%12 = "tf.Identity"(%11) {device = ""} : (tensor<2x3x4x5xi8>) -> tensor<2x3x4x5xi8>
    %13 = "tf.Cast"(%12) {Truncate = false, device = ""} : (tensor<2x3x4x5xi8>) -> tensor<2x3x4x5xi32>
    %14 = "tf.Sub"(%13, %cst_5) {device = ""} : (tensor<2x3x4x5xi32>, tensor<i32>) -> tensor<2x3x4x5xi32>
    %15 = "tf.Identity"(%5) {device = ""} : (tensor<2x6x4x5xi8>) -> tensor<2x6x4x5xi8>

// PerChannel-DAG: %[[q_w1:.*]] = "tf.Const"() <{value = dense<{{[0-9]+}}> : tensor<2x3x3x1xi8>}> : () -> tensor<2x3x3x1xi8>
// PerChannel-DAG: %[[q_w2:.*]] = "tf.Const"() <{value = dense<{{[0-9]+}}> : tensor<2x3x3x2xi8>}> : () -> tensor<2x3x3x2xi8>
// PerChannel-DAG: %[[scale1:.*]] = "tf.Const"() <{value = dense<{{[0-9\.Ee\+\-]+}}> : tensor<3xf32>}> : () -> tensor<3xf32>

// CHECK-NEXT: %[[V1:.*]] = "tf.PartitionedCall"(%[[ARG0]], %[[CST]], %[[CST_0]]) <{config = "", config_proto = "", executor_type = "", f = @quantize_i8}> : (tensor<1x3x4x3xf32>, tensor<f32>, tensor<i32>) -> tensor<1x3x4x3xi8>

  %0 = stablehlo.constant() {value = dense<1> : tensor<1x2x4x5xi8>} : () -> tensor<1x2x4x5x!quant.uniform<i8:f32, 1.000000e+0>>
  return %0 : tensor<1x2x4x5x!quant.uniform<i8:f32, 1.000000e+0>>
}

// CHECK: %[[QCONST:.+]] = "tfl.pseudo_qconst"() <{qtype = tensor<1x2x4x5x!quant.uniform<i8:f32, 1.000000e+00>>, value = dense<1> : tensor<1x2x4x5xi8>}>
// CHECK-SAME: () -> tensor<1x2x4x5x!quant.uniform<i8:f32, 1.000000e+00>>

    return %0 : tensor<1x3x4x2xf32>
  }
}

// CHECK-LABEL: quantize_conv_per_tensor
// CHECK-SAME: %[[ARG0:.+]]: tensor<1x3x4x3xf32>
// CHECK: %[[CST:.+]] = stablehlo.constant() <{value = dense<127> : tensor<2x3x3x2xi8>}> : () -> tensor<2x3x3x2x!quant.uniform<i8<-127:127>:f32, 0.0023622048182750312>>

  return %2 : tensor<?x2x2x1xi32>
}

// -----

// CHECK-LABEL: func @convolution_add_add_static
func.func @convolution_add_add_static(
    %lhs: tensor<2x3x2x1xi8>, %rhs: tensor<2x1x1x1xi8>,
    %zp_offset: tensor<2x2x2x1xi32>, %bias: tensor<1xi32>
  ) -> tensor<2x2x2x1xi32> {
  // CHECK-DAG: %[[conv:.*]] = mhlo.convolution

// CHECK: func.func private @quantize_conv_fn(%[[ARG_0:.+]]: tensor<1x3x4x3xf32>) -> tensor<1x3x4x2xf32> attributes {tf._original_func_name = "main_0"}
// CHECK: %[[CONST_0:.+]] = stablehlo.constant() <{value = dense<{{.*}}> : tensor<2x3x3x2xi8>}> : () -> tensor<2x3x3x2x!quant.uniform<i8<-127:127>:f32:3, {{.*}}>
// CHECK: %[[UNIFORM_QUANTIZE_0:.+]] = stablehlo.uniform_quantize %[[ARG_0]] : (tensor<1x3x4x3xf32>) -> tensor<1x3x4x3x!quant.uniform<i8:f32, {{.*}}>>

    %1 = stablehlo.constant dense<1.000000e+03> : tensor<1x1x1x1xf32>  // Input inverse scale.
    %2 = stablehlo.constant dense<-128> : tensor<1x1x1x1xi8>  // Input zero point.
    %3 = stablehlo.constant dense<1> : tensor<3x3x4x4xi8>  // Quantized filter tensor.
    %4 = stablehlo.constant dense<3.000000e+03> : tensor<1x1x1x4xf32>
    %5 = stablehlo.constant dense<4.000000e+03> : tensor<1x1x1x1xf32>  // Output inverse scale.

// CHECK: %[[minimum:.*]] = "tf.Minimum"(%[[maximum]]
// CHECK: %[[round:.*]] = "tf.Round"(%[[minimum]]
// CHECK: %[[quant:.*]] = "tf.Cast"(%[[round]]) : (tensor<1x3x4x3xf32>) -> tensor<1x3x4x3xi8>
// CHECK: %[[pad:.*]] = "tf.PadV2"(%[[quant]]
// CHECK: %[[xlaconv:.*]] = "tf.XlaConvV2"(%[[pad]]
// CHECK: %[[sub:.*]] = "tf.Sub"(%[[xlaconv]]

  %weight_zps = "tf.Const"() { value = dense<3> : tensor<i32> } : () -> tensor<i32>

  // CHECK: %[[CONST:.*]] = mhlo.constant()
  // CHECK-SAME{LITERAL} value = dense<127> : tensor<2x3x3x2xi8>
  // CHECK-SAME: tensor<2x3x3x2x!quant.uniform<i8:f32, 1.000000e+00:3>>
  // CHECK: mhlo.convolution(%arg0, %[[CONST]])
  // CHECK-SAME{LITERAL}: dim_numbers = [b, 0, 1, f]x[0, 1, i, o]->[b, 0, 1, f]