}

func.func @gatherV2VectorIndices(%arg0 : tensor<1x2x20xf32>, %arg1 : tensor<3x5xi32>) -> tensor<1x3x5x20xf32> {
  %0 = "tf.Const"() { value = dense<[1]> : tensor<1xi32> } : () -> tensor<1xi32>
  %1 = "tf.GatherV2"(%arg0, %arg1, %0) : (tensor<1x2x20xf32>, tensor<3x5xi32>, tensor<1xi32>) -> tensor<1x3x5x20xf32>
  func.return %1 : tensor<1x3x5x20xf32>

// CHECK-LABEL:gatherV2VectorIndices

  func.return %0 : tensor<2x4x128xf32>
}

//===----------------------------------------------------------------------===//
// tf.GatherV2 legalization
//===----------------------------------------------------------------------===//

// -----

// CHECK-LABEL: @gather_v2
//  CHECK-SAME: %[[PARAMS:[a-zA-Z0-9_]+]]

            quant_opts_pb2.UnitWiseQuantizationSpec(
                unit=[
                    quant_opts_pb2.UnitWiseQuantizationSpec.QuantizationUnit(
                        op_type='GatherV2',
                    ),
                ],
                quantization_method=quant_opts_pb2.QuantizationMethod(
                    preset_method=quant_opts_pb2.QuantizationMethod.METHOD_NO_QUANTIZE

  func.return %0 : tensor<8xui8>
}

// CHECK-LABEL:   func @torch_index_select(
// CHECK:       %[[AXIS:.+]] = "tf.Const"() <{value = dense<0> : tensor<i64>}> : () -> tensor<i64>
// CHECK:       %[[RES:.+]] = "tf.GatherV2"(%arg0, %arg1, %[[AXIS]]) <{batch_dims = 0 : i64}>
// CHECK:       return %[[RES]]

func.func @torch_index_select(%arg0: tensor<2x1xf32>, %arg1: tensor<2xi32>) -> tensor<2x1xf32> {

  );

  TF_DerivedOperandTypeAttr Tindices = TF_DerivedOperandTypeAttr<1>;
  TF_DerivedOperandTypeAttr Tparams = TF_DerivedOperandTypeAttr<0>;
}

def TF_GatherV2Op : TF_Op<"GatherV2", [Pure]> {
  let summary = [{
Gather slices from `params` axis `axis` according to `indices`.
  }];

  let description = [{
`indices` must be an integer tensor of any dimension (usually 0-D or 1-D).