func.func @softmax(%arg0: tensor<8x16xf32>) -> tensor<8x16xf32> {
  %0 = "tf.Softmax"(%arg0) : (tensor<8x16xf32>) -> tensor<8x16xf32>
  func.return %0 : tensor<8x16xf32>
// CHECK: %[[CUSTOM_0:.*]] = "tfl.custom"(%arg0) <{custom_code = "FlexSoftmax", custom_option = #tfl<const_bytes : "0x07536F66746D617800161207536F66746D61781A002A070A0154120230013200000221191414042801">}> : (tensor<8x16xf32>) -> tensor<8x16xf32>

  %1 = "tfl.fully_connected"(%0, %arg1, %cst_1) {asymmetric_quantize_inputs = false, fused_activation_function = "NONE", keep_num_dims = false, weights_format = "DEFAULT"} : (tensor<4x1024xf32>, tensor<8x1024xf32>, none) -> tensor<4x8xf32>

func.func @move_transpose_handle_broadcast(%arg0:tensor<8x64xf32>, %arg1:tensor<8x64x64xf32>) -> tensor<512x64xf32> {
  %cst = "tf.Const"() {value = dense<3> : tensor<i32>} : () -> tensor<i32>
  %cst_1 = "tf.Const"() {value = dense<[2, 0, 1]> : tensor<3xi32>} : () -> tensor<3xi32>
  %cst_2 = "tf.Const"() {value = dense<[512, 64]> : tensor<2xi32>} : () -> tensor<2xi32>
  %0 = "tf.ExpandDims"(%arg0, %cst) {device = ""} : (tensor<8x64xf32>, tensor<i32>) -> tensor<8x64x1xf32>

// -----

// CHECK-LABEL: @duplicate_const_for_shape_determining_operand_at_idx_2
// CHECK-SAME: (%[[ARG_0:.*]]: tensor<16x4xf32>, %[[ARG_1:.*]]: tensor<16xi32>)
func.func private @duplicate_const_for_shape_determining_operand_at_idx_2(%arg0: tensor<16x4xf32>, %arg1: tensor<16xi32>) -> tensor<16xf32> {
  %cst = "tf.Const"() {device = "", value = dense<[1]> : tensor<1xi32>} : () -> tensor<1xi32>

  func.return %0 : tensor<8x16xf32>
}

// -----

func.func @testCumprod(%arg: tensor<8x16xf32>) -> tensor<8x16xf32> {
  %axis = arith.constant dense<-3> : tensor<i32>
  // expected-error @+1 {{axis operand should be within range [-2, 2)}}
  %0 = "tf.Cumprod"(%arg, %axis) : (tensor<8x16xf32>, tensor<i32>) -> tensor<8x16xf32>
  func.return %0 : tensor<8x16xf32>
}

    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-SAME:    lhs_batching_dimensions = #vhlo.tensor_v1<dense<0> : tensor<1xi64>>,

  %cst_0 = arith.constant dense<[1, 1, 64]> : tensor<3xi32>
  %0 = "tfl.reshape"(%arg0, %cst_0) : (tensor<1x64xf32>, tensor<3xi32>) -> tensor<1x1x64xf32>
  %1 = "tfl.batch_matmul"(%0, %arg1) {adj_x = false, adj_y = false} : (tensor<1x1x64xf32>, tensor<1x64x1024xf32>) -> tensor<1x1x1024xf32>
  %2 = "tfl.reshape"(%1, %cst) : (tensor<1x1x1024xf32>, tensor<2xi32>) -> tensor<1x1024xf32>
  return %2 : tensor<1x1024xf32>

  // CHECK: [[RESHAPE:%.*]] = mhlo.reshape [[SLICE]] : (tensor<2x16x2xf32>) -> tensor<2x16x2xf32>
  %0 = "tf.StridedSlice"(%input, %begin, %end, %strides) {Index = i32, T = f32} : (tensor<10x16x2xf32>, tensor<1xi32>, tensor<1xi32>, tensor<1xi32>) -> tensor<2x16x2xf32>
  // CHECK: return [[RESHAPE]] : tensor<2x16x2xf32>
  func.return %0 : tensor<2x16x2xf32>
}

// -----

// 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

  %1 = stablehlo.transpose %arg0, dims = [1, 0] : (tensor<16x8xf32>) -> tensor<8x16xf32>
  %2 = "stablehlo.reduce_window"(%1, %0) ({
  ^bb0(%arg1: tensor<f32>, %arg2: tensor<f32>):
      %3 = stablehlo.maximum %arg1, %arg2 : tensor<f32>
      stablehlo.return %3 : tensor<f32>
  }) {window_dimensions = array<i64: 2, 2>, window_strides = array<i64: 2, 2>} : (tensor<8x16xf32>, tensor<f32>) -> tensor<4x8xf32>
  return %2 : tensor<4x8xf32>
}