//
// Examples:
//   * If `xla_gather_op_output_type` == tensor<*xf32>, then it returns:
//     tensor<*xf32>.
//   * If `xla_gather_op_output_type` == tensor<3x5xi32> and `collapsed_dims` ==
//     {0}, then it returns: tensor<1x3x5xi32>.
//   * If `xla_gather_op_output_type` == tensor<3x5xf32> and `collapsed_dims` ==
//     {1, 3}, then it returns: tensor<3x1x5x1xf32>.
Type GetSliceOpOutputType(Type xla_gather_op_output_type,

    %8 = stablehlo.convert %7 : (tensor<1x2xi32>) -> tensor<1x2xf32>
    %9 = stablehlo.convert %2 : (tensor<2x3xi8>) -> tensor<2x3xf32>
    %10 = stablehlo.dot_general %8, %9, contracting_dims = [1] x [0] : (tensor<1x2xf32>, tensor<2x3xf32>) -> tensor<1x3xf32>
    %11 = stablehlo.convert %3 : (tensor<1x3xi32>) -> tensor<1x3xf32>

module attributes {tf.versions = {producer = 888 : i32}} {
func.func @testReshapeShapeInference(%arg0: tensor<3x4xi32>) -> tensor<*xi32> {
  %cst = arith.constant dense<[1, 6, 2]> : tensor<3xi32>
  // CHECK: "tfl.reshape"(%arg0, %cst) : (tensor<3x4xi32>, tensor<3xi32>) -> tensor<1x6x2xi32>
  %0 = "tfl.reshape"(%arg0, %cst) : (tensor<3x4xi32>, tensor<3xi32>) -> tensor<*xi32>
  func.return %0 : tensor<*xi32>
}
}

// -----

func.func @add_dense_dense_int_mixing_1_n() -> tensor<2x2xi32> {
  %cst_0 = arith.constant dense<[[1, 2]]> : tensor<1x2xi32>
  %cst_1 = arith.constant dense<[[3], [4]]> : tensor<2x1xi32>

  %0 = "tfl.add"(%cst_0, %cst_1) {fused_activation_function = "NONE"} : (tensor<1x2xi32>, tensor<2x1xi32>) -> tensor<2x2xi32>

  func.return %0 : tensor<2x2xi32>

func.func @RemoveRedundantUnpackPack(%arg0: tensor<2x5xf32>) -> tensor<2x5xf32> {
  %0:2 = "tfl.unpack"(%arg0) {axis = 0 : i32, num = 2 : i32} : (tensor<2x5xf32>) -> (tensor<5xf32>, tensor<5xf32>)
  %1 = "tfl.pack"(%0#0, %0#1) {axis = 0 : i32, values_count = 2 : i32} : (tensor<5xf32>, tensor<5xf32>) -> (tensor<2x5xf32>)
  func.return %1: tensor<2x5xf32>
  // CHECK-NOT: pack
  // CHECK: return %arg0 : tensor<2x5xf32>
}

// -----

  // CHECK-SAME: -> tensor<3x2x!quant.uniform<i32:f32, 2.000000e+00:4>>
  // CHECK: %[[RES_INT:.*]] = mhlo.bitcast_convert %[[RES]] : (tensor<3x2x!quant.uniform<i32:f32, 2.000000e+00:4>>) -> tensor<3x2xi32>
  // CHECK: return %[[RES_INT]] : tensor<3x2xi32>

  %0 = "tf.UniformQuantizedAdd"(
    %arg0, %bias,
    %input_scales, %input_zps,
    %bias_scales, %bias_zps,
    %output_scales, %output_zps) {

  func.return %0: tensor<3x2xf32>
}

func.func @sparse_f16() -> tensor<3x2xf16> {
  // CHECK-LABEL: @sparse_f16

  func.return %0: tensor<3x2xf32>
}

func.func @sparse_f16() -> tensor<3x2xf16> {
  // CHECK-LABEL: @sparse_f16

        }) : () -> tensor<f32>
        "tf_device.cluster"() ({
          %group = "tf.Const"() <{value = dense<[[0, 1]]> : tensor<1x2xi32>}> : () -> tensor<1x2xi32>
          %arg1_reduced = "tf.XlaAllReduce"(%arg1_id, %group) <{mode = "CrossReplica", reduce_op = "Add"}> : (tensor<f32>, tensor<1x2xi32>) -> tensor<f32>
          "tf.OpA"(%arg1_reduced) : (tensor<f32>) -> ()
          tf_device.return
        }) : () -> ()

  %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>