%i8_max = "tf.Const"() {value = dense<127.0> : tensor<f32>} : () -> tensor<f32>
    %act_max =  "tf.Const"() {value = dense<6.0> : tensor<f32>} : () -> tensor<f32>
    %i8_act_max_0 = "tf.PartitionedCall"(%act_max, %out_scale, %out_zp) {
        config = "", config_proto = "", executor_type = "", f=@quantize_i8
      } : (tensor<f32>, tensor<*xf32>, tensor<*xi32>) -> tensor<*xi8>

    ^bb0(%arg23: tensor<f32>, %arg24: tensor<f32>):
      %1112 = "vhlo.add_v1" (%arg23, %arg24) : (tensor<f32>, tensor<f32>) -> tensor<f32>
      "vhlo.return_v1" (%1112) : (tensor<f32>) -> ()
    }) : (tensor<1x160x1xf32>, tensor<f32>) -> tensor<1x160x1xf32>
  return %0 : tensor<1x160x1xf32>
}

//CHECK:func.func private @reduce_window(%arg0: tensor<1x160x1xf32>, %arg1: tensor<f32>) -> tensor<1x160x1xf32> {

// PerTensor-DAG: %[[zp:.*]] = "tf.Const"() <{value = dense<0> : tensor<i32>}> : () -> tensor<i32>
// PerTensor: %[[out:.*]] = "tf.PartitionedCall"(%arg0, %[[q_w]], %[[scale]], %[[zp]]) <{config = "", config_proto = "", executor_type = "",
// PerTensor-SAME: f = @quantized_matmul_fn_0}> : (tensor<2x12xf32>, tensor<12x2xi8>, tensor<f32>, tensor<i32>) -> tensor<*xf32>

		case tf32*100 + ti32:
			f32 = float32(v); i32 = int32(f32); w = big(i32)
		case tf32*100 + tu32:
			f32 = float32(v); u32 = uint32(f32); w = big(u32)
		case tf32*100 + ti64:
			f32 = float32(v); i64 = int64(f32); w = big(i64)
		case tf32*100 + tu64:
			f32 = float32(v); u64 = uint64(f32); w = big(u64)
		case tf32*100 + tf32:
			f32 = float32(v); f32 = float32(f32); w = big(f32)
		case tf32*100 + tf64:

    %input_zps = "tf.Const"() { value = dense<4> : tensor<i32> } : () -> tensor<i32>
    %bias_scales = "tf.Const"() { value = dense<2.0> : tensor<f32> } : () -> tensor<f32>
    %bias_zps = "tf.Const"() { value = dense<4> : tensor<i32> } : () -> tensor<i32>
    %output_scales = "tf.Const"() { value = dense<2.0> : tensor<f32> } : () -> tensor<f32>

  %1 = "tf.Relu"(%0) {__op_key = 1: i32, device = "/device:CPU:0"} : (tensor<f32>) -> (tensor<f32>)
  %2 = "tf.TPUCompileMlirAndExecute"(%1, %0) {metadata = "metadata", mlir_module = "mlir_module", operandSegmentSizes = array<i32: 2, 0>, producer_name = "producer_name"} : (tensor<f32>, tensor<f32>) -> tensor<i32>
  func.return %2 : tensor<i32>
}

// -----

// CHECK-DAG: %[[CONST_0:.*]] = stablehlo.constant dense<[{{.*}}]> : tensor<2xf32>
// CHECK-DAG: %[[CONST_1:.*]] = stablehlo.constant dense<6.000000e+00> : tensor<f32>
// CHECK-DAG: %[[CONST_2:.*]] = stablehlo.constant dense<0.000000e+00> : tensor<f32>
// CHECK-DAG: %[[CONST_3:.*]] = stablehlo.constant dense<[{{.*}}]> : tensor<2x3x3x2xf32>

  %1 = "tfl.quantize"(%arg0) {qtype = tensor<2x2x!quant.uniform<u8:f32, 1.0:128>>} : (tensor<2x2xf32>) -> tensor<2x2x!quant.uniform<u8:f32, 1.0:128>>
  %2 = "tfl.dequantize"(%1) : (tensor<2x2x!quant.uniform<u8:f32, 1.0:128>>) -> tensor<2x2xf32>
  func.return %2 : tensor<2x2xf32>

// CHECK-NEXT: %[[q:.*]] = "tfl.quantize"(%arg0) <{qtype = tensor<2x2x!quant.uniform<i8:f32, 1.000000e+00>>}> : (tensor<2x2xf32>)

  %5 = "tfl.reshape"(%4, %0) : (tensor<1x112x112x32x!quant.uniform<u8:f32, 0.023528476789885875>>, tensor<2xi32>) -> tensor<1x401408x!quant.uniform<u8:f32, 0.023528476789885875>>
  %6 = "tfl.softmax"(%5) {beta = 1.000000e+00 : f32} : (tensor<1x401408x!quant.uniform<u8:f32, 0.023528476789885875>>) -> tensor<1x401408x!quant.uniform<u8:f32, 3.906250e-03>>

    "tfl.yield"(%barg0, %cst) : (tensor<f32>, tensor<f32>) -> ()
  }) : (tensor<f32>, tensor<f32>) -> (tensor<f32>, tensor<f32>)
  func.return %0#1 : tensor<f32>
}

// -----

// Test case to test While op with resources that are not read-only variables.
// Do not remove resource arugments if they are not read-only variables to keep
// the graph's control dependency.
// CHECK-LABEL: WhileWithNonReadOnlyVariableResources