// CHECK-NOT: "tf._FusedConv2D"
  %0 = "tf.Conv2D"(%arg2, %arg1) <{data_format = "NHWC", dilations = [1, 1, 1, 1], explicit_paddings = [], padding = "SAME", strides = [1, 1, 1, 1], use_cudnn_on_gpu = true}> : (tensor<8x32x32x3xf32>, tensor<1x1x3x128xf32>) -> tensor<*xf32>
  // The result of the conv must be the first input to BiasAdd to be fusable.

// CHECK-DAG: %[[cst:.*]] = "tf.Const{{.*}} dense<8.000000e+00> : tensor<3x3x3x16xf32>
// CHECK-DAG: %[[cst_0:.*]] = "tf.Const{{.*}} dense<1.200000e+01> : tensor<16xf32>
// CHECK-NEXT: %[[conv:.*]] = "tf.Conv2D"(%arg0, %[[cst]])
// CHECK-NEXT: %[[bias:.*]] = "tf.AddV2"(%[[conv]], %[[cst_0]])
// CHECK-NEXT: return %[[bias]] : tensor<256x8x7x16xf32>
}

// CHECK-LABEL: convaddv2mul

  // CHECK: %[[ARG_PERM:.*]] = "tf.Const"() <{value = dense<[0, 2, 3, 1]> : tensor<4xi64>}>
  // CHECK: %[[ARG_TRANSPOSE:[0-9]*]] = "tf.Transpose"(%arg0, %[[ARG_PERM]])

  // CHECK: %[[CONV2D:[0-9]*]] = "tf.Conv2D"(%[[ARG_TRANSPOSE]], %arg1)
  // CHECK-SAME: data_format = "NHWC"
  // CHECK-SAME: dilations = [1, 3, 4, 2]
  // CHECK-SAME: explicit_paddings = [1, 2, 5, 6, 7, 8, 3, 4]
  // CHECK-SAME: padding = "EXPLICIT"

      type: DT_INT32
    }
  }
  attr {
    key: "keep_dims"
    value {
      b: false
    }
  }
}
node {
  name: "conv1_1/Conv2D"
  op: "Conv2D"
  input: "conv1_pad_1/Pad"
  input: "conv1_1/Conv2D/Cast"
  device: "/job:localhost/replica:0/task:0/device:GPU:0"
  attr {
    key: "T"
    value {
      type: DT_HALF
    }
  }
  attr {

// CHECK-SAME:      [[ARG:%.+]]: tensor<2x12x21x7xf16>
// CHECK:           [[CONV32:%.+]] = mhlo.convert %arg0 : (tensor<2x12x21x7xf16>) -> tensor<2x12x21x7xf32>
// CHECK:           [[ZERO:%.+]] = mhlo.constant dense<0.000000e+00> : tensor<f32>
// CHECK:           [[DIVIDEND:%.+]] = "mhlo.reduce_window"([[CONV32]], [[ZERO]])
// CHECK-SAME:        window_dimensions = dense<[1, 2, 2, 1]>

// CHECK: %[[CONV2D:.+]] = "tfl.conv_2d"(%[[ARG]], %[[QCONST_0]], %[[QCONST_1]]) <{dilation_h_factor = 1 : i32, dilation_w_factor = 1 : i32, fused_activation_function = "NONE", padding = "SAME", stride_h = 1 : i32, stride_w = 1 : i32}> : (tensor<1x3x3x4x!quant.uniform<i8:f32,...

// CHECK: %[[cast:.*]] = "tf.Cast"(%[[cst]]) <{Truncate = false}> : (tensor<2x3x3x2xbf16>) -> tensor<2x3x3x2xf32>
// CHECK: %[[conv:.*]] = "tf.Conv2D"(%arg0, %[[cast]])
// CHECK: %[[identity:.*]] = "tf.IdentityN"(%[[conv]]) {device = ""} : (tensor<1x3x2x2xf32>) -> tensor<1x3x2x2xf32>
// CHECK: return %[[identity]] : tensor<1x3x2x2xf32>

// -----

func.func @conv_with_non_constant_filter(%arg0: tensor<1x3x4x3xf32>, %arg1: tensor<2x3x3x2xf32>) -> tensor<*xf32> {
  %cst = "tf.Const"() {value = dense<0.000000e+00> : tensor<2xf32>} : () -> tensor<2xf32>
  %0 = "tf.Conv2D"(%arg0, %arg1) {data_format = "NHWC", dilations = [1, 1, 2, 1], explicit_paddings = [], padding = "SAME", strides = [1, 1, 2, 1], use_cudnn_on_gpu = true} : (tensor<1x3x4x3xf32>, tensor<2x3x3x2xf32>) -> tensor<*xf32>

// CHECK: return %0 : tensor<*xf32>
}

// -----

module {
  func.func @conv2d(%arg0: tensor<1x3x4x512xf32>) -> (tensor<*xf32>) {
    %cst_0 = "tf.Const"() {value = dense<0.000000e+00> : tensor<2xf32>} : () -> tensor<2xf32>

      %outputs_2, %control_3 = tf_executor.island(%control_1) wraps "tf.Const"() {value = dense<0.000000e+00> : tensor<5x5x32x16xf32>} : () -> tensor<5x5x32x16xf32>
      %outputs_4, %control_5 = tf_executor.island wraps "tf.Conv2D"(%outputs, %outputs_2) {data_format = "NHWC", device = "", dilations = [1, 1, 1, 1], explicit_paddings = [], padding = "VALID", strides = [1, 1, 1, 1], use_cudnn_on_gpu = true} : (tensor<*xf32>, tensor<5x5x32x16xf32>) -> tensor<*xf32>