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/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce1.hpp

@@ -14,7 +14,7 @@ namespace device {
 namespace instance {
 
 void add_device_softmax_f32_f32_rank4_reduce1_instances(
-    std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 4>>& instances);
+    std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 4, 1>>& instances);
 
 } // namespace instance
 } // namespace device

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

@@ -14,7 +14,7 @@ namespace device {
 namespace instance {
 
 void add_device_softmax_f32_f32_rank4_reduce2_instances(
-    std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 4>>& instances);
+    std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 4, 2>>& instances);
 
 } // namespace instance
 } // namespace device

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

@@ -14,7 +14,7 @@ namespace device {
 namespace instance {
 
 void add_device_softmax_f32_f32_rank4_reduce3_instances(
-    std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 4>>& instances);
+    std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 4, 3>>& instances);
 
 } // namespace instance
 } // namespace device

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

@@ -14,7 +14,7 @@ namespace device {
 namespace instance {
 
 void add_device_softmax_f32_f32_rank4_reduce4_instances(
-    std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 4>>& instances);
+    std::vector<DeviceSoftmaxPtr<F32, F32, F32, PassThrough, PassThrough, 4, 4>>& instances);
 
 } // namespace instance
 } // namespace device

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

@@ -16,7 +16,7 @@ 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,               1,               1>,
     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>,
@@ -32,6 +32,13 @@ using device_softmax_f32_f32_instances = std::tuple<
     // clang-format on
     >;
 
+template <index_t Rank, index_t Reduce>
+using device_softmax_f32_f32_generic_instance = std::tuple<
+    // clang-format off
+    DeviceSoftmaxImpl<       F32,         F32,         F32,     PassThrough,      PassThrough, Rank,       Reduce,        64,                  8,                  8,                1,                1,              1,               1,               1>
+    // clang-format on
+    >;
+
 } // namespace instance
 } // namespace device
 } // namespace tensor_operation

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

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

@@ -3,6 +3,17 @@
 
 #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"
+#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"
+#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"

library/src/tensor_operation_instance/gpu/elementwise/device_normalize_instance.cpp

@@ -30,7 +30,12 @@ using device_normalize_from_mean_squaremean_f16_f32_f32_f16_f16_instances = std:
     //###################|<in, mean, square_mean, gamma, beta>| <out>|  functor| NDim| MPerThread| <in, mean, square_mean, gamma, beta ScalarPerVector>| <out ScalarPerVector>|
     DeviceElementwiseImpl<Tuple<F16, F32, F32, F16, F16>,  Tuple<F16>,  Normalize,  2,   8,       Sequence<8, 1, 1, 8, 8>,      Sequence<8>                >,
     DeviceElementwiseImpl<Tuple<F16, F32, F32, F16, F16>,  Tuple<F16>,  Normalize,  2,   4,       Sequence<4, 1, 1, 4, 4>,      Sequence<4>                >,
-    DeviceElementwiseImpl<Tuple<F16, F32, F32, F16, F16>,  Tuple<F16>,  Normalize,  2,   2,       Sequence<2, 1, 1, 2, 2>,      Sequence<2>                >,
+    DeviceElementwiseImpl<Tuple<F16, F32, F32, F16, F16>,  Tuple<F16>,  Normalize,  2,   2,       Sequence<2, 1, 1, 2, 2>,      Sequence<2>                >
+    // clang-format on
+    >;
+
+using device_normalize_from_mean_squaremean_f16_f32_f32_f16_f16_generic_instance = std::tuple<
+    // clang-format off
     DeviceElementwiseImpl<Tuple<F16, F32, F32, F16, F16>,  Tuple<F16>,  Normalize,  2,   1,       Sequence<1, 1, 1, 1, 1>,      Sequence<1>                >
     // clang-format on
     >;
@@ -39,6 +44,9 @@ void add_device_normalize_from_mean_squaremean_f16_f32_f32_f16_f16_instances(
     std::vector<DeviceElementwisePtr<Tuple<F16, F32, F32, F16, F16>, Tuple<F16>, Normalize, 2>>&
         instances)
 {
+    add_device_operation_instances(
+        instances, device_normalize_from_mean_squaremean_f16_f32_f32_f16_f16_generic_instance{});
+
     add_device_operation_instances(
         instances, device_normalize_from_mean_squaremean_f16_f32_f32_f16_f16_instances{});
 }

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

@@ -14,6 +14,8 @@ 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_generic_instance<Pass, 5, 3>{});
     add_device_operation_instances(instances, device_normalization_f16_instances<Pass, 5, 3>{});
 }
 

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

@@ -14,6 +14,8 @@ 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_generic_instance<Pass, 5, 3>{});
     add_device_operation_instances(instances, device_normalization_f32_instances<Pass, 5, 3>{});
 }
 

library/src/tensor_operation_instance/gpu/normalization/device_groupnorm_swish_f16_f32_f32_f16_instance.cpp

@@ -14,6 +14,8 @@ void add_device_normalization_rank_5_3_swish_f16_f32_f32_f16_instances(
     std::vector<std::unique_ptr<DeviceNormalization<F16, F32, F32, F32, F16, Swish, 5, 3>>>&
         instances)
 {
+    add_device_operation_instances(
+        instances, device_normalization_f16_f32_f32_f16_generic_instance<Swish, 5, 3>{});
     add_device_operation_instances(instances,
                                    device_normalization_f16_f32_f32_f16_instances<Swish, 5, 3>{});
 }

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

@@ -14,6 +14,8 @@ 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_generic_instance<Swish, 5, 3>{});
     add_device_operation_instances(instances, device_normalization_f16_instances<Swish, 5, 3>{});
 }