Groupnorm + swish external api (#668)

* Rename to proper naming

* Add example of groupnorm + swish

* Extract duplicate code in example

* Add groupnorm + swish instances

* Ractor instance generation, split into multiple cpp file

* Add external api and client example

* Refine profiler message

* Use ck math version of exp

* Refine problem size in example

* Add host version of exp

library/src/tensor_operation_instance/gpu/normalization/CMakeLists.txt

@@ -1,4 +1,10 @@
 add_instance_library(device_normalization_instance
-    device_normalization_f16_instance.cpp
-    device_normalization_f32_instance.cpp
+    device_layernorm2d_f16_instance.cpp
+    device_layernorm2d_f32_instance.cpp
+    device_layernorm4d_f16_instance.cpp
+    device_layernorm4d_f32_instance.cpp
+    device_groupnorm_f16_instance.cpp
+    device_groupnorm_f32_instance.cpp
+    device_groupnorm_swish_f16_instance.cpp
+    device_groupnorm_swish_f32_instance.cpp
 )

library/src/tensor_operation_instance/gpu/normalization/device_groupnorm_f16_instance.cpp

@@ -0,0 +1,23 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "normalization_instance_common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using Pass = ck::tensor_operation::element_wise::PassThrough;
+
+void add_device_normalization_rank_5_3_f16_instances(
+    std::vector<std::unique_ptr<DeviceNormalization<F16, F16, F16, F32, F16, Pass, 5, 3>>>&
+        instances)
+{
+    add_device_operation_instances(instances, device_normalization_f16_instances<Pass, 5, 3>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/normalization/device_groupnorm_f32_instance.cpp

@@ -0,0 +1,23 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "normalization_instance_common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using Pass = ck::tensor_operation::element_wise::PassThrough;
+
+void add_device_normalization_rank_5_3_f32_instances(
+    std::vector<std::unique_ptr<DeviceNormalization<F32, F32, F32, F32, F32, Pass, 5, 3>>>&
+        instances)
+{
+    add_device_operation_instances(instances, device_normalization_f32_instances<Pass, 5, 3>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/normalization/device_groupnorm_swish_f16_instance.cpp

@@ -0,0 +1,23 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "normalization_instance_common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using Swish = ck::tensor_operation::element_wise::Swish;
+
+void add_device_normalization_rank_5_3_swish_f16_instances(
+    std::vector<std::unique_ptr<DeviceNormalization<F16, F16, F16, F32, F16, Swish, 5, 3>>>&
+        instances)
+{
+    add_device_operation_instances(instances, device_normalization_f16_instances<Swish, 5, 3>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

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

@@ -0,0 +1,23 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "normalization_instance_common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using Swish = ck::tensor_operation::element_wise::Swish;
+
+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_instances<Swish, 5, 3>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

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

@@ -0,0 +1,23 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "normalization_instance_common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using Pass = ck::tensor_operation::element_wise::PassThrough;
+
+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_instances<Pass, 2, 1>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

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

@@ -0,0 +1,23 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "normalization_instance_common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using Pass = ck::tensor_operation::element_wise::PassThrough;
+
+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_instances<Pass, 2, 1>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

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

@@ -0,0 +1,23 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "normalization_instance_common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using Pass = ck::tensor_operation::element_wise::PassThrough;
+
+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_instances<Pass, 4, 3>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

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

@@ -0,0 +1,23 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "normalization_instance_common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using Pass = ck::tensor_operation::element_wise::PassThrough;
+
+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_instances<Pass, 4, 3>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/normalization/device_normalization_f16_instance.cpp

@@ -1,70 +0,0 @@
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
-
-#include "ck/ck.hpp"
-#include "ck/tensor_operation/gpu/device/impl/device_normalization_impl.hpp"
-#include "ck/utility/data_type.hpp"
-
-#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
-
-namespace ck {
-namespace tensor_operation {
-namespace device {
-namespace instance {
-
-using F16 = ck::half_t;
-using F32 = float;
-
-using Pass = ck::tensor_operation::element_wise::PassThrough;
-
-template <typename OutElementwise, index_t Rank, index_t Reduce>
-// clang-format off
-using device_normalization_f16_instances =
-    std::tuple <
-        // XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize>
-        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
-        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
-        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
-        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
-        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 2, 1, 2, 1, 2, 1, 2, 2>,   // irregular size
-        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 4, 1, 4, 1, 4, 1, 4, 4>,   // irregular size
-        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 64, 1, 64, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
-        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
-        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 16, 1, 8, 1, 8, 1, 8, 8>,
-        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 32, 1, 8, 1, 8, 1, 8, 8>,
-        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
-        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 16, 1, 8, 1, 8, 1, 8, 8>,
-        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 2, 16, 1, 8, 1, 8, 1, 8, 8>,
-        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 8, 1, 8, 1, 8, 8>,
-        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
-        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 16, 1, 8, 1, 8, 1, 8, 8>,
-        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
-        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 16, 1, 8, 1, 8, 1, 8, 8>
-    >;
-// clang-format on
-
-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_instances<Pass, 2, 1>{});
-}
-
-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_instances<Pass, 4, 3>{});
-}
-
-void add_device_normalization_rank_5_3_f16_instances(
-    std::vector<std::unique_ptr<DeviceNormalization<F16, F16, F16, F32, F16, Pass, 5, 3>>>&
-        instances)
-{
-    add_device_operation_instances(instances, device_normalization_f16_instances<Pass, 5, 3>{});
-}
-
-} // namespace instance
-} // namespace device
-} // namespace tensor_operation
-} // namespace ck

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

@@ -1,6 +1,8 @@
 // SPDX-License-Identifier: MIT
 // Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
 
+#pragma once
+
 #include "ck/ck.hpp"
 #include "ck/tensor_operation/gpu/device/impl/device_normalization_impl.hpp"
 #include "ck/utility/data_type.hpp"
@@ -12,12 +14,37 @@ namespace tensor_operation {
 namespace device {
 namespace instance {
 
+using F16 = ck::half_t;
 using F32 = float;
 
-using Pass = ck::tensor_operation::element_wise::PassThrough;
+template <typename OutElementwise, index_t Rank, index_t Reduce>
+using device_normalization_f16_instances =
+    // clang-format off
+    std::tuple <
+        // XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize>
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 2, 1, 2, 1, 2, 1, 2, 2>,   // irregular size
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 4, 1, 4, 1, 4, 1, 4, 4>,   // irregular size
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 64, 1, 64, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 16, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 32, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 16, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 2, 16, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 16, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 16, 1, 8, 1, 8, 1, 8, 8>
+        // clang-format on
+        >;
 
 template <typename OutElementwise, index_t Rank, index_t Reduce>
-using device_layernorm_f32_instances = std::tuple<
+using device_normalization_f32_instances = std::tuple<
     // clang-format off
         // XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorSize, BetaSrcVectorSize, YDstVectorSize>
         DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
@@ -42,27 +69,6 @@ using device_layernorm_f32_instances = std::tuple<
     // clang-format on
     >;
 
-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_layernorm_f32_instances<Pass, 2, 1>{});
-}
-
-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_layernorm_f32_instances<Pass, 4, 3>{});
-}
-
-void add_device_normalization_rank_5_3_f32_instances(
-    std::vector<std::unique_ptr<DeviceNormalization<F32, F32, F32, F32, F32, Pass, 5, 3>>>&
-        instances)
-{
-    add_device_operation_instances(instances, device_layernorm_f32_instances<Pass, 5, 3>{});
-}
-
 } // namespace instance
 } // namespace device
 } // namespace tensor_operation