func.return %1 : tensor<128x32xf32>
  // CHECK: %[[FULLY_CONNECTED:.*]] = "tfl.fully_connected"(%arg0, %arg1, %arg2) <{fused_activation_function = "NONE", keep_num_dims = false, weights_format = "DEFAULT"}> : (tensor<128x64xf32>, tensor<32x64xf32>, tensor<32xf32>) -> tensor<128x32xf32>
  // CHECK: return %[[FULLY_CONNECTED]] : tensor<128x32xf32>
}

  // CHECK: %[[RESHAPE:.*]] = "tf.Reshape"(%[[ADD]], %[[CST_2]]) {device = ""} : (tensor<64x8x64xf32>, tensor<2xi32>) -> tensor<512x64xf32>
  // CHECK: return %[[RESHAPE]] : tensor<512x64xf32>
}

// CHECK-LABEL: dont_move_transpose_different_ranks

// DISABLE-ALL-DOT-GENERAL: @main
func.func @main(%arg0: tensor<1x1x167xf32>) -> tensor<1x1x64xf32> {
  %0 = stablehlo.constant dense<2.000000e+00> : tensor<167x64xf32>
  %1 = stablehlo.dot_general %arg0, %0, contracting_dims = [2] x [0], precision = [DEFAULT, DEFAULT] : (tensor<1x1x167xf32>, tensor<167x64xf32>) -> tensor<1x1x64xf32>
  return %1 : tensor<1x1x64xf32>
}

// DISABLE-ALL-DOT-GENERAL: %[[CONST:.+]] = stablehlo.constant dense<2.000000e+00>

func.func @dot_general_with_relu_fn(%arg0: tensor<1x1x167xf32>, %arg1: tensor<167x64xf32>) -> tensor<1x1x64xf32> {
  %0 = stablehlo.constant dense<2.000000e+00> : tensor<167x64xf32>
  %1 = stablehlo.constant dense<0.000000e+00> : tensor<1x1x64xf32>
  %2 = stablehlo.dot_general %arg0, %0, contracting_dims = [2] x [0], precision = [DEFAULT, DEFAULT] : (tensor<1x1x167xf32>, tensor<167x64xf32>) -> tensor<1x1x64xf32>
  %3 = stablehlo.maximum %2, %1 : tensor<1x1x64xf32>

// dimensions into a single dimension. For example,
//
//   %shape = arith.constant dense<[1, 128, 64]> : tensor<3xi32>
//   %reshape = tfl.reshape(%input, %shape) // %input: tensor<128x64xf32>
//   %fc = tfl.fully_connected(%reshape, %filter, %bias)
//           {keep_num_dims = false, weights_format = "DEFAULT"}
//
// can be canonicalized to
//

    precision_config = #vhlo.array_v1<[#vhlo<precision_v1 DEFAULT>, #vhlo<precision_v1 DEFAULT>]>}> : (tensor<1x1x167xf32>, tensor<167x64xf32>) -> tensor<1x1x64xf32>
  return %0 : tensor<1x1x64xf32>
}

//CHECK:func.func private @dot_general(%arg0: tensor<1x1x167xf32>, %arg1: tensor<167x64xf32>) -> tensor<1x1x64xf32> {
//CHECK-NEXT: %0 = "vhlo.dot_general_v1"(%arg0, %arg1) <{

// CHECK: return %[[v4]] : tensor<?x36x32xf32>
}

func.func @einsum_with_runtime_shape2(%arg0 : tensor<?x?x8x64xf32>, %arg1 : tensor<8x8x64xf32>) -> tensor<?x?x8xf32> {
  %0 = "tf.Einsum"(%arg0, %arg1) {device = "", equation = "ABNH,DNH->ABD"} : (tensor<?x?x8x64xf32>, tensor<8x8x64xf32>) -> tensor<?x?x8xf32>
  func.return %0 : tensor<?x?x8xf32>
// CHECK-LABEL: einsum_with_runtime_shape2

// CHECK-LABEL: matmulBiasAdd
func.func @matmulBiasAdd(%arg0: tensor<64xf32>, %arg1: tensor<8x32xf32>, %arg2: tensor<32x64xf32>) -> (tensor<*xf32>) {
  // CHECK: %[[VAL_3:.*]] = "tf._FusedMatMul"(%arg1, %arg2, %arg0) <{epsilon = 0.000000e+00 : f32, fused_ops = ["BiasAdd"], transpose_a = false, transpose_b = false}> : (tensor<8x32xf32>, tensor<32x64xf32>, tensor<64xf32>) -> tensor<*xf32>

  // CHECK-LABEL: func @parallel_execute_with_tiled_input
  // CHECK-SAME: (%[[ARG_0:[a-z0-9]*]]: tensor<128x10xf32>, %[[ARG_1:[a-z0-9]*]]: tensor<128x10xf32>, %[[ARG_2:[a-z0-9]*]]: tensor<*xi32>, %[[ARG_3:[a-z0-9]*]]: tensor<*xi32>)
  func.func @parallel_execute_with_tiled_input(%arg0: tensor<128x10xf32>, %arg1: tensor<128x10xf32>, %arg2: tensor<*xi32>, %arg3: tensor<*xi32>) -> (tensor<*xi32>, tensor<*xi1>) {
    // CHECK: tf_device.replicate

  func.func @main(%arg0: tensor<128x10xf32>, %arg1: tensor<10x1024xf32>, %arg2: tensor<128x1024xf32>) -> (tensor<128x10xf32> {mhlo.sharding = "\08\03\1A\02\01\02\22\02\00\01"}, tensor<10x1024xf32> {mhlo.sharding = "\08\01\1A\01\01\22\01\00"}, tensor<128x1024xf32> {mhlo.sharding = ""}) {
    func.return %arg0, %arg1, %arg2 : tensor<128x10xf32>, tensor<10x1024xf32>, tensor<128x1024xf32>
  }
}