func.return %1 : tensor<1024x1x128xf32>

  // CHECK-DAG: %cst = arith.constant dense<2.000000e+00> : tensor<128xf32>
  // CHECK: %[[fc:.*]] = "tfl.fully_connected"(%arg0, %arg1, %cst)
  // CHECK: return %[[fc]]
}

// CHECK-LABEL: @FuseFullyConnectedAddWithExistingBias

  %custom_1 = "tfl.custom"(%arg0, %dq_w) {custom_code = "CustomTestOp", custom_option = #tfl<const_bytes : "0x">} : (tensor<1x1x1x1xf32>, tensor<1024x1x1x1xf32>) -> tensor<*xf32>
  %custom_2 = "tfl.custom"(%arg0, %dq_w) {custom_code = "CustomTestOp", custom_option = #tfl<const_bytes : "0x">} : (tensor<1x1x1x1xf32>, tensor<1024x1x1x1xf32>) -> tensor<*xf32>

// CHECK-LABEL: FuseTransposeFCRhsToBatchMatmul
func.func @FuseTransposeFCRhsToBatchMatmul(%arg0: tensor<16x1024xf32>, %arg1: tensor<1024x128xf32>, %arg2: none) -> tensor<16x128xf32> {
  %cst = arith.constant dense<[1, 0]> : tensor<2xi32>
  %0 = "tfl.transpose"(%arg1, %cst) : (tensor<1024x128xf32>, tensor<2xi32>) -> tensor<128x1024xf32>
  // CHECK: "tfl.batch_matmul"(%arg0, %arg1)

  %w = arith.constant dense<127.0> : tensor<1024x1x1x1xf32>
  %custom = "tfl.custom"(%0, %w) {custom_code = "CustomTestOp", custom_option = #tfl<const_bytes : "0x">} : (tensor<1x1x1x1xf32>, tensor<1024x1x1x1xf32>) -> tensor<*xf32>
  func.return %custom : tensor<*xf32>

// CHECK: %[[w:.*]] = arith.constant dense<1.270000e+02> : tensor<1024x1x1x1xf32>

// CHECK: return %[[v8]] : tensor<?x36x?xf32>
}

func.func @einsum_with_runtime_outputshape2(%arg0 : tensor<?x?x1024xf32>, %arg1 : tensor<1024x8x128xf32>) -> tensor<?x?x8x128xf32> {
  %0 = "tf.Einsum"(%arg0, %arg1) {device = "", equation = "ABD,DNH->ABNH"} : (tensor<?x?x1024xf32>, tensor<1024x8x128xf32>) -> tensor<?x?x8x128xf32>
  func.return %0 : tensor<?x?x8x128xf32>
// CHECK-LABEL: einsum_with_runtime_outputshape2