%10 = "quantfork.qcast"(%9) {volatile} : (tensor<4x5xf32>) -> tensor<4x5x!quant.uniform<i8:f32, 1.000000e-03:-3>>
    %11 = "quantfork.dcast"(%10) : (tensor<4x5x!quant.uniform<i8:f32, 1.000000e-03:-3>>) -> tensor<4x5xf32>
    return %11 : tensor<4x5xf32>
  }

  // CHECK: quantized_dot_general_fn_1

    %9 = "tfl.fully_connected"(%8, %cst_10, %cst_0) {fused_activation_function = "NONE", keep_num_dims = false, weights_format = "DEFAULT"} : (tensor<4x5xf32>, tensor<8x5xf32>, tensor<8xf32>) -> tensor<4x8xf32>
    %10:4 = "tfl.split"(%cst_5, %9) {num_splits = 4 : i32} : (tensor<i32>, tensor<4x8xf32>) -> (tensor<4x2xf32>, tensor<4x2xf32>, tensor<4x2xf32>, tensor<4x2xf32>)

  %1 = "tf.AddV2"(%arg0, %0) : (tensor<4x1xf32>, tensor<1x2xf32>) -> tensor<4x2xf32>
  %2 = "tf.AddV2"(%0, %arg0) : (tensor<1x2xf32>, tensor<4x1xf32>) -> tensor<4x2xf32>

  // If operand has the same shape as a result, we can fold it.
  %3 = "tf.AddV2"(%arg1, %0) : (tensor<4x2xf32>, tensor<1x2xf32>) -> tensor<4x2xf32>
  %4 = "tf.AddV2"(%0, %arg1) : (tensor<1x2xf32>, tensor<4x2xf32>) -> tensor<4x2xf32>

  // CHECK: %[[CONST:.*]] = "tf.Const"()

func.func @einsum(%arg0: tensor<2x3xf32>, %arg1: tensor<3x4xf32>) -> tensor<2x4xf32> {
  // CHECK:  mhlo.einsum
  %0 = "tf.Einsum"(%arg0, %arg1) {equation = "ab,bc->ac"} : (tensor<2x3xf32>, tensor<3x4xf32>) -> tensor<2x4xf32>
  func.return %0: tensor<2x4xf32>
}

// -----

// CHECK-LABEL: func @unary_einsum
func.func @unary_einsum(%arg0: tensor<2x3xf32>) -> tensor<2x2xf32> {
  // CHECK:  mhlo.unary_einsum

// `tfl.concatenation`.

func.func @concatenate_float(%arg0: tensor<3x2xf32>, %arg1: tensor<1x2xf32>) -> tensor<4x2xf32> {
  %0 = "stablehlo.concatenate"(%arg0, %arg1) {dimension = 0 : i64} : (tensor<3x2xf32>, tensor<1x2xf32>) -> tensor<4x2xf32>
  return %0 : tensor<4x2xf32>
}
// CHECK-LABEL: concatenate_float
// CHECK-NOT: tfl.concatenation
// CHECK: stablehlo.concatenate

// -----

  %1 = "tfl.add"(%0, %cst2) {fused_activation_function = "NONE"} : (tensor<4x4xf32>, tensor<4x4xf32>) -> tensor<4x4xf32>

  func.return %1 : tensor<4x4xf32>

  // CHECK-DAG: %[[bias:.*]] = arith.constant dense<{{.*}}> : tensor<4xf32>
  // CHECK-DAG: %[[filter:.*]] = arith.constant dense<{{.*}}> : tensor<4x4xf32>

        tensor<1x1x5xf32>,
        tensor<2x5xf32>, tensor<2x5xf32>, tensor<2x5xf32>, tensor<2x5xf32>,
        tensor<2x4xf32>, tensor<2x4xf32>, tensor<2x4xf32>, tensor<2x4xf32>,
        tensor<2xf32>, tensor<2xf32>, tensor<2xf32>,
        tensor<2xf32>, tensor<2xf32>, tensor<2xf32>, tensor<2xf32>,
        tensor<4x2xf32>, tensor<4xf32>,
        tensor<1x4xf32>, tensor<1x2xf32>,
        none, none, none, none) -> tensor<*xf32>

    %0 = stablehlo.dot_general %arg0, %arg1, contracting_dims = [1] x [0], precision = [DEFAULT, DEFAULT] : (tensor<1x4xf32>, tensor<4x3xf32>) -> tensor<1x3xf32>
    %1 = stablehlo.add %0, %arg2 : tensor<1x3xf32>
    return %1 : tensor<1x3xf32>
  }

    %9 = stablehlo.convert %2 : (tensor<2x3xi8>) -> tensor<2x3xf32>
    %10 = stablehlo.dot_general %8, %9, contracting_dims = [1] x [0] : (tensor<1x2xf32>, tensor<2x3xf32>) -> tensor<1x3xf32>
    %11 = stablehlo.convert %3 : (tensor<1x3xi32>) -> tensor<1x3xf32>
    %12 = stablehlo.subtract %10, %11 : tensor<1x3xf32>  // q1 * q2 - z1 * q2
    %13 = stablehlo.multiply %12, %4 : tensor<1x3xf32>  // s1 * s2

// CHECK-SAME:        rhs_contracting_dimensions = [1]
  %0 = "tf.BatchMatMulV2"(%arg0, %arg1) {adj_x = true, adj_y = true, device = ""} : (tensor<2x5xf32>, tensor<4x2xf32>) -> tensor<5x4xf32>
  func.return %0 : tensor<5x4xf32>
}

func.func @batchmatmulv2_adj_complex(%arg0: tensor<2x5xcomplex<f32>>, %arg1: tensor<4x2xcomplex<f32>>) -> tensor<5x4xcomplex<f32>> {
// CHECK-LABEL:   func @batchmatmulv2_adj_complex(