}

  // For weight-only
  func.func @dequantize_i8(%input : tensor<*xi8>, %scale : tensor<*xf32>, %zp : tensor<*xi32>) -> tensor<*xf32> {
    // Use identity op to avoid the weight being constant-folded.
    %identity = "tf.Identity"(%input) : (tensor<*xi8>) -> tensor<*xi8>
    %input_i32 = "tf.Cast"(%identity) : (tensor<*xi8>) -> tensor<*xi32>

    }
  }
}
versions {
  producer: 216
}

# NONE: func @main(%arg0: tensor<i1>, %arg1: tensor<1xi32>) -> (tensor<i1>, tensor<1xi32>)

    %i8 = "tf.Cast"(%round) : (tensor<*xf32>) -> tensor<*xi8>
    func.return %i8 : tensor<*xi8>
  }

  func.func @dequantize_i8(%input : tensor<*xi8>, %scale : tensor<*xf32>, %zp : tensor<*xi32>) -> tensor<*xf32> {
    // Use identity op to avoid the weight being constant-folded.
    %identity = "tf.Identity"(%input) : (tensor<*xi8>) -> tensor<*xi8>
    %input_i32 = "tf.Cast"(%identity) : (tensor<*xi8>) -> tensor<*xi32>

    return %3 : tensor<1x2x2x1024xbf16>
  }

  func.func private @composite_dequantize_uniform(%arg0: tensor<*xi8>) -> tensor<*xbf16> {
    %cst = "tf.Const"() {value = dense<1.574710e-02> : tensor<bf16>} : () -> tensor<bf16>
    %0 = "tf.Cast"(%arg0) {Truncate = false} : (tensor<*xi8>) -> tensor<*xbf16>
    %1 = "tf.Mul"(%0, %cst) : (tensor<*xbf16>, tensor<bf16>) -> tensor<*xbf16>
    return %1 : tensor<*xbf16>
  }

tensor<i8>, tensor<f64>, tensor<!tf_type.resource>)

# CHECK-SUBTYPE: func @main(%arg0: tensor<10xi32>, %arg1: tensor<!tf_type.variant<tensor<10xf32>>>, %arg2: tensor<10xi8>, %arg3: tensor<10xf64>, %arg4: tensor<!tf_type.resource<tensor<10xi32>>>) -> (tensor<10xi32>, tensor<!tf_type.variant<tensor<10xf32>>>, tensor<10xi8>, tensor<10xf64>, tensor<!tf_type.resource<tensor<10xi32>>>)...

// CHECK: %[[maxpool_i8:.*]] = "tf.MaxPool"(%[[sc1]])
// CHECK-SAME: (tensor<*xi8>) -> tensor<*xi8>
// CHECK: %[[reshape_i8:.*]] = "tf.Reshape"(%[[maxpool_i8]]
// CHECK-SAME: (tensor<*xi8>, tensor<2xi32>) -> tensor<*xi8>
// CHECK: %[[sc2:.*]] = "quantfork.scast"(%[[reshape_i8]])
// CHECK: %[[dq:.*]] = "quantfork.dcast"(%[[sc2]]) : (tensor<*x!quant.uniform<i8:f32, 5.000000e-02:-10>>)

    %19 = "tf.BroadcastTo"(%4, %17) : (tensor<2x?x?xi8>, tensor<3xi64>) -> tensor<2x?x?xi8>
    %20 = "tf.BroadcastTo"(%9, %18) : (tensor<2x?x?xi8>, tensor<3xi64>) -> tensor<2x?x?xi8>
    %21 = "tf.XlaDotV2"(%19, %20) {dimension_numbers = "\22\01\00\1A\01\00\12\01\01\0A\01\02", precision_config = ""} : (tensor<2x?x?xi8>, tensor<2x?x?xi8>) -> tensor<2x?x?xi32>
    %22 = "tf.Cast"(%19) {Truncate = false} : (tensor<2x?x?xi8>) -> tensor<2x?x?xi32>

}

func.func private @testIf3Then(tensor<*xf32>) -> (tensor<*xf32>, tensor<*xi8>, tensor<*xbf16>)
func.func private @testIf3Else(tensor<*xf32>) -> (tensor<*xf32>, tensor<*xi8>, tensor<*xbf16>)

// CHECK-LABEL: func @testIf3Result(%arg0: tensor<i1>, %arg1: tensor<*xf32>)
func.func @testIf3Result(tensor<i1>, tensor<*xf32>) -> (tensor<*xf32>, tensor<*xi8>, tensor<*xbf16>) {
^bb0(%arg0: tensor<i1>, %arg1: tensor<*xf32>):

// CHECK-SAME: (tensor<1x2x2x3xi8>, tensor<2x2x3x2xi8>, tensor<2xi32>, tensor<f32>, tensor<i32>, tensor<2xf32>, tensor<2xi32>, tensor<2xf32>, tensor<2xi32>, tensor<f32>, tensor<i32>) -> tensor<*xi8>
// CHECK: %[[cast_1:.*]] = "tf.Cast"(%[[conv_quant]]) <{Truncate = false}> : (tensor<*xi8>) -> tensor<*xf32>

  func.return %11 : tensor<*xf32>
}

// CHECK: %[[maxpool_i8:.*]] = "tf.MaxPool"
// CHECK-SAME: (tensor<*xi8>) -> tensor<*xi8>
// CHECK: %[[reshape_i8:.*]] = "tf.Reshape"(%[[maxpool_i8]]
// CHECK-SAME: (tensor<*xi8>, tensor<2xi32>) -> tensor<*xi8>
// CHECK: %[[scast:.*]] = "quantfork.scast"(%[[reshape_i8]]
// CHECK: %[[matmul:.*]] = "tf.PartitionedCall"(%[[scast]]
// CHECK-SAME: f = @composite_matmul_fn_1