Padded Generic Kernel Instance (#730)

* Add NumReduceDim template parameter to DeviceSoftmax and Softmax client API to simplify instances collecting

* Move the generic kernel instance to be the first of the instance list for elementwise op of normalization

* Add GetGenericInstance() interface for DeviceOperationInstanceFactory class of DeviceSoftmax

* Add testing of GetGenericInstance() in client_example of Softmax

* Revert "Add testing of GetGenericInstance() in client_example of Softmax"

This reverts commit f629cd9a93ce38dfed4886d849f3c38d2e5379c8.

* Revert "Add GetGenericInstance() interface for DeviceOperationInstanceFactory class of DeviceSoftmax"

This reverts commit a9f0d000eb9fd240404112a526ef125429a351df.

* Support generic kernel instance to be the first instance returned by GetInstances() for GroupNorm

* Move generic kernel instance to separate tuple for elementwise op of normalization

* Remove un-used files for softmax instance

* Store generic kernel instance to separate tuple for softmax

* Add IsSupported checking for generic instance to client example of softmax

* Replace the get_device_normalize_from_mean_meansquare_instances() by the DeviceOperationInstanceFactory class for elementwise-normalization

* clang-format fix

* Remove int8 from softmax instances

---------
Co-authored-by: zjing14 <zhangjing14@gmail.com>

library/src/tensor_operation_instance/gpu/normalization/device_groupnorm_swish_f32_instance.cpp

@@ -14,6 +14,8 @@ void add_device_normalization_rank_5_3_swish_f32_instances(
     std::vector<std::unique_ptr<DeviceNormalization<F32, F32, F32, F32, F32, Swish, 5, 3>>>&
         instances)
 {
+    add_device_operation_instances(instances,
+                                   device_normalization_f32_generic_instance<Swish, 5, 3>{});
     add_device_operation_instances(instances, device_normalization_f32_instances<Swish, 5, 3>{});
 }
 

library/src/tensor_operation_instance/gpu/normalization/device_layernorm2d_f16_instance.cpp

@@ -14,6 +14,8 @@ void add_device_normalization_rank_2_1_f16_instances(
     std::vector<std::unique_ptr<DeviceNormalization<F16, F16, F16, F32, F16, Pass, 2, 1>>>&
         instances)
 {
+    add_device_operation_instances(instances,
+                                   device_normalization_f16_generic_instance<Pass, 2, 1>{});
     add_device_operation_instances(instances, device_normalization_f16_instances<Pass, 2, 1>{});
 }
 

library/src/tensor_operation_instance/gpu/normalization/device_layernorm2d_f32_instance.cpp

@@ -14,6 +14,8 @@ void add_device_normalization_rank_2_1_f32_instances(
     std::vector<std::unique_ptr<DeviceNormalization<F32, F32, F32, F32, F32, Pass, 2, 1>>>&
         instances)
 {
+    add_device_operation_instances(instances,
+                                   device_normalization_f32_generic_instance<Pass, 2, 1>{});
     add_device_operation_instances(instances, device_normalization_f32_instances<Pass, 2, 1>{});
 }
 

library/src/tensor_operation_instance/gpu/normalization/device_layernorm4d_f16_instance.cpp

@@ -14,6 +14,8 @@ void add_device_normalization_rank_4_3_f16_instances(
     std::vector<std::unique_ptr<DeviceNormalization<F16, F16, F16, F32, F16, Pass, 4, 3>>>&
         instances)
 {
+    add_device_operation_instances(instances,
+                                   device_normalization_f16_generic_instance<Pass, 4, 3>{});
     add_device_operation_instances(instances, device_normalization_f16_instances<Pass, 4, 3>{});
 }
 

library/src/tensor_operation_instance/gpu/normalization/device_layernorm4d_f32_instance.cpp

@@ -14,6 +14,8 @@ void add_device_normalization_rank_4_3_f32_instances(
     std::vector<std::unique_ptr<DeviceNormalization<F32, F32, F32, F32, F32, Pass, 4, 3>>>&
         instances)
 {
+    add_device_operation_instances(instances,
+                                   device_normalization_f32_generic_instance<Pass, 4, 3>{});
     add_device_operation_instances(instances, device_normalization_f32_instances<Pass, 4, 3>{});
 }
 

library/src/tensor_operation_instance/gpu/normalization/normalization_instance_common.hpp

@@ -43,6 +43,13 @@ using device_normalization_f16_instances =
         // clang-format on
         >;
 
+template <typename OutElementwise, index_t Rank, index_t Reduce>
+using device_normalization_f16_generic_instance = std::tuple<
+    // clang-format off
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 64, 1, 64, 1, 1, 1, 1, 1, 1, 1, 1, 1>
+    // clang-format on
+    >;
+
 template <typename OutElementwise, index_t Rank, index_t Reduce>
 using device_normalization_f32_instances = std::tuple<
     // clang-format off
@@ -69,6 +76,13 @@ using device_normalization_f32_instances = std::tuple<
     // clang-format on
     >;
 
+template <typename OutElementwise, index_t Rank, index_t Reduce>
+using device_normalization_f32_generic_instance = std::tuple<
+    // clang-format off
+        DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 64, 1, 64, 1, 1, 1, 1, 1, 1, 1, 1, 1>
+    // clang-format on
+    >;
+
 template <typename OutElementwise, index_t Rank, index_t Reduce>
 using device_normalization_f16_f32_f32_f16_instances = std::tuple<
     // clang-format off
@@ -95,6 +109,13 @@ using device_normalization_f16_f32_f32_f16_instances = std::tuple<
     // clang-format on
     >;
 
+template <typename OutElementwise, index_t Rank, index_t Reduce>
+using device_normalization_f16_f32_f32_f16_generic_instance = std::tuple<
+    // clang-format off
+        DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 64, 1, 64, 1, 1, 1, 1, 1, 1, 1, 1, 1>
+    // clang-format on
+    >;
+
 } // namespace instance
 } // namespace device
 } // namespace tensor_operation

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
@@ -15,7 +6,6 @@ add_instance_library(device_softmax_instance
     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

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

-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2023, 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

@@ -13,12 +13,11 @@ 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)
+    std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, 3, 1>>& instances)
 {
-    add_device_operation_instances(instances, device_softmax_f16_f16_instances<RANK, 1>{});
+    add_device_operation_instances(instances, device_softmax_f16_f16_generic_instance<3, 1>{});
+    add_device_operation_instances(instances, device_softmax_f16_f16_instances<3, 1>{});
 }
 
 } // namespace instance

library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce2.cpp

@@ -13,12 +13,11 @@ 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)
+    std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, 3, 2>>& instances)
 {
-    add_device_operation_instances(instances, device_softmax_f16_f16_instances<RANK, 2>{});
+    add_device_operation_instances(instances, device_softmax_f16_f16_generic_instance<3, 2>{});
+    add_device_operation_instances(instances, device_softmax_f16_f16_instances<3, 2>{});
 }
 
 } // namespace instance

library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce3.cpp

@@ -13,12 +13,11 @@ 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)
+    std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, 3, 3>>& instances)
 {
-    add_device_operation_instances(instances, device_softmax_f16_f16_instances<RANK, 3>{});
+    add_device_operation_instances(instances, device_softmax_f16_f16_generic_instance<3, 3>{});
+    add_device_operation_instances(instances, device_softmax_f16_f16_instances<3, 3>{});
 }
 
 } // namespace instance

library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce1.cpp

@@ -13,12 +13,11 @@ 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)
+    std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, 4, 1>>& instances)
 {
-    add_device_operation_instances(instances, device_softmax_f16_f16_instances<RANK, 1>{});
+    add_device_operation_instances(instances, device_softmax_f16_f16_generic_instance<4, 1>{});
+    add_device_operation_instances(instances, device_softmax_f16_f16_instances<4, 1>{});
 }
 
 } // namespace instance

library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce2.cpp

@@ -13,12 +13,11 @@ namespace tensor_operation {
 namespace device {
 namespace instance {
 
-static constexpr index_t RANK = 4;
-
 void add_device_softmax_f16_f16_rank4_reduce2_instances(
-    std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, RANK>>& instances)
+    std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, 4, 2>>& instances)
 {
-    add_device_operation_instances(instances, device_softmax_f16_f16_instances<RANK, 2>{});
+    add_device_operation_instances(instances, device_softmax_f16_f16_generic_instance<4, 2>{});
+    add_device_operation_instances(instances, device_softmax_f16_f16_instances<4, 2>{});
 }
 
 } // namespace instance

library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce3.cpp

@@ -13,12 +13,11 @@ namespace tensor_operation {
 namespace device {
 namespace instance {
 
-static constexpr index_t RANK = 4;
-
 void add_device_softmax_f16_f16_rank4_reduce3_instances(
-    std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, RANK>>& instances)
+    std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, 4, 3>>& instances)
 {
-    add_device_operation_instances(instances, device_softmax_f16_f16_instances<RANK, 3>{});
+    add_device_operation_instances(instances, device_softmax_f16_f16_generic_instance<4, 3>{});
+    add_device_operation_instances(instances, device_softmax_f16_f16_instances<4, 3>{});
 }
 
 } // namespace instance

library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce4.cpp

@@ -13,12 +13,11 @@ namespace tensor_operation {
 namespace device {
 namespace instance {
 
-static constexpr index_t RANK = 4;
-
 void add_device_softmax_f16_f16_rank4_reduce4_instances(
-    std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, RANK>>& instances)
+    std::vector<DeviceSoftmaxPtr<F16, F32, F16, PassThrough, PassThrough, 4, 4>>& instances)
 {
-    add_device_operation_instances(instances, device_softmax_f16_f16_instances<RANK, 4>{});
+    add_device_operation_instances(instances, device_softmax_f16_f16_generic_instance<4, 4>{});
+    add_device_operation_instances(instances, device_softmax_f16_f16_instances<4, 4>{});
 }
 
 } // namespace instance

library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance.cpp

-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
-
-#include <vector>
-
-#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce1.hpp"
-#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce2.hpp"
-#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce3.hpp"
-
-#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce1.hpp"
-#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce2.hpp"
-#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce3.hpp"
-#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce4.hpp"
-
-namespace ck {
-namespace tensor_operation {
-namespace device {
-namespace instance {
-
-void add_device_softmax_f32_f32_rank3_instances(
-    std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 3>>& instances)
-{
-    add_device_softmax_f32_f32_rank3_reduce1_instances(instances);
-    add_device_softmax_f32_f32_rank3_reduce2_instances(instances);
-    add_device_softmax_f32_f32_rank3_reduce3_instances(instances);
-}
-
-void add_device_softmax_f32_f32_rank4_instances(
-    std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 4>>& instances)
-{
-    add_device_softmax_f32_f32_rank4_reduce1_instances(instances);
-    add_device_softmax_f32_f32_rank4_reduce2_instances(instances);
-    add_device_softmax_f32_f32_rank4_reduce3_instances(instances);
-    add_device_softmax_f32_f32_rank4_reduce4_instances(instances);
-}
-
-} // namespace instance
-} // namespace device
-} // namespace tensor_operation
-} // namespace ck

library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce1.cpp

@@ -13,12 +13,11 @@ namespace tensor_operation {
 namespace device {
 namespace instance {
 
-static constexpr index_t RANK = 3;
-
 void add_device_softmax_f32_f32_rank3_reduce1_instances(
-    std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, RANK>>& instances)
+    std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 3, 1>>& instances)
 {
-    add_device_operation_instances(instances, device_softmax_f32_f32_instances<RANK, 1>{});
+    add_device_operation_instances(instances, device_softmax_f32_f32_generic_instance<3, 1>{});
+    add_device_operation_instances(instances, device_softmax_f32_f32_instances<3, 1>{});
 }
 
 } // namespace instance

library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce2.cpp

@@ -13,12 +13,11 @@ namespace tensor_operation {
 namespace device {
 namespace instance {
 
-static constexpr index_t RANK = 3;
-
 void add_device_softmax_f32_f32_rank3_reduce2_instances(
-    std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, RANK>>& instances)
+    std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 3, 2>>& instances)
 {
-    add_device_operation_instances(instances, device_softmax_f32_f32_instances<RANK, 2>{});
+    add_device_operation_instances(instances, device_softmax_f32_f32_generic_instance<3, 2>{});
+    add_device_operation_instances(instances, device_softmax_f32_f32_instances<3, 2>{});
 }
 
 } // namespace instance

library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce3.cpp

@@ -13,12 +13,11 @@ namespace tensor_operation {
 namespace device {
 namespace instance {
 
-static constexpr index_t RANK = 3;
-
 void add_device_softmax_f32_f32_rank3_reduce3_instances(
-    std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, RANK>>& instances)
+    std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 3, 3>>& instances)
 {
-    add_device_operation_instances(instances, device_softmax_f32_f32_instances<RANK, 3>{});
+    add_device_operation_instances(instances, device_softmax_f32_f32_generic_instance<3, 3>{});
+    add_device_operation_instances(instances, device_softmax_f32_f32_instances<3, 3>{});
 }
 
 } // namespace instance

library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce1.cpp

@@ -13,12 +13,11 @@ namespace tensor_operation {
 namespace device {
 namespace instance {
 
-static constexpr index_t RANK = 4;
-
 void add_device_softmax_f32_f32_rank4_reduce1_instances(
-    std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, RANK>>& instances)
+    std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 4, 1>>& instances)
 {
-    add_device_operation_instances(instances, device_softmax_f32_f32_instances<RANK, 1>{});
+    add_device_operation_instances(instances, device_softmax_f32_f32_generic_instance<4, 1>{});
+    add_device_operation_instances(instances, device_softmax_f32_f32_instances<4, 1>{});
 }
 
 } // namespace instance