%1 = "tf.Sub"(%0, %cst_0) {data_format = "NHWC"} : (tensor<1x3x2x2xf32>, tensor<1x1x1x2xf32>) -> tensor<1x3x2x2xf32>
  %2 = "tf.Mul"(%1, %cst_1) : (tensor<1x3x2x2xf32>, tensor<1x1x1x2xf32>) -> tensor<1x3x2x2xf32>
  %3 = "tf.AddV2"(%2, %cst_2) : (tensor<1x3x2x2xf32>, tensor<1x1x1x2xf32>) -> tensor<1x3x2x2xf32>
  func.return %3 : tensor<1x3x2x2xf32>
}

    %6 = "quantfork.qcast"(%5) : (tensor<1x3x2x2xf32>) -> tensor<1x3x2x2x!quant.uniform<i8:f32, 0.0054901962186775953:-19>>
    %7 = "quantfork.dcast"(%6) : (tensor<1x3x2x2x!quant.uniform<i8:f32, 0.0054901962186775953:-19>>) -> tensor<1x3x2x2xf32>
    return %7 : tensor<1x3x2x2xf32>
  }

  %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>

    : (tensor<?x3x2x1xf32>, tensor<2x1x1x1xf32>) -> tensor<?x2x2x1xf32>
  %1 = chlo.broadcast_add %0, %zp_offset : (
      tensor<?x2x2x1xf32>, tensor<?x2x2x1xf32>) -> tensor<?x2x2x1xf32>
  %2 = chlo.broadcast_add %1, %bias : (
      tensor<?x2x2x1xf32>, tensor<1xf32>) ->tensor<?x2x2x1xf32>
  return %2 : tensor<?x2x2x1xf32>
}

// -----

// CHECK-LABEL: func @dot_general_add_add

  %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>

// CHECK{LITERAL}: %cst_1 = arith.constant dense<[[[[1.000000e+00], [1.250000e+00]]]]> : tensor<1x1x2x1xf32>
// CHECK: %3 = tfl.mul %2, %cst_1 {fused_activation_function = "NONE"} : tensor<1x1x2x1xf32>
// CHECK: %cst_2 = arith.constant dense<[0, 3, 1, 2]> : tensor<4xi32>
// CHECK: %4 = "tfl.transpose"(%3, %cst_2) : (tensor<1x1x2x1xf32>, tensor<4xi32>) -> tensor<1x1x1x2xf32>

    has_token_input_output = false, module = "", platforms = [],
    version = 5 : i64
  } : (tensor<1x3x2x3xf32>, tensor<2x3x3x2xf32>) -> tensor<1x2x2x2xf32>
  return %0 : tensor<1x2x2x2xf32>
}

// CHECK-LABEL: func.func @qdq_for_conv_weight_empty
// CHECK-SAME: (%[[ARG_0:.+]]: tensor<1x3x2x3xf32>) -> tensor<1x2x2x2xf32>

  %w = arith.constant dense<127.0> : tensor<3x3x3x3xf32>
  %b = arith.constant dense<0.0> : tensor<3xf32>
  %conv = "tfl.conv_2d"(%arg0, %w, %b) {dilation_h_factor = 1 : i32, dilation_w_factor = 1 : i32, fused_activation_function = "RELU", padding = "SAME", stride_h = 1 : i32, stride_w = 1 : i32} : (tensor<1x5x5x3xf32>, tensor<3x3x3x3xf32>, tensor<3xf32>) -> tensor<1x5x5x3xf32>
  func.return %conv : tensor<1x5x5x3xf32>

        %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>

  %3 = stablehlo.add %2, %1 : tensor<1x4x5x5xf32>
  return %3 : tensor<1x4x5x5xf32>
}
// CHECK-DAG: %[[WEIGHT_CONST:.+]] = stablehlo.constant {{.*}} : tensor<3x3x2x4xf32>
// CHECK-DAG: %[[BIAS_CONST:.+]] = stablehlo.constant {{.*}} : tensor<1x5x5x4xf32>
// CHECK-DAG: %[[TRANSPOSE_0:.+]] = stablehlo.transpose %[[ARG]], dims = [0, 2, 3, 1] : (tensor<1x2x5x5xf32>) -> tensor<1x5x5x2xf32>