Softmax unit-test reduction across all and non innermost dims cases. (#406)

* Add reduction across all dims cases.

* host softmax: handle all reduce

* Test cases when reduced dim is not innermost axis.

* Fix syntax.

* Test non innermost dim for fp32 and int8

* Group test suites wrt NumReduceDim.

* Additionally test failing cases.

* Throw error when Rank or NumReduceDims doesn't match arguments.

* Check reducedDims has correct values

* Move don't reuse DeviceReduceMultiblock IsSupportedArgument method.
Instead implement own. (in fact just get rid of one check to enable
reduction across inner dimensions).

* Reorganize unit tests to better cover use scenarios.

* Test input validation
* Test reduction of inner dimensions with custom op instances.

* Refactor fp32 and int8 unit tests.

* Fix FP32 instance template parameters.

* Add more instances.

* Instances with InSrcVectorDim=0.

* Do not initialize and copy data when arg not supported.

* ckProfiler Softmax use instance factory.

* Refactor device softmax IsSupported.

* Additionally add non-polymorphic api functions

* Split softmax instances into multiple files.

* Fix profiler.

* Reorganize tests to reuse profiler and cover edge cases.

* Clang-format

* I8 Softmax instances along with UT.

* Reuse type alias definitions from instance factory header.

* Clean included headers

* Fix variable names.

* Add missing checks in Argument constructor.
Co-authored-by: Adam Osewski <aosewski@amd.com>
Co-authored-by: Anthony Chang <ac.chang@outlook.com>

library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce2.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_f32_f32_rank4_reduce2_instances(
+    std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 4>>& instances);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce3.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_f32_f32_rank4_reduce3_instances(
+    std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 4>>& instances);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce4.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_f32_f32_rank4_reduce4_instances(
+    std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 4>>& instances);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_type.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <tuple>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_softmax_impl.hpp"
+#include "ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+template <index_t Rank, index_t Reduce>
+using device_softmax_f32_f32_instances = std::tuple<
+    // clang-format off
+    //                InDataType, AccDataType, OutDataType, InElementwiseOp, AccElementwiseOp, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, InSrcVectorDim, InSrcVectorSize, OutDstVectorSize>
+    DeviceSoftmaxImpl<       F32,         F32,         F32,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  8,                 32,                1,                8,              1,               1,               1>, // fallback kernel
+    DeviceSoftmaxImpl<       F32,         F32,         F32,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  8,                 32,                1,                8,              1,               4,               4>,
+    DeviceSoftmaxImpl<       F32,         F32,         F32,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  4,                 64,                1,                8,              1,               4,               4>,
+    DeviceSoftmaxImpl<       F32,         F32,         F32,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  2,                128,                1,                8,              1,               4,               4>,
+    DeviceSoftmaxImpl<       F32,         F32,         F32,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  2,                128,                1,               16,              1,               4,               4>,
+    DeviceSoftmaxImpl<       F32,         F32,         F32,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  2,                128,                1,               32,              1,               4,               4>,
+    DeviceSoftmaxImpl<       F32,         F32,         F32,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  1,                256,                1,                8,              1,               4,               4>,
+    DeviceSoftmaxImpl<       F32,         F32,         F32,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  1,                256,                1,               16,              1,               4,               4>,
+    DeviceSoftmaxImpl<       F32,         F32,         F32,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  1,                256,                1,               32,              1,               4,               4>,
+    // Reduction on middle dimensions
+    // InSrcVectorDim is 0 since we want to coalesce reads on M dimension
+    DeviceSoftmaxImpl<       F32,         F32,         F32,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  8,                 32,                8,                4,              0,               1,               1>,
+    DeviceSoftmaxImpl<       F32,         F32,         F32,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  8,                 32,                8,                4,              0,               4,               4>
+    // clang-format on
+    >;
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_i8_i8_rank3_instances(
+    std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, 3>>& instances);
+void add_device_softmax_i8_i8_rank4_instances(
+    std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, 4>>& instances);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank3_reduce1.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_i8_i8_rank3_reduce1_instances(
+    std::vector<DeviceSoftmaxPtr<I8, F32, I8, PassThrough, PassThrough, 3>>& instances);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank3_reduce2.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_i8_i8_rank3_reduce2_instances(
+    std::vector<DeviceSoftmaxPtr<I8, F32, I8, PassThrough, PassThrough, 3>>& instances);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank3_reduce3.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_i8_i8_rank3_reduce3_instances(
+    std::vector<DeviceSoftmaxPtr<I8, F32, I8, PassThrough, PassThrough, 3>>& instances);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank4_reduce1.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_i8_i8_rank4_reduce1_instances(
+    std::vector<DeviceSoftmaxPtr<I8, F32, I8, PassThrough, PassThrough, 4>>& instances);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank4_reduce2.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_i8_i8_rank4_reduce2_instances(
+    std::vector<DeviceSoftmaxPtr<I8, F32, I8, PassThrough, PassThrough, 4>>& instances);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank4_reduce3.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_i8_i8_rank4_reduce3_instances(
+    std::vector<DeviceSoftmaxPtr<I8, F32, I8, PassThrough, PassThrough, 4>>& instances);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank4_reduce4.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_softmax_i8_i8_rank4_reduce4_instances(
+    std::vector<DeviceSoftmaxPtr<I8, F32, I8, PassThrough, PassThrough, 4>>& instances);
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_type.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <tuple>
+
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_softmax_impl.hpp"
+#include "ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+template <index_t Rank, index_t Reduce>
+using device_softmax_i8_i8_instances = std::tuple<
+    // clang-format off
+    //                InDataType, AccDataType, OutDataType, InElementwiseOp, AccElementwiseOp, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, InSrcVectorDim, InSrcVectorSize, OutDstVectorSize>
+    // fallback kernel
+    DeviceSoftmaxImpl<        I8,         F32,          I8,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  8,                 32,                1,               16,              1,               1,              1>,
+    DeviceSoftmaxImpl<        I8,         F32,          I8,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  8,                 32,                1,               16,              1,              16,             16>,
+    DeviceSoftmaxImpl<        I8,         F32,          I8,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  4,                 64,                1,               16,              1,              16,             16>,
+    DeviceSoftmaxImpl<        I8,         F32,          I8,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  2,                128,                1,               16,              1,              16,             16>,
+    DeviceSoftmaxImpl<        I8,         F32,          I8,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  2,                128,                1,               32,              1,              16,             16>,
+    DeviceSoftmaxImpl<        I8,         F32,          I8,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  2,                128,                1,               64,              1,              16,             16>,
+    DeviceSoftmaxImpl<        I8,         F32,          I8,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  1,                256,                1,               16,              1,              16,             16>,
+    DeviceSoftmaxImpl<        I8,         F32,          I8,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  1,                256,                1,               32,              1,              16,             16>,
+    DeviceSoftmaxImpl<        I8,         F32,          I8,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  1,                256,                1,               64,              1,              16,             16>,
+    // Reduction on middle dimensions
+    // InSrcVectorDim is 0 since we want to coalesce reads on M dimension
+    DeviceSoftmaxImpl<        I8,         F32,          I8,     PassThrough,      PassThrough, Rank,       Reduce,       256,                  8,                 32,                8,                8,              0,               1,              1>,
+    DeviceSoftmaxImpl<        I8,         F32,          I8,     PassThrough,      PassThrough, Rank,       Reduce,       256,                 32,                  8,               32,                8,              0,              16,              8>
+    // clang-format on
+    >;
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_instance.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance.hpp"

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 <tuple>
 #include <vector>
 
-#include "ck/ck.hpp"
-#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
-#include "ck/tensor_operation/gpu/device/impl/device_softmax_impl.hpp"
-#include "ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp"
-#include "ck/utility/data_type.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_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 {
 
-namespace {
-using F16  = ck::half_t;
-using F32  = float;
-using Pass = ck::tensor_operation::element_wise::PassThrough;
-} // namespace
-
-template <index_t Rank, index_t Reduce>
-using device_softmax_f16_f16_instances = std::tuple<
-    // clang-format off
-    //                InDataType, AccDataType, OutDataType, InElementwiseOp, AccElementwiseOp, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, InSrcVectorDim, InSrcVectorSize, OutDstVectorSize>
-    DeviceSoftmaxImpl<       F16,         F32,         F16,            Pass,             Pass, Rank,       Reduce,       256,                  8,                 32,                1,                8,              1,               1,              1>, // fallback kernel
-    DeviceSoftmaxImpl<       F16,         F32,         F16,            Pass,             Pass, Rank,       Reduce,       256,                  8,                 32,                1,                8,              1,               8,              8>,
-    DeviceSoftmaxImpl<       F16,         F32,         F16,            Pass,             Pass, Rank,       Reduce,       256,                  4,                 64,                1,                8,              1,               8,              8>,
-    DeviceSoftmaxImpl<       F16,         F32,         F16,            Pass,             Pass, Rank,       Reduce,       256,                  2,                128,                1,                8,              1,               8,              8>,
-    DeviceSoftmaxImpl<       F16,         F32,         F16,            Pass,             Pass, Rank,       Reduce,       256,                  2,                128,                1,               16,              1,               8,              8>,
-    DeviceSoftmaxImpl<       F16,         F32,         F16,            Pass,             Pass, Rank,       Reduce,       256,                  2,                128,                1,               32,              1,               8,              8>,
-    DeviceSoftmaxImpl<       F16,         F32,         F16,            Pass,             Pass, Rank,       Reduce,       256,                  1,                256,                1,                8,              1,               8,              8>,
-    DeviceSoftmaxImpl<       F16,         F32,         F16,            Pass,             Pass, Rank,       Reduce,       256,                  1,                256,                1,               16,              1,               8,              8>,
-    DeviceSoftmaxImpl<       F16,         F32,         F16,            Pass,             Pass, Rank,       Reduce,       256,                  1,                256,                1,               32,              1,               8,              8>
-    // clang-format on
-    >;
-
 void add_device_softmax_f16_f16_rank3_instances(
-    std::vector<DeviceSoftmaxPtr<F16, F32, F16, Pass, Pass, 3>>& instances)
+    std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, 3>>& instances)
 {
-    add_device_operation_instances(instances, device_softmax_f16_f16_instances<3, 1>{});
-    add_device_operation_instances(instances, device_softmax_f16_f16_instances<3, 2>{});
+    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, Pass, Pass, 4>>& instances)
+    std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, 4>>& instances)
 {
-    add_device_operation_instances(instances, device_softmax_f16_f16_instances<4, 1>{});
-    add_device_operation_instances(instances, device_softmax_f16_f16_instances<4, 2>{});
-    add_device_operation_instances(instances, device_softmax_f16_f16_instances<4, 3>{});
+    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

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

library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce2.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_reduce2.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_reduce2_instances(
+    std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, RANK>>& instances)
+{
+    add_device_operation_instances(instances, device_softmax_f16_f16_instances<RANK, 2>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce3.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_reduce3.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_reduce3_instances(
+    std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, RANK>>& instances)
+{
+    add_device_operation_instances(instances, device_softmax_f16_f16_instances<RANK, 3>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_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_rank4_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 = 4;
+
+void add_device_softmax_f16_f16_rank4_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