Merge branch 'develop' into gemm_layernorm_welford

library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_norm2.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#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 
+template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 3, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 3, UnarySquare, UnarySqrt>>&); 
+template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 4, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 4, UnarySquare, UnarySqrt>>&); 
+template void add_device_reduce_instance_threadwise<F32, F32, F32, 4, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 1, UnarySquare, UnarySqrt>>&); 
+template void add_device_reduce_instance_threadwise<F32, F32, F32, 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/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_add.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#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 
+template void add_device_reduce_instance_threadwise<F32, F64, F32, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&); 
+template void add_device_reduce_instance_threadwise<F32, F64, F32, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&); 
+template void add_device_reduce_instance_threadwise<F32, F64, F32, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&); 
+template void add_device_reduce_instance_threadwise<F32, F64, 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/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_avg.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#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 
+template void add_device_reduce_instance_threadwise<F32, F64, F32, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, UnaryDivide>>&); 
+template void add_device_reduce_instance_threadwise<F32, F64, F32, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, UnaryDivide>>&); 
+template void add_device_reduce_instance_threadwise<F32, F64, F32, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, UnaryDivide>>&); 
+template void add_device_reduce_instance_threadwise<F32, F64, F32, 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/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_norm2.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#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 
+template void add_device_reduce_instance_threadwise<F32, F64, F32, 4, 3, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 3, UnarySquare, UnarySqrt>>&); 
+template void add_device_reduce_instance_threadwise<F32, F64, F32, 4, 4, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 4, UnarySquare, UnarySqrt>>&); 
+template void add_device_reduce_instance_threadwise<F32, F64, F32, 4, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 1, UnarySquare, UnarySqrt>>&); 
+template void add_device_reduce_instance_threadwise<F32, F64, F32, 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/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64.cpp

-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
-
-#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_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
-
-} // namespace instance
-} // namespace device
-} // namespace tensor_operation
-} // namespace ck

library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_add.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#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 
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&); 
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&); 
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&); 
+template void add_device_reduce_instance_threadwise<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/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_amax.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#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 
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 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/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_avg.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#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 
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, UnaryDivide>>&); 
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, UnaryDivide>>&); 
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, UnaryDivide>>&); 
+template void add_device_reduce_instance_threadwise<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/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_max.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#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 
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 3, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 4, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 2, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 3, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 4, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 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/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_min.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#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 
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 3, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 4, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 2, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 3, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 4, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 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/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_norm2.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#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 
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 3, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 3, UnarySquare, UnarySqrt>>&); 
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 4, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 4, UnarySquare, UnarySqrt>>&); 
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 4, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 1, UnarySquare, UnarySqrt>>&); 
+template void add_device_reduce_instance_threadwise<F64, F64, F64, 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/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8.cpp → library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8_add.cpp

 // SPDX-License-Identifier: MIT
 // Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
 
+#include "ck/utility/reduction_enums.hpp"
 #include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp"
 
 namespace ck {
@@ -9,15 +10,11 @@ namespace device {
 namespace instance {
 
 // clang-format off
-// InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | NumReduceDim
-ADD_THREADWISE_INST_BY_ID(int8_t, int32_t, int8_t, 0, 0, 0, 4, 3); // for ADD
-ADD_THREADWISE_INST_BY_ID(int8_t, int32_t, int8_t, 0, 0, 0, 4, 4);
-ADD_THREADWISE_INST_BY_ID(int8_t, int32_t, int8_t, 0, 0, 0, 4, 1);
-ADD_THREADWISE_INST_BY_ID(int8_t, int32_t, int8_t, 0, 0, 0, 2, 1);
-ADD_THREADWISE_INST_BY_ID(int8_t, int32_t, int8_t, 5, 0, 0, 4, 3); // for AVG
-ADD_THREADWISE_INST_BY_ID(int8_t, int32_t, int8_t, 5, 0, 0, 4, 4);
-ADD_THREADWISE_INST_BY_ID(int8_t, int32_t, int8_t, 5, 0, 0, 4, 1);
-ADD_THREADWISE_INST_BY_ID(int8_t, int32_t, int8_t, 5, 0, 0, 2, 1);
+// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
+template void add_device_reduce_instance_threadwise<I8, I32, I8, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&); 
+template void add_device_reduce_instance_threadwise<I8, I32, I8, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&); 
+template void add_device_reduce_instance_threadwise<I8, I32, I8, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&); 
+template void add_device_reduce_instance_threadwise<I8, I32, I8, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
 // clang-format on
 // clang-format on
 

library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8_avg.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#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 
+template void add_device_reduce_instance_threadwise<I8, I32, I8, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, UnaryDivide>>&); 
+template void add_device_reduce_instance_threadwise<I8, I32, I8, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, UnaryDivide>>&); 
+template void add_device_reduce_instance_threadwise<I8, I32, I8, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, UnaryDivide>>&); 
+template void add_device_reduce_instance_threadwise<I8, I32, I8, 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/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8.cpp

-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
-
-#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_BY_ID(int8_t, int8_t, int8_t, 2, 0, 0, 4, 3); // for MIN
-ADD_THREADWISE_INST_BY_ID(int8_t, int8_t, int8_t, 2, 0, 0, 4, 4);       
-ADD_THREADWISE_INST_BY_ID(int8_t, int8_t, int8_t, 2, 0, 0, 4, 1);       
-ADD_THREADWISE_INST_BY_ID(int8_t, int8_t, int8_t, 2, 0, 0, 2, 1);       
-ADD_THREADWISE_INST_BY_ID(int8_t, int8_t, int8_t, 3, 0, 0, 4, 3); // for MAX
-ADD_THREADWISE_INST_BY_ID(int8_t, int8_t, int8_t, 3, 0, 0, 4, 4);       
-ADD_THREADWISE_INST_BY_ID(int8_t, int8_t, int8_t, 3, 0, 0, 4, 1);       
-ADD_THREADWISE_INST_BY_ID(int8_t, int8_t, int8_t, 3, 0, 0, 2, 1);       
-ADD_THREADWISE_INST_BY_ID(int8_t, int8_t, int8_t, 4, 0, 0, 4, 3); // for AMAX
-ADD_THREADWISE_INST_BY_ID(int8_t, int8_t, int8_t, 4, 0, 0, 4, 4);       
-ADD_THREADWISE_INST_BY_ID(int8_t, int8_t, int8_t, 4, 0, 0, 4, 1);       
-ADD_THREADWISE_INST_BY_ID(int8_t, int8_t, int8_t, 4, 0, 0, 2, 1);       
-ADD_THREADWISE_INST_BY_ID(int8_t, int8_t, int8_t, 2, 0, 1, 4, 3); // for MIN
-ADD_THREADWISE_INST_BY_ID(int8_t, int8_t, int8_t, 2, 0, 1, 4, 4);       
-ADD_THREADWISE_INST_BY_ID(int8_t, int8_t, int8_t, 2, 0, 1, 4, 1);       
-ADD_THREADWISE_INST_BY_ID(int8_t, int8_t, int8_t, 2, 0, 1, 2, 1);       
-ADD_THREADWISE_INST_BY_ID(int8_t, int8_t, int8_t, 3, 0, 1, 4, 3); // for MAX
-ADD_THREADWISE_INST_BY_ID(int8_t, int8_t, int8_t, 3, 0, 1, 4, 4);       
-ADD_THREADWISE_INST_BY_ID(int8_t, int8_t, int8_t, 3, 0, 1, 4, 1);       
-ADD_THREADWISE_INST_BY_ID(int8_t, int8_t, int8_t, 3, 0, 1, 2, 1);       
-ADD_THREADWISE_INST_BY_ID(int8_t, int8_t, int8_t, 4, 0, 1, 4, 3); // for AMAX
-ADD_THREADWISE_INST_BY_ID(int8_t, int8_t, int8_t, 4, 0, 1, 4, 4);       
-ADD_THREADWISE_INST_BY_ID(int8_t, int8_t, int8_t, 4, 0, 1, 4, 1);       
-ADD_THREADWISE_INST_BY_ID(int8_t, int8_t, int8_t, 4, 0, 1, 2, 1);
-// clang-format on
-
-} // namespace instance
-} // namespace device
-} // namespace tensor_operation
-
-} // namespace ck

library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_amax.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#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 
+template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<I8, I8, I8, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, UnaryAbs, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<I8, I8, I8, 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/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_max.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#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 
+template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 3, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 4, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<I8, I8, I8, 2, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 3, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 4, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<I8, I8, I8, 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/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_min.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#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 
+template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 3, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 4, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<I8, I8, I8, 2, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 3, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 4, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
+template void add_device_reduce_instance_threadwise<I8, I8, I8, 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/src/tensor_operation_instance/gpu/softmax/CMakeLists.txt

+add_instance_library(device_softmax_instance
+    device_softmax_i8_i8_instance.cpp
+    device_softmax_i8_i8_instance_rank3_reduce1.cpp
+    device_softmax_i8_i8_instance_rank3_reduce2.cpp
+    device_softmax_i8_i8_instance_rank3_reduce3.cpp
+    device_softmax_i8_i8_instance_rank4_reduce1.cpp
+    device_softmax_i8_i8_instance_rank4_reduce2.cpp
+    device_softmax_i8_i8_instance_rank4_reduce3.cpp
+    device_softmax_i8_i8_instance_rank4_reduce4.cpp
+    device_softmax_f16_f16_instance.cpp
+    device_softmax_f16_f16_instance_rank3_reduce1.cpp
+    device_softmax_f16_f16_instance_rank3_reduce2.cpp
+    device_softmax_f16_f16_instance_rank3_reduce3.cpp
+    device_softmax_f16_f16_instance_rank4_reduce1.cpp
+    device_softmax_f16_f16_instance_rank4_reduce2.cpp
+    device_softmax_f16_f16_instance_rank4_reduce3.cpp
+    device_softmax_f16_f16_instance_rank4_reduce4.cpp
+    device_softmax_f32_f32_instance.cpp
+    device_softmax_f32_f32_instance_rank3_reduce1.cpp
+    device_softmax_f32_f32_instance_rank3_reduce2.cpp
+    device_softmax_f32_f32_instance_rank3_reduce3.cpp
+    device_softmax_f32_f32_instance_rank4_reduce1.cpp
+    device_softmax_f32_f32_instance_rank4_reduce2.cpp
+    device_softmax_f32_f32_instance_rank4_reduce3.cpp
+    device_softmax_f32_f32_instance_rank4_reduce4.cpp
+)

library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce1.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce2.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce3.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce1.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce2.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce3.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce4.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_f16_f16_rank3_instances(
+    std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, 3>>& instances)
+{
+    add_device_softmax_f16_f16_rank3_reduce1_instances(instances);
+    add_device_softmax_f16_f16_rank3_reduce2_instances(instances);
+    add_device_softmax_f16_f16_rank3_reduce3_instances(instances);
+}
+
+void add_device_softmax_f16_f16_rank4_instances(
+    std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, 4>>& instances)
+{
+    add_device_softmax_f16_f16_rank4_reduce1_instances(instances);
+    add_device_softmax_f16_f16_rank4_reduce2_instances(instances);
+    add_device_softmax_f16_f16_rank4_reduce3_instances(instances);
+    add_device_softmax_f16_f16_rank4_reduce4_instances(instances);
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce1.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce1.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_type.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+static constexpr index_t RANK = 3;
+
+void add_device_softmax_f16_f16_rank3_reduce1_instances(
+    std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, RANK>>& instances)
+{
+    add_device_operation_instances(instances, device_softmax_f16_f16_instances<RANK, 1>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck