```

    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"> {

          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();

  // CHECK: "mhlo.dot"(%[[UPDATED_A]], %[[UPDATED_B]])
  %0 = "tf.MatMul"(%a, %b) {transpose_a = true, transpose_b = true} : (tensor<7x5xf32>, tensor<11x7xf32>) -> tensor<5x11xf32>

  func.return %0 : tensor<5x11xf32>
}

// Verify that MatMul with ranked inputs are lowered to HLO.
// CHECK-LABEL: matmul_ranked

  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>

  // - rhs: [RHSBATCHDIMS..., RHSROWS, RHSCOLS]
  // - result: [broadcast(LHSBATCHDIMS, RHSBATCHDIMS)..., LHSROWS, RHSCOLS]
  // To perform the matmul, we need to first broadcast lhs and rhs to a common
  // set of leading dimensions before doing the actual matmul.
  // That's what the code below does.
  // In particular, we populate out_lhs and out_rhs to have dimension structure:

* `tf.config.experimental.enable_tensor_float_32_execution`

}

def TF__FusedMatMulOp : TF_Op<"_FusedMatMul", [Pure, TF_SameOperandsAndResultElementTypeResolveRef]> {
  let summary = [{
Performs a MatMul followed by a specified series of operations.
  }];

  let description = [{
The inputs to the MatMul are specified by `a` and `b`. The series of operations
that follows is specified by the `fused_ops` attribute, which is a list of TF op