func.func @testConv(%arg0: tensor<256x32x32x3xf32>, %arg1: tensor<16x3x3x3xf32>, %arg2: tensor<16xf32>) -> tensor<256x30x30x16xf32> {
  // CHECK: tac.device = "GPU", tac.inference_type = "FLOAT"
  %0 = "tfl.conv_2d"(%arg0, %arg1, %arg2) {dilation_h_factor = 1 : i32, dilation_w_factor = 1 : i32, fused_activation_function = "NONE", padding = "VALID", stride_h = 1 : i32, stride_w = 1 : i32} : (tensor<256x32x32x3xf32>, tensor<16x3x3x3xf32>, tensor<16xf32>) -> tensor<256x30x30x16xf32>

      } : (tensor<1x3x4x3xf32>, tensor<2x3x3x2xf32>) -> tensor<1x3x4x2xf32>
    return %0 : tensor<1x3x4x2xf32>
  }

  // CHECK: func.func private @qdq_for_conv_weight_per_channel(%[[ARG0:.+]]: tensor<1x3x4x3xf32>)
  // CHECK: %[[CST:.+]] = "tf.Const"() <{value = dense<3.000000e-01> : tensor<2x3x3x2xf32>}> : () -> tensor<2x3x3x2xf32>

  %cst_1 = "tf.Const"() {value = dense<[0.1, 0.2]> : tensor<2xf32>} : () -> tensor<2xf32>
  %cst_2 = "tf.Const"() {value = dense<[0.3, 0.4]> : tensor<2xf32>} : () -> tensor<2xf32>

        %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>

  func.func private @composite_conv2d_with_bias_and_relu6_fn_10(%arg0: tensor<1x3x2x3xf32>, %arg1: tensor<2x3x3x2xf32>, %arg2: tensor<2xf32>) -> tensor<1x2x2x2xf32> attributes {tf.tf_quant.composite_function} {
    %0 = "quantfork.stats"(%arg1) {layerStats = dense<[-3.54062747, 0.54742622]> : tensor<2xf32>} : (tensor<2x3x3x2xf32>) -> tensor<2x3x3x2xf32>

    return %1 : tensor<1x3x4x2xf32>
  }

  func.func private @composite_conv_fn(%arg0: tensor<1x3x4x3xf32>, %arg1: tensor<2x3x3x2xf32>) -> tensor<1x3x4x2xf32> attributes {_from_xla_call_module} {

    %12 = "quantfork.qcast"(%11) {volatile} : (tensor<2x3x1x3xf32>) -> tensor<2x3x1x3x!quant.uniform<i8:f32, 3.000000e-01:1>>
    %13 = "quantfork.dcast"(%12) : (tensor<2x3x1x3x!quant.uniform<i8:f32, 3.000000e-01:1>>) -> tensor<2x3x1x3xf32>
    return %13 : tensor<2x3x1x3xf32>
  }

  // CHECK: quantized_dot_general_fn_1

    %0 = "tf.PartitionedCall"(%arg0, %cst_1) {_tfl_quant_trait = "fully_quantizable", config = "", config_proto = "", executor_type = "", f = @composite_depthwise_conv2d_fn} : (tensor<1x3x4x3xf32>, tensor<2x3x3x2xf32>) -> tensor<*xf32>

      stablehlo.return %2 : tensor<f32>
  }) {padding = dense<[[0, 0], [1, 1], [1, 1], [0, 0]]> : tensor<4x2xi64>, window_dimensions = array<i64: 1, 3, 3, 1>} : (tensor<2x3x1x3xf32>, tensor<f32>) -> tensor<2x3x1x3xf32>
  return %1 : tensor<2x3x1x3xf32>
}

// -----

// CHECK-LABEL: @bitcast_convert_i32_f32(%arg0: tensor<1x256128xi32>) -> tensor<1x256128xbf16>

  return %4 : tensor<1x3x2x2xf32>
}

// CHECK: func @cast_bf16_conv_with_bias_to_fp32
// CHECK-DAG: %[[cst:.*]] = "tf.Const"() <{value = dense<1.000000e+00> : tensor<2x3x3x2xf32>}> : () -> tensor<2x3x3x2xf32>
// CHECK-DAG: %[[cst_0:.*]] = "tf.Const"() <{value = dense<1.000000e+00> : tensor<2xf32>}> : () -> tensor<2xf32>
// CHECK: %[[conv:.*]] = "tf.Conv2D"(%arg0, %[[cst]])