Merge branch 'develop' into conv_quant_int8

library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_add.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_multiblock_atomic_add<F64, F64, F64, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&); 
+extern template void add_device_reduce_instance_multiblock_atomic_add<F64, F64, F64, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&); 
+extern template void add_device_reduce_instance_multiblock_atomic_add<F64, F64, F64, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&); 
+extern template void add_device_reduce_instance_multiblock_atomic_add<F64, F64, F64, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_avg.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_multiblock_atomic_add<F64, F64, F64, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, UnaryDivide>>&); 
+extern template void add_device_reduce_instance_multiblock_atomic_add<F64, F64, F64, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, UnaryDivide>>&); 
+extern template void add_device_reduce_instance_multiblock_atomic_add<F64, F64, F64, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, UnaryDivide>>&); 
+extern template void add_device_reduce_instance_multiblock_atomic_add<F64, F64, F64, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, UnaryDivide>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp

@@ -5,6 +5,8 @@
 
 #include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
 #include "ck/tensor_operation/gpu/device/impl/device_reduce_threadwise.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
 #include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_impl_common.hpp"
 
 namespace ck {
@@ -49,33 +51,20 @@ using reduce_configuration_2_instances_threadwise = std::tuple<
     >;
 #endif
 
-template <ReduceTensorOp ReduceOpId>
-using deviceReduceThreadWisePtrType = DeviceReducePtr<
-    typename reduce_unary_operator<ReduceOpId, true, true>::InElementwiseOperation,
-    typename reduce_unary_operator<ReduceOpId, true, true>::AccElementwiseOperation>;
-
 template <typename InDataType,
           typename AccDataType,
           typename OutDataType,
           int Rank,
           int NumReduceDim,
-          ReduceTensorOp ReduceOpId,
+          typename ReduceOperation,
+          typename InElementwiseOp,
+          typename AccElementwiseOp,
           bool PropagateNan,
-          bool UseIndex>
+          bool OutputIndex>
 void add_device_reduce_instance_threadwise(
-    std::vector<deviceReduceThreadWisePtrType<ReduceOpId>>& device_op_instances)
+    std::vector<DeviceReducePtr<Rank, NumReduceDim, InElementwiseOp, AccElementwiseOp>>&
+        device_op_instances)
 {
-    using ReduceOperation = typename reduce_binary_operator<ReduceOpId>::opType;
-    using InElementwiseOperation =
-        typename reduce_unary_operator<ReduceOpId, true, true>::InElementwiseOperation;
-    using AccElementwiseOperation =
-        typename reduce_unary_operator<ReduceOpId, true, true>::AccElementwiseOperation;
-
-    constexpr bool Indexable =
-        (ReduceOpId == ReduceTensorOp::MIN || ReduceOpId == ReduceTensorOp::MAX ||
-         ReduceOpId == ReduceTensorOp::AMAX);
-    constexpr bool OutputIndex = Indexable && UseIndex;
-
     using cfg1 = ReductionConfiguration_1<256, 256, 1>;
 
     static_for<0, std::tuple_size<reduce_configuration_2_instances_threadwise>::value, 1>{}(
@@ -89,8 +78,8 @@ void add_device_reduce_instance_threadwise(
                                                             Rank,
                                                             NumReduceDim,
                                                             ReduceOperation,
-                                                            InElementwiseOperation,
-                                                            AccElementwiseOperation,
+                                                            InElementwiseOp,
+                                                            AccElementwiseOp,
                                                             PropagateNan,
                                                             OutputIndex,
                                                             false, // HaveIndexInputIfOutputIndex
@@ -105,52 +94,6 @@ void add_device_reduce_instance_threadwise(
         });
 };
 
-#define ADD_THREADWISE_INST_BY_TYPE(                                          \
-    inT, compT, outT, ReduceOpId, PropagateNan, UseIndex, Rank, NumReduceDim) \
-    template void add_device_reduce_instance_threadwise<inT,                  \
-                                                        compT,                \
-                                                        outT,                 \
-                                                        Rank,                 \
-                                                        NumReduceDim,         \
-                                                        ReduceOpId,           \
-                                                        PropagateNan,         \
-                                                        UseIndex>(            \
-        std::vector<deviceReduceThreadWisePtrType<ReduceOpId>> & device_op_instances)
-
-#define ADD_THREADWISE_INST_BY_ID(                                        \
-    inT, compT, outT, ReduceOpId, NanOpt, IndicesOpt, Rank, NumReduceDim) \
-    ADD_THREADWISE_INST_BY_TYPE(inT,                                      \
-                                compT,                                    \
-                                outT,                                     \
-                                static_cast<ReduceTensorOp>(ReduceOpId),  \
-                                static_cast<bool>(NanOpt),                \
-                                static_cast<bool>(IndicesOpt),            \
-                                Rank,                                     \
-                                NumReduceDim)
-
-#define ADD_THREADWISE_INST_REF_BY_TYPE(                                      \
-    inT, compT, outT, ReduceOpId, PropagateNan, UseIndex, Rank, NumReduceDim) \
-    extern template void add_device_reduce_instance_threadwise<inT,           \
-                                                               compT,         \
-                                                               outT,          \
-                                                               Rank,          \
-                                                               NumReduceDim,  \
-                                                               ReduceOpId,    \
-                                                               PropagateNan,  \
-                                                               UseIndex>(     \
-        std::vector<deviceReduceThreadWisePtrType<ReduceOpId>> & device_op_instances)
-
-#define ADD_THREADWISE_INST_REF_BY_ID(                                       \
-    inT, compT, outT, ReduceOpId, NanOpt, IndicesOpt, Rank, NumReduceDim)    \
-    ADD_THREADWISE_INST_REF_BY_TYPE(inT,                                     \
-                                    compT,                                   \
-                                    outT,                                    \
-                                    static_cast<ReduceTensorOp>(ReduceOpId), \
-                                    static_cast<bool>(NanOpt),               \
-                                    static_cast<bool>(IndicesOpt),           \
-                                    Rank,                                    \
-                                    NumReduceDim)
-
 } // namespace instance
 } // namespace device
 } // namespace tensor_operation

library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16.hpp

-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
-
-#pragma once
-
-#include "ck/utility/data_type.hpp"
-
-#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
-
-namespace ck {
-namespace tensor_operation {
-namespace device {
-namespace instance {
-
-// clang-format off
-// InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | NumReduceDim 
-ADD_THREADWISE_INST_REF_BY_ID(bhalf_t, float, bhalf_t, 0, 0, 0, 4, 3); // for ADD
-ADD_THREADWISE_INST_REF_BY_ID(bhalf_t, float, bhalf_t, 0, 0, 0, 4, 4);
-ADD_THREADWISE_INST_REF_BY_ID(bhalf_t, float, bhalf_t, 0, 0, 0, 4, 1);
-ADD_THREADWISE_INST_REF_BY_ID(bhalf_t, float, bhalf_t, 0, 0, 0, 2, 1);
-ADD_THREADWISE_INST_REF_BY_ID(bhalf_t, float, bhalf_t, 5, 0, 0, 4, 3); // for AVG
-ADD_THREADWISE_INST_REF_BY_ID(bhalf_t, float, bhalf_t, 5, 0, 0, 4, 4);       
-ADD_THREADWISE_INST_REF_BY_ID(bhalf_t, float, bhalf_t, 5, 0, 0, 4, 1);       
-ADD_THREADWISE_INST_REF_BY_ID(bhalf_t, float, bhalf_t, 5, 0, 0, 2, 1);       
-ADD_THREADWISE_INST_REF_BY_ID(bhalf_t, float, bhalf_t, 7, 0, 0, 4, 3); // for NORM2
-ADD_THREADWISE_INST_REF_BY_ID(bhalf_t, float, bhalf_t, 7, 0, 0, 4, 4);       
-ADD_THREADWISE_INST_REF_BY_ID(bhalf_t, float, bhalf_t, 7, 0, 0, 4, 1);       
-ADD_THREADWISE_INST_REF_BY_ID(bhalf_t, float, bhalf_t, 7, 0, 0, 2, 1);
-
-ADD_THREADWISE_INST_REF_BY_ID(bhalf_t, float, bhalf_t, 2, 0, 0, 4, 3); // for MIN
-ADD_THREADWISE_INST_REF_BY_ID(bhalf_t, float, bhalf_t, 2, 0, 0, 4, 4);
-ADD_THREADWISE_INST_REF_BY_ID(bhalf_t, float, bhalf_t, 2, 0, 0, 4, 1);
-ADD_THREADWISE_INST_REF_BY_ID(bhalf_t, float, bhalf_t, 2, 0, 0, 2, 1);
-ADD_THREADWISE_INST_REF_BY_ID(bhalf_t, float, bhalf_t, 3, 0, 0, 4, 3); // for MAX
-ADD_THREADWISE_INST_REF_BY_ID(bhalf_t, float, bhalf_t, 3, 0, 0, 4, 4);
-ADD_THREADWISE_INST_REF_BY_ID(bhalf_t, float, bhalf_t, 3, 0, 0, 4, 1);
-ADD_THREADWISE_INST_REF_BY_ID(bhalf_t, float, bhalf_t, 3, 0, 0, 2, 1);
-ADD_THREADWISE_INST_REF_BY_ID(bhalf_t, float, bhalf_t, 4, 0, 0, 4, 3); // for AMAX
-ADD_THREADWISE_INST_REF_BY_ID(bhalf_t, float, bhalf_t, 4, 0, 0, 4, 4);
-ADD_THREADWISE_INST_REF_BY_ID(bhalf_t, float, bhalf_t, 4, 0, 0, 4, 1);
-ADD_THREADWISE_INST_REF_BY_ID(bhalf_t, float, bhalf_t, 4, 0, 0, 2, 1);
-ADD_THREADWISE_INST_REF_BY_ID(bhalf_t, float, bhalf_t, 2, 0, 1, 4, 3); // for MIN
-ADD_THREADWISE_INST_REF_BY_ID(bhalf_t, float, bhalf_t, 2, 0, 1, 4, 4);
-ADD_THREADWISE_INST_REF_BY_ID(bhalf_t, float, bhalf_t, 2, 0, 1, 4, 1);
-ADD_THREADWISE_INST_REF_BY_ID(bhalf_t, float, bhalf_t, 2, 0, 1, 2, 1);
-ADD_THREADWISE_INST_REF_BY_ID(bhalf_t, float, bhalf_t, 3, 0, 1, 4, 3); // for MAX
-ADD_THREADWISE_INST_REF_BY_ID(bhalf_t, float, bhalf_t, 3, 0, 1, 4, 4);
-ADD_THREADWISE_INST_REF_BY_ID(bhalf_t, float, bhalf_t, 3, 0, 1, 4, 1);
-ADD_THREADWISE_INST_REF_BY_ID(bhalf_t, float, bhalf_t, 3, 0, 1, 2, 1);
-ADD_THREADWISE_INST_REF_BY_ID(bhalf_t, float, bhalf_t, 4, 0, 1, 4, 3); // for AMAX
-ADD_THREADWISE_INST_REF_BY_ID(bhalf_t, float, bhalf_t, 4, 0, 1, 4, 4);
-ADD_THREADWISE_INST_REF_BY_ID(bhalf_t, float, bhalf_t, 4, 0, 1, 4, 1);
-ADD_THREADWISE_INST_REF_BY_ID(bhalf_t, float, bhalf_t, 4, 0, 1, 2, 1);
-// clang-format on
-
-} // namespace instance
-} // namespace device
-} // namespace tensor_operation
-} // namespace ck

library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32.hpp → library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_add.hpp

@@ -4,6 +4,7 @@
 #pragma once
 
 #include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
 
 #include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
 
@@ -13,19 +14,11 @@ namespace device {
 namespace instance {
 
 // clang-format off
-// InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | NumReduceDim 
-ADD_THREADWISE_INST_REF_BY_ID(float, double, float, 0, 0, 0, 4, 3); // for ADD
-ADD_THREADWISE_INST_REF_BY_ID(float, double, float, 0, 0, 0, 4, 4);
-ADD_THREADWISE_INST_REF_BY_ID(float, double, float, 0, 0, 0, 4, 1);
-ADD_THREADWISE_INST_REF_BY_ID(float, double, float, 0, 0, 0, 2, 1);
-ADD_THREADWISE_INST_REF_BY_ID(float, double, float, 5, 0, 0, 4, 3); // for AVG
-ADD_THREADWISE_INST_REF_BY_ID(float, double, float, 5, 0, 0, 4, 4);       
-ADD_THREADWISE_INST_REF_BY_ID(float, double, float, 5, 0, 0, 4, 1);       
-ADD_THREADWISE_INST_REF_BY_ID(float, double, float, 5, 0, 0, 2, 1);       
-ADD_THREADWISE_INST_REF_BY_ID(float, double, float, 7, 0, 0, 4, 3); // for NORM2
-ADD_THREADWISE_INST_REF_BY_ID(float, double, float, 7, 0, 0, 4, 4);       
-ADD_THREADWISE_INST_REF_BY_ID(float, double, float, 7, 0, 0, 4, 1);       
-ADD_THREADWISE_INST_REF_BY_ID(float, double, float, 7, 0, 0, 2, 1);
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&); 
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&); 
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&); 
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
 // clang-format on
 
 } // namespace instance

library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_amax.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, UnaryAbs, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16.hpp → library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_avg.hpp

@@ -4,6 +4,7 @@
 #pragma once
 
 #include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
 
 #include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
 
@@ -13,19 +14,11 @@ namespace device {
 namespace instance {
 
 // clang-format off
-// InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | NumReduceDim 
-ADD_THREADWISE_INST_REF_BY_ID(half_t, float, half_t, 0, 0, 0, 4, 3); // for ADD
-ADD_THREADWISE_INST_REF_BY_ID(half_t, float, half_t, 0, 0, 0, 4, 4);
-ADD_THREADWISE_INST_REF_BY_ID(half_t, float, half_t, 0, 0, 0, 4, 1);
-ADD_THREADWISE_INST_REF_BY_ID(half_t, float, half_t, 0, 0, 0, 2, 1);
-ADD_THREADWISE_INST_REF_BY_ID(half_t, float, half_t, 5, 0, 0, 4, 3); // for AVG
-ADD_THREADWISE_INST_REF_BY_ID(half_t, float, half_t, 5, 0, 0, 4, 4);       
-ADD_THREADWISE_INST_REF_BY_ID(half_t, float, half_t, 5, 0, 0, 4, 1);       
-ADD_THREADWISE_INST_REF_BY_ID(half_t, float, half_t, 5, 0, 0, 2, 1);       
-ADD_THREADWISE_INST_REF_BY_ID(half_t, float, half_t, 7, 0, 0, 4, 3); // for NORM2
-ADD_THREADWISE_INST_REF_BY_ID(half_t, float, half_t, 7, 0, 0, 4, 4);       
-ADD_THREADWISE_INST_REF_BY_ID(half_t, float, half_t, 7, 0, 0, 4, 1);       
-ADD_THREADWISE_INST_REF_BY_ID(half_t, float, half_t, 7, 0, 0, 2, 1);
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, UnaryDivide>>&); 
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, UnaryDivide>>&); 
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, UnaryDivide>>&); 
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, UnaryDivide>>&);
 // clang-format on
 
 } // namespace instance

library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_max.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 3, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 4, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 2, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 3, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 4, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 2, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_min.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 3, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 4, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 2, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 3, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 4, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 2, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_norm2.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 3, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 3, UnarySquare, UnarySqrt>>&); 
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 4, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 4, UnarySquare, UnarySqrt>>&); 
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 4, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 1, UnarySquare, UnarySqrt>>&); 
+extern template void add_device_reduce_instance_threadwise<BF16, F32, BF16, 2, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<2, 1, UnarySquare, UnarySqrt>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16.hpp

-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
-
-#pragma once
-
-#include "ck/utility/data_type.hpp"
-
-#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
-
-namespace ck {
-namespace tensor_operation {
-namespace device {
-namespace instance {
-
-// clang-format off
-// InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | NumReduceDim 
-ADD_THREADWISE_INST_REF_BY_ID(half_t, half_t, half_t, 2, 0, 0, 4, 3); // for MIN
-ADD_THREADWISE_INST_REF_BY_ID(half_t, half_t, half_t, 2, 0, 0, 4, 4);       
-ADD_THREADWISE_INST_REF_BY_ID(half_t, half_t, half_t, 2, 0, 0, 4, 1);       
-ADD_THREADWISE_INST_REF_BY_ID(half_t, half_t, half_t, 2, 0, 0, 2, 1);       
-ADD_THREADWISE_INST_REF_BY_ID(half_t, half_t, half_t, 3, 0, 0, 4, 3); // for MAX
-ADD_THREADWISE_INST_REF_BY_ID(half_t, half_t, half_t, 3, 0, 0, 4, 4);       
-ADD_THREADWISE_INST_REF_BY_ID(half_t, half_t, half_t, 3, 0, 0, 4, 1);       
-ADD_THREADWISE_INST_REF_BY_ID(half_t, half_t, half_t, 3, 0, 0, 2, 1);       
-ADD_THREADWISE_INST_REF_BY_ID(half_t, half_t, half_t, 4, 0, 0, 4, 3); // for AMAX
-ADD_THREADWISE_INST_REF_BY_ID(half_t, half_t, half_t, 4, 0, 0, 4, 4);       
-ADD_THREADWISE_INST_REF_BY_ID(half_t, half_t, half_t, 4, 0, 0, 4, 1);       
-ADD_THREADWISE_INST_REF_BY_ID(half_t, half_t, half_t, 4, 0, 0, 2, 1);       
-ADD_THREADWISE_INST_REF_BY_ID(half_t, half_t, half_t, 2, 0, 1, 4, 3); // for MIN
-ADD_THREADWISE_INST_REF_BY_ID(half_t, half_t, half_t, 2, 0, 1, 4, 4);       
-ADD_THREADWISE_INST_REF_BY_ID(half_t, half_t, half_t, 2, 0, 1, 4, 1);       
-ADD_THREADWISE_INST_REF_BY_ID(half_t, half_t, half_t, 2, 0, 1, 2, 1);       
-ADD_THREADWISE_INST_REF_BY_ID(half_t, half_t, half_t, 3, 0, 1, 4, 3); // for MAX
-ADD_THREADWISE_INST_REF_BY_ID(half_t, half_t, half_t, 3, 0, 1, 4, 4);       
-ADD_THREADWISE_INST_REF_BY_ID(half_t, half_t, half_t, 3, 0, 1, 4, 1);       
-ADD_THREADWISE_INST_REF_BY_ID(half_t, half_t, half_t, 3, 0, 1, 2, 1);       
-ADD_THREADWISE_INST_REF_BY_ID(half_t, half_t, half_t, 4, 0, 1, 4, 3); // for AMAX
-ADD_THREADWISE_INST_REF_BY_ID(half_t, half_t, half_t, 4, 0, 1, 4, 4);       
-ADD_THREADWISE_INST_REF_BY_ID(half_t, half_t, half_t, 4, 0, 1, 4, 1);       
-ADD_THREADWISE_INST_REF_BY_ID(half_t, half_t, half_t, 4, 0, 1, 2, 1);
-// clang-format on
-
-} // namespace instance
-} // namespace device
-} // namespace tensor_operation
-} // namespace ck

library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_amax.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F16, F16, F16, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F16, F16, F16, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, UnaryAbs, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_max.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 3, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 4, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F16, F16, F16, 2, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 3, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 4, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F16, F16, F16, 2, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_min.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 3, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 4, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F16, F16, F16, 2, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 3, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 4, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F16, F16, F16, 4, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F16, F16, F16, 2, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_add.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_threadwise<F16, F32, F16, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&); 
+extern template void add_device_reduce_instance_threadwise<F16, F32, F16, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&); 
+extern template void add_device_reduce_instance_threadwise<F16, F32, F16, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&); 
+extern template void add_device_reduce_instance_threadwise<F16, F32, F16, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_avg.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_threadwise<F16, F32, F16, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, UnaryDivide>>&); 
+extern template void add_device_reduce_instance_threadwise<F16, F32, F16, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, UnaryDivide>>&); 
+extern template void add_device_reduce_instance_threadwise<F16, F32, F16, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, UnaryDivide>>&); 
+extern template void add_device_reduce_instance_threadwise<F16, F32, F16, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, UnaryDivide>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8.hpp → library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_norm2.hpp

@@ -4,6 +4,7 @@
 #pragma once
 
 #include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
 
 #include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
 
@@ -13,15 +14,11 @@ namespace device {
 namespace instance {
 
 // clang-format off
-// InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | NumReduceDim 
-ADD_THREADWISE_INST_REF_BY_ID(int8_t, int32_t, int8_t, 0, 0, 0, 4, 3); // for ADD
-ADD_THREADWISE_INST_REF_BY_ID(int8_t, int32_t, int8_t, 0, 0, 0, 4, 4);
-ADD_THREADWISE_INST_REF_BY_ID(int8_t, int32_t, int8_t, 0, 0, 0, 4, 1);
-ADD_THREADWISE_INST_REF_BY_ID(int8_t, int32_t, int8_t, 0, 0, 0, 2, 1);
-ADD_THREADWISE_INST_REF_BY_ID(int8_t, int32_t, int8_t, 5, 0, 0, 4, 3); // for AVG
-ADD_THREADWISE_INST_REF_BY_ID(int8_t, int32_t, int8_t, 5, 0, 0, 4, 4);
-ADD_THREADWISE_INST_REF_BY_ID(int8_t, int32_t, int8_t, 5, 0, 0, 4, 1);
-ADD_THREADWISE_INST_REF_BY_ID(int8_t, int32_t, int8_t, 5, 0, 0, 2, 1);
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_threadwise<F16, F32, F16, 4, 3, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 3, UnarySquare, UnarySqrt>>&); 
+extern template void add_device_reduce_instance_threadwise<F16, F32, F16, 4, 4, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 4, UnarySquare, UnarySqrt>>&); 
+extern template void add_device_reduce_instance_threadwise<F16, F32, F16, 4, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 1, UnarySquare, UnarySqrt>>&); 
+extern template void add_device_reduce_instance_threadwise<F16, F32, F16, 2, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<2, 1, UnarySquare, UnarySqrt>>&);
 // clang-format on
 
 } // namespace instance

library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32.hpp

-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
-
-#pragma once
-
-#include "ck/utility/data_type.hpp"
-
-#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
-
-namespace ck {
-namespace tensor_operation {
-namespace device {
-namespace instance {
-
-// clang-format off
-// InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | NumReduceDim 
-ADD_THREADWISE_INST_REF_BY_ID(float, float, float, 0, 0, 0, 4, 3); // for ADD
-ADD_THREADWISE_INST_REF_BY_ID(float, float, float, 0, 0, 0, 4, 4);
-ADD_THREADWISE_INST_REF_BY_ID(float, float, float, 0, 0, 0, 4, 1);
-ADD_THREADWISE_INST_REF_BY_ID(float, float, float, 0, 0, 0, 2, 1);
-ADD_THREADWISE_INST_REF_BY_ID(float, float, float, 5, 0, 0, 4, 3); // for AVG
-ADD_THREADWISE_INST_REF_BY_ID(float, float, float, 5, 0, 0, 4, 4);       
-ADD_THREADWISE_INST_REF_BY_ID(float, float, float, 5, 0, 0, 4, 1);       
-ADD_THREADWISE_INST_REF_BY_ID(float, float, float, 5, 0, 0, 2, 1);       
-ADD_THREADWISE_INST_REF_BY_ID(float, float, float, 7, 0, 0, 4, 3); // for NORM2
-ADD_THREADWISE_INST_REF_BY_ID(float, float, float, 7, 0, 0, 4, 4);       
-ADD_THREADWISE_INST_REF_BY_ID(float, float, float, 7, 0, 0, 4, 1);       
-ADD_THREADWISE_INST_REF_BY_ID(float, float, float, 7, 0, 0, 2, 1);       
-ADD_THREADWISE_INST_REF_BY_ID(float, float, float, 2, 0, 0, 4, 3); // for MIN
-ADD_THREADWISE_INST_REF_BY_ID(float, float, float, 2, 0, 0, 4, 4);       
-ADD_THREADWISE_INST_REF_BY_ID(float, float, float, 2, 0, 0, 4, 1);       
-ADD_THREADWISE_INST_REF_BY_ID(float, float, float, 2, 0, 0, 2, 1);       
-ADD_THREADWISE_INST_REF_BY_ID(float, float, float, 3, 0, 0, 4, 3); // for MAX
-ADD_THREADWISE_INST_REF_BY_ID(float, float, float, 3, 0, 0, 4, 4);       
-ADD_THREADWISE_INST_REF_BY_ID(float, float, float, 3, 0, 0, 4, 1);       
-ADD_THREADWISE_INST_REF_BY_ID(float, float, float, 3, 0, 0, 2, 1);       
-ADD_THREADWISE_INST_REF_BY_ID(float, float, float, 4, 0, 0, 4, 3); // for AMAX
-ADD_THREADWISE_INST_REF_BY_ID(float, float, float, 4, 0, 0, 4, 4);       
-ADD_THREADWISE_INST_REF_BY_ID(float, float, float, 4, 0, 0, 4, 1);       
-ADD_THREADWISE_INST_REF_BY_ID(float, float, float, 4, 0, 0, 2, 1);       
-ADD_THREADWISE_INST_REF_BY_ID(float, float, float, 2, 0, 1, 4, 3); // for MIN
-ADD_THREADWISE_INST_REF_BY_ID(float, float, float, 2, 0, 1, 4, 4);       
-ADD_THREADWISE_INST_REF_BY_ID(float, float, float, 2, 0, 1, 4, 1);       
-ADD_THREADWISE_INST_REF_BY_ID(float, float, float, 2, 0, 1, 2, 1);       
-ADD_THREADWISE_INST_REF_BY_ID(float, float, float, 3, 0, 1, 4, 3); // for MAX
-ADD_THREADWISE_INST_REF_BY_ID(float, float, float, 3, 0, 1, 4, 4);       
-ADD_THREADWISE_INST_REF_BY_ID(float, float, float, 3, 0, 1, 4, 1);       
-ADD_THREADWISE_INST_REF_BY_ID(float, float, float, 3, 0, 1, 2, 1);       
-ADD_THREADWISE_INST_REF_BY_ID(float, float, float, 4, 0, 1, 4, 3); // for AMAX
-ADD_THREADWISE_INST_REF_BY_ID(float, float, float, 4, 0, 1, 4, 4);       
-ADD_THREADWISE_INST_REF_BY_ID(float, float, float, 4, 0, 1, 4, 1);       
-ADD_THREADWISE_INST_REF_BY_ID(float, float, float, 4, 0, 1, 2, 1);
-// clang-format on
-
-} // namespace instance
-} // namespace device
-} // namespace tensor_operation
-} // namespace ck

library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_add.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&); 
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&); 
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&); 
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_amax.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_threadwise<F32, F32, F32, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, UnaryAbs, PassThrough>>&);
+// clang-format on
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck