]
  )
  def matmul(self, matmul_input: core.Tensor) -> Mapping[str, core.Tensor]:
    """Performs a matrix multiplication.

    Args:
      matmul_input: Input tensor to matmul with the filter.

    Returns:
      A map of: output key -> output result.
    """
    out = math_ops.matmul(matmul_input, self.matmul_filters)

    return {'output': out}

}

func.func @matmul(%arg0: tensor<40x37xf32>, %arg1: tensor<37x40xf32>) -> tensor<40x40xf32> {
  %0 = "tf.MatMul"(%arg0, %arg1) {T = "tfdtype$DT_FLOAT", device = "/device:CPU:0", name = "MatMul", transpose_a = false, transpose_b = false} :
(tensor<40x37xf32>, tensor<37x40xf32>) -> tensor<40x40xf32>
  func.return %0 : tensor<40x40xf32>
// CHECK-LABEL: matmul

  auto matmul = rewriter.create<TF::BatchMatMulV3Op>(
      loc, RankedTensorType::get(matmul_shape, result_type.getElementType()),
      lhs_flattend, rhs_flattend);

  if (result_type.hasStaticShape()) {
    auto reshaped =
        rewriter.create<mhlo::ReshapeOp>(loc, result_type, matmul.getResult());
    return reshaped.getResult();
  }

    ```

    The pass also works across control flow and functional calls.
  }];
}



def UnrollBatchMatMulPass : Pass<"tf-unroll-batch-matmul", "mlir::func::FuncOp"> {
  let summary = "Unroll TF BatchMatMul op into Reshape, Slice, MatMul, Pack ops.";
  let constructor = "TF::CreateUnrollBatchMatMulPassPass()";
}

def ClusterFormationPass : Pass<"tf-device-cluster-formation", "mlir::ModuleOp"> {

        "node": ">=8"
      },
      "funding": {
        "url": "https://github.com/sponsors/sindresorhus"
      }
    },
    "node_modules/mathml-tag-names": {
      "version": "2.1.3",
      "resolved": "https://registry.npmjs.org/mathml-tag-names/-/mathml-tag-names-2.1.3.tgz",
      "integrity": "sha512-APMBEanjybaPzUrfqU0IMU5I0AswKMH7k8OTLs0vvV4KZpExkTkY87nR/zpbuTPj+gARop7aGUbl11pnDfW6xg==",
      "dev": true,

          UniformQuantizedStableHloToTflPass> {
 private:
  void runOnOperation() override;
};

// TODO: b/323645515 - Refactor reference functions.
// Bias scales for matmul-like ops should be input scale * filter scale. Here it
// is assumed that the input is per-tensor quantized and filter is per-channel
// quantized.
SmallVector<double> GetBiasScales(const double input_scale,

    if (fc_op.getFusedActivationFunction() != "NONE") return failure();

    // Only fuse multiplier if all dimensions other than the depth dimension
    // are equal to 1 since otherwise
    // `matmul(x, filter) * cst != matmul(x, filter * cst)`
    // even if `filter` and `cst` are be broadcastable.
    auto shape = cst.getType().getShape();
    if (!IsDimensionsDegenerateExceptLastOne(shape)) return failure();

  func.return %0: tensor<2x3x7xf32>
}

// CHECK-LABEL: testBatchMatMulToMatMul
func.func @testBatchMatMulToMatMul(%arg0: tensor<2x3xf32>, %arg1: tensor<3x2xf32>) -> tensor<2x2xf32> {
  // CHECK: %0 = "tf.MatMul"(%arg0, %arg1) <{grad_a = false, grad_b = false, transpose_a = false, transpose_b = false}> {device = "/job:localhost/replica:0/task:0/device:GPU:0"} : (tensor<2x3xf32>, tensor<3x2xf32>) -> tensor<2x2xf32>