Merge branch 'develop' into amd-develop

library/include/ck/library/tensor_operation_instance/gpu/layernorm_bwd_data.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+#include <memory>
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_normalization_bwd_data.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+#ifdef CK_ENABLE_FP16
+// FP16
+void add_device_layernorm2d_bwd_data_f16_instances(
+    std::vector<std::unique_ptr<DeviceNormalizationBwdData<F16, F16, F16, F16, F16, 2, 1>>>&);
+#endif
+#ifdef CK_ENABLE_FP32
+// FP32
+void add_device_layernorm2d_bwd_data_f32_instances(
+    std::vector<std::unique_ptr<DeviceNormalizationBwdData<F32, F32, F32, F32, F32, 2, 1>>>&);
+#endif
+template <typename DYDataType,
+          typename XDataType,
+          typename GammaDataType,
+          typename MeanInvStdDataType,
+          typename DXDataType,
+          index_t Rank,
+          index_t NumReduceDim>
+struct DeviceOperationInstanceFactory<
+    ck::tensor_operation::device::DeviceNormalizationBwdData<DYDataType,
+                                                             XDataType,
+                                                             GammaDataType,
+                                                             MeanInvStdDataType,
+                                                             DXDataType,
+                                                             Rank,
+                                                             NumReduceDim>>
+{
+    using DeviceOp = DeviceNormalizationBwdData<DYDataType,
+                                                XDataType,
+                                                GammaDataType,
+                                                MeanInvStdDataType,
+                                                DXDataType,
+                                                Rank,
+                                                NumReduceDim>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+#ifdef CK_ENABLE_FP16
+        if constexpr(is_same_v<DYDataType, F16> && is_same_v<XDataType, F16> &&
+                     is_same_v<GammaDataType, F16> && is_same_v<MeanInvStdDataType, F16> &&
+                     is_same_v<DXDataType, F16>)
+        {
+            if constexpr(Rank == 2 && NumReduceDim == 1)
+            {
+                add_device_layernorm2d_bwd_data_f16_instances(op_ptrs);
+            }
+        }
+#endif
+#ifdef CK_ENABLE_FP32
+        if constexpr(is_same_v<DYDataType, F32> && is_same_v<XDataType, F32> &&
+                     is_same_v<GammaDataType, F32> && is_same_v<MeanInvStdDataType, F32> &&
+                     is_same_v<DXDataType, F32>)
+        {
+            if constexpr(Rank == 2 && NumReduceDim == 1)
+            {
+                add_device_layernorm2d_bwd_data_f32_instances(op_ptrs);
+            }
+        }
+#endif
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/include/ck/library/tensor_operation_instance/gpu/normalization_fwd.hpp

@@ -20,15 +20,15 @@ namespace instance {
 // FP16
 void add_device_normalization_fwd_rank_2_1_f16_instances(
     std::vector<
-        std::unique_ptr<DeviceNormalizationFwd<F16, F16, F16, F16, F32, PassThrough, 2, 1>>>&);
+        std::unique_ptr<DeviceNormalizationFwd<F16, F16, F16, F16, F16, PassThrough, 2, 1>>>&);
 
 void add_device_normalization_fwd_rank_4_3_f16_instances(
     std::vector<
-        std::unique_ptr<DeviceNormalizationFwd<F16, F16, F16, F16, F32, PassThrough, 4, 3>>>&);
+        std::unique_ptr<DeviceNormalizationFwd<F16, F16, F16, F16, F16, PassThrough, 4, 3>>>&);
 
 void add_device_normalization_fwd_rank_5_3_f16_instances(
     std::vector<
-        std::unique_ptr<DeviceNormalizationFwd<F16, F16, F16, F16, F32, PassThrough, 5, 3>>>&);
+        std::unique_ptr<DeviceNormalizationFwd<F16, F16, F16, F16, F16, PassThrough, 5, 3>>>&);
 #endif
 #ifdef CK_ENABLE_FP32
 // FP32
@@ -76,7 +76,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceNormal
 #ifdef CK_ENABLE_FP16
         if constexpr(is_same_v<XDataType, F16> && is_same_v<GammaDataType, F16> &&
                      is_same_v<BetaDataType, F16> && is_same_v<YDataType, F16> &&
-                     is_same_v<SaveMeanInvStdDataType, F32>)
+                     is_same_v<SaveMeanInvStdDataType, F16>)
         {
             if constexpr(Rank == 2 && NumReduceDim == 1)
             {

library/include/ck/library/tensor_operation_instance/gpu/normalization_fwd_swish.hpp

@@ -19,7 +19,7 @@ namespace instance {
 
 // FP16
 void add_device_normalization_fwd_rank_5_3_swish_f16_instances(
-    std::vector<std::unique_ptr<DeviceNormalizationFwd<F16, F16, F16, F16, F32, Swish, 5, 3>>>&);
+    std::vector<std::unique_ptr<DeviceNormalizationFwd<F16, F16, F16, F16, F16, Swish, 5, 3>>>&);
 
 // FP32
 void add_device_normalization_fwd_rank_5_3_swish_f32_instances(
@@ -61,7 +61,7 @@ struct DeviceOperationInstanceFactory<
 
         if constexpr(is_same_v<XDataType, F16> && is_same_v<GammaDataType, F16> &&
                      is_same_v<BetaDataType, F16> && is_same_v<YDataType, F16> &&
-                     is_same_v<SaveMeanInvStdDataType, F32>)
+                     is_same_v<SaveMeanInvStdDataType, F16>)
         {
             if constexpr(Rank == 5 && NumReduceDim == 3)
             {

library/include/ck/library/tensor_operation_instance/gpu/permute_scale.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_elementwise_scale.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_permute_scale_f16_instances(
+    std::vector<std::unique_ptr<DeviceElementwise<ck::Tuple<F16>,
+                                                  ck::Tuple<F16>,
+                                                  PassThrough,
+                                                  element_wise::UnarySquare,
+                                                  Scale,
+                                                  4>>>&);
+
+void add_device_permute_scale_f32_instances(
+    std::vector<std::unique_ptr<DeviceElementwise<ck::Tuple<F32>,
+                                                  ck::Tuple<F32>,
+                                                  PassThrough,
+                                                  element_wise::UnarySquare,
+                                                  Scale,
+                                                  4>>>&);
+
+template <typename InDataTypeTuple,
+          typename OutDataTypeTuple,
+          typename ElementwiseOperation,
+          typename UnaryOperation,
+          typename Scale,
+          index_t NumDim>
+struct DeviceOperationInstanceFactory<
+    ck::tensor_operation::device::DeviceElementwise<InDataTypeTuple,
+                                                    OutDataTypeTuple,
+                                                    ElementwiseOperation,
+                                                    UnaryOperation,
+                                                    Scale,
+                                                    NumDim>>
+{
+    using DeviceOp = DeviceElementwise<InDataTypeTuple,
+                                       OutDataTypeTuple,
+                                       ElementwiseOperation,
+                                       UnaryOperation,
+                                       Scale,
+                                       NumDim>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+        if constexpr(is_same_v<InDataTypeTuple, ck::Tuple<F32>> &&
+                     is_same_v<OutDataTypeTuple, ck::Tuple<F32>>)
+        {
+            add_device_permute_scale_f32_instances(op_ptrs);
+        }
+        else if constexpr(is_same_v<InDataTypeTuple, ck::Tuple<F16>> &&
+                          is_same_v<OutDataTypeTuple, ck::Tuple<F16>>)
+        {
+            add_device_permute_scale_f16_instances(op_ptrs);
+        }
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_lds_direct_load_f16_f16_f16_mk_nk_mn_instance.cpp

@@ -34,22 +34,19 @@ using device_gemm_xdl_c_shuffle_lds_direct_load_f16_f16_f16_mk_nk_mn_instances =
     // ##################################|        |        |        |  Type|  Type|  Type|    Type| DataType| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster| SrcAccessOrder|   SrcVectorDim|         Scalar| AddExtraM|   ThreadCluster| SrcAccessOrder|  SrcVectorDim|         Scalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
     // ##################################|        |        |        |      |      |      |        |         |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|               |               |      PerVector|          | Lengths_K0_N_K1|               |              |      PerVector|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
     // ##################################|        |        |        |      |      |      |        |         |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |          |                |               |              |               |          |            |            |                             |                |
-    DeviceGemm_Xdl_CShuffle_LdsDirectLoad<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,    64,    64,    32,   8,   8,   32,   32,    1,    1,     S<4, 16, 4>,     S<1, 0, 2>,              2,              2,         1,     S<4, 16, 4>,     S<1, 0, 2>,             2,              2,         1,           1,           1,                S<1, 8, 1, 8>,               4>,
-    DeviceGemm_Xdl_CShuffle_LdsDirectLoad<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 4>,     S<1, 0, 2>,              2,              2,         1,     S<4, 16, 4>,     S<1, 0, 2>,             2,              2,         1,           1,           1,                S<1, 8, 1, 8>,               4>,
-    DeviceGemm_Xdl_CShuffle_LdsDirectLoad<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 4>,     S<1, 0, 2>,              2,              2,         1,     S<4, 16, 4>,     S<1, 0, 2>,             2,              2,         1,           1,           1,                S<1, 8, 1, 8>,               4>,
-    DeviceGemm_Xdl_CShuffle_LdsDirectLoad<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,    64,    32,    32,    64,   8,   8,   32,   32,    1,    1,     S<1, 16, 4>,     S<1, 0, 2>,              2,              2,         1,     S<1, 16, 4>,     S<1, 0, 2>,             2,              2,         1,           1,           1,                S<1, 8, 1, 8>,               4>,
-    DeviceGemm_Xdl_CShuffle_LdsDirectLoad<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,     S<1, 16, 4>,     S<1, 0, 2>,              2,              2,         1,     S<1, 16, 4>,     S<1, 0, 2>,             2,              2,         1,           1,           1,                S<1, 8, 1, 8>,               4>,
-    DeviceGemm_Xdl_CShuffle_LdsDirectLoad<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    64,    32,    32,   8,   8,   32,   32,    1,    1,     S<2, 16, 4>,     S<1, 0, 2>,              2,              2,         1,     S<2, 16, 4>,     S<1, 0, 2>,             2,              2,         1,           1,           1,                S<1, 8, 1, 8>,               4>,
-
-    DeviceGemm_Xdl_CShuffle_LdsDirectLoad<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,   256,    64,    64,    32,   8,   8,   32,   32,    1,    1,     S<4, 16, 4>,     S<1, 0, 2>,              2,              2,         1,     S<4, 16, 4>,     S<1, 0, 2>,             2,              2,         1,           1,           1,                S<1, 8, 1, 8>,               4>,
-    DeviceGemm_Xdl_CShuffle_LdsDirectLoad<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 16, 4>,     S<1, 0, 2>,              2,              2,         0,     S<4, 16, 4>,     S<1, 0, 2>,             2,              2,         0,           1,           1,                S<1, 8, 1, 8>,               4>,
-    DeviceGemm_Xdl_CShuffle_LdsDirectLoad<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2,     S<4, 16, 4>,     S<1, 0, 2>,              2,              2,         0,     S<4, 16, 4>,     S<1, 0, 2>,             2,              2,         0,           1,           1,                S<1, 8, 1, 8>,               4>,
-    DeviceGemm_Xdl_CShuffle_LdsDirectLoad<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,   256,    32,   128,    32,   8,   8,   32,   32,    1,    1,     S<4, 16, 4>,     S<1, 0, 2>,              2,              2,         1,     S<4, 16, 4>,     S<1, 0, 2>,             2,              2,         1,           1,           1,                S<1, 8, 1, 8>,               4>,
-    DeviceGemm_Xdl_CShuffle_LdsDirectLoad<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,    64,    32,    32,    64,   8,   8,   32,   32,    1,    1,     S<1, 16, 4>,     S<1, 0, 2>,              2,              2,         1,     S<1, 16, 4>,     S<1, 0, 2>,             2,              2,         1,           1,           1,                S<1, 8, 1, 8>,               4>,
-    DeviceGemm_Xdl_CShuffle_LdsDirectLoad<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,   128,    64,    32,    32,   8,   8,   32,   32,    1,    1,     S<2, 16, 4>,     S<1, 0, 2>,              2,              2,         1,     S<2, 16, 4>,     S<1, 0, 2>,             2,              2,         1,           1,           1,                S<1, 8, 1, 8>,               4>,
-
-    DeviceGemm_Xdl_CShuffle_LdsDirectLoad<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F32, PassThrough, PassThrough, PassThrough,  GemmMNPadding,        1,   256,    64,    64,    32,   8,   8,   32,   32,    1,    1,     S<4, 16, 4>,     S<1, 0, 2>,              2,              2,         1,     S<4, 16, 4>,     S<1, 0, 2>,             2,              2,         1,           1,           1,                S<1, 8, 1, 8>,               4>,
-    DeviceGemm_Xdl_CShuffle_LdsDirectLoad<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F32, PassThrough, PassThrough, PassThrough,  GemmMNPadding,        2,   256,    64,    64,    32,   8,   8,   32,   32,    1,    1,     S<4, 16, 4>,     S<1, 0, 2>,              2,              2,         1,     S<4, 16, 4>,     S<1, 0, 2>,             2,              2,         1,           1,           1,                S<1, 8, 1, 8>,               4>
+    DeviceGemm_Xdl_CShuffle_LdsDirectLoad<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,   256,    64,    64,    64,  16,  16,   32,   32,    1,    1,      S<4, 8, 8>,     S<1, 0, 2>,              2,              2,         0,      S<4, 8, 8>,     S<1, 0, 2>,             2,              2,         0,           1,           1,                S<1, 8, 1, 8>,               4>,
+    DeviceGemm_Xdl_CShuffle_LdsDirectLoad<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,   128,    16,    32,    32,   8,   8,   16,   16,    1,    1,     S<2, 16, 4>,     S<1, 0, 2>,              2,              2,         0,     S<2, 16, 4>,     S<1, 0, 2>,             2,              2,         0,           1,           1,                S<1, 8, 1, 8>,               4>,
+    DeviceGemm_Xdl_CShuffle_LdsDirectLoad<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        1,   128,    16,    32,    64,  16,  16,   16,   16,    1,    1,      S<2, 8, 8>,     S<1, 0, 2>,              2,              2,         0,      S<2, 8, 8>,     S<1, 0, 2>,             2,              2,         0,           1,           1,                S<1, 8, 1, 8>,               4>,
+    DeviceGemm_Xdl_CShuffle_LdsDirectLoad<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F32, PassThrough, PassThrough, PassThrough,    GemmDefault,        2,   128,    16,    32,    64,  16,  16,   16,   16,    1,    1,      S<2, 8, 8>,     S<1, 0, 2>,              2,              2,         0,      S<2, 8, 8>,     S<1, 0, 2>,             2,              2,         0,           1,           1,                S<1, 8, 1, 8>,               4>,
+    DeviceGemm_Xdl_CShuffle_LdsDirectLoad<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F32, PassThrough, PassThrough, PassThrough,  GemmMNPadding,        1,   256,    64,    64,    64,  16,  16,   16,   16,    2,    2,      S<4, 8, 8>,     S<1, 0, 2>,              2,              2,         0,      S<4, 8, 8>,     S<1, 0, 2>,             2,              2,         0,           1,           1,                S<1, 8, 1, 8>,               4>,
+    DeviceGemm_Xdl_CShuffle_LdsDirectLoad<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F32, PassThrough, PassThrough, PassThrough,  GemmMNPadding,        2,   128,    64,    32,    32,   8,   8,   32,   32,    1,    1,     S<2, 16, 4>,     S<1, 0, 2>,              2,              2,         1,     S<2, 16, 4>,     S<1, 0, 2>,             2,              2,         1,           1,           1,                S<1, 8, 1, 8>,               4>,
+    DeviceGemm_Xdl_CShuffle_LdsDirectLoad<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F32, PassThrough, PassThrough, PassThrough,  GemmMNPadding,        1,   128,    16,    32,    64,  16,  16,   16,   16,    1,    1,      S<2, 8, 8>,     S<1, 0, 2>,              2,              2,         0,      S<2, 8, 8>,     S<1, 0, 2>,             2,              2,         0,           1,           1,                S<1, 8, 1, 8>,               4>,
+    DeviceGemm_Xdl_CShuffle_LdsDirectLoad<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F32, PassThrough, PassThrough, PassThrough,  GemmMNPadding,        2,   128,    16,    32,    64,  16,  16,   16,   16,    1,    1,      S<2, 8, 8>,     S<1, 0, 2>,              2,              2,         0,      S<2, 8, 8>,     S<1, 0, 2>,             2,              2,         0,           1,           1,                S<1, 8, 1, 8>,               4>,
+    DeviceGemm_Xdl_CShuffle_LdsDirectLoad<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F32, PassThrough, PassThrough, PassThrough,  GemmMNPadding,        2,   128,    16,    32,    32,   8,   8,   16,   16,    1,    1,     S<2, 16, 4>,     S<1, 0, 2>,              2,              2,         0,     S<2, 16, 4>,     S<1, 0, 2>,             2,              2,         0,           1,           1,                S<1, 8, 1, 8>,               4>,
+    DeviceGemm_Xdl_CShuffle_LdsDirectLoad<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F32, PassThrough, PassThrough, PassThrough,  GemmMNPadding,        1,   128,    32,    16,    64,  16,  16,   16,   16,    1,    1,      S<2, 8, 8>,     S<1, 0, 2>,              2,              2,         0,      S<2, 8, 8>,     S<1, 0, 2>,             2,              2,         0,           1,           1,               S<1, 16, 1, 4>,               4>,
+    DeviceGemm_Xdl_CShuffle_LdsDirectLoad<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F32, PassThrough, PassThrough, PassThrough,  GemmMNPadding,        2,   128,    32,    16,    64,  16,  16,   16,   16,    1,    1,      S<2, 8, 8>,     S<1, 0, 2>,              2,              2,         0,      S<2, 8, 8>,     S<1, 0, 2>,             2,              2,         0,           1,           1,               S<1, 16, 1, 4>,               4>,
+    DeviceGemm_Xdl_CShuffle_LdsDirectLoad<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F32, PassThrough, PassThrough, PassThrough,  GemmMNPadding,        1,    64,    16,    16,   128,  32,  32,   16,   16,    1,    1,     S<1, 4, 16>,     S<1, 0, 2>,              2,              2,         0,     S<1, 4, 16>,     S<1, 0, 2>,             2,              2,         0,           1,           1,               S<1, 16, 1, 4>,               4>,
+    DeviceGemm_Xdl_CShuffle_LdsDirectLoad<     Row,     Col,     Row,   F16,   F16,   F16,     F32,      F32, PassThrough, PassThrough, PassThrough,  GemmMNPadding,        2,    64,    16,    16,   128,  32,  32,   16,   16,    1,    1,     S<1, 4, 16>,     S<1, 0, 2>,              2,              2,         0,     S<1, 4, 16>,     S<1, 0, 2>,             2,              2,         0,           1,           1,               S<1, 16, 1, 4>,               4>
     // clang-format on
     >;
 

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

+set(DEVICE_NORMALIZATION_bwd_data_INSTANCES)
+
+list(APPEND DEVICE_NORMALIZATION_bwd_data_INSTANCES
+    device_groupnorm_bwd_data_f32_instance.cpp
+    device_layernorm2d_bwd_data_f16_instance.cpp
+    device_layernorm2d_bwd_data_f32_instance.cpp)
+
+add_instance_library(device_normalization_bwd_data_instance ${DEVICE_NORMALIZATION_bwd_data_INSTANCES})

library/src/tensor_operation_instance/gpu/normalization_bwd_data/device_groupnorm_bwd_data_f32_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "normalization_bwd_data_instance_common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_groupnorm_bwd_data_f32_instances(
+    std::vector<std::unique_ptr<DeviceNormalizationBwdData<F32, F32, F32, F32, F32, 5, 3>>>&
+        instances)
+{
+    add_device_operation_instances(instances, device_groupnorm_bwd_data_f32_generic_instance{});
+    add_device_operation_instances(instances, device_groupnorm_bwd_data_f32_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/normalization_bwd_data/device_layernorm2d_bwd_data_f16_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "normalization_bwd_data_instance_common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_layernorm2d_bwd_data_f16_instances(
+    std::vector<std::unique_ptr<DeviceNormalizationBwdData<F16, F16, F16, F16, F16, 2, 1>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_layernorm_bwd_data_f16_generic_instance<2, 1>{});
+    add_device_operation_instances(instances, device_layernorm_bwd_data_f16_instances<2, 1>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/normalization_bwd_data/device_layernorm2d_bwd_data_f32_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "normalization_bwd_data_instance_common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_layernorm2d_bwd_data_f32_instances(
+    std::vector<std::unique_ptr<DeviceNormalizationBwdData<F32, F32, F32, F32, F32, 2, 1>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_layernorm_bwd_data_f32_generic_instance<2, 1>{});
+    add_device_operation_instances(instances, device_layernorm_bwd_data_f32_instances<2, 1>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/normalization_bwd_data/normalization_bwd_data_instance_common.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_normalization_bwd_data_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;
+
+template <index_t Rank, index_t Reduce>
+using device_layernorm_bwd_data_f16_instances =
+    // clang-format off
+    std::tuple <
+        // DYDataType, XDataType, GammaDataType, MeanInvStdDataType, ComputeDataType, DXDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, IsDYFastestDimReduced, DYSrcVectorSize, IsXFastestDimReduced, XSrcVectorSize, IsGammaFastestDimReduced, GammaSrcVectorSize, IsMeanInvStdFastestDimReduced, MeanInvStdSrcVectorSize, IsDXFastestDimReduced, DXDstVectorSize>
+        DeviceNormalizationBwdDataImpl<F16, F16, F16, F16, F32, F16, Rank, Reduce, 256, 1, 256, 1, 2, true, 2, true, 2, true, 2, false, 1, true, 2>,
+        DeviceNormalizationBwdDataImpl<F16, F16, F16, F16, F32, F16, Rank, Reduce, 256, 1, 256, 1, 4, true, 4, true, 4, true, 4, false, 1, true, 4>,
+        DeviceNormalizationBwdDataImpl<F16, F16, F16, F16, F32, F16, Rank, Reduce, 256, 1, 256, 1, 8, true, 8, true, 8, true, 8, false, 1, true, 8>
+        // clang-format on
+        >;
+
+template <index_t Rank, index_t Reduce>
+using device_layernorm_bwd_data_f16_generic_instance = std::tuple<
+    // clang-format off
+        DeviceNormalizationBwdDataImpl<F16, F16, F16, F16, F32, F16, Rank, Reduce, 64, 1, 64, 1, 1, true, 1, true, 1, true, 1, false, 1, true, 1>
+    // clang-format on
+    >;
+
+template <index_t Rank, index_t Reduce>
+using device_layernorm_bwd_data_f32_instances =
+    // clang-format off
+    std::tuple <
+        // DYDataType, XDataType, GammaDataType, MeanInvStdDataType, ComputeDataType, DXDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, IsDYFastestDimReduced, DYSrcVectorSize, IsXFastestDimReduced, XSrcVectorSize, IsGammaFastestDimReduced, GammaSrcVectorSize, IsMeanInvStdFastestDimReduced, MeanInvStdSrcVectorSize, IsDXFastestDimReduced, DXDstVectorSize>
+        DeviceNormalizationBwdDataImpl<F32, F32, F32, F32, F32, F32, Rank, Reduce, 256, 1, 256, 1, 2, true, 2, true, 2, true, 2, false, 1, true, 2>,
+        DeviceNormalizationBwdDataImpl<F32, F32, F32, F32, F32, F32, Rank, Reduce, 256, 1, 256, 1, 4, true, 4, true, 4, true, 4, false, 1, true, 4>
+        // clang-format on
+        >;
+
+template <index_t Rank, index_t Reduce>
+using device_layernorm_bwd_data_f32_generic_instance = std::tuple<
+    // clang-format off
+        DeviceNormalizationBwdDataImpl<F32, F32, F32, F32, F32, F32, Rank, Reduce, 64, 1, 64, 1, 1, true, 1, true, 1, true, 1, false, 1, true, 1>
+    // clang-format on
+    >;
+
+using device_groupnorm_bwd_data_f32_instances =
+    // clang-format off
+    std::tuple <
+        // DYDataType, XDataType, GammaDataType, MeanInvStdDataType, ComputeDataType, DXDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, IsDYFastestDimReduced, DYSrcVectorSize, IsXFastestDimReduced, XSrcVectorSize, IsGammaFastestDimReduced, GammaSrcVectorSize, IsMeanInvStdFastestDimReduced, MeanInvStdSrcVectorSize, IsDXFastestDimReduced, DXDstVectorSize>
+        DeviceNormalizationBwdDataImpl<F32, F32, F32, F32, F32, F32, 5, 3, 256, 1, 256, 1, 2, true, 2, true, 2, true, 2, false, 1, true, 2>,
+        DeviceNormalizationBwdDataImpl<F32, F32, F32, F32, F32, F32, 5, 3, 256, 1, 256, 1, 4, true, 4, true, 4, true, 4, false, 1, true, 4>
+        // clang-format on
+        >;
+
+using device_groupnorm_bwd_data_f32_generic_instance = std::tuple<
+    // clang-format off
+        DeviceNormalizationBwdDataImpl<F32, F32, F32, F32, F32, F32, 5, 3, 64, 1, 64, 1, 1, true, 1, true, 1, true, 1, false, 1, true, 1>
+    // clang-format on
+    >;
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

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

+set(DEVICE_NORMALIZATION_BWD_GAMMA_BETA_INSTANCES)
+
+list(APPEND DEVICE_NORMALIZATION_BWD_GAMMA_BETA_INSTANCES
+    device_groupnorm_bwd_gamma_beta_f32_instance.cpp
+    device_layernorm2d_bwd_gamma_beta_f16_instance.cpp
+    device_layernorm2d_bwd_gamma_beta_f32_instance.cpp)
+
+add_instance_library(device_normalization_bwd_gamma_beta_instance ${DEVICE_NORMALIZATION_BWD_GAMMA_BETA_INSTANCES})

library/src/tensor_operation_instance/gpu/normalization_bwd_gamma_beta/device_groupnorm_bwd_gamma_beta_f32_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "normalization_bwd_gamma_beta_instance_common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_groupnorm_bwd_gamma_beta_f32_instances(
+    std::vector<std::unique_ptr<DeviceNormalizationBwdGammaBeta<F32, F32, F32, F32, F32, 5, 3>>>&
+        instances)
+{
+    add_device_operation_instances(instances, device_groupnorm_bwd_gamma_beta_f32_instances{});
+    add_device_operation_instances(instances,
+                                   device_groupnorm_bwd_gamma_beta_f32_generic_instance{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/normalization_bwd_gamma_beta/device_layernorm2d_bwd_gamma_beta_f16_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "normalization_bwd_gamma_beta_instance_common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_layernorm2d_bwd_gamma_beta_rank_2_1_f16_instances(
+    std::vector<std::unique_ptr<DeviceNormalizationBwdGammaBeta<F16, F16, F16, F16, F16, 2, 1>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_layernorm_bwd_gamma_beta_f16_generic_instance<2, 1>{});
+    add_device_operation_instances(instances,
+                                   device_layernorm_bwd_gamma_beta_f16_instances<2, 1>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/normalization_bwd_gamma_beta/device_layernorm2d_bwd_gamma_beta_f32_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "normalization_bwd_gamma_beta_instance_common.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_layernorm2d_bwd_gamma_beta_rank_2_1_f32_instances(
+    std::vector<std::unique_ptr<DeviceNormalizationBwdGammaBeta<F32, F32, F32, F32, F32, 2, 1>>>&
+        instances)
+{
+    add_device_operation_instances(instances,
+                                   device_layernorm_bwd_gamma_beta_f32_generic_instance<2, 1>{});
+    add_device_operation_instances(instances,
+                                   device_layernorm_bwd_gamma_beta_f32_instances<2, 1>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/normalization_bwd_gamma_beta/normalization_bwd_gamma_beta_instance_common.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_normalization_bwd_gamma_beta_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;
+
+template <index_t Rank, index_t Reduce>
+using device_layernorm_bwd_gamma_beta_f16_instances =
+    // clang-format off
+    std::tuple <
+        // DYDataType, XDataType, MeanInvStdDataType, ComputeDataType, DGammaDataType, DBetaDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, IsDYFastestDimReduced, DYSrcVectorSize, IsXFastestDimReduced, XSrcVectorSize, IsMeanInvStdFastestDimReduced, MeanInvStdSrcVectorSize, DGammaDstVectorSize, DBetaDstVectorSize>
+        DeviceNormalizationBwdGammaBetaImpl<F16, F16, F16, F32, F16, F16, Rank, Reduce, 256, 1, 256, 2, 1, false, 2, false, 2, true, 1, 2, 2>,
+        DeviceNormalizationBwdGammaBetaImpl<F16, F16, F16, F32, F16, F16, Rank, Reduce, 256, 1, 256, 4, 1, false, 4, false, 4, true, 1, 4, 4>,
+        DeviceNormalizationBwdGammaBetaImpl<F16, F16, F16, F32, F16, F16, Rank, Reduce, 256, 1, 256, 8, 1, false, 8, false, 8, true, 1, 8, 8>
+        // clang-format on
+        >;
+
+template <index_t Rank, index_t Reduce>
+using device_layernorm_bwd_gamma_beta_f16_generic_instance = std::tuple<
+    // clang-format off
+        DeviceNormalizationBwdGammaBetaImpl<F16, F16, F16, F32, F16, F16, Rank, Reduce, 64, 1, 64, 1, 1, false, 1, false, 1, true, 1, 1, 1>
+    // clang-format on
+    >;
+
+template <index_t Rank, index_t Reduce>
+using device_layernorm_bwd_gamma_beta_f32_instances =
+    // clang-format off
+    std::tuple <
+        // DYDataType, XDataType, MeanInvStdDataType, ComputeDataType, DGammaDataType, DBetaDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, IsDYFastestDimReduced, DYSrcVectorSize, IsXFastestDimReduced, XSrcVectorSize, IsMeanInvStdFastestDimReduced, MeanInvStdSrcVectorSize, DGammaDstVectorSize, DBetaDstVectorSize>
+        DeviceNormalizationBwdGammaBetaImpl<F32, F32, F32, F32, F32, F32, Rank, Reduce, 256, 1, 256, 2, 1, false, 2, false, 2, true, 1, 2, 2>,
+        DeviceNormalizationBwdGammaBetaImpl<F32, F32, F32, F32, F32, F32, Rank, Reduce, 256, 1, 256, 4, 1, false, 4, false, 4, true, 1, 4, 4>
+        // clang-format on
+        >;
+
+template <index_t Rank, index_t Reduce>
+using device_layernorm_bwd_gamma_beta_f32_generic_instance = std::tuple<
+    // clang-format off
+        DeviceNormalizationBwdGammaBetaImpl<F32, F32, F32, F32, F32, F32, Rank, Reduce, 64, 1, 64, 1, 1, false, 1, false, 1, true, 1, 1, 1>
+    // clang-format on
+    >;
+
+using device_groupnorm_bwd_gamma_beta_f32_instances =
+    // clang-format off
+    std::tuple <
+        // DYDataType, XDataType, MeanInvStdDataType, ComputeDataType, DGammaDataType, DBetaDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, IsDYFastestDimReduced, DYSrcVectorSize, IsXFastestDimReduced, XSrcVectorSize, IsMeanInvStdFastestDimReduced, MeanInvStdSrcVectorSize, DGammaDstVectorSize, DBetaDstVectorSize>
+        DeviceNormalizationBwdGammaBetaImpl<F32, F32, F32, F32, F32, F32, 5, 3, 256, 1, 256, 2, 1, false, 2, false, 2, false, 1, 2, 2>,
+        DeviceNormalizationBwdGammaBetaImpl<F32, F32, F32, F32, F32, F32, 5, 3, 256, 1, 256, 4, 1, false, 4, false, 4, false, 1, 4, 4>
+        // clang-format on
+        >;
+
+using device_groupnorm_bwd_gamma_beta_f32_generic_instance = std::tuple<
+    // clang-format off
+        DeviceNormalizationBwdGammaBetaImpl<F32, F32, F32, F32, F32, F32, 5, 3, 64, 1, 64, 1, 1, false, 1, false, 1, false, 1, 1, 1>
+    // clang-format on
+    >;
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/normalization_fwd/device_groupnorm_fwd_f16_instance.cpp

@@ -11,7 +11,7 @@ namespace instance {
 using Pass = ck::tensor_operation::element_wise::PassThrough;
 
 void add_device_normalization_fwd_rank_5_3_f16_instances(
-    std::vector<std::unique_ptr<DeviceNormalizationFwd<F16, F16, F16, F16, F32, Pass, 5, 3>>>&
+    std::vector<std::unique_ptr<DeviceNormalizationFwd<F16, F16, F16, F16, F16, Pass, 5, 3>>>&
         instances)
 {
     add_device_operation_instances(instances,

library/src/tensor_operation_instance/gpu/normalization_fwd/device_groupnorm_fwd_swish_f16_instance.cpp

@@ -11,7 +11,7 @@ namespace instance {
 using Swish = ck::tensor_operation::element_wise::Swish;
 
 void add_device_normalization_fwd_rank_5_3_swish_f16_instances(
-    std::vector<std::unique_ptr<DeviceNormalizationFwd<F16, F16, F16, F16, F32, Swish, 5, 3>>>&
+    std::vector<std::unique_ptr<DeviceNormalizationFwd<F16, F16, F16, F16, F16, Swish, 5, 3>>>&
         instances)
 {
     add_device_operation_instances(instances,

library/src/tensor_operation_instance/gpu/normalization_fwd/device_layernorm2d_fwd_f16_instance.cpp

@@ -11,7 +11,7 @@ namespace instance {
 using Pass = ck::tensor_operation::element_wise::PassThrough;
 
 void add_device_normalization_fwd_rank_2_1_f16_instances(
-    std::vector<std::unique_ptr<DeviceNormalizationFwd<F16, F16, F16, F16, F32, Pass, 2, 1>>>&
+    std::vector<std::unique_ptr<DeviceNormalizationFwd<F16, F16, F16, F16, F16, Pass, 2, 1>>>&
         instances)
 {
     add_device_operation_instances(instances,

library/src/tensor_operation_instance/gpu/normalization_fwd/device_layernorm4d_fwd_f16_instance.cpp

@@ -11,7 +11,7 @@ namespace instance {
 using Pass = ck::tensor_operation::element_wise::PassThrough;
 
 void add_device_normalization_fwd_rank_4_3_f16_instances(
-    std::vector<std::unique_ptr<DeviceNormalizationFwd<F16, F16, F16, F16, F32, Pass, 4, 3>>>&
+    std::vector<std::unique_ptr<DeviceNormalizationFwd<F16, F16, F16, F16, F16, Pass, 4, 3>>>&
         instances)
 {
     add_device_operation_instances(instances,

library/src/tensor_operation_instance/gpu/normalization_fwd/normalization_fwd_instance_common.hpp

@@ -23,24 +23,24 @@ using device_normalization_f16_instances =
     // clang-format off
     std::tuple <
         // XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, SaveMeanInvStdDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize, SaveMeanInvStdScalarPerVector>
-        DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
-        DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
-        DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
-        DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
-        DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 2, 1, 2, 1, 2, 1, 2, 2, 1>,   // irregular size
-        DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>,   // irregular size
-        DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 64, 1, 64, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
-        DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
-        DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 16, 1, 8, 1, 8, 1, 8, 8, 1>,
-        DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 32, 1, 8, 1, 8, 1, 8, 8, 1>,
-        DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
-        DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 16, 1, 8, 1, 8, 1, 8, 8, 1>,
-        DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 2, 16, 1, 8, 1, 8, 1, 8, 8, 2>,
-        DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 8, 1, 8, 1, 8, 8, 1>,
-        DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
-        DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 16, 1, 8, 1, 8, 1, 8, 8, 1>,
-        DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
-        DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 16, 1, 8, 1, 8, 1, 8, 8, 1>
+        DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
+        DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
+        DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
+        DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
+        DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 2, 1, 2, 1, 2, 1, 2, 2, 1>,   // irregular size
+        DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>,   // irregular size
+        DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F16, OutElementwise, Rank, Reduce, 64, 1, 64, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
+        DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
+        DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 16, 1, 8, 1, 8, 1, 8, 8, 1>,
+        DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 32, 1, 8, 1, 8, 1, 8, 8, 1>,
+        DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
+        DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 16, 1, 8, 1, 8, 1, 8, 8, 1>,
+        DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 2, 16, 1, 8, 1, 8, 1, 8, 8, 2>,
+        DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 8, 1, 8, 1, 8, 8, 1>,
+        DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
+        DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 16, 1, 8, 1, 8, 1, 8, 8, 1>,
+        DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
+        DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 16, 1, 8, 1, 8, 1, 8, 8, 1>
         // clang-format on
         >;
 
@@ -49,31 +49,31 @@ using device_normalization_splitk_f16_instances =
     // clang-format off
     std::tuple <
         // XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, SaveMeanInvStdDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize, SaveMeanInvStdScalarPerVector>
-        DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
-        DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
-        DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
-        DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
-        DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 2, 1, 2, 1, 2, 1, 2, 2, 1>,   // irregular size
-        DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>,   // irregular size
-        DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 64, 1, 64, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
-        DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
-        DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 16, 1, 8, 1, 8, 1, 8, 8, 1>,
-        DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 32, 1, 8, 1, 8, 1, 8, 8, 1>,
-        DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
-        DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 16, 1, 8, 1, 8, 1, 8, 8, 1>,
-        DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 2, 16, 1, 8, 1, 8, 1, 8, 8, 2>,
-        DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 8, 1, 8, 1, 8, 8, 1>,
-        DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
-        DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 16, 1, 8, 1, 8, 1, 8, 8, 1>,
-        DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
-        DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 16, 1, 8, 1, 8, 1, 8, 8, 1>
+        DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
+        DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
+        DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
+        DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
+        DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 2, 1, 2, 1, 2, 1, 2, 2, 1>,   // irregular size
+        DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>,   // irregular size
+        DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F16, OutElementwise, Rank, Reduce, 64, 1, 64, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
+        DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
+        DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 16, 1, 8, 1, 8, 1, 8, 8, 1>,
+        DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 32, 1, 8, 1, 8, 1, 8, 8, 1>,
+        DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
+        DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 16, 1, 8, 1, 8, 1, 8, 8, 1>,
+        DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 2, 16, 1, 8, 1, 8, 1, 8, 8, 2>,
+        DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 8, 1, 8, 1, 8, 8, 1>,
+        DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
+        DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 16, 1, 8, 1, 8, 1, 8, 8, 1>,
+        DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
+        DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 16, 1, 8, 1, 8, 1, 8, 8, 1>
         // clang-format on
         >;
 
 template <typename OutElementwise, index_t Rank, index_t Reduce>
 using device_normalization_f16_generic_instance = std::tuple<
     // clang-format off
-        DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 64, 1, 64, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>
+        DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F16, OutElementwise, Rank, Reduce, 64, 1, 64, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>
     // clang-format on
     >;