ones = array_ops.ones_like(out)
          out = array_ops.concat([out, ones], 0)
        elif self.same_scale_op == 'gather':
          out = array_ops.gather(out, indices=[0], axis=0)
        elif self.same_scale_op == 'max_pool':
          out = nn_ops.max_pool(out, ksize=3, strides=1, padding='SAME')
        elif self.same_scale_op == 'pad':
          paddings = array_ops.ones(

    %0:3 = "tf.Unpack"(%arg0) {axis = 0 : i64} : (tensor<3xi64>) -> (tensor<i64>, tensor<i64>, tensor<i64>)
    return
  }
})";

class LegalizationOpConfigTest : public ::testing::Test {
 public:

  func.return %1 : tensor<*xf32>

// CHECK-DAG:  [[AXIS:%.*]] = arith.constant dense<0> : tensor<i32>
// CHECK-DAG:  [[EXPAND_DIM:%[0-9]+]] = "tf.ExpandDims"(%arg1, [[AXIS]]) : (tensor<i32>, tensor<i32>) -> tensor<1xi32>
// CHECK-DAG:  [[AXIS_1:%.*]] = arith.constant dense<0> : tensor<i32>

  %1 = "quantfork.stats"(%0) {
    layerStats = dense<[-1.0, 1.0]> : tensor<2xf32>,
    axisStats = dense<[
      [-1.0, 1.0],
      [-8.0, 8.0],
      [-0.5, 0.5]
    ]> : tensor<3x2xf32>, axis = 2 : i64
  } : (tensor<8x4x3xf32>) -> tensor<8x4x3xf32>
  func.return %1 : tensor<8x4x3xf32>

// CHECK: %[[q1:.*]] = "tfl.quantize"(%arg0) <{qtype = tensor<8x4x3x!quant.uniform<i8:f32, 0.0078431372549019607:-1>>}> {volatile}

  const double clamp_max_double_;

  const uint32_t storage_bit_width_;
  const bool is_signed_;
  const llvm::APFloat::roundingMode round_mode_;
};

// An utility class to quantize an attribute by the per-axis quantization
// parameters. The size of the quantization dim in the converted elements
// attribute should match the size of of scales/zero_points vectors in the
// quantization parameters.

TFE_Op* MinOp(TFE_Context* ctx, TFE_TensorHandle* input,
              TFE_TensorHandle* axis) {
  TF_Status* status = TF_NewStatus();

  TFE_Op* op = TFE_NewOp(ctx, "Min", status);
  CHECK_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  TFE_OpAddInput(op, input, status);
  CHECK_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  TFE_OpAddInput(op, axis, status);
  CHECK_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

      %outputs_10, %control_11 = tf_executor.island wraps "tf.Pack"(%outputs_6, %outputs_8) {axis = 0 : i64, device = ""} : (tensor<0xi32>, tensor<0xi32>) -> tensor<*xi32>

			j++
			y++
		}
		j &= 7                               // octant modulo 2Pi radians (360 degrees)
		z = ((x - y*PI4A) - y*PI4B) - y*PI4C // Extended precision modular arithmetic
	}
	// reflect in x axis
	if j > 3 {
		sign = !sign
		j -= 4
	}
	zz := z * z
	if j == 1 || j == 2 {
		y = 1.0 - 0.5*zz + zz*zz*((((((_cos[0]*zz)+_cos[1])*zz+_cos[2])*zz+_cos[3])*zz+_cos[4])*zz+_cos[5])
	} else {

func.func @QuantizeConcat(tensor<1x2xf32>, tensor<1x2xf32>) -> tensor<2x2x!quant.uniform<u8:f32, 1.000000e-01:128>> {
^bb0(%arg0: tensor<1x2xf32>, %arg1: tensor<1x2xf32>):
  %0 = "tfl.concatenation"(%arg0, %arg1) {axis = 0 : i32, fused_activation_function = "NONE"} : (tensor<1x2xf32>, tensor<1x2xf32>) -> tensor<2x2xf32>

// PerTensor-LABEL: QuantizeGatherWeightOnly
func.func @QuantizeGatherWeightOnly(%arg0: tensor<3xi32>) -> tensor<3x3x3x3xf32> {
  %w = arith.constant dense<1.270000e+02> : tensor<64x3x3x3xf32>
  %emb = "tfl.gather"(%w, %arg0) {axis = 0 : i32, batch_dims = 0 : i32} : (tensor<64x3x3x3xf32>, tensor<3xi32>) -> tensor<3x3x3x3xf32>