Merge remote-tracking branch 'origin/develop' into cpu_avx2

library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_second_call_i32_i32_i8.cpp

+#include "device_reduce_instance_blockwise_second_call.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace device_reduce_instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | NumReduceDim
+ADD_BLOCKWISE_SECOND_CALL_INST_BY_ID(int32_t, int32_t, int8_t, 0, 0, 0, 4, 3); // for ADD
+ADD_BLOCKWISE_SECOND_CALL_INST_BY_ID(int32_t, int32_t, int8_t, 0, 0, 0, 4, 4);
+ADD_BLOCKWISE_SECOND_CALL_INST_BY_ID(int32_t, int32_t, int8_t, 0, 0, 0, 4, 1);
+ADD_BLOCKWISE_SECOND_CALL_INST_BY_ID(int32_t, int32_t, int8_t, 0, 0, 0, 2, 1);
+ADD_BLOCKWISE_SECOND_CALL_INST_BY_ID(int32_t, int32_t, int8_t, 5, 0, 0, 4, 3); // for AVG
+ADD_BLOCKWISE_SECOND_CALL_INST_BY_ID(int32_t, int32_t, int8_t, 5, 0, 0, 4, 4);
+ADD_BLOCKWISE_SECOND_CALL_INST_BY_ID(int32_t, int32_t, int8_t, 5, 0, 0, 4, 1);
+ADD_BLOCKWISE_SECOND_CALL_INST_BY_ID(int32_t, int32_t, int8_t, 5, 0, 0, 2, 1);
+// clang-format on
+
+} // namespace device_reduce_instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck

library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_second_call_i8_i8_i8.cpp

+#include "device_reduce_instance_blockwise_second_call.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace device_reduce_instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | NumReduceDim
+ADD_BLOCKWISE_SECOND_CALL_INST_BY_ID(int8_t, int8_t, int8_t, 2, 0, 0, 4, 3); // for MIN
+ADD_BLOCKWISE_SECOND_CALL_INST_BY_ID(int8_t, int8_t, int8_t, 2, 0, 0, 4, 4);       
+ADD_BLOCKWISE_SECOND_CALL_INST_BY_ID(int8_t, int8_t, int8_t, 2, 0, 0, 4, 1);       
+ADD_BLOCKWISE_SECOND_CALL_INST_BY_ID(int8_t, int8_t, int8_t, 2, 0, 0, 2, 1);       
+ADD_BLOCKWISE_SECOND_CALL_INST_BY_ID(int8_t, int8_t, int8_t, 3, 0, 0, 4, 3); // for MAX
+ADD_BLOCKWISE_SECOND_CALL_INST_BY_ID(int8_t, int8_t, int8_t, 3, 0, 0, 4, 4);       
+ADD_BLOCKWISE_SECOND_CALL_INST_BY_ID(int8_t, int8_t, int8_t, 3, 0, 0, 4, 1);       
+ADD_BLOCKWISE_SECOND_CALL_INST_BY_ID(int8_t, int8_t, int8_t, 3, 0, 0, 2, 1);       
+ADD_BLOCKWISE_SECOND_CALL_INST_BY_ID(int8_t, int8_t, int8_t, 4, 0, 0, 4, 3); // for AMAX
+ADD_BLOCKWISE_SECOND_CALL_INST_BY_ID(int8_t, int8_t, int8_t, 4, 0, 0, 4, 4);       
+ADD_BLOCKWISE_SECOND_CALL_INST_BY_ID(int8_t, int8_t, int8_t, 4, 0, 0, 4, 1);       
+ADD_BLOCKWISE_SECOND_CALL_INST_BY_ID(int8_t, int8_t, int8_t, 4, 0, 0, 2, 1);       
+ADD_BLOCKWISE_SECOND_CALL_INST_BY_ID(int8_t, int8_t, int8_t, 2, 0, 1, 4, 3); // for MIN
+ADD_BLOCKWISE_SECOND_CALL_INST_BY_ID(int8_t, int8_t, int8_t, 2, 0, 1, 4, 4);       
+ADD_BLOCKWISE_SECOND_CALL_INST_BY_ID(int8_t, int8_t, int8_t, 2, 0, 1, 4, 1);       
+ADD_BLOCKWISE_SECOND_CALL_INST_BY_ID(int8_t, int8_t, int8_t, 2, 0, 1, 2, 1);       
+ADD_BLOCKWISE_SECOND_CALL_INST_BY_ID(int8_t, int8_t, int8_t, 3, 0, 1, 4, 3); // for MAX
+ADD_BLOCKWISE_SECOND_CALL_INST_BY_ID(int8_t, int8_t, int8_t, 3, 0, 1, 4, 4);       
+ADD_BLOCKWISE_SECOND_CALL_INST_BY_ID(int8_t, int8_t, int8_t, 3, 0, 1, 4, 1);       
+ADD_BLOCKWISE_SECOND_CALL_INST_BY_ID(int8_t, int8_t, int8_t, 3, 0, 1, 2, 1);       
+ADD_BLOCKWISE_SECOND_CALL_INST_BY_ID(int8_t, int8_t, int8_t, 4, 0, 1, 4, 3); // for AMAX
+ADD_BLOCKWISE_SECOND_CALL_INST_BY_ID(int8_t, int8_t, int8_t, 4, 0, 1, 4, 4);       
+ADD_BLOCKWISE_SECOND_CALL_INST_BY_ID(int8_t, int8_t, int8_t, 4, 0, 1, 4, 1);       
+ADD_BLOCKWISE_SECOND_CALL_INST_BY_ID(int8_t, int8_t, int8_t, 4, 0, 1, 2, 1);
+// clang-format on
+
+} // namespace device_reduce_instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck

library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_b16_f32_f32.cpp

+#include "device_reduce_instance_multiblock_atomic_add.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace device_reduce_instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | NumReduceDim
+ADD_MULTIBLOCK_ATOMIC_ADD_INST_BY_ID(bhalf_t, float, float, 0, 0, 0, 4, 3); // for ADD
+ADD_MULTIBLOCK_ATOMIC_ADD_INST_BY_ID(bhalf_t, float, float, 0, 0, 0, 4, 4);
+ADD_MULTIBLOCK_ATOMIC_ADD_INST_BY_ID(bhalf_t, float, float, 0, 0, 0, 4, 1);
+ADD_MULTIBLOCK_ATOMIC_ADD_INST_BY_ID(bhalf_t, float, float, 0, 0, 0, 2, 1);
+ADD_MULTIBLOCK_ATOMIC_ADD_INST_BY_ID(bhalf_t, float, float, 5, 0, 0, 4, 3); // for AVG
+ADD_MULTIBLOCK_ATOMIC_ADD_INST_BY_ID(bhalf_t, float, float, 5, 0, 0, 4, 4);       
+ADD_MULTIBLOCK_ATOMIC_ADD_INST_BY_ID(bhalf_t, float, float, 5, 0, 0, 4, 1);       
+ADD_MULTIBLOCK_ATOMIC_ADD_INST_BY_ID(bhalf_t, float, float, 5, 0, 0, 2, 1);
+// clang-format on
+
+} // namespace device_reduce_instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck

library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f16_f32_f32.cpp

@@ -8,9 +8,11 @@ namespace device_reduce_instance {
 // clang-format off
 // InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | NumReduceDim
 ADD_MULTIBLOCK_ATOMIC_ADD_INST_BY_ID(half_t, float, float, 0, 0, 0, 4, 3); // for ADD
+ADD_MULTIBLOCK_ATOMIC_ADD_INST_BY_ID(half_t, float, float, 0, 0, 0, 4, 4);
 ADD_MULTIBLOCK_ATOMIC_ADD_INST_BY_ID(half_t, float, float, 0, 0, 0, 4, 1);
 ADD_MULTIBLOCK_ATOMIC_ADD_INST_BY_ID(half_t, float, float, 0, 0, 0, 2, 1);
 ADD_MULTIBLOCK_ATOMIC_ADD_INST_BY_ID(half_t, float, float, 5, 0, 0, 4, 3); // for AVG
+ADD_MULTIBLOCK_ATOMIC_ADD_INST_BY_ID(half_t, float, float, 5, 0, 0, 4, 4);       
 ADD_MULTIBLOCK_ATOMIC_ADD_INST_BY_ID(half_t, float, float, 5, 0, 0, 4, 1);       
 ADD_MULTIBLOCK_ATOMIC_ADD_INST_BY_ID(half_t, float, float, 5, 0, 0, 2, 1);
 // clang-format on

library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f32_f32.cpp

@@ -8,9 +8,11 @@ namespace device_reduce_instance {
 // clang-format off
 // InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | NumReduceDim
 ADD_MULTIBLOCK_ATOMIC_ADD_INST_BY_ID(float, float, float, 0, 0, 0, 4, 3); // for ADD
+ADD_MULTIBLOCK_ATOMIC_ADD_INST_BY_ID(float, float, float, 0, 0, 0, 4, 4);
 ADD_MULTIBLOCK_ATOMIC_ADD_INST_BY_ID(float, float, float, 0, 0, 0, 4, 1);
 ADD_MULTIBLOCK_ATOMIC_ADD_INST_BY_ID(float, float, float, 0, 0, 0, 2, 1);
 ADD_MULTIBLOCK_ATOMIC_ADD_INST_BY_ID(float, float, float, 5, 0, 0, 4, 3); // for AVG
+ADD_MULTIBLOCK_ATOMIC_ADD_INST_BY_ID(float, float, float, 5, 0, 0, 4, 4);       
 ADD_MULTIBLOCK_ATOMIC_ADD_INST_BY_ID(float, float, float, 5, 0, 0, 4, 1);       
 ADD_MULTIBLOCK_ATOMIC_ADD_INST_BY_ID(float, float, float, 5, 0, 0, 2, 1);
 // clang-format on

library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f64_f32.cpp

@@ -8,9 +8,11 @@ namespace device_reduce_instance {
 // clang-format off
 // InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | NumReduceDim
 ADD_MULTIBLOCK_ATOMIC_ADD_INST_BY_ID(float, double, float, 0, 0, 0, 4, 3); // for ADD
+ADD_MULTIBLOCK_ATOMIC_ADD_INST_BY_ID(float, double, float, 0, 0, 0, 4, 4);
 ADD_MULTIBLOCK_ATOMIC_ADD_INST_BY_ID(float, double, float, 0, 0, 0, 4, 1);
 ADD_MULTIBLOCK_ATOMIC_ADD_INST_BY_ID(float, double, float, 0, 0, 0, 2, 1);
 ADD_MULTIBLOCK_ATOMIC_ADD_INST_BY_ID(float, double, float, 5, 0, 0, 4, 3); // for AVG
+ADD_MULTIBLOCK_ATOMIC_ADD_INST_BY_ID(float, double, float, 5, 0, 0, 4, 4);       
 ADD_MULTIBLOCK_ATOMIC_ADD_INST_BY_ID(float, double, float, 5, 0, 0, 4, 1);       
 ADD_MULTIBLOCK_ATOMIC_ADD_INST_BY_ID(float, double, float, 5, 0, 0, 2, 1);
 // clang-format on

library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_partial_reduce_b16_f32_b16.cpp

+#include "device_reduce_instance_multiblock_partial_reduce.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace device_reduce_instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | NumReduceDim
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(bhalf_t, float, bhalf_t, 0, 0, 0, 4, 3); // for ADD
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(bhalf_t, float, bhalf_t, 0, 0, 0, 4, 4);
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(bhalf_t, float, bhalf_t, 0, 0, 0, 4, 1);
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(bhalf_t, float, bhalf_t, 0, 0, 0, 2, 1);
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(bhalf_t, float, bhalf_t, 5, 0, 0, 4, 3); // for AVG
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(bhalf_t, float, bhalf_t, 5, 0, 0, 4, 4);       
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(bhalf_t, float, bhalf_t, 5, 0, 0, 4, 1);       
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(bhalf_t, float, bhalf_t, 5, 0, 0, 2, 1);       
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(bhalf_t, float, bhalf_t, 7, 0, 0, 4, 3); // for NORM2
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(bhalf_t, float, bhalf_t, 7, 0, 0, 4, 4);       
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(bhalf_t, float, bhalf_t, 7, 0, 0, 4, 1);       
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(bhalf_t, float, bhalf_t, 7, 0, 0, 2, 1);
+
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(bhalf_t, float, bhalf_t, 2, 0, 0, 4, 3); // for MIN
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(bhalf_t, float, bhalf_t, 2, 0, 0, 4, 4);
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(bhalf_t, float, bhalf_t, 2, 0, 0, 4, 1);
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(bhalf_t, float, bhalf_t, 2, 0, 0, 2, 1);
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(bhalf_t, float, bhalf_t, 3, 0, 0, 4, 3); // for MAX
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(bhalf_t, float, bhalf_t, 3, 0, 0, 4, 4);
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(bhalf_t, float, bhalf_t, 3, 0, 0, 4, 1);
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(bhalf_t, float, bhalf_t, 3, 0, 0, 2, 1);
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(bhalf_t, float, bhalf_t, 4, 0, 0, 4, 3); // for AMAX
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(bhalf_t, float, bhalf_t, 4, 0, 0, 4, 4);
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(bhalf_t, float, bhalf_t, 4, 0, 0, 4, 1);
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(bhalf_t, float, bhalf_t, 4, 0, 0, 2, 1);
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(bhalf_t, float, bhalf_t, 2, 0, 1, 4, 3); // for MIN
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(bhalf_t, float, bhalf_t, 2, 0, 1, 4, 4);
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(bhalf_t, float, bhalf_t, 2, 0, 1, 4, 1);
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(bhalf_t, float, bhalf_t, 2, 0, 1, 2, 1);
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(bhalf_t, float, bhalf_t, 3, 0, 1, 4, 3); // for MAX
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(bhalf_t, float, bhalf_t, 3, 0, 1, 4, 4);
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(bhalf_t, float, bhalf_t, 3, 0, 1, 4, 1);
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(bhalf_t, float, bhalf_t, 3, 0, 1, 2, 1);
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(bhalf_t, float, bhalf_t, 4, 0, 1, 4, 3); // for AMAX
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(bhalf_t, float, bhalf_t, 4, 0, 1, 4, 4);
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(bhalf_t, float, bhalf_t, 4, 0, 1, 4, 1);
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(bhalf_t, float, bhalf_t, 4, 0, 1, 2, 1);
+// clang-format on
+
+} // namespace device_reduce_instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck

library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_partial_reduce_f16_f16_f16.cpp

@@ -8,21 +8,27 @@ namespace device_reduce_instance {
 // clang-format off
 // InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | NumReduceDim
 ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(half_t, half_t, half_t, 2, 0, 0, 4, 3); // for MIN
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(half_t, half_t, half_t, 2, 0, 0, 4, 4);       
 ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(half_t, half_t, half_t, 2, 0, 0, 4, 1);       
 ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(half_t, half_t, half_t, 2, 0, 0, 2, 1);       
 ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(half_t, half_t, half_t, 3, 0, 0, 4, 3); // for MAX
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(half_t, half_t, half_t, 3, 0, 0, 4, 4);       
 ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(half_t, half_t, half_t, 3, 0, 0, 4, 1);       
 ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(half_t, half_t, half_t, 3, 0, 0, 2, 1);       
 ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(half_t, half_t, half_t, 4, 0, 0, 4, 3); // for AMAX
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(half_t, half_t, half_t, 4, 0, 0, 4, 4);       
 ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(half_t, half_t, half_t, 4, 0, 0, 4, 1);       
 ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(half_t, half_t, half_t, 4, 0, 0, 2, 1);       
 ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(half_t, half_t, half_t, 2, 0, 1, 4, 3); // for MIN
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(half_t, half_t, half_t, 2, 0, 1, 4, 4);       
 ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(half_t, half_t, half_t, 2, 0, 1, 4, 1);       
 ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(half_t, half_t, half_t, 2, 0, 1, 2, 1);       
 ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(half_t, half_t, half_t, 3, 0, 1, 4, 3); // for MAX
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(half_t, half_t, half_t, 3, 0, 1, 4, 4);       
 ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(half_t, half_t, half_t, 3, 0, 1, 4, 1);       
 ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(half_t, half_t, half_t, 3, 0, 1, 2, 1);       
 ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(half_t, half_t, half_t, 4, 0, 1, 4, 3); // for AMAX
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(half_t, half_t, half_t, 4, 0, 1, 4, 4);       
 ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(half_t, half_t, half_t, 4, 0, 1, 4, 1);       
 ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(half_t, half_t, half_t, 4, 0, 1, 2, 1);
 // clang-format on

library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_partial_reduce_f16_f32_f16.cpp

@@ -8,12 +8,15 @@ namespace device_reduce_instance {
 // clang-format off
 // InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | NumReduceDim
 ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(half_t, float, half_t, 0, 0, 0, 4, 3); // for ADD
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(half_t, float, half_t, 0, 0, 0, 4, 4);
 ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(half_t, float, half_t, 0, 0, 0, 4, 1);
 ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(half_t, float, half_t, 0, 0, 0, 2, 1);
 ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(half_t, float, half_t, 5, 0, 0, 4, 3); // for AVG
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(half_t, float, half_t, 5, 0, 0, 4, 4);       
 ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(half_t, float, half_t, 5, 0, 0, 4, 1);       
 ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(half_t, float, half_t, 5, 0, 0, 2, 1);       
 ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(half_t, float, half_t, 7, 0, 0, 4, 3); // for NORM2
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(half_t, float, half_t, 7, 0, 0, 4, 4);       
 ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(half_t, float, half_t, 7, 0, 0, 4, 1);       
 ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(half_t, float, half_t, 7, 0, 0, 2, 1);
 // clang-format on

library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_partial_reduce_f32_f32_f32.cpp

@@ -8,25 +8,32 @@ namespace device_reduce_instance {
 // clang-format off
 // InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | NumReduceDim
 ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(float, float, float, 2, 0, 0, 4, 3); // for MIN
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(float, float, float, 2, 0, 0, 4, 4);       
 ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(float, float, float, 2, 0, 0, 4, 1);       
 ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(float, float, float, 2, 0, 0, 2, 1);       
 ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(float, float, float, 3, 0, 0, 4, 3); // for MAX
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(float, float, float, 3, 0, 0, 4, 4);       
 ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(float, float, float, 3, 0, 0, 4, 1);       
 ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(float, float, float, 3, 0, 0, 2, 1);       
 ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(float, float, float, 4, 0, 0, 4, 3); // for AMAX
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(float, float, float, 4, 0, 0, 4, 4);       
 ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(float, float, float, 4, 0, 0, 4, 1);       
 ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(float, float, float, 4, 0, 0, 2, 1);       
 ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(float, float, float, 2, 0, 1, 4, 3); // for MIN
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(float, float, float, 2, 0, 1, 4, 4);       
 ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(float, float, float, 2, 0, 1, 4, 1);       
 ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(float, float, float, 2, 0, 1, 2, 1);       
 ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(float, float, float, 3, 0, 1, 4, 3); // for MAX
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(float, float, float, 3, 0, 1, 4, 4);       
 ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(float, float, float, 3, 0, 1, 4, 1);       
 ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(float, float, float, 3, 0, 1, 2, 1);       
 ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(float, float, float, 4, 0, 1, 4, 3); // for AMAX
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(float, float, float, 4, 0, 1, 4, 4);       
 ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(float, float, float, 4, 0, 1, 4, 1);       
 ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(float, float, float, 4, 0, 1, 2, 1);       
 
 ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(float, float, float, 7, 0, 0, 4, 3); // for NORM2
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(float, float, float, 7, 0, 0, 4, 4);       
 ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(float, float, float, 7, 0, 0, 4, 1);       
 ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(float, float, float, 7, 0, 0, 2, 1);
 // clang-format on

library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_partial_reduce_f32_f64_f32.cpp

@@ -8,6 +8,7 @@ namespace device_reduce_instance {
 // clang-format off
 // InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | NumReduceDim
 ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(float, double, float, 7, 0, 0, 4, 3); // for NORM2
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(float, double, float, 7, 0, 0, 4, 4);       
 ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(float, double, float, 7, 0, 0, 4, 1);       
 ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(float, double, float, 7, 0, 0, 2, 1);
 // clang-format on

library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_partial_reduce_f64_f64_f64.cpp

@@ -8,33 +8,42 @@ namespace device_reduce_instance {
 // clang-format off
 // InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | NumReduceDim
 ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 2, 0, 0, 4, 3); // for MIN
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 2, 0, 0, 4, 4);       
 ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 2, 0, 0, 4, 1);       
 ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 2, 0, 0, 2, 1);       
 ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 3, 0, 0, 4, 3); // for MAX
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 3, 0, 0, 4, 4);       
 ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 3, 0, 0, 4, 1);       
 ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 3, 0, 0, 2, 1);       
 ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 4, 0, 0, 4, 3); // for AMAX
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 4, 0, 0, 4, 4);       
 ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 4, 0, 0, 4, 1);       
 ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 4, 0, 0, 2, 1);       
 ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 2, 0, 1, 4, 3); // for MIN
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 2, 0, 1, 4, 4);       
 ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 2, 0, 1, 4, 1);       
 ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 2, 0, 1, 2, 1);       
 ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 3, 0, 1, 4, 3); // for MAX
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 3, 0, 1, 4, 4);       
 ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 3, 0, 1, 4, 1);       
 ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 3, 0, 1, 2, 1);       
 ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 4, 0, 1, 4, 3); // for AMAX
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 4, 0, 1, 4, 4);       
 ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 4, 0, 1, 4, 1);       
 ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 4, 0, 1, 2, 1);       
 
 ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 7, 0, 0, 4, 3); // for NORM2
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 7, 0, 0, 4, 4);       
 ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 7, 0, 0, 4, 1);       
 ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 7, 0, 0, 2, 1);       
 
 // Will be moved to use MultiBlockAtomicAdd
 ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 0, 0, 0, 4, 3); // for ADD
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 0, 0, 0, 4, 4);       
 ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 0, 0, 0, 4, 1);       
 ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 0, 0, 0, 2, 1);       
 ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 5, 0, 0, 4, 3); // for AVG
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 5, 0, 0, 4, 4);       
 ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 5, 0, 0, 4, 1);       
 ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(double, double, double, 5, 0, 0, 2, 1);
 // clang-format on

library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_partial_reduce_i8_i32_i8.cpp

+#include "device_reduce_instance_multiblock_partial_reduce.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace device_reduce_instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | NumReduceDim
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int32_t, int8_t, 0, 0, 0, 4, 3); // for ADD
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int32_t, int8_t, 0, 0, 0, 4, 4);
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int32_t, int8_t, 0, 0, 0, 4, 1);
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int32_t, int8_t, 0, 0, 0, 2, 1);
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int32_t, int8_t, 5, 0, 0, 4, 3); // for AVG
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int32_t, int8_t, 5, 0, 0, 4, 4);
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int32_t, int8_t, 5, 0, 0, 4, 1);
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int32_t, int8_t, 5, 0, 0, 2, 1);
+// clang-format on
+
+} // namespace device_reduce_instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck

library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_partial_reduce_i8_i8_i8.cpp

+#include "device_reduce_instance_multiblock_partial_reduce.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace device_reduce_instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | NumReduceDim
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 2, 0, 0, 4, 3); // for MIN
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 2, 0, 0, 4, 4);       
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 2, 0, 0, 4, 1);       
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 2, 0, 0, 2, 1);       
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 3, 0, 0, 4, 3); // for MAX
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 3, 0, 0, 4, 4);       
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 3, 0, 0, 4, 1);       
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 3, 0, 0, 2, 1);       
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 4, 0, 0, 4, 3); // for AMAX
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 4, 0, 0, 4, 4);       
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 4, 0, 0, 4, 1);       
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 4, 0, 0, 2, 1);       
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 2, 0, 1, 4, 3); // for MIN
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 2, 0, 1, 4, 4);       
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 2, 0, 1, 4, 1);       
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 2, 0, 1, 2, 1);       
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 3, 0, 1, 4, 3); // for MAX
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 3, 0, 1, 4, 4);       
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 3, 0, 1, 4, 1);       
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 3, 0, 1, 2, 1);       
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 4, 0, 1, 4, 3); // for AMAX
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 4, 0, 1, 4, 4);       
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 4, 0, 1, 4, 1);       
+ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_BY_ID(int8_t, int8_t, int8_t, 4, 0, 1, 2, 1);
+// clang-format on
+
+} // namespace device_reduce_instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck

library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16.cpp

+#include "device_reduce_instance_threadwise.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace device_reduce_instance {
+
+// clang-format off
+// InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | NumReduceDim
+ADD_THREADWISE_INST_BY_ID(bhalf_t, float, bhalf_t, 0, 0, 0, 4, 3); // for ADD
+ADD_THREADWISE_INST_BY_ID(bhalf_t, float, bhalf_t, 0, 0, 0, 4, 4);
+ADD_THREADWISE_INST_BY_ID(bhalf_t, float, bhalf_t, 0, 0, 0, 4, 1);
+ADD_THREADWISE_INST_BY_ID(bhalf_t, float, bhalf_t, 0, 0, 0, 2, 1);
+ADD_THREADWISE_INST_BY_ID(bhalf_t, float, bhalf_t, 5, 0, 0, 4, 3); // for AVG
+ADD_THREADWISE_INST_BY_ID(bhalf_t, float, bhalf_t, 5, 0, 0, 4, 4);       
+ADD_THREADWISE_INST_BY_ID(bhalf_t, float, bhalf_t, 5, 0, 0, 4, 1);       
+ADD_THREADWISE_INST_BY_ID(bhalf_t, float, bhalf_t, 5, 0, 0, 2, 1);       
+ADD_THREADWISE_INST_BY_ID(bhalf_t, float, bhalf_t, 7, 0, 0, 4, 3); // for NORM2
+ADD_THREADWISE_INST_BY_ID(bhalf_t, float, bhalf_t, 7, 0, 0, 4, 4);       
+ADD_THREADWISE_INST_BY_ID(bhalf_t, float, bhalf_t, 7, 0, 0, 4, 1);       
+ADD_THREADWISE_INST_BY_ID(bhalf_t, float, bhalf_t, 7, 0, 0, 2, 1);
+
+ADD_THREADWISE_INST_BY_ID(bhalf_t, float, bhalf_t, 2, 0, 0, 4, 3); // for MIN
+ADD_THREADWISE_INST_BY_ID(bhalf_t, float, bhalf_t, 2, 0, 0, 4, 4);
+ADD_THREADWISE_INST_BY_ID(bhalf_t, float, bhalf_t, 2, 0, 0, 4, 1);
+ADD_THREADWISE_INST_BY_ID(bhalf_t, float, bhalf_t, 2, 0, 0, 2, 1);
+ADD_THREADWISE_INST_BY_ID(bhalf_t, float, bhalf_t, 3, 0, 0, 4, 3); // for MAX
+ADD_THREADWISE_INST_BY_ID(bhalf_t, float, bhalf_t, 3, 0, 0, 4, 4);
+ADD_THREADWISE_INST_BY_ID(bhalf_t, float, bhalf_t, 3, 0, 0, 4, 1);
+ADD_THREADWISE_INST_BY_ID(bhalf_t, float, bhalf_t, 3, 0, 0, 2, 1);
+ADD_THREADWISE_INST_BY_ID(bhalf_t, float, bhalf_t, 4, 0, 0, 4, 3); // for AMAX
+ADD_THREADWISE_INST_BY_ID(bhalf_t, float, bhalf_t, 4, 0, 0, 4, 4);
+ADD_THREADWISE_INST_BY_ID(bhalf_t, float, bhalf_t, 4, 0, 0, 4, 1);
+ADD_THREADWISE_INST_BY_ID(bhalf_t, float, bhalf_t, 4, 0, 0, 2, 1);
+ADD_THREADWISE_INST_BY_ID(bhalf_t, float, bhalf_t, 2, 0, 1, 4, 3); // for MIN
+ADD_THREADWISE_INST_BY_ID(bhalf_t, float, bhalf_t, 2, 0, 1, 4, 4);
+ADD_THREADWISE_INST_BY_ID(bhalf_t, float, bhalf_t, 2, 0, 1, 4, 1);
+ADD_THREADWISE_INST_BY_ID(bhalf_t, float, bhalf_t, 2, 0, 1, 2, 1);
+ADD_THREADWISE_INST_BY_ID(bhalf_t, float, bhalf_t, 3, 0, 1, 4, 3); // for MAX
+ADD_THREADWISE_INST_BY_ID(bhalf_t, float, bhalf_t, 3, 0, 1, 4, 4);
+ADD_THREADWISE_INST_BY_ID(bhalf_t, float, bhalf_t, 3, 0, 1, 4, 1);
+ADD_THREADWISE_INST_BY_ID(bhalf_t, float, bhalf_t, 3, 0, 1, 2, 1);
+ADD_THREADWISE_INST_BY_ID(bhalf_t, float, bhalf_t, 4, 0, 1, 4, 3); // for AMAX
+ADD_THREADWISE_INST_BY_ID(bhalf_t, float, bhalf_t, 4, 0, 1, 4, 4);
+ADD_THREADWISE_INST_BY_ID(bhalf_t, float, bhalf_t, 4, 0, 1, 4, 1);
+ADD_THREADWISE_INST_BY_ID(bhalf_t, float, bhalf_t, 4, 0, 1, 2, 1);
+// clang-format on
+
+} // namespace device_reduce_instance
+} // namespace device
+} // namespace tensor_operation
+
+} // namespace ck

library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16.cpp

@@ -8,21 +8,27 @@ namespace device_reduce_instance {
 // clang-format off
 // InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | NumReduceDim
 ADD_THREADWISE_INST_BY_ID(half_t, half_t, half_t, 2, 0, 0, 4, 3); // for MIN
+ADD_THREADWISE_INST_BY_ID(half_t, half_t, half_t, 2, 0, 0, 4, 4);       
 ADD_THREADWISE_INST_BY_ID(half_t, half_t, half_t, 2, 0, 0, 4, 1);       
 ADD_THREADWISE_INST_BY_ID(half_t, half_t, half_t, 2, 0, 0, 2, 1);       
 ADD_THREADWISE_INST_BY_ID(half_t, half_t, half_t, 3, 0, 0, 4, 3); // for MAX
+ADD_THREADWISE_INST_BY_ID(half_t, half_t, half_t, 3, 0, 0, 4, 4);       
 ADD_THREADWISE_INST_BY_ID(half_t, half_t, half_t, 3, 0, 0, 4, 1);       
 ADD_THREADWISE_INST_BY_ID(half_t, half_t, half_t, 3, 0, 0, 2, 1);       
 ADD_THREADWISE_INST_BY_ID(half_t, half_t, half_t, 4, 0, 0, 4, 3); // for AMAX
+ADD_THREADWISE_INST_BY_ID(half_t, half_t, half_t, 4, 0, 0, 4, 4);       
 ADD_THREADWISE_INST_BY_ID(half_t, half_t, half_t, 4, 0, 0, 4, 1);       
 ADD_THREADWISE_INST_BY_ID(half_t, half_t, half_t, 4, 0, 0, 2, 1);       
 ADD_THREADWISE_INST_BY_ID(half_t, half_t, half_t, 2, 0, 1, 4, 3); // for MIN
+ADD_THREADWISE_INST_BY_ID(half_t, half_t, half_t, 2, 0, 1, 4, 4);       
 ADD_THREADWISE_INST_BY_ID(half_t, half_t, half_t, 2, 0, 1, 4, 1);       
 ADD_THREADWISE_INST_BY_ID(half_t, half_t, half_t, 2, 0, 1, 2, 1);       
 ADD_THREADWISE_INST_BY_ID(half_t, half_t, half_t, 3, 0, 1, 4, 3); // for MAX
+ADD_THREADWISE_INST_BY_ID(half_t, half_t, half_t, 3, 0, 1, 4, 4);       
 ADD_THREADWISE_INST_BY_ID(half_t, half_t, half_t, 3, 0, 1, 4, 1);       
 ADD_THREADWISE_INST_BY_ID(half_t, half_t, half_t, 3, 0, 1, 2, 1);       
 ADD_THREADWISE_INST_BY_ID(half_t, half_t, half_t, 4, 0, 1, 4, 3); // for AMAX
+ADD_THREADWISE_INST_BY_ID(half_t, half_t, half_t, 4, 0, 1, 4, 4);       
 ADD_THREADWISE_INST_BY_ID(half_t, half_t, half_t, 4, 0, 1, 4, 1);       
 ADD_THREADWISE_INST_BY_ID(half_t, half_t, half_t, 4, 0, 1, 2, 1);
 // clang-format on

library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16.cpp

@@ -8,12 +8,15 @@ namespace device_reduce_instance {
 // clang-format off
 // InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | NumReduceDim
 ADD_THREADWISE_INST_BY_ID(half_t, float, half_t, 0, 0, 0, 4, 3); // for ADD
+ADD_THREADWISE_INST_BY_ID(half_t, float, half_t, 0, 0, 0, 4, 4);
 ADD_THREADWISE_INST_BY_ID(half_t, float, half_t, 0, 0, 0, 4, 1);
 ADD_THREADWISE_INST_BY_ID(half_t, float, half_t, 0, 0, 0, 2, 1);
 ADD_THREADWISE_INST_BY_ID(half_t, float, half_t, 5, 0, 0, 4, 3); // for AVG
+ADD_THREADWISE_INST_BY_ID(half_t, float, half_t, 5, 0, 0, 4, 4);       
 ADD_THREADWISE_INST_BY_ID(half_t, float, half_t, 5, 0, 0, 4, 1);       
 ADD_THREADWISE_INST_BY_ID(half_t, float, half_t, 5, 0, 0, 2, 1);       
 ADD_THREADWISE_INST_BY_ID(half_t, float, half_t, 7, 0, 0, 4, 3); // for NORM2
+ADD_THREADWISE_INST_BY_ID(half_t, float, half_t, 7, 0, 0, 4, 4);       
 ADD_THREADWISE_INST_BY_ID(half_t, float, half_t, 7, 0, 0, 4, 1);       
 ADD_THREADWISE_INST_BY_ID(half_t, float, half_t, 7, 0, 0, 2, 1);
 // clang-format on

library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32.cpp

@@ -8,30 +8,39 @@ namespace device_reduce_instance {
 // clang-format off
 // InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | NumReduceDim
 ADD_THREADWISE_INST_BY_ID(float, float, float, 0, 0, 0, 4, 3); // for ADD
+ADD_THREADWISE_INST_BY_ID(float, float, float, 0, 0, 0, 4, 4);
 ADD_THREADWISE_INST_BY_ID(float, float, float, 0, 0, 0, 4, 1);
 ADD_THREADWISE_INST_BY_ID(float, float, float, 0, 0, 0, 2, 1);
 ADD_THREADWISE_INST_BY_ID(float, float, float, 5, 0, 0, 4, 3); // for AVG
+ADD_THREADWISE_INST_BY_ID(float, float, float, 5, 0, 0, 4, 4);       
 ADD_THREADWISE_INST_BY_ID(float, float, float, 5, 0, 0, 4, 1);       
 ADD_THREADWISE_INST_BY_ID(float, float, float, 5, 0, 0, 2, 1);       
 ADD_THREADWISE_INST_BY_ID(float, float, float, 7, 0, 0, 4, 3); // for NORM2
+ADD_THREADWISE_INST_BY_ID(float, float, float, 7, 0, 0, 4, 4);       
 ADD_THREADWISE_INST_BY_ID(float, float, float, 7, 0, 0, 4, 1);       
 ADD_THREADWISE_INST_BY_ID(float, float, float, 7, 0, 0, 2, 1);       
 ADD_THREADWISE_INST_BY_ID(float, float, float, 2, 0, 0, 4, 3); // for MIN
+ADD_THREADWISE_INST_BY_ID(float, float, float, 2, 0, 0, 4, 4);       
 ADD_THREADWISE_INST_BY_ID(float, float, float, 2, 0, 0, 4, 1);       
 ADD_THREADWISE_INST_BY_ID(float, float, float, 2, 0, 0, 2, 1);       
 ADD_THREADWISE_INST_BY_ID(float, float, float, 3, 0, 0, 4, 3); // for MAX
+ADD_THREADWISE_INST_BY_ID(float, float, float, 3, 0, 0, 4, 4);       
 ADD_THREADWISE_INST_BY_ID(float, float, float, 3, 0, 0, 4, 1);       
 ADD_THREADWISE_INST_BY_ID(float, float, float, 3, 0, 0, 2, 1);       
 ADD_THREADWISE_INST_BY_ID(float, float, float, 4, 0, 0, 4, 3); // for AMAX
+ADD_THREADWISE_INST_BY_ID(float, float, float, 4, 0, 0, 4, 4);       
 ADD_THREADWISE_INST_BY_ID(float, float, float, 4, 0, 0, 4, 1);       
 ADD_THREADWISE_INST_BY_ID(float, float, float, 4, 0, 0, 2, 1);       
 ADD_THREADWISE_INST_BY_ID(float, float, float, 2, 0, 1, 4, 3); // for MIN
+ADD_THREADWISE_INST_BY_ID(float, float, float, 2, 0, 1, 4, 4);       
 ADD_THREADWISE_INST_BY_ID(float, float, float, 2, 0, 1, 4, 1);       
 ADD_THREADWISE_INST_BY_ID(float, float, float, 2, 0, 1, 2, 1);       
 ADD_THREADWISE_INST_BY_ID(float, float, float, 3, 0, 1, 4, 3); // for MAX
+ADD_THREADWISE_INST_BY_ID(float, float, float, 3, 0, 1, 4, 4);       
 ADD_THREADWISE_INST_BY_ID(float, float, float, 3, 0, 1, 4, 1);       
 ADD_THREADWISE_INST_BY_ID(float, float, float, 3, 0, 1, 2, 1);       
 ADD_THREADWISE_INST_BY_ID(float, float, float, 4, 0, 1, 4, 3); // for AMAX
+ADD_THREADWISE_INST_BY_ID(float, float, float, 4, 0, 1, 4, 4);       
 ADD_THREADWISE_INST_BY_ID(float, float, float, 4, 0, 1, 4, 1);       
 ADD_THREADWISE_INST_BY_ID(float, float, float, 4, 0, 1, 2, 1);
 // clang-format on

library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32.cpp

@@ -8,12 +8,15 @@ namespace device_reduce_instance {
 // clang-format off
 // InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | NumReduceDim
 ADD_THREADWISE_INST_BY_ID(float, double, float, 0, 0, 0, 4, 3); // for ADD
+ADD_THREADWISE_INST_BY_ID(float, double, float, 0, 0, 0, 4, 4);
 ADD_THREADWISE_INST_BY_ID(float, double, float, 0, 0, 0, 4, 1);
 ADD_THREADWISE_INST_BY_ID(float, double, float, 0, 0, 0, 2, 1);
 ADD_THREADWISE_INST_BY_ID(float, double, float, 5, 0, 0, 4, 3); // for AVG
+ADD_THREADWISE_INST_BY_ID(float, double, float, 5, 0, 0, 4, 4);       
 ADD_THREADWISE_INST_BY_ID(float, double, float, 5, 0, 0, 4, 1);       
 ADD_THREADWISE_INST_BY_ID(float, double, float, 5, 0, 0, 2, 1);       
 ADD_THREADWISE_INST_BY_ID(float, double, float, 7, 0, 0, 4, 3); // for NORM2
+ADD_THREADWISE_INST_BY_ID(float, double, float, 7, 0, 0, 4, 4);       
 ADD_THREADWISE_INST_BY_ID(float, double, float, 7, 0, 0, 4, 1);       
 ADD_THREADWISE_INST_BY_ID(float, double, float, 7, 0, 0, 2, 1);
 // clang-format on

library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64.cpp

@@ -8,30 +8,39 @@ namespace device_reduce_instance {
 // clang-format off
 // InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | NumReduceDim
 ADD_THREADWISE_INST_BY_ID(double, double, double, 0, 0, 0, 4, 3); // for ADD
+ADD_THREADWISE_INST_BY_ID(double, double, double, 0, 0, 0, 4, 4);
 ADD_THREADWISE_INST_BY_ID(double, double, double, 0, 0, 0, 4, 1);
 ADD_THREADWISE_INST_BY_ID(double, double, double, 0, 0, 0, 2, 1);
 ADD_THREADWISE_INST_BY_ID(double, double, double, 5, 0, 0, 4, 3); // for AVG
+ADD_THREADWISE_INST_BY_ID(double, double, double, 5, 0, 0, 4, 4);       
 ADD_THREADWISE_INST_BY_ID(double, double, double, 5, 0, 0, 4, 1);       
 ADD_THREADWISE_INST_BY_ID(double, double, double, 5, 0, 0, 2, 1);       
 ADD_THREADWISE_INST_BY_ID(double, double, double, 7, 0, 0, 4, 3); // for NORM2
+ADD_THREADWISE_INST_BY_ID(double, double, double, 7, 0, 0, 4, 4);       
 ADD_THREADWISE_INST_BY_ID(double, double, double, 7, 0, 0, 4, 1);       
 ADD_THREADWISE_INST_BY_ID(double, double, double, 7, 0, 0, 2, 1);       
 ADD_THREADWISE_INST_BY_ID(double, double, double, 2, 0, 0, 4, 3); // for MIN
+ADD_THREADWISE_INST_BY_ID(double, double, double, 2, 0, 0, 4, 4);       
 ADD_THREADWISE_INST_BY_ID(double, double, double, 2, 0, 0, 4, 1);       
 ADD_THREADWISE_INST_BY_ID(double, double, double, 2, 0, 0, 2, 1);       
 ADD_THREADWISE_INST_BY_ID(double, double, double, 3, 0, 0, 4, 3); // for MAX
+ADD_THREADWISE_INST_BY_ID(double, double, double, 3, 0, 0, 4, 4);       
 ADD_THREADWISE_INST_BY_ID(double, double, double, 3, 0, 0, 4, 1);       
 ADD_THREADWISE_INST_BY_ID(double, double, double, 3, 0, 0, 2, 1);       
 ADD_THREADWISE_INST_BY_ID(double, double, double, 4, 0, 0, 4, 3); // for AMAX
+ADD_THREADWISE_INST_BY_ID(double, double, double, 4, 0, 0, 4, 4);       
 ADD_THREADWISE_INST_BY_ID(double, double, double, 4, 0, 0, 4, 1);       
 ADD_THREADWISE_INST_BY_ID(double, double, double, 4, 0, 0, 2, 1);       
 ADD_THREADWISE_INST_BY_ID(double, double, double, 2, 0, 1, 4, 3); // for MIN
+ADD_THREADWISE_INST_BY_ID(double, double, double, 2, 0, 1, 4, 4);       
 ADD_THREADWISE_INST_BY_ID(double, double, double, 2, 0, 1, 4, 1);       
 ADD_THREADWISE_INST_BY_ID(double, double, double, 2, 0, 1, 2, 1);       
 ADD_THREADWISE_INST_BY_ID(double, double, double, 3, 0, 1, 4, 3); // for MAX
+ADD_THREADWISE_INST_BY_ID(double, double, double, 3, 0, 1, 4, 4);       
 ADD_THREADWISE_INST_BY_ID(double, double, double, 3, 0, 1, 4, 1);       
 ADD_THREADWISE_INST_BY_ID(double, double, double, 3, 0, 1, 2, 1);       
 ADD_THREADWISE_INST_BY_ID(double, double, double, 4, 0, 1, 4, 3); // for AMAX
+ADD_THREADWISE_INST_BY_ID(double, double, double, 4, 0, 1, 4, 4);       
 ADD_THREADWISE_INST_BY_ID(double, double, double, 4, 0, 1, 4, 1);       
 ADD_THREADWISE_INST_BY_ID(double, double, double, 4, 0, 1, 2, 1);
 // clang-format on