Add outputindex to base class

include/ck/tensor_operation/gpu/device/device_pool_fwd.hpp

@@ -13,7 +13,7 @@ namespace ck {
 namespace tensor_operation {
 namespace device {
 
-template <index_t InOutRank, index_t WindowRank, ReduceTensorOp ReduceOpId>
+template <index_t InOutRank, index_t WindowRank, ReduceTensorOp ReduceOpId, bool OuputIndex>
 struct DevicePoolFwd : public BaseOperator
 {
     virtual std::unique_ptr<BaseArgument>

include/ck/tensor_operation/gpu/device/impl/device_pool2d_fwd_nhwc_nhwc.hpp

@@ -29,7 +29,8 @@ template <typename InDataType,
           ck::index_t ReduceMThreadSliceSize,
           ck::index_t ReduceKThreadSliceSize,
           ck::index_t InSrcOutDstVectorSize>
-struct DevicePool2dFwd_Input_N_Hi_Wi_C_Output_N_Ho_Wo_C : public DevicePoolFwd<5, 2, ReduceOpId>
+struct DevicePool2dFwd_Input_N_Hi_Wi_C_Output_N_Ho_Wo_C
+    : public DevicePoolFwd<5, 2, ReduceOpId, OuputIndex>
 {
     static constexpr auto I0 = Number<0>{};
     static constexpr auto I1 = Number<1>{};

include/ck/tensor_operation/gpu/device/impl/device_pool3d_fwd_ndhwc_ndhwc.hpp

@@ -30,7 +30,7 @@ template <typename InDataType,
           ck::index_t KThreadSliceSize,
           ck::index_t InSrcOutDstVectorSize>
 struct DevicePool3dFwd_Input_N_Di_Hi_Wi_C_Output_N_Do_Ho_Wo_C
-    : public DevicePoolFwd<5, 3, ReduceOpId>
+    : public DevicePoolFwd<5, 3, ReduceOpId, OuputIndex>
 {
     static constexpr auto I0 = Number<0>{};
     static constexpr auto I1 = Number<1>{};