//   %items0 = "tf.Unpack"(%[[INP0]]) {axis = 0 : i64}
//     : (tensor<1x2xf32>) -> tensor<2xf32>
//   %items1:4 = "tf.Unpack"(%[[INP1]]) {axis = 0 : i64}
//     : (tensor<4x2xf32>) -> (tensor<2xf32>, tensor<2xf32>, tensor<2xf32>,
//     tensor<2xf32>)
//   %axis = "tf.Const"() {value = dense<0> : tensor<i64>}
//   %0 = "tf.Pack"(items1#3, items1#2, items1#1, items1#0, %items0, %axis)

    // CHECK-DAG: [[VAR:%.+]] = "tf.VarHandleOp"
    %resource = "tf.VarHandleOp"() {container = "c", shared_name = "v"} : () -> tensor<*x!tf_type.resource<tensor<*xi32>>>

    // CHECK-DAG: [[READVAR:%.+]] = "tf.ReadVariableOp"([[VAR]])

}

// TODO(lyandy): Investigate supported dtypes (`minval`, `maxval`, `output`) for
// `tf.StatefulUniformInt`. tf2xla kernels support i32, i64, ui32, and ui64
// while TensorFlow CPU/GPU kernels only support i32 and i64.
def TF_StatefulUniformIntOp : TF_Op<"StatefulUniformInt", []> {
  let summary = "Outputs random integers from a uniform distribution.";

  let description = [{

      (returnType (GetEmbeddingLookupShape $indices, $filter))
    ),
    (Arith_ConstantOp (GetShape (GetIthValue<0> $outputs)))),
  [(I32ElementsVal $indices),     // lookup is not implemented for i64
   (IsNoneType $bias)]>;          // Maybe folded into the lookup matrix later

// Same as above, but if indices are already 1D, then no reshape is needed.
def ReplaceOneHotDynamicShapeFullyConnectedWithLookup : Pat<

  %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<u8:f32, 0.0078431372549019607:128>>}> {volatile}