%21 = "tf.Cast"(%20) {Truncate = false, device = ""} : (tensor<1x1xi32>) -> tensor<1x1xi64>
  %22 = "tf.Reshape"(%21, %12) {device = ""} : (tensor<1x1xi64>, tensor<1xi64>) -> tensor<1xi64>
  %23 = "tf.Reshape"(%arg0, %5) {device = ""} : (tensor<1x!tf_type.string>, tensor<1xi32>) -> tensor<1x!tf_type.string>

  %3 = "quantfork.qcast"(%2) {volatile} : (tensor<*xf32>) -> tensor<*x!quant.uniform<i8:f32, 5.000000e-02:-10>>
  %4 = "quantfork.dcast"(%3) : (tensor<*x!quant.uniform<i8:f32, 5.000000e-02:-10>>) -> tensor<*xf32>
  %5 = "tf.Reshape"(%4, %cst) {device = ""} : (tensor<*xf32>, tensor<2xi32>) -> tensor<*xf32>
  %6 = "quantfork.qcast"(%5) {volatile} : (tensor<*xf32>) -> tensor<*x!quant.uniform<i8:f32, 5.000000e-02:-10>>

//
// %iota_r1 = "mhlo.iota"(){iota_dimension = 0 : i32} : () -> tensor<44xi32>
// %iota = "mhlo.reshape"(%iota_r1): (tensor<44xi32>) -> tensor<1x1x44xi32>
//
// Where $dimensions is of size 1 and $dimensions[0] = 2.
//
// In general matches a 1-D Iota with multiple dimensions of size 1 added
// through a reshape.
bool MatchReshapedIota(DenseIntElementsAttr dimensions, Value iota) {

    %2 = "tfl.reshape"(%0, %cst) {tac.device = "GPU", tac.inference_type = "FLOAT"} : (tensor<100xf32>, tensor<4xi32>) -> tensor<1x1x1x100xf32>
    %3 = "tfl.reshape"(%1, %cst) {tac.device = "GPU", tac.inference_type = "FLOAT"} : (tensor<100xf32>, tensor<4xi32>) -> tensor<1x1x1x100xf32>

      buffer_type.getShape().drop_front(), buffer_type.getElementType());
  auto reshape = builder.create<TF::ReshapeOp>(
      loc, ArrayRef<Type>{element_type},
      ArrayRef<Value>{slice,
                      GetR1Const(element_type.getShape(), builder, loc)});
  return reshape.getOutput();
}

Value SetElement(Value index, Value buffer, Value element, OpBuilder builder,

// masks will complicate the strided_slice computation logic, we can simplify
// the logic by inserting a reshape op to pad the inputs so strided_slice can
// be easier to handle.
//
// So the graph may looks like below:
//   original_input -> strided_slice -> output
//      (transforms)
//   original_input -> reshape -> strided_slice -> output
//
// And the new shape is computed based on the masks.
//

  // CHECK-DAG: %[[NEW_DIMS:.*]] = arith.constant dense<[6, 1, 7, 8, 1]> : tensor<5xi32>
  // CHECK: %[[RESHAPE:.*]] = "tf.Reshape"(%arg0, %[[NEW_DIMS]]) : (tensor<6x7x8xf32>, tensor<5xi32>) -> tensor<6x1x7x8x1xf32>

//     -> tensor<5x2xf32>
//
// is lowered to
//
//   %shape = "tf.Const"() {value = dense<[-1, 2]> : tensor<2xi64>}
//   %inp0 = "tf.Reshape"(%arg0, %shape)
//     : (tensor<2xf32>, tensor<2xi64>) -> tensor<1x2xf32>
//   %inp1 = "tf.Reshape"(%arg1, %shape)
//     : (tensor<2x2x2xf32>, tensor<2xi64>) -> tensor<4x2xf32>
//   %items0 = "tf.Unpack"(%[[INP0]]) {axis = 0 : i64}

  Output reshape_input =
      ops::Placeholder(root.WithOpName("reshape_input"), DT_FLOAT,
                       ops::Placeholder::Shape(TensorShape({500, 500})));
  Output reshape =
      ops::Reshape(root.WithOpName("reshape"), reshape_input, shape);

  std::unique_ptr<Graph> graph(new Graph(OpRegistry::Global()));

  TF_ASSERT_OK(root.ToGraph(graph.get()));

  ):
    n = 5
    x_shape = [v if v is not None else n for v in shapes[0]]
    y_shape = [v if v is not None else n for v in shapes[1]]

    class MatmulModel(module.Module):

      def __init__(self, bias: Optional[core.Tensor]):
        self._bias = bias
        self._kernel = np.random.uniform(size=y_shape).astype('f4')
        self._min = (-0.8, -0.8, -0.9)
        self._max = (0.9, 0.9, 1.0)