Merge branch 'develop' into hip_tensor_permute

library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_i8_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_wmma_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+void add_device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_i8_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
+                                                           GNDHWC,
+                                                           GKZYXC,
+                                                           GNDHWK,
+                                                           int8_t,
+                                                           int8_t,
+                                                           int8_t,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_bwd_weight_wmma_i8_instances<3,
+                                                         GNDHWC,
+                                                         GKZYXC,
+                                                         GNDHWK,
+                                                         ConvBwdWeightDefault>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_f16_1x1s1p0_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_wmma_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+void add_device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_f16_1x1s1p0_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
+                                                           NDHWGC,
+                                                           GKZYXC,
+                                                           NDHWGK,
+                                                           F16,
+                                                           F16,
+                                                           F16,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_bwd_weight_wmma_f16_instances<3,
+                                                          NDHWGC,
+                                                          GKZYXC,
+                                                          NDHWGK,
+                                                          ConvBwdWeightFilter1x1Stride1Pad0>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_wmma_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+void add_device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_f16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
+                                                           NDHWGC,
+                                                           GKZYXC,
+                                                           NDHWGK,
+                                                           F16,
+                                                           F16,
+                                                           F16,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_bwd_weight_wmma_f16_instances<3,
+                                                          NDHWGC,
+                                                          GKZYXC,
+                                                          NDHWGK,
+                                                          ConvBwdWeightDefault>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_i8_1x1s1p0_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_wmma_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+void add_device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_i8_1x1s1p0_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
+                                                           NDHWGC,
+                                                           GKZYXC,
+                                                           NDHWGK,
+                                                           int8_t,
+                                                           int8_t,
+                                                           int8_t,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_bwd_weight_wmma_i8_instances<3,
+                                                         NDHWGC,
+                                                         GKZYXC,
+                                                         NDHWGK,
+                                                         ConvBwdWeightFilter1x1Stride1Pad0>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_i8_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_wmma_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+void add_device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_i8_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
+                                                           NDHWGC,
+                                                           GKZYXC,
+                                                           NDHWGK,
+                                                           int8_t,
+                                                           int8_t,
+                                                           int8_t,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_bwd_weight_wmma_i8_instances<3,
+                                                         NDHWGC,
+                                                         GKZYXC,
+                                                         NDHWGK,
+                                                         ConvBwdWeightDefault>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/CMakeLists.txt

-add_instance_library(device_grouped_conv3d_fwd_instance
+set(GROUPED_CONV3D_FWD
    xdl/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp
    xdl/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_f16_instance.cpp
    xdl/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp
    xdl/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_int8_instance.cpp
-
-
    xdl/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp
    xdl/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
    xdl/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp
    xdl/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_int8_instance.cpp
-   xdl/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f16_comp_fp8_instance.cpp
-
    wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_instance.cpp
    wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_instance.cpp
    wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
    wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_instance.cpp
-
    wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_1x1p0_instance.cpp
    wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_1x1p0_instance.cpp
    wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_1x1p0_instance.cpp
    wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_1x1p0_instance.cpp
-
    wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_1x1s1p0_instance.cpp
    wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_1x1s1p0_instance.cpp
    wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_1x1s1p0_instance.cpp
    wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_1x1s1p0_instance.cpp
-
    wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_oddc_instance.cpp
    wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_oddc_instance.cpp
    wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_oddc_instance.cpp
-   wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_oddc_instance.cpp
-)
+   wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_oddc_instance.cpp)
+
+if((DTYPES MATCHES "fp8" AND DTYPES MATCHES "fp16") OR NOT DEFINED DTYPES)
+    list(APPEND GROUPED_CONV3D_FWD
+      xdl/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f16_comp_fp8_instance.cpp)
+endif()
+
+add_instance_library(device_grouped_conv3d_fwd_instance ${GROUPED_CONV3D_FWD})

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

 set(GROUPED_GEMM_FIXED_NK_INSTANCES)
 
-if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
-  list(APPEND GROUPED_GEMM_FIXED_NK_INSTANCES device_grouped_gemm_xdl_fixed_nk_f16_f16_f16_mk_kn_mn_instance.cpp)
-  list(APPEND GROUPED_GEMM_FIXED_NK_INSTANCES device_grouped_gemm_xdl_fixed_nk_f16_f16_f16_mk_nk_mn_instance.cpp)
-endif()
-
-if((DTYPES MATCHES "fp8" AND DTYPES MATCHES "fp16") OR NOT DEFINED DTYPES)
-  list(APPEND GROUPED_GEMM_FIXED_NK_INSTANCES device_grouped_gemm_xdl_fixed_nk_f16_fp8_f16_mk_kn_mn_instance.cpp)
-  list(APPEND GROUPED_GEMM_FIXED_NK_INSTANCES device_grouped_gemm_xdl_fixed_nk_f16_fp8_f16_mk_nk_mn_instance.cpp)
-endif()
-
-if((DTYPES MATCHES "int8" AND DTYPES MATCHES "fp16") OR NOT DEFINED DTYPES)
-  list(APPEND GROUPED_GEMM_FIXED_NK_INSTANCES device_grouped_gemm_xdl_fixed_nk_f16_i8_f16_mk_kn_mn_instance.cpp)
-  list(APPEND GROUPED_GEMM_FIXED_NK_INSTANCES device_grouped_gemm_xdl_fixed_nk_f16_i8_f16_mk_nk_mn_instance.cpp)
-endif()
+list(APPEND GROUPED_GEMM_FIXED_NK_INSTANCES device_grouped_gemm_xdl_fixed_nk_f16_f16_f16_mk_kn_mn_instance.cpp
+                                            device_grouped_gemm_xdl_fixed_nk_f16_f16_f16_mk_nk_mn_instance.cpp
+                                            device_grouped_gemm_xdl_fixed_nk_f16_fp8_f16_mk_kn_mn_instance.cpp
+                                            device_grouped_gemm_xdl_fixed_nk_f16_fp8_f16_mk_nk_mn_instance.cpp
+                                            device_grouped_gemm_xdl_fixed_nk_f16_i8_f16_mk_kn_mn_instance.cpp
+                                            device_grouped_gemm_xdl_fixed_nk_f16_i8_f16_mk_nk_mn_instance.cpp)
 
 add_instance_library(device_grouped_gemm_fixed_nk_instance ${GROUPED_GEMM_FIXED_NK_INSTANCES})

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

@@ -11,7 +11,7 @@ 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>>>&
+    std::vector<std::unique_ptr<DeviceNormalization<F16, F16, F16, F16, F32, Pass, 5, 3>>>&
         instances)
 {
     add_device_operation_instances(instances,

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

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

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

@@ -11,7 +11,7 @@ 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>>>&
+    std::vector<std::unique_ptr<DeviceNormalization<F16, F16, F16, F16, F32, Swish, 5, 3>>>&
         instances)
 {
     add_device_operation_instances(instances,

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

@@ -11,7 +11,7 @@ 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>>>&
+    std::vector<std::unique_ptr<DeviceNormalization<F16, F16, F16, F16, F32, Pass, 2, 1>>>&
         instances)
 {
     add_device_operation_instances(instances,

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

@@ -11,7 +11,7 @@ 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>>>&
+    std::vector<std::unique_ptr<DeviceNormalization<F16, F16, F16, F16, F32, Pass, 4, 3>>>&
         instances)
 {
     add_device_operation_instances(instances,

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

@@ -22,25 +22,25 @@ 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>
+        // XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, SaveMeanInvStdDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize, SaveMeanInvStdScalarPerVector>
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 2, 1, 2, 1, 2, 1, 2, 2, 1>,   // irregular size
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>,   // irregular size
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 64, 1, 64, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 16, 1, 8, 1, 8, 1, 8, 8, 1>,
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 32, 1, 8, 1, 8, 1, 8, 8, 1>,
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 16, 1, 8, 1, 8, 1, 8, 8, 1>,
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 2, 16, 1, 8, 1, 8, 1, 8, 8, 2>,
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 8, 1, 8, 1, 8, 8, 1>,
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 16, 1, 8, 1, 8, 1, 8, 8, 1>,
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 16, 1, 8, 1, 8, 1, 8, 8, 1>
         // clang-format on
         >;
 
@@ -48,150 +48,150 @@ template <typename OutElementwise, index_t Rank, index_t Reduce>
 using device_normalization_splitk_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>
-        DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
-        DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
-        DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
-        DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
-        DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 2, 1, 2, 1, 2, 1, 2, 2>,   // irregular size
-        DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 4, 1, 4, 1, 4, 1, 4, 4>,   // irregular size
-        DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 64, 1, 64, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
-        DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
-        DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 16, 1, 8, 1, 8, 1, 8, 8>,
-        DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 32, 1, 8, 1, 8, 1, 8, 8>,
-        DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
-        DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 16, 1, 8, 1, 8, 1, 8, 8>,
-        DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 2, 16, 1, 8, 1, 8, 1, 8, 8>,
-        DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 8, 1, 8, 1, 8, 8>,
-        DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
-        DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 16, 1, 8, 1, 8, 1, 8, 8>,
-        DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
-        DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 16, 1, 8, 1, 8, 1, 8, 8>
+        // XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, SaveMeanInvStdDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize, SaveMeanInvStdScalarPerVector>
+        DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
+        DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
+        DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
+        DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
+        DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 2, 1, 2, 1, 2, 1, 2, 2, 1>,   // irregular size
+        DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>,   // irregular size
+        DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 64, 1, 64, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
+        DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
+        DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 16, 1, 8, 1, 8, 1, 8, 8, 1>,
+        DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 32, 1, 8, 1, 8, 1, 8, 8, 1>,
+        DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
+        DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 16, 1, 8, 1, 8, 1, 8, 8, 1>,
+        DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 2, 16, 1, 8, 1, 8, 1, 8, 8, 2>,
+        DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 8, 1, 8, 1, 8, 8, 1>,
+        DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
+        DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 16, 1, 8, 1, 8, 1, 8, 8, 1>,
+        DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
+        DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, F32, 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
-        DeviceNormalizationImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 64, 1, 64, 1, 1, 1, 1, 1, 1, 1, 1, 1>
+        DeviceNormalizationImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 64, 1, 64, 1, 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
-        // 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
-        DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
-        DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
-        DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
-        DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 2, 1, 2, 1, 2, 1, 2, 2>,   // irregular size
-        DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 4, 1, 4, 1, 4, 1, 4, 4>,
-        DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 8, 1, 4, 1, 4, 1, 4, 4>,
-        DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 16, 1, 4, 1, 4, 1, 4, 4>,
-        DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 32, 1, 4, 1, 4, 1, 4, 4>,
-        DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 4, 1, 4, 1, 4, 1, 4, 4>,
-        DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 4, 1, 4, 1, 4, 4>,
-        DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 16, 1, 4, 1, 4, 1, 4, 4>,
-        DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 2, 16, 1, 4, 1, 4, 1, 4, 4>,
-        DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 4, 1, 4, 1, 4, 4>,
-        DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 4, 1, 4, 1, 4, 1, 4, 4>,
-        DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 8, 1, 4, 1, 4, 1, 4, 4>,
-        DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 2, 8, 1, 4, 1, 4, 1, 4, 4>,
-        DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 4, 1, 4, 1, 4, 1, 4, 4>,
-        DeviceNormalizationImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 8, 1, 4, 1, 4, 1, 4, 4>
+        // XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, SaveMeanInvStdDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize, SaveMeanInvStdScalarPerVector>
+        DeviceNormalizationImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
+        DeviceNormalizationImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
+        DeviceNormalizationImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
+        DeviceNormalizationImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
+        DeviceNormalizationImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 2, 1, 2, 1, 2, 1, 2, 2, 1>,   // irregular size
+        DeviceNormalizationImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>,
+        DeviceNormalizationImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 8, 1, 4, 1, 4, 1, 4, 4, 1>,
+        DeviceNormalizationImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 16, 1, 4, 1, 4, 1, 4, 4, 1>,
+        DeviceNormalizationImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 32, 1, 4, 1, 4, 1, 4, 4, 1>,
+        DeviceNormalizationImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>,
+        DeviceNormalizationImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 4, 1, 4, 1, 4, 4, 1>,
+        DeviceNormalizationImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 16, 1, 4, 1, 4, 1, 4, 4, 1>,
+        DeviceNormalizationImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 2, 16, 1, 4, 1, 4, 1, 4, 4, 2>,
+        DeviceNormalizationImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 4, 1, 4, 1, 4, 4, 1>,
+        DeviceNormalizationImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>,
+        DeviceNormalizationImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 8, 1, 4, 1, 4, 1, 4, 4, 1>,
+        DeviceNormalizationImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 2, 8, 1, 4, 1, 4, 1, 4, 4, 2>,
+        DeviceNormalizationImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>,
+        DeviceNormalizationImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 8, 1, 4, 1, 4, 1, 4, 4, 1>
     // clang-format on
     >;
 
 template <typename OutElementwise, index_t Rank, index_t Reduce>
 using device_normalization_splitk_f32_instances = std::tuple<
     // clang-format off
-        // XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorSize, BetaSrcVectorSize, YDstVectorSize>
-        DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
-        DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
-        DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
-        DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
-        DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 2, 1, 2, 1, 2, 1, 2, 2>,   // irregular size
-        DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 4, 1, 4, 1, 4, 1, 4, 4>,
-        DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 8, 1, 4, 1, 4, 1, 4, 4>,
-        DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 16, 1, 4, 1, 4, 1, 4, 4>,
-        DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 32, 1, 4, 1, 4, 1, 4, 4>,
-        DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 4, 1, 4, 1, 4, 1, 4, 4>,
-        DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 4, 1, 4, 1, 4, 4>,
-        DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 16, 1, 4, 1, 4, 1, 4, 4>,
-        DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 2, 16, 1, 4, 1, 4, 1, 4, 4>,
-        DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 4, 1, 4, 1, 4, 4>,
-        DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 4, 1, 4, 1, 4, 1, 4, 4>,
-        DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 8, 1, 4, 1, 4, 1, 4, 4>,
-        DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 2, 8, 1, 4, 1, 4, 1, 4, 4>,
-        DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 4, 1, 4, 1, 4, 1, 4, 4>,
-        DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 8, 1, 4, 1, 4, 1, 4, 4>
+        // XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, SaveMeanInvStdDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize, SaveMeanInvStdScalarPerVector>
+        DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
+        DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
+        DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
+        DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
+        DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 2, 1, 2, 1, 2, 1, 2, 2, 1>,   // irregular size
+        DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>,
+        DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 8, 1, 4, 1, 4, 1, 4, 4, 1>,
+        DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 16, 1, 4, 1, 4, 1, 4, 4, 1>,
+        DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 32, 1, 4, 1, 4, 1, 4, 4, 1>,
+        DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>,
+        DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 4, 1, 4, 1, 4, 4, 1>,
+        DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 16, 1, 4, 1, 4, 1, 4, 4, 1>,
+        DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 2, 16, 1, 4, 1, 4, 1, 4, 4, 2>,
+        DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 4, 1, 4, 1, 4, 4, 1>,
+        DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>,
+        DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 8, 1, 4, 1, 4, 1, 4, 4, 1>,
+        DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 2, 8, 1, 4, 1, 4, 1, 4, 4, 2>,
+        DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>,
+        DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 8, 1, 4, 1, 4, 1, 4, 4, 1>
     // 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>
+        DeviceNormalizationImpl<F32, F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 64, 1, 64, 1, 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
-        // XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorSize, BetaSrcVectorSize, YDstVectorSize>
-        DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
-        DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
-        DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
-        DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
-        DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 2, 1, 2, 1, 2, 1, 2, 2>,   // irregular size
-        DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 4, 1, 4, 1, 4, 1, 4, 4>,
-        DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 8, 1, 4, 1, 4, 1, 4, 4>,
-        DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 16, 1, 4, 1, 4, 1, 4, 4>,
-        DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 32, 1, 4, 1, 4, 1, 4, 4>,
-        DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 4, 1, 4, 1, 4, 1, 4, 4>,
-        DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 4, 1, 4, 1, 4, 4>,
-        DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 16, 1, 4, 1, 4, 1, 4, 4>,
-        DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 2, 16, 1, 4, 1, 4, 1, 4, 4>,
-        DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 4, 1, 4, 1, 4, 4>,
-        DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 4, 1, 4, 1, 4, 1, 4, 4>,
-        DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 8, 1, 4, 1, 4, 1, 4, 4>,
-        DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 2, 8, 1, 4, 1, 4, 1, 4, 4>,
-        DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 4, 1, 4, 1, 4, 1, 4, 4>,
-        DeviceNormalizationImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 8, 1, 4, 1, 4, 1, 4, 4>
+        // XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, SaveMeanInvStdDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize, SaveMeanInvStdScalarPerVector>
+        DeviceNormalizationImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
+        DeviceNormalizationImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
+        DeviceNormalizationImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
+        DeviceNormalizationImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
+        DeviceNormalizationImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 2, 1, 2, 1, 2, 1, 2, 2, 1>,   // irregular size
+        DeviceNormalizationImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>,
+        DeviceNormalizationImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 8, 1, 4, 1, 4, 1, 4, 4, 1>,
+        DeviceNormalizationImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 16, 1, 4, 1, 4, 1, 4, 4, 1>,
+        DeviceNormalizationImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 32, 1, 4, 1, 4, 1, 4, 4, 1>,
+        DeviceNormalizationImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>,
+        DeviceNormalizationImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 4, 1, 4, 1, 4, 4, 1>,
+        DeviceNormalizationImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 16, 1, 4, 1, 4, 1, 4, 4, 1>,
+        DeviceNormalizationImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 2, 16, 1, 4, 1, 4, 1, 4, 4, 2>,
+        DeviceNormalizationImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 4, 1, 4, 1, 4, 4, 1>,
+        DeviceNormalizationImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>,
+        DeviceNormalizationImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 8, 1, 4, 1, 4, 1, 4, 4, 1>,
+        DeviceNormalizationImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 2, 8, 1, 4, 1, 4, 1, 4, 4, 2>,
+        DeviceNormalizationImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>,
+        DeviceNormalizationImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 8, 1, 4, 1, 4, 1, 4, 4, 1>
     // clang-format on
     >;
 
 template <typename OutElementwise, index_t Rank, index_t Reduce>
 using device_normalization_splitk_f16_f32_f32_f16_instances = std::tuple<
     // clang-format off
-        // XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorSize, BetaSrcVectorSize, YDstVectorSize>
-        DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
-        DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
-        DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
-        DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
-        DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 2, 1, 2, 1, 2, 1, 2, 2>,   // irregular size
-        DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 4, 1, 4, 1, 4, 1, 4, 4>,
-        DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 8, 1, 4, 1, 4, 1, 4, 4>,
-        DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 16, 1, 4, 1, 4, 1, 4, 4>,
-        DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 32, 1, 4, 1, 4, 1, 4, 4>,
-        DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 4, 1, 4, 1, 4, 1, 4, 4>,
-        DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 4, 1, 4, 1, 4, 4>,
-        DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 16, 1, 4, 1, 4, 1, 4, 4>,
-        DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 2, 16, 1, 4, 1, 4, 1, 4, 4>,
-        DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 4, 1, 4, 1, 4, 4>,
-        DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 4, 1, 4, 1, 4, 1, 4, 4>,
-        DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 8, 1, 4, 1, 4, 1, 4, 4>,
-        DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 2, 8, 1, 4, 1, 4, 1, 4, 4>,
-        DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 4, 1, 4, 1, 4, 1, 4, 4>,
-        DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 8, 1, 4, 1, 4, 1, 4, 4>
+        // XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, SaveMeanInvStdDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize, SaveMeanInvStdScalarPerVector>
+        DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
+        DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
+        DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
+        DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
+        DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 2, 1, 2, 1, 2, 1, 2, 2, 1>,   // irregular size
+        DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>,
+        DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 8, 1, 4, 1, 4, 1, 4, 4, 1>,
+        DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 16, 1, 4, 1, 4, 1, 4, 4, 1>,
+        DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 32, 1, 4, 1, 4, 1, 4, 4, 1>,
+        DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>,
+        DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 4, 1, 4, 1, 4, 4, 1>,
+        DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 16, 1, 4, 1, 4, 1, 4, 4, 1>,
+        DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 2, 16, 1, 4, 1, 4, 1, 4, 4, 2>,
+        DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 4, 1, 4, 1, 4, 4, 1>,
+        DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>,
+        DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 8, 1, 4, 1, 4, 1, 4, 4, 1>,
+        DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 2, 8, 1, 4, 1, 4, 1, 4, 4, 2>,
+        DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>,
+        DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 8, 1, 4, 1, 4, 1, 4, 4, 1>
     // 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>
+        DeviceNormalizationImpl<F16, F32, F32, F32, F16, F32, OutElementwise, Rank, Reduce, 64, 1, 64, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>
     // clang-format on
     >;