>, %b0_mask : vector<[4]xi1>,
+    %a1_mask : vector<[4]xi1>, %b1_mask : vector<[4]xi1>) -> vector<[4]x[4]xi32> {
+  %a0_ext = arith.extui %a0 : vector<[4]xi16> to vector<[4]xi32>
+  %b0_ext = arith.extui %b0 : vector<[4]xi16> to vector<[4]xi32>
+  %a1_ext = arith.extui %a1 : vector<[4]xi16> to vector<[4]xi32>
+  %b1_ext = arith.extui %b1 : vector<[4]xi16> to vector<[4]xi32>
+
+  %acc = arith.constant dense<0> : vector<[4]x[4]xi32>
+
+  %0 = arm_sme.outerproduct %a0_ext, %b0_ext kind<sub> acc(%acc) masks(%a0_mask, %b0_mask) : vector<[4]xi32>, vector<[4]xi32>
+  %1 = arm_sme.outerproduct %a1_ext, %b1_ext kind<sub> acc(%0) masks(%a1_mask, %b1_mask) : vector<[4]xi32>, vector<[4]xi32>
+
+  return %1 : vector<[4]x[4]xi32>
+}
diff --git a/mlir/test/Dialect/ArmSME/roundtrip.mlir b/mlir/test/Dialect/ArmSME/roundtrip.mlir
index 2ad742493408b0..a96756f4d34268 100644
--- a/mlir/test/Dialect/ArmSME/roundtrip.mlir
+++ b/mlir/test/Dialect/ArmSME/roundtrip.mlir
@@ -1131,3 +1131,115 @@ func.func @arm_sme_streaming_vl_double_words() -> index {
   %svl_d = arm_sme.streaming_vl <double>
   return %svl_d : index
 }
+
+//===----------------------------------------------------------------------===//
+// arm_sme.fmopa_wide_2way
+//===----------------------------------------------------------------------===//
+
+// -----
+
+func.func @arm_sme_fmopa_wide_2way_f16f16_to_f32(%vecA: vector<[8]xf16>, %vecB: vector<[8]xf16>) -> vector<[4]x[4]xf32> {
+  // CHECK: arm_sme.fmopa_wide_2way {{.*}}, {{.*}} : vector<[8]xf16>, vector<[8]xf16> into vector<[4]x[4]xf32>
+  %result = arm_sme.fmopa_wide_2way %vecA, %vecB : vector<[8]xf16>, vector<[8]xf16> into vector<[4]x[4]xf32>
+  return %result : vector<[4]x[4]xf32>
+}
+
+// -----
+
+func.func @arm_sme_fmopa_wide_2way_bf16bf16_to_f32(%vecA: vector<[8]xbf16>, %vecB: vector<[8]xbf16>) -> vector<[4]x[4]xf32> {
+  // CHECK: arm_sme.fmopa_wide_2way {{.*}}, {{.*}} : vector<[8]xbf16>, vector<[8]xbf16> into vector<[4]x[4]xf32>
+  %result = arm_sme.fmopa_wide_2way %vecA, %vecB : vector<[8]xbf16>, vector<[8]xbf16> into vector<[4]x[4]xf32>
+  return %result : vector<[4]x[4]xf32>
+}
+
+// -----
+
+func.func @arm_sme_fmopa_wide_2way_with_masking(%vecA: vector<[8]xf16>, %vecB: vector<[8]xf16>, %maskA: vector<[8]xi1>, %maskB: vector<[8]xi1>) -> vector<[4]x[4]xf32> {
+  // CHECK: arm_sme.fmopa_wide_2way {{.*}}, {{.*}} masks({{.*}}, {{.*}}) : vector<[8]xf16>, vector<[8]xf16> into vector<[4]x[4]xf32>
+  %result = arm_sme.fmopa_wide_2way %vecA, %vecB masks(%maskA, %maskB) : vector<[8]xf16>, vector<[8]xf16> into vector<[4]x[4]xf32>
+  return %result : vector<[4]x[4]xf32>
+}
+
+// -----
+
+func.func @arm_sme_fmopa_wide_2way_with_acc(%vecA: vector<[8]xf16>, %vecB: vector<[8]xf16>, %acc : vector<[4]x[4]xf32>) -> vector<[4]x[4]xf32> {
+  // CHECK: arm_sme.fmopa_wide_2way {{.*}}, {{.*}} acc({{.*}}) : vector<[8]xf16>, vector<[8]xf16> into vector<[4]x[4]xf32>
+  %result = arm_sme.fmopa_wide_2way %vecA, %vecB acc(%acc) : vector<[8]xf16>, vector<[8]xf16> into vector<[4]x[4]xf32>
+  return %result : vector<[4]x[4]xf32>
+}
+
+// -----
+
+func.func @arm_sme_fmopa_wide_2way_with_everything(%vecA: vector<[8]xf16>, %vecB: vector<[8]xf16>, %acc : vector<[4]x[4]xf32>, %maskA: vector<[8]xi1>, %maskB: vector<[8]xi1>) -> vector<[4]x[4]xf32> {
+  // CHECK: arm_sme.fmopa_wide_2way {{.*}}, {{.*}} acc({{.*}}) masks({{.*}}, {{.*}}) : vector<[8]xf16>, vector<[8]xf16> into vector<[4]x[4]xf32>
+  %result = arm_sme.fmopa_wide_2way %vecA, %vecB acc(%acc) masks(%maskA, %maskB) : vector<[8]xf16>, vector<[8]xf16> into vector<[4]x[4]xf32>
+  return %result : vector<[4]x[4]xf32>
+}
+
+//===----------------------------------------------------------------------===//
+// arm_sme.fmops_wide_2way
+//===----------------------------------------------------------------------===//
+
+// -----
+
+func.func @arm_sme_fmops_wide_2way_f16f16_to_f32(%vecA: vector<[8]xf16>, %vecB: vector<[8]xf16>) -> vector<[4]x[4]xf32> {
+  // CHECK: arm_sme.fmops_wide_2way {{.*}}, {{.*}} : vector<[8]xf16>, vector<[8]xf16> into vector<[4]x[4]xf32>
+  %result = arm_sme.fmops_wide_2way %vecA, %vecB : vector<[8]xf16>, vector<[8]xf16> into vector<[4]x[4]xf32>
+  return %result : vector<[4]x[4]xf32>
+}
+
+// -----
+
+func.func @arm_sme_fmops_wide_2way_bf16bf16_to_f32(%vecA: vector<[8]xbf16>, %vecB: vector<[8]xbf16>) -> vector<[4]x[4]xf32> {
+  // CHECK: arm_sme.fmops_wide_2way {{.*}}, {{.*}} : vector<[8]xbf16>, vector<[8]xbf16> into vector<[4]x[4]xf32>
+  %result = arm_sme.fmops_wide_2way %vecA, %vecB : vector<[8]xbf16>, vector<[8]xbf16> into vector<[4]x[4]xf32>
+  return %result : vector<[4]x[4]xf32>
+}
+
+//===----------------------------------------------------------------------===//
+// arm_sme.smopa_wide_2way
+//===----------------------------------------------------------------------===//
+
+// -----
+
+func.func @arm_sme_smopa_wide_2way_i16i16_to_i32(%vecA: vector<[8]xi16>, %vecB: vector<[8]xi16>) -> vector<[4]x[4]xi32> {
+  // CHECK: arm_sme.smopa_wide_2way {{.*}}, {{.*}} : vector<[8]xi16>, vector<[8]xi16> into vector<[4]x[4]xi32>
+  %result = arm_sme.smopa_wide_2way %vecA, %vecB : vector<[8]xi16>, vector<[8]xi16> into vector<[4]x[4]xi32>
+  return %result : vector<[4]x[4]xi32>
+}
+
+//===----------------------------------------------------------------------===//
+// arm_sme.smops_wide_2way
+//===----------------------------------------------------------------------===//
+
+// -----
+
+func.func @arm_sme_smops_wide_2way_i16i16_to_i32(%vecA: vector<[8]xi16>, %vecB: vector<[8]xi16>) -> vector<[4]x[4]xi32> {
+  // CHECK: arm_sme.smops_wide_2way {{.*}}, {{.*}} : vector<[8]xi16>, vector<[8]xi16> into vector<[4]x[4]xi32>
+  %result = arm_sme.smops_wide_2way %vecA, %vecB : vector<[8]xi16>, vector<[8]xi16> into vector<[4]x[4]xi32>
+  return %result : vector<[4]x[4]xi32>
+}
+
+//===----------------------------------------------------------------------===//
+// arm_sme.umopa_wide_2way
+//===----------------------------------------------------------------------===//
+
+// -----
+
+func.func @arm_sme_umopa_wide_2way_i16i16_to_i32(%vecA: vector<[8]xi16>, %vecB: vector<[8]xi16>) -> vector<[4]x[4]xi32> {
+  // CHECK: arm_sme.umopa_wide_2way {{.*}}, {{.*}} : vector<[8]xi16>, vector<[8]xi16> into vector<[4]x[4]xi32>
+  %result = arm_sme.umopa_wide_2way %vecA, %vecB : vector<[8]xi16>, vector<[8]xi16> into vector<[4]x[4]xi32>
+  return %result : vector<[4]x[4]xi32>
+}
+
+//===----------------------------------------------------------------------===//
+// arm_sme.umops_wide_2way
+//===----------------------------------------------------------------------===//
+
+// -----
+
+func.func @arm_sme_umops_wide_2way_i16i16_to_i32(%vecA: vector<[8]xi16>, %vecB: vector<[8]xi16>) -> vector<[4]x[4]xi32> {
+  // CHECK: arm_sme.umops_wide_2way {{.*}}, {{.*}} : vector<[8]xi16>, vector<[8]xi16> into vector<[4]x[4]xi32>
+  %result = arm_sme.umops_wide_2way %vecA, %vecB : vector<[8]xi16>, vector<[8]xi16> into vector<[4]x[4]xi32>
+  return %result : vector<[4]x[4]xi32>
+}
diff --git a/mlir/test/Integration/Dialect/Vector/CPU/ArmSME/test-outerproduct-f16f16f32.mlir b/mlir/test/Integration/Dialect/Vector/CPU/ArmSME/test-outerproduct-f16f16f32.mlir
new file mode 100644
index 00000000000000..8fbdf5d0011ce2
--- /dev/null
+++ b/mlir/test/Integration/Dialect/Vector/CPU/ArmSME/test-outerproduct-f16f16f32.mlir
@@ -0,0 +1,100 @@
+// DEFINE: %{entry} = test_outerproduct_f16f16f32
+// DEFINE: %{widening_opts} = -arm-sme-outer-product-widening
+// DEFINE: %{compile} = mlir-opt %s \
+// DEFINE:   -enable-arm-streaming="streaming-mode=streaming-locally za-mode=new-za" \
+// DEFINE:   -convert-vector-to-arm-sme -convert-arith-to-arm-sme %{widening_opts} \
+// DEFINE:   -convert-arm-sme-to-scf -allocate-arm-sme-tiles \
+// DEFINE:   -convert-arm-sme-to-llvm -cse -canonicalize \
+// DEFINE:   -test-lower-to-llvm -o %t
+// DEFINE: %{run} = %mcr_aarch64_cmd %t \
+// DEFINE:   -march=aarch64 -mattr=+sve,+sme \
+// DEFINE:   -e %{entry} -entry-point-result=void \
+// DEFINE:   -shared-libs=%mlir_runner_utils,%mlir_c_runner_utils,%arm_sme_abi_shlib
+
+// RUN: %{compile}
+
+// RUN: %{run} | FileCheck %s
+
+// Check result is the same when outerproducts are not combined into widening
+// variant.
+
+// REDEFINE: %{widening_opts} =
+// RUN: %{run} | FileCheck %s
+
+func.func @test_outerproduct_f16f16f32() {
+  %undef = llvm.mlir.undef : vector<[4]xf16>
+
+  %a0_data = arith.constant dense<[0., 2., 4., 6.]> : vector<4xf16>
+  %b0_data = arith.constant dense<[1., 3., 5., 7.]> : vector<4xf16>
+  %a1_data = arith.constant dense<[8., 10., 12., 14.]> : vector<4xf16>
+  %b1_data = arith.constant dense<[9., 11., 13., 15.]> : vector<4xf16>
+
+  %a0 = vector.scalable.insert %a0_data, %undef[0] : vector<4xf16> into vector<[4]xf16>
+  %b0 = vector.scalable.insert %b0_data, %undef[0] : vector<4xf16> into vector<[4]xf16>
+  %a1 = vector.scalable.insert %a1_data, %undef[0] : vector<4xf16> into vector<[4]xf16>
+  %b1 = vector.scalable.insert %b1_data, %undef[0] : vector<4xf16> into vector<[4]xf16>
+
+  %a0_ext = arith.extf %a0 : vector<[4]xf16> to vector<[4]xf32>
+  %b0_ext = arith.extf %b0 : vector<[4]xf16> to vector<[4]xf32>
+  %a1_ext = arith.extf %a1 : vector<[4]xf16> to vector<[4]xf32>
+  %b1_ext = arith.extf %b1 : vector<[4]xf16> to vector<[4]xf32>
+
+  %acc = arith.constant dense<7.0> : vector<[4]x[4]xf32>
+  %0 = vector.outerproduct %a0_ext, %b0_ext, %acc : vector<[4]xf32>, vector<[4]xf32>
+  %1 = vector.outerproduct %a1_ext, %b1_ext, %0 : vector<[4]xf32>, vector<[4]xf32>
+
+  // CHECK:      (  79,  95, 111, 127
+  // CHECK-NEXT: (  99, 123, 147, 171
+  // CHECK-NEXT: ( 119, 151, 183, 215
+  // CHECK-NEXT: ( 139, 179, 219, 259
+  vector.print %1 : vector<[4]x[4]xf32>
+
+  return
+}
+
+// TODO: A bug in QEMU causes masked FMOPAs to hang [1]. Should be fixed in
+// 8.2.0, this test currently isn't run, once this version is available in CI
+// it can be run. The check lines here are correct and have been verified on a
+// version with the fix.
+// [1] https://gitlab.com/qemu-project/qemu/-/issues/1985
+func.func @test_masked_outerproduct_f16f16f32() {
+  %undef = llvm.mlir.undef : vector<[4]xf16>
+
+  %a0_data = arith.constant dense<[0., 2., 4., 6.]> : vector<4xf16>
+  %b0_data = arith.constant dense<[1., 3., 5., 7.]> : vector<4xf16>
+  %a1_data = arith.constant dense<[8., 10., 12., 14.]> : vector<4xf16>
+  %b1_data = arith.constant dense<[9., 11., 13., 15.]> : vector<4xf16>
+
+  %a0 = vector.scalable.insert %a0_data, %undef[0] : vector<4xf16> into vector<[4]xf16>
+  %b0 = vector.scalable.insert %b0_data, %undef[0] : vector<4xf16> into vector<[4]xf16>
+  %a1 = vector.scalable.insert %a1_data, %undef[0] : vector<4xf16> into vector<[4]xf16>
+  %b1 = vector.scalable.insert %b1_data, %undef[0] : vector<4xf16> into vector<[4]xf16>
+
+  %a0_ext = arith.extf %a0 : vector<[4]xf16> to vector<[4]xf32>
+  %b0_ext = arith.extf %b0 : vector<[4]xf16> to vector<[4]xf32>
+  %a1_ext = arith.extf %a1 : vector<[4]xf16> to vector<[4]xf32>
+  %b1_ext = arith.extf %b1 : vector<[4]xf16> to vector<[4]xf32>
+
+  %acc = arith.constant dense<7.0> : vector<[4]x[4]xf32>
+
+  %c2 = arith.constant 2 : index
+  %c3 = arith.constant 3 : index
+  %mask0 = vector.create_mask %c2, %c3 : vector<[4]x[4]xi1>
+  %mask1 = vector.create_mask %c3, %c2 : vector<[4]x[4]xi1>
+
+  %0 = vector.mask %mask0 {
+    vector.outerproduct %a0_ext, %b0_ext, %acc : vector<[4]xf32>, vector<[4]xf32>
+  } : vector<[4]x[4]xi1> -> vector<[4]x[4]xf32>
+
+  %1 = vector.mask %mask1 {
+    vector.outerproduct %a1_ext, %b1_ext, %0 : vector<[4]xf32>, vector<[4]xf32>
+  } : vector<[4]x[4]xi1> -> vector<[4]x[4]xf32>
+
+  // MASKED:      (  79,  95,  7, 7
+  // MASKED-NEXT: (  99, 123, 17, 7
+  // MASKED-NEXT: ( 115, 139,  7, 7
+  // MASKED-NEXT: (   7,   7,  7, 7
+  vector.print %1 : vector<[4]x[4]xf32>
+
+  return
+}
diff --git a/mlir/test/Target/LLVMIR/arm-sme.mlir b/mlir/test/Target/LLVMIR/arm-sme.mlir
index 7a42033dc04bc0..aedb6730b06bb3 100644
--- a/mlir/test/Target/LLVMIR/arm-sme.mlir
+++ b/mlir/test/Target/LLVMIR/arm-sme.mlir
@@ -63,6 +63,12 @@ llvm.func @arm_sme_imopa(%nxv8i16 : vector<[8]xi16>,
   // CHECK: call void @llvm.aarch64.sme.usmopa.wide.nxv16i8
   "arm_sme.intr.usmopa.wide"(%nxv16i1, %nxv16i1, %nxv16i8, %nxv16i8) <{tile_id = 0 : i32}> :
     (vector<[16]xi1>, vector<[16]xi1>, vector<[16]xi8>, vector<[16]xi8>) -> ()
+  // CHECK: call void @llvm.aarch64.sme.smopa.za32.nxv8i16
+  "arm_sme.intr.smopa.za32"(%nxv8i1, %nxv8i1, %nxv8i16, %nxv8i16) <{tile_id = 0 : i32}> :
+    (vector<[8]xi1>, vector<[8]xi1>, vector<[8]xi16>, vector<[8]xi16>) -> ()
+  // CHECK: call void @llvm.aarch64.sme.umopa.za32.nxv8i16
+  "arm_sme.intr.umopa.za32"(%nxv8i1, %nxv8i1, %nxv8i16, %nxv8i16) <{tile_id = 0 : i32}> :
+    (vector<[8]xi1>, vector<[8]xi1>, vector<[8]xi16>, vector<[8]xi16>) -> ()
   llvm.return
 }
 
@@ -122,6 +128,12 @@ llvm.func @arm_sme_imops(%nxv8i16 : vector<[8]xi16>,
   // CHECK: call void @llvm.aarch64.sme.usmops.wide.nxv16i8
   "arm_sme.intr.usmops.wide"(%nxv16i1, %nxv16i1, %nxv16i8, %nxv16i8) <{tile_id = 0 : i32}> :
     (vector<[16]xi1>, vector<[16]xi1>, vector<[16]xi8>, vector<[16]xi8>) -> ()
+  // CHECK: call void @llvm.aarch64.sme.smops.za32.nxv8i16
+  "arm_sme.intr.smops.za32"(%nxv8i1, %nxv8i1, %nxv8i16, %nxv8i16) <{tile_id = 0 : i32}> :
+    (vector<[8]xi1>, vector<[8]xi1>, vector<[8]xi16>, vector<[8]xi16>) -> ()
+  // CHECK: call void @llvm.aarch64.sme.umops.za32.nxv8i16
+  "arm_sme.intr.umops.za32"(%nxv8i1, %nxv8i1, %nxv8i16, %nxv8i16) <{tile_id = 0 : i32}> :
+    (vector<[8]xi1>, vector<[8]xi1>, vector<[8]xi16>, vector<[8]xi16>) -> ()
   llvm.return
 }
 



More information about the Mlir-commits mailing list