// CHECK: func private @quantized_conv2d_with_relu6_fn
// CHECK: func private @quantized_depthwise_conv2d_with_bias_and_relu_float_output_fn
// CHECK-SAME: tf_quant.quantized_ops = ["DepthwiseConv2D", "BiasAdd", "Relu"]
// CHECK: func private @quantized_matmul_with_bias_fn
// CHECK: func private @quantized_matmul_with_bias_and_relu_fn
// CHECK: func private @quantized_matmul_with_bias_and_relu6_fn
// CHECK: func private @quantized_matmul_fn

    t1 = constant_op.constant([[1.0, 2.0], [3.0, 4.0]])
    t2 = constant_op.constant([[1.0, 2.0], [3.0, 4.0]])
    t3 = constant_op.constant([[-10.0, -10.0], [-10.0, -10.0]])
    sq = biased_dense(t1, t2, t3, act='relu')
    self.assertAllEqual(sq.numpy().reshape(-1), [0, 0, 5, 12])

  def testWithKnownKernel(self):

    @def_function.function
    def biasd_dense_elu(x, y, z):
      dot = gen_composite_ops.my_biased_dense(x, y, z)

    t1 = constant_op.constant([[1.0, 2.0], [3.0, 4.0]])
    t2 = constant_op.constant([[1.0, 2.0], [3.0, 4.0]])
    t3 = constant_op.constant([[-10.0, -10.0], [-10.0, -10.0]])
    sq = gen_composite_ops.my_biased_dense(t1, t2, t3, act='relu')
    self.assertAllEqual(sq.numpy().reshape(-1), [0, 0, 5, 12])

  def testWithKnownKernel(self):

    def biasd_dense_elu(x, y, z):
      dot = gen_composite_ops.my_biased_dense(x, y, z)

  // tf.MyCustomOp is the result of conversion to a Custom op
  %2 = "tf.MyCustomOp"(%1, %0) {fused_activation_function = "RELU", int_attr = 2 : i32}  : (tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32> loc("MyCustomOp")
  %3 = "tfl.exp"(%2)  : (tensor<4xf32>) -> tensor<4xf32> loc("exp")
  func.return %3 : tensor<4xf32>

  %cst = arith.constant dense<0.0> : tensor<2x3xbf16>
  %0 = "tfl.maximum"(%arg0, %cst) : (tensor<2x3xbf16>, tensor<2x3xbf16>) -> tensor<2x3xbf16>
  func.return %0 : tensor<2x3xbf16>

  // CHECK: %[[RESULT:.*]] = "tfl.relu"(%arg0)
  // CHECK: return %[[RESULT]]
}

// CHECK-LABEL: fuseScalarAddIntoConv2dBf16
func.func @fuseScalarAddIntoConv2dBf16(%arg0: tensor<256x32x32x3xbf16>, %arg1: tensor<16x3x3x3xbf16>) -> tensor<256x8x7x16xbf16> {

}

// Allowed activation function cases
// These should match the ActivationFunctionType enum in TFLite schema.
def TFL_AFEnum_None  : I32EnumAttrCase<"NONE", 0>;
def TFL_AFEnum_Relu  : I32EnumAttrCase<"RELU", 1>;
def TFL_AFEnum_Relu1 : I32EnumAttrCase<"RELU_N1_TO_1", 2>;
def TFL_AFEnum_Relu6 : I32EnumAttrCase<"RELU6", 3>;
def TFL_AFEnum_Tanh  : I32EnumAttrCase<"TANH", 4>;
def TFL_AFEnum_Sign  : I32EnumAttrCase<"SIGN_BIT", 5>;

// tfl.Abs / tfl.Average_pool_2d / tfl.Cos / tfl.div / tfl.exp / tfl.hardswish /
// tfl.log / tfl.logistic / tfl.max_pool_2d / tfl.mirror_pad / tfl.maximum /
// tfl.custom / tfl.mean / tfl.minimum / tfl.pad / tfl.pow / tfl.prelu /
// tfl.relu / tfl.relu6 / tfl.rsqrt / tfl.sin / tfl.slice / tfl.softmax /
// tfl.space_to_depth / tfl.sqrt / tfl.square / tfl.squared_difference /
// tfl.strided_slice / tfl.tanh / tfl.transpose / tfl.transpose_conv

    func_op.eraseResult(0);
    func_op.insertResult(0, new_call_op.getResult(0).getType(),
                         /*resultAttrs=*/nullptr);

    // Modify the quantized fused function to do dequantize+relu(6).
    rewriter.setInsertionPoint(req_op);
    Value new_result = rewriter.create<mlir::stablehlo::UniformDequantizeOp>(
        req_op.getLoc(), func_op.getResultTypes()[0], req_op.getOperand());

  %3 = "tf.BiasAdd"(%2, %cst_0) {data_format = "NHWC", device = ""} : (tensor<1x3x2x2xf32>, tensor<2xf32>) -> tensor<1x3x2x2xf32>
  %4 = "tf.Relu"(%3) {device = ""} : (tensor<1x3x2x2xf32>) -> tensor<1x3x2x2xf32>
  %5 = "quantfork.qcast"(%4) : (tensor<1x3x2x2xf32>) -> tensor<1x3x2x2x!quant.uniform<i8:f32, 0.0027450981093387976:-19>>

      %outputs_12, %control_13 = tf_executor.island wraps "tf.Relu"(%outputs_10) {device = ""} : (tensor<16x112x112x?xf32>) -> tensor<16x112x112x?xf32>