%filter = arith.constant dense<2.0> : tensor<3x3x3x16xf32>
  %bias = arith.constant dense<3.0> : tensor<16xf32>
  %value = arith.constant dense<4.0> : tensor<16xf32>
  %0 = "tf.Conv2D"(%arg, %filter) {T = "tfdtype$DT_FLOAT", data_format = "NHWC", dilations = [1, 2, 3, 1], padding = "SAME", strides = [1, 4, 5, 1]} : (tensor<256x32x32x3xf32>, tensor<3x3x3x16xf32>) -> tensor<256x8x7x16xf32>

  %fq = "tf.FakeQuantWithMinMaxVarsPerChannel"(%in, %mini, %maxi) {num_bits = 3, narrow_range = false} : (tensor<3x3x3x16xf32>, tensor<16xf32>, tensor<16xf32>) -> tensor<3x3x3x16xf32>
  %rst = "tf.Conv2D"(%arg, %fq) {T = "tfdtype$DT_FLOAT", data_format = "NHWC", dilations = [1, 2, 3, 1], padding = "SAME", strides = [1, 4, 5, 1]} : (tensor<256x32x32x3xf32>, tensor<3x3x3x16xf32>) -> tensor<256x8x7x16xf32>
  func.return %rst : tensor<256x8x7x16xf32>

  %fq = "tf.FakeQuantWithMinMaxVarsPerChannel"(%in, %mini, %maxi) {num_bits = 5, narrow_range = false} : (tensor<3x3x3x16xf32>, tensor<16xf32>, tensor<16xf32>) -> tensor<3x3x3x16xf32>
  %rst = "tf.Conv2D"(%arg, %fq) {T = "tfdtype$DT_FLOAT", data_format = "NHWC", dilations = [1, 2, 3, 1], padding = "SAME", strides = [1, 4, 5, 1]} : (tensor<256x32x32x3xf32>, tensor<3x3x3x16xf32>) -> tensor<256x8x7x16xf32>
  func.return %rst : tensor<256x8x7x16xf32>

}

func.func @depthwiseConv2D(tensor<256x32x32x3xf32>, tensor<3x3x3x4xf32>, tensor<256x3x32x32xf32>) -> (tensor<256x30x30x12xf32>, tensor<256x12x30x30xf32>, tensor<256x30x30x12xf32>, tensor<256x30x30x12xf32>) {
^bb0(%arg0: tensor<256x32x32x3xf32>, %arg1: tensor<3x3x3x4xf32>, %arg2: tensor<256x3x32x32xf32>) :
   // OK

  // CHECK-LABEL: func @conv2d_unranked_input
  func.func @conv2d_unranked_input(%arg0: tensor<*xf32>, %arg1: tensor<3x3x3x16xf32>) -> tensor<*xf32> {
    // CHECK: "tf.Conv2D"
    // CHECK-SAME: -> tensor<?x?x?x16xf32>
    %0 = "tf.Conv2D"(%arg0, %arg1) {padding = "SAME", strides = [1, 1, 1, 1]} : (tensor<*xf32>, tensor<3x3x3x16xf32>) -> tensor<*xf32>
    func.return %0 : tensor<*xf32>
  }

  %w = arith.constant dense<2.0> : tensor<3x3x3x3xf32>
  %q = "tfl.quantize"(%w) {qtype = tensor<3x3x3x3x!quant.uniform<i8<-127:127>:f32:0,{1.0,2.0,3.0}>>} : (tensor<3x3x3x3xf32>) -> tensor<3x3x3x3x!quant.uniform<i8<-127:127>:f32:0,{1.0,2.0,3.0}>>
  %dq = "tfl.dequantize"(%q) : (tensor<3x3x3x3x!quant.uniform<i8<-127:127>:f32:0,{1.0,2.0,3.0}>>) -> tensor<3x3x3x3xf32>

    strides = [1, 1, 1, 1],
    use_cudnn_on_gpu = true
  } : (tensor<4xi32>, tensor<2x2x5x21xf32>, tensor<5x2x2x21xf32>) -> tensor<5x3x3x15xf32>
  func.return %result : tensor<5x3x3x15xf32>
}


// CHECK-LABEL: @conv3d_backprop_input
func.func @conv3d_backprop_input(%filter: tensor<3x3x3x1x6xf32>, %out_backprop: tensor<2x8x8x8x6xf32>) -> tensor<2x8x8x8x1xf32> {

  %cst_0 = arith.constant dense<[1.0, 2.0, 3.0]> : tensor<3xf32>
  %w = arith.constant dense<2.0> : tensor<3x3x3x3xf32>
  %q = "tfl.quantize"(%w) {qtype = tensor<3x3x3x3x!quant.uniform<i8:f32, 0.1:1>>} : (tensor<3x3x3x3xf32>) -> tensor<3x3x3x3x!quant.uniform<i8:f32, 0.1:1>>
  %dq = "tfl.dequantize"(%q) : (tensor<3x3x3x3x!quant.uniform<i8:f32, 0.1:1>>) -> tensor<3x3x3x3xf32>

        %cst = "tf.Const"() {value = dense<2.000000e+00> : tensor<2x3x3x1xf32>} : () -> tensor<2x3x3x1xf32>
        %cst_0 = "tf.Const"() {value = dense<0.400000e+00> : tensor<3xf32>} : () -> tensor<3xf32>
        %cst_1 = "tf.Const"() {value = dense<0.500000e+00> : tensor<3xf32>} : () -> tensor<3xf32>
        %w = "tf.Mul"(%cst, %arg1) : (tensor<2x3x3x1xf32>, tensor<f32>) -> tensor<2x3x3x1xf32>

  %cst = "tf.Const"() {value = dense<2.000000e+00> : tensor<2x3x3x1xf32>} : () -> tensor<2x3x3x1xf32>
  %cst_0 = "tf.Const"() {value = dense<0.400000e+00> : tensor<3xf32>} : () -> tensor<3xf32>
  %cst_1 = "tf.Const"() {value = dense<0.500000e+00> : tensor<3xf32>} : () -> tensor<3xf32>
  %identity = "tf.Identity"(%cst) : (tensor<2x3x3x1xf32>) -> tensor<2x3x3x1xf32>