// Casts the `target` type to a quantized type by using the quantization
// parameters from the type in the `source` type attribute.
// Examples:
//   f32 -> !quant.uniform<i8:f32, 1.0>
//   tensor<4xf32> -> tensor<4x!quant.uniform<i8:f32, 1.0>>
// The result is wrapped by a type attribute. Returns nullptr if the cast
// isn't valid.
//
// `axis` is to specify the quantization dimension in the `target` and only

// MinElement-LABEL: QuantizeCustomOp
func.func @QuantizeCustomOp(%arg0: tensor<1x1x1x1xf32>) -> (tensor<*xf32>, tensor<*xf32>, tensor<*xf32>) attributes {tf.entry_function = {inputs = "input", outputs = "custom_op"}} {
  %0 = "quantfork.stats"(%arg0) {layerStats = dense<[0.000000e+00, 2.550000e+02]> : tensor<2xf32>} : (tensor<1x1x1x1xf32>) -> tensor<1x1x1x1xf32>
  %w_1 = arith.constant dense<127.0> : tensor<4096x1x1x1xf32>

  %cst = arith.constant dense<1> : tensor<4xi32>
  %cst_0 = arith.constant dense<[0, 3, 0, 1]> : tensor<4xi32>
  %cst_1 = arith.constant dense<0> : tensor<4xi32>
  %cst_2 = arith.constant dense<0.000000e+00> : tensor<1xf32>
  %0 = "tfl.quantize"(%cst_2) {qtype = tensor<1x!quant.uniform<i32:f32, 1.0>>, volatile} : (tensor<1xf32>) -> tensor<1x!quant.uniform<i32:f32, 1.0>>

module {
  func.func @simpleTest(%arg0: tensor<1xf32>, %arg1: tensor<1xf32>, %arg2: tensor<1xf32>, %arg3: tensor<1xf32>) -> tensor<2x1xf32> {
    %0 = func.call @func_0_GPU_FLOAT(%arg0, %arg1, %arg2) {tac.interface_name = "func_0"} : (tensor<1xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>
    %1 = func.call @func_1_GPU_FLOAT(%arg0, %arg3) {tac.interface_name = "func_1"} : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>

    return
  }
  // CHECK-LABEL: func private @foo.impl
  func.func private @foo.impl(%arg0: tensor<1xf32>, %arg1: tensor<2xf32>) -> (tensor<2xf32>, tensor<1xf32>) {
    return %arg1, %arg0 : tensor<2xf32>, tensor<1xf32>
  }

  func.return %add : tensor<*xf32>
}
// CHECK: func @decompose_batch_norm
// CHECK-DAG: %[[CONST:.*]] = "tf.Const"() <{value = dense<2.49743462E-5> : tensor<2xf32>}> : () -> tensor<2xf32>

module {
func.func @main(%arg0: tensor<1xf32>, %arg1: tensor<1xf32>, %arg2: tensor<1xf32>, %arg3: tensor<1xf32>) -> tensor<2x1xf32> attributes {tf.entry_function = {inputs = "input0,input1,input2,input3", outputs = "output"}} {
  %0 = "tfl.add"(%arg0, %arg1) {fused_activation_function = "RELU6"} : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>
  %1 = "tfl.mul"(%0, %arg2) {fused_activation_function = "RELU6"} : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>

    %1 = "tfl.abs"(%arg2) {fused_activation_function = "NONE"} : (tensor<5xf32>) -> tensor<5xf32>
    %2 = stablehlo.add %1, %arg2 : tensor<5xf32>
    %3 = "tfl.abs"(%2) {fused_activation_function = "NONE"} : (tensor<5xf32>) -> tensor<5xf32>
    "stablehlo.return"(%3) : (tensor<5xf32>) -> ()
  }) {dimensions = array<i64: 0>} : (tensor<7x5xf32>, tensor<5xf32>) -> tensor<5xf32>
  func.return %0: tensor<5xf32>
}

// -----

///      : (tensor<f32>, tensor<f32>, tensor<f32>) -> tensor<5xf32>
///
/// Output would be:
///   %iota = "mhlo.iota"() {iota_dimension = 0 : i64} : () -> tensor<5xf32>
///   %scaled = "mhlo.multiply"(%iota, %delta)
///       {broadcast_dimensions = dense<[]> : tensor<0xi64>} :
///       (tensor<5xf32>, tensor<f32>) -> tensor<5xf32>
///   %result = "mhlo.add"(%scaled, %offset)

    %cst = "tf.Const"() {value = dense<3.00000000e-1> : tensor<2x3xf32>} : () -> tensor<2x3xf32>
    %cst_0 = "tf.Const"() {value = dense<4.00000000e-1> : tensor<3xf32>} : () -> tensor<3xf32>
    %0 = "quantfork.stats"(%arg0) {layerStats = dense<[6.00000000e-6, 9.00000000e-1]> : tensor<2xf32>} : (tensor<?x2xf32>) -> tensor<?x2xf32>