('none', None, False, False, quant_opts_pb2.TF, False, 'SAME'),
      ('relu', nn_ops.relu, False, False, quant_opts_pb2.TF, False, 'SAME'),
      ('relu6', nn_ops.relu6, False, False, quant_opts_pb2.TF, False, 'SAME'),
      ('with_bias', None, True, False, quant_opts_pb2.TF, False, 'SAME'),
      (
          'with_bias_and_relu',
          nn_ops.relu,
          True,
          False,
          quant_opts_pb2.TF,

     (HasRankAtMost<4> $a),
     (HasRankAtMost<4> $b)]>;
}

// We can eliminate Relu from Relu(SquaredDifference(x, y)),
// since the result of SquaredDifference is always non-negative.
// TFLite interpreter doesn't support Relu+int32 for now. So the test cases
// are failing without the following pattern to optimize Relu away fixes
// the problem.
def OptimizeReluSquaredDifference : Pat<

// 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

// Fusing:  %[[relu:[0-9].*]] = "tfl.relu"(%arg0) : (tensor<1xf32>) -> tensor<1xf32>
// Fusing:  %[[add2:[0-9].*]] = tfl.add %[[relu]], %[[add1]] {fused_activation_function = "RELU6"} : tensor<1xf32>
// Fusing:  %[[add3:[0-9].*]] = tfl.add %[[add2]], %[[relu]] {fused_activation_function = "RELU6"} : tensor<1xf32>
// Fusing:  return

// 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>;

def TFL_AFAttr : TFL_AnyStrAttrOf<[

  %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> {

  %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>>

    %14 = "tfl.relu"(%10#1) : (tensor<4x2xf32>) -> tensor<4x2xf32>
    %15 = "tfl.logistic"(%10#0) : (tensor<4x2xf32>) -> tensor<4x2xf32>
    %16 = tfl.mul %15, %14 {fused_activation_function = "NONE"} : tensor<4x2xf32>
    %17 = tfl.add %13, %16 {fused_activation_function = "NONE"} : tensor<4x2xf32>
    %18 = "tfl.relu"(%17) : (tensor<4x2xf32>) -> tensor<4x2xf32>

    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());

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