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

// -----

module {
  func.func @conv2d(%arg0: tensor<1x3x4x3xf32>) -> (tensor<*xf32>) {
    %cst_0 = "tf.Const"() {value = dense<0.000000e+00> : tensor<2xf32>} : () -> tensor<2xf32>

  %cst2 = arith.constant dense<[1.0, 2.0]> : tensor<2xf32>
  %0 = "tf.Conv2D"(%arg0, %cst0) {T = "tfdtype$DT_FLOAT", data_format = "NHWC", dilations = [1, 2, 3, 1], padding = "SAME", strides = [1, 4, 5, 1]} : (tensor<1x112x112x3xf32>, tensor<1x3x3x2xf32>) -> tensor<1x28x23x2xf32>

def MultiplyFakeQuantValue : NativeCodeCall<
  "MultiplyFakeQuantValue($_builder, $_loc, $0...)">;

// Convert AddV2Op following an AffineOp to BiasAddOp.
// For Conv3D, even though the Conv3D op has "NDHWC" data format, the BiasAdd
// will still has the data format of "NHWC".
def ConvertAddToBiasAdd : Pat<
  (TF_AddV2Op
    (SupportedAffineOpMatcher $conv_out, $input, $weight),

  %fq = "tf.FakeQuantWithMinMaxVars"(%in, %mini, %maxi) {num_bits = 5, narrow_range = false} : (tensor<3x3x3x16xf32>, tensor<f32>, tensor<f32>) -> 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>

  %4 = "tf.Transpose"(%3, %cst_1) {_tpu_replicate = "cluster", device = ""} : (tensor<1x3x3x4xbf16>, tensor<4xi32>) -> tensor<1x3x4x3xbf16>

  %fq = "tf.FakeQuantWithMinMaxVars"(%in, %mini, %maxi) {num_bits = 3, narrow_range = false} : (tensor<3x3x3x16xf32>, tensor<f32>, tensor<f32>) -> 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>

  %2 = "tfl.pseudo_qconst"() {qtype = tensor<32x!quant.uniform<i32:f32, 1.7052092479439231E-4>>, value = dense<0> : tensor<32xi32>} : () -> tensor<32x!quant.uniform<i32:f32, 1.7052092479439231E-4>>

func.func @not_lift_float_conv(%arg0: tensor<1x3x4x512xf32>) -> (tensor<*xf32>) {
  %cst = "tf.Const"() {value = dense<3.000000e+00> : tensor<2x3x512x512xf32>} : () -> tensor<2x3x512x512xf32>
  %0 = "tf.Conv2D"(%arg0, %cst) {
    data_format = "NHWC", device = "", dilations = [1, 1, 1, 1], explicit_paddings = [],
    padding = "SAME", strides = [1, 1, 2, 1], use_cudnn_on_gpu = true

  %6 = "tf.Cast"(%5) {Truncate = false} : (tensor<1x100x100x1xi8>) -> tensor<1x100x100x1xf32>

  %7 = "tf.Sub"(%6, %cst_4) : (tensor<1x100x100x1xf32>, tensor<f32>) -> tensor<1x100x100x1xf32>
  %8 = "tf.Conv2D"(%7, %cst_3) {dilations = [1, 1, 1, 1], padding = "VALID", strides = [1, 1, 1, 1]} : (tensor<1x100x100x1xf32>, tensor<3x3x1x1xf32>) -> tensor<1x98x98x1xf32>

      if (!input_type || !input_type.hasStaticShape()) {
        return false;
      }
      total_count += input_type.getNumElements();
    }
    *count = total_count;
    return true;
  }

  // For conv2d/depthwise_conv/fully_connected ops.
  // This algorithm actually comes from TOCO tooling_util.cc
  static bool GetArithmeticCountForConvAndFullyconnectedOp(mlir::Operation* op,