update layernorm (#1570)

* port layernorm

* change warp_welford.hpp

* Update warpshuffle

* 1. Add save mean and save std back
2. Move construction of tensor_view and tile_window to operator()

* refine welford max count calculation

* unify layernorm api

* Rename file

* Remove save mean and inv std

* Revert "refine welford max count calculation"

This reverts commit 02236580

.

* Fix order of parameter

* refine welford max count calculation again

* Remove fp32 instances

* Fix bug of padding

* refactor api

* Support bf16

* Extract common function

* Refine arg of operator()

* Add kMThreadPerBlock to template parameter

* clang format

* Refine variable name

* Refine file name

* remove redundant line

* refactor layernorm2d pipeline and add block-per-block utility

* fix name

* rename more

* add more block-per-tile instance

* remove duplicated define

* update instance for 2048, 1024 case

* support up to 2048 now

* opt loading

* add n1536

* Add two pass pipeline

* format

* Fix incorrect type

* parallel compilation

* Use smaller N

* fix 2p pass

* Support Repeat_M in distribution

* Refine nameing

* Add reduce example

---------
Co-authored-by: letaoqin <letaoqin@amd.com>
Co-authored-by: aska-0096 <haocwang@amd.com>
Co-authored-by: rocking <ChunYu.Lai@amd.com>
Co-authored-by: carlushuang <carlus.huang@amd.com>

example/66_complex_contraction_bilinear/run_complex_contraction_bilinear_example.inc

@@ -127,44 +127,47 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[])
 
     switch(init_method)
     {
-        case 0: break;
-        case 1:
+    case 0: break;
+    case 1:
 
-            a_ms_ks_re.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
-            b_ns_ks_re.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
-            d_ms_ns_re.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        a_ms_ks_re.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_ns_ks_re.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        d_ms_ns_re.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
 
-            a_ms_ks_img.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
-            b_ns_ks_img.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
-            d_ms_ns_img.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
-            break;
+        a_ms_ks_img.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_ns_ks_img.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        d_ms_ns_img.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        break;
 
-        default:
-            a_ms_ks_re.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
-            b_ns_ks_re.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
-            d_ms_ns_re.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+    default:
+        a_ms_ks_re.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_ns_ks_re.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        d_ms_ns_re.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
 
-            a_ms_ks_img.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
-            b_ns_ks_img.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
-            d_ms_ns_img.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        a_ms_ks_img.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_ns_ks_img.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        d_ms_ns_img.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
 
-            break;
+        break;
     }
 
     DeviceMem a_device_buf_re(sizeof(ADataType) * a_ms_ks_re.mDesc.GetElementSpaceSize());
     DeviceMem b_device_buf_re(sizeof(BDataType) * b_ns_ks_re.mDesc.GetElementSpaceSize());
     DeviceMem d_device_buf_re(sizeof(DDataType) * d_ms_ns_re.mDesc.GetElementSpaceSize());
-    DeviceMem e_device_buf_re(sizeof(EDataType) * e_ms_ns_device_result_re.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf_re(sizeof(EDataType) *
+                              e_ms_ns_device_result_re.mDesc.GetElementSpaceSize());
 
     DeviceMem a_device_buf_img(sizeof(ADataType) * a_ms_ks_img.mDesc.GetElementSpaceSize());
     DeviceMem b_device_buf_img(sizeof(BDataType) * b_ns_ks_img.mDesc.GetElementSpaceSize());
     DeviceMem d_device_buf_img(sizeof(DDataType) * d_ms_ns_img.mDesc.GetElementSpaceSize());
-    DeviceMem e_device_buf_img(sizeof(EDataType) * e_ms_ns_device_result_img.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf_img(sizeof(EDataType) *
+                               e_ms_ns_device_result_img.mDesc.GetElementSpaceSize());
 
     // Intermediate Value For E Real and Img
-    DeviceMem e_device_buf_re1(sizeof(EDataType) * e_ms_ns_device_result_re.mDesc.GetElementSpaceSize());
-    DeviceMem e_device_buf_img1(sizeof(EDataType) * e_ms_ns_device_result_img.mDesc.GetElementSpaceSize());
-
+    DeviceMem e_device_buf_re1(sizeof(EDataType) *
+                               e_ms_ns_device_result_re.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf_img1(sizeof(EDataType) *
+                                e_ms_ns_device_result_img.mDesc.GetElementSpaceSize());
 
     a_device_buf_re.ToDevice(a_ms_ks_re.mData.data());
     b_device_buf_re.ToDevice(b_ns_ks_re.mData.data());
@@ -181,7 +184,7 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[])
     // set zero for intermediate values
     e_device_buf_re1.SetZero();
     e_device_buf_img1.SetZero();
- 
+
     auto a_element_op   = AElementOp{};
     auto b_element_op   = BElementOp{};
     auto cde_element_op = CDEElementOp{alpha, beta};
@@ -189,23 +192,24 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[])
     // device operation
     // For real Intermediate Value re_1
 
-    auto op       = DeviceOpInstance{};
-    auto invoker  = op.MakeInvoker();
-    auto argument_re1 = op.MakeArgument(a_device_buf_re.GetDeviceBuffer(),
-                                    b_device_buf_re.GetDeviceBuffer(),
-                                    std::array<const void*, 1>{d_device_buf_re.GetDeviceBuffer()},
-                                    e_device_buf_re1.GetDeviceBuffer(),
-                                    a_ms_ks_lengths,
-                                    a_ms_ks_strides,
-                                    b_ns_ks_lengths,
-                                    b_ns_ks_strides,
-                                    std::array<std::vector<ck::index_t>, 1>{d_ms_ns_lengths},
-                                    std::array<std::vector<ck::index_t>, 1>{d_ms_ns_strides},
-                                    e_ms_ns_lengths,
-                                    e_ms_ns_strides,
-                                    a_element_op,
-                                    b_element_op,
-                                    cde_element_op);
+    auto op      = DeviceOpInstance{};
+    auto invoker = op.MakeInvoker();
+    auto argument_re1 =
+        op.MakeArgument(a_device_buf_re.GetDeviceBuffer(),
+                        b_device_buf_re.GetDeviceBuffer(),
+                        std::array<const void*, 1>{d_device_buf_re.GetDeviceBuffer()},
+                        e_device_buf_re1.GetDeviceBuffer(),
+                        a_ms_ks_lengths,
+                        a_ms_ks_strides,
+                        b_ns_ks_lengths,
+                        b_ns_ks_strides,
+                        std::array<std::vector<ck::index_t>, 1>{d_ms_ns_lengths},
+                        std::array<std::vector<ck::index_t>, 1>{d_ms_ns_strides},
+                        e_ms_ns_lengths,
+                        e_ms_ns_strides,
+                        a_element_op,
+                        b_element_op,
+                        cde_element_op);
 
     if(!op.IsSupportedArgument(argument_re1))
     {
@@ -216,7 +220,6 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[])
 
     float ave_time_re1 = invoker.Run(argument_re1, StreamConfig{nullptr, time_kernel});
 
-
     alpha = -1.f;
     beta  = 1.f;
 
@@ -228,21 +231,22 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[])
     // For real Intermediate Value re_2
     // auto op       = DeviceOpInstance{};
     // auto invoker  = op.MakeInvoker();
-    auto argument_re2 = op.MakeArgument(a_device_buf_img.GetDeviceBuffer(),
-                                    b_device_buf_img.GetDeviceBuffer(),
-                                    std::array<const void*, 1>{e_device_buf_re1.GetDeviceBuffer()},
-                                    e_device_buf_re.GetDeviceBuffer(),
-                                    a_ms_ks_lengths,
-                                    a_ms_ks_strides,
-                                    b_ns_ks_lengths,
-                                    b_ns_ks_strides,
-                                    std::array<std::vector<ck::index_t>, 1>{d_ms_ns_lengths},
-                                    std::array<std::vector<ck::index_t>, 1>{d_ms_ns_strides},
-                                    e_ms_ns_lengths,
-                                    e_ms_ns_strides,
-                                    a_element_op,
-                                    b_element_op,
-                                    cde_element_op);
+    auto argument_re2 =
+        op.MakeArgument(a_device_buf_img.GetDeviceBuffer(),
+                        b_device_buf_img.GetDeviceBuffer(),
+                        std::array<const void*, 1>{e_device_buf_re1.GetDeviceBuffer()},
+                        e_device_buf_re.GetDeviceBuffer(),
+                        a_ms_ks_lengths,
+                        a_ms_ks_strides,
+                        b_ns_ks_lengths,
+                        b_ns_ks_strides,
+                        std::array<std::vector<ck::index_t>, 1>{d_ms_ns_lengths},
+                        std::array<std::vector<ck::index_t>, 1>{d_ms_ns_strides},
+                        e_ms_ns_lengths,
+                        e_ms_ns_strides,
+                        a_element_op,
+                        b_element_op,
+                        cde_element_op);
 
     if(!op.IsSupportedArgument(argument_re2))
     {
@@ -253,7 +257,6 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[])
 
     float ave_time_re2 = invoker.Run(argument_re2, StreamConfig{nullptr, time_kernel});
 
-    
     alpha = 1.f;
     beta  = 1.f;
 
@@ -261,22 +264,22 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[])
     b_element_op   = BElementOp{};
     cde_element_op = CDEElementOp{alpha, beta};
 
-    auto argument_img1 = op.MakeArgument(a_device_buf_re.GetDeviceBuffer(),
-                                b_device_buf_img.GetDeviceBuffer(),
-                                std::array<const void*, 1>{d_device_buf_img.GetDeviceBuffer()},
-                                e_device_buf_img1.GetDeviceBuffer(),
-                                a_ms_ks_lengths,
-                                a_ms_ks_strides,
-                                b_ns_ks_lengths,
-                                b_ns_ks_strides,
-                                std::array<std::vector<ck::index_t>, 1>{d_ms_ns_lengths},
-                                std::array<std::vector<ck::index_t>, 1>{d_ms_ns_strides},
-                                e_ms_ns_lengths,
-                                e_ms_ns_strides,
-                                a_element_op,
-                                b_element_op,
-                                cde_element_op);
-
+    auto argument_img1 =
+        op.MakeArgument(a_device_buf_re.GetDeviceBuffer(),
+                        b_device_buf_img.GetDeviceBuffer(),
+                        std::array<const void*, 1>{d_device_buf_img.GetDeviceBuffer()},
+                        e_device_buf_img1.GetDeviceBuffer(),
+                        a_ms_ks_lengths,
+                        a_ms_ks_strides,
+                        b_ns_ks_lengths,
+                        b_ns_ks_strides,
+                        std::array<std::vector<ck::index_t>, 1>{d_ms_ns_lengths},
+                        std::array<std::vector<ck::index_t>, 1>{d_ms_ns_strides},
+                        e_ms_ns_lengths,
+                        e_ms_ns_strides,
+                        a_element_op,
+                        b_element_op,
+                        cde_element_op);
 
     if(!op.IsSupportedArgument(argument_img1))
     {
@@ -290,23 +293,22 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[])
     alpha = 1.f;
     beta  = 1.f;
 
-    auto argument_img2 = op.MakeArgument(a_device_buf_img.GetDeviceBuffer(),
-                                b_device_buf_re.GetDeviceBuffer(),
-                                std::array<const void*, 1>{e_device_buf_img1.GetDeviceBuffer()},
-                                e_device_buf_img.GetDeviceBuffer(),
-                                a_ms_ks_lengths,
-                                a_ms_ks_strides,
-                                b_ns_ks_lengths,
-                                b_ns_ks_strides,
-                                std::array<std::vector<ck::index_t>, 1>{d_ms_ns_lengths},
-                                std::array<std::vector<ck::index_t>, 1>{d_ms_ns_strides},
-                                e_ms_ns_lengths,
-                                e_ms_ns_strides,
-                                a_element_op,
-                                b_element_op,
-                                cde_element_op);
-
-
+    auto argument_img2 =
+        op.MakeArgument(a_device_buf_img.GetDeviceBuffer(),
+                        b_device_buf_re.GetDeviceBuffer(),
+                        std::array<const void*, 1>{e_device_buf_img1.GetDeviceBuffer()},
+                        e_device_buf_img.GetDeviceBuffer(),
+                        a_ms_ks_lengths,
+                        a_ms_ks_strides,
+                        b_ns_ks_lengths,
+                        b_ns_ks_strides,
+                        std::array<std::vector<ck::index_t>, 1>{d_ms_ns_lengths},
+                        std::array<std::vector<ck::index_t>, 1>{d_ms_ns_strides},
+                        e_ms_ns_lengths,
+                        e_ms_ns_strides,
+                        a_element_op,
+                        b_element_op,
+                        cde_element_op);
 
     if(!op.IsSupportedArgument(argument_img2))
     {
@@ -317,7 +319,6 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[])
 
     float ave_time_img2 = invoker.Run(argument_img2, StreamConfig{nullptr, time_kernel});
 
-
     ck::index_t M =
         ck::accumulate_n<ck::index_t>(e_ms_ns_lengths.begin(), NumDimM, 1, std::multiplies<>{});
 
@@ -331,9 +332,9 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[])
     std::size_t num_btype = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
                             sizeof(DDataType) * M * N + sizeof(EDataType) * M * N * 2;
 
-    float ave_time = ave_time_img2 + ave_time_img1 + ave_time_re2 + ave_time_re1 ; 
+    float ave_time = ave_time_img2 + ave_time_img1 + ave_time_re2 + ave_time_re1;
 
-    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+    float tflops     = static_cast<float>(flop) / 1.E9 / ave_time;
     float gb_per_sec = num_btype / 1.E6 / ave_time;
 
     std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
@@ -343,7 +344,7 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[])
     e_device_buf_img.FromDevice(e_ms_ns_device_result_img.mData.data());
 
     auto isRealOk = 0;
-    auto isImgOk = 0;
+    auto isImgOk  = 0;
 
     if(do_verification)
     {
@@ -366,17 +367,16 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[])
         auto ref_op      = ReferenceOpInstance{};
         auto ref_invoker = ref_op.MakeInvoker();
 
-        auto ref_argument_re =
-            ref_op.MakeArgument(a_ms_ks_re, b_ns_ks_re, c_ms_ns_host_result_re, a_element_op, b_element_op);
+        auto ref_argument_re = ref_op.MakeArgument(
+            a_ms_ks_re, b_ns_ks_re, c_ms_ns_host_result_re, a_element_op, b_element_op);
 
         ref_invoker.Run(ref_argument_re);
 
         alpha = 1.f;
         beta  = 1.f;
-   
+
         cde_element_op = CDEElementOp{alpha, beta};
 
-       
         for(size_t m0 = 0; m0 < e_ms_ns_host_result_re.mDesc.GetLengths()[0]; ++m0)
         {
             for(size_t m1 = 0; m1 < e_ms_ns_host_result_re.mDesc.GetLengths()[1]; ++m1)
@@ -395,11 +395,11 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[])
 
         alpha = 1.f;
         beta  = -1.f;
-   
+
         cde_element_op = CDEElementOp{alpha, beta};
 
-        auto ref_argument_re1 =
-            ref_op.MakeArgument(a_ms_ks_img, b_ns_ks_img, c_ms_ns_host_result_re1, a_element_op, b_element_op);
+        auto ref_argument_re1 = ref_op.MakeArgument(
+            a_ms_ks_img, b_ns_ks_img, c_ms_ns_host_result_re1, a_element_op, b_element_op);
 
         ref_invoker.Run(ref_argument_re1);
 
@@ -419,23 +419,20 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[])
             }
         }
 
-        isRealOk =  ck::utils::check_err(e_ms_ns_device_result_re, e_ms_ns_host_result_re) ? 0 : 1;
-
-        
-
+        isRealOk = ck::utils::check_err(e_ms_ns_device_result_re, e_ms_ns_host_result_re) ? 0 : 1;
 
         // Img Part Verification
         Tensor<CShuffleDataType> c_ms_ns_host_result_img(e_ms_ns_lengths, e_ms_ns_strides);
         Tensor<CShuffleDataType> c_ms_ns_host_result_img1(e_ms_ns_lengths, e_ms_ns_strides);
 
-        auto ref_argument_img =
-            ref_op.MakeArgument(a_ms_ks_re, b_ns_ks_img, c_ms_ns_host_result_img, a_element_op, b_element_op);
-            
+        auto ref_argument_img = ref_op.MakeArgument(
+            a_ms_ks_re, b_ns_ks_img, c_ms_ns_host_result_img, a_element_op, b_element_op);
+
         ref_invoker.Run(ref_argument_img);
 
         alpha = 1.f;
         beta  = 1.f;
-   
+
         cde_element_op = CDEElementOp{alpha, beta};
 
         for(size_t m0 = 0; m0 < e_ms_ns_host_result_img.mDesc.GetLengths()[0]; ++m0)
@@ -454,9 +451,9 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[])
             }
         }
 
-        auto ref_argument_img1 =
-            ref_op.MakeArgument(a_ms_ks_img, b_ns_ks_re, c_ms_ns_host_result_img1, a_element_op, b_element_op);
-            
+        auto ref_argument_img1 = ref_op.MakeArgument(
+            a_ms_ks_img, b_ns_ks_re, c_ms_ns_host_result_img1, a_element_op, b_element_op);
+
         ref_invoker.Run(ref_argument_img1);
 
         for(size_t m0 = 0; m0 < e_ms_ns_host_result_img.mDesc.GetLengths()[0]; ++m0)
@@ -475,7 +472,7 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[])
             }
         }
 
-        isImgOk =  ck::utils::check_err(e_ms_ns_device_result_re, e_ms_ns_host_result_re) ? 0 : 1;
+        isImgOk = ck::utils::check_err(e_ms_ns_device_result_re, e_ms_ns_host_result_re) ? 0 : 1;
 
         return (isRealOk && isImgOk);
     }

example/ck_tile/02_layernorm2d/CMakeLists.txt

+set(EXAMPLE_LAYERNORM2D_FWD "tile_example_layernorm2d_fwd")
 # not using add_example_executable() to add this target, since we don't want this to have
 # to be included in "make all/install/check"
-add_executable(tile_example_layernorm2d_fwd EXCLUDE_FROM_ALL layernorm2d_fwd.cpp)
-target_compile_options(tile_example_layernorm2d_fwd PRIVATE -DSAVE_MEAN_INV_STD)
\ No newline at end of file
+message("adding example ${EXAMPLE_LAYERNORM2D_FWD}")
+file(GLOB INSTANCE_SRCS instances/*.cpp)
+add_executable(${EXAMPLE_LAYERNORM2D_FWD} EXCLUDE_FROM_ALL layernorm2d_fwd.cpp)
+target_include_directories(${EXAMPLE_LAYERNORM2D_FWD} PRIVATE ${CMAKE_CURRENT_LIST_DIR})
+target_sources(${EXAMPLE_LAYERNORM2D_FWD} PRIVATE ${INSTANCE_SRCS})
+
+set(EXAMPLE_LAYERNORM2D_FWD_COMPILE_OPTIONS)
+
+# NOTE: we turn off undefined-func-template to let source compile without explicit declare function specializations
+list(APPEND EXAMPLE_LAYERNORM2D_FWD_COMPILE_OPTIONS -Wno-undefined-func-template -Wno-float-equal)
+
+target_compile_options(${EXAMPLE_LAYERNORM2D_FWD} PRIVATE ${EXAMPLE_LAYERNORM2D_FWD_COMPILE_OPTIONS})
+
+# TODO: we have to turn off this global prop, otherwise the progress bar generated
+# by cmake will print too many files, execvp: /bin/sh: Argument list too long
+# however, this property may affect global
+# TODO: consider codegen a makefile by us
+set_property(GLOBAL PROPERTY RULE_MESSAGES OFF)

example/ck_tile/02_layernorm2d/README.md

@@ -6,8 +6,7 @@ This folder contains example for Layernorm2D forward using ck_tile tile-programm
 ```
 # in the root of ck_tile
 mkdir build && cd build
-# you can replace <arch> with the appropriate architecture (for example gfx90a or gfx942) or leave it blank
-sh ../script/cmake-ck-dev.sh  ../ <arch>
+sh ../script/cmake-ck-dev.sh  ../ <arch>  # you can replace this <arch> to gfx90a, gfx942...
 make tile_example_layernorm2d_fwd -j
 ```
 This will result in an executable `build/bin/tile_example_layernorm2d_fwd`
@@ -20,4 +19,4 @@ args:
           -e    epsilon (default:1e-5)
           -v    cpu validation or not (default:1)
        -prec    precision (default:fp16)
-```
\ No newline at end of file
+```

example/ck_tile/02_layernorm2d/instances/layernorm2d_fwd_api.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <ck_tile/core.hpp>
+#include "layernorm2d_fwd.hpp"
+
+template <typename DataType_,
+          ck_tile::index_t Repeat_M_,         // each thread repeat along M
+          ck_tile::index_t Repeat_N_,         // each thread repeat along N
+          ck_tile::index_t ThreadPerBlock_M_, // num threads along M
+          ck_tile::index_t ThreadPerBlock_N_, // num threads along N
+          ck_tile::index_t Vector_N_,         // vector size along N
+          bool kPadN_,
+          bool kSaveMeanInvStd_,
+          bool kTwoPass_>
+using trait_ = layernorm2d_fwd_traits_<DataType_,
+                                       Repeat_M_,
+                                       Repeat_N_,
+                                       ThreadPerBlock_M_,
+                                       ThreadPerBlock_N_,
+                                       Vector_N_,
+                                       kPadN_,
+                                       kSaveMeanInvStd_,
+                                       kTwoPass_>;
+
+template <typename data_type>
+float layernorm2d_fwd_b16_(layernorm2d_fwd_traits /*t*/,
+                           layernorm2d_fwd_args a,
+                           const ck_tile::stream_config& s)
+{
+#if 1
+    float r = -1;
+    // clang-format off
+    //                                            rm  rn  tm   tn  vn  pd     mv     2p
+    if(a.n <= 64) {
+            r = layernorm2d_fwd_<trait_<data_type, 1,  1,  4,  64, 1,  true,  false, false>>(s, a);
+    }
+    else if(a.n <= 128) {
+        if (a.n % 2 == 0)
+            r = layernorm2d_fwd_<trait_<data_type, 1,  1,  4,  64, 2,  true,  false, false>>(s, a);
+        else
+            r = layernorm2d_fwd_<trait_<data_type, 1,  2,  4,  64, 1,  true,  false, false>>(s, a);
+    }
+    else if(a.n <= 256) {
+        if (a.n % 4 == 0)
+            r = layernorm2d_fwd_<trait_<data_type,  1, 1,  4,  64, 4,  true,  false, false>>(s, a);
+        else if (a.n % 2 == 0)
+            r = layernorm2d_fwd_<trait_<data_type,  1, 2,  4,  64, 2,  true,  false, false>>(s, a);
+        else
+            r = layernorm2d_fwd_<trait_<data_type,  1, 4,  4,  64, 1,  true,  false, false>>(s, a);
+    }
+    else if(a.n <= 512) {
+        if (a.n % 8 == 0)
+            r = layernorm2d_fwd_<trait_<data_type,  1, 1,  4,  64, 8,  true,  false, false>>(s, a);
+        else if (a.n % 4 == 0)
+            r = layernorm2d_fwd_<trait_<data_type,  1, 2,  4,  64, 4,  true,  false, false>>(s, a);
+        else if (a.n % 2 == 0)
+            r = layernorm2d_fwd_<trait_<data_type,  1, 4,  4,  64, 2,  true,  false, false>>(s, a);
+        else
+            r = layernorm2d_fwd_<trait_<data_type,  1, 8,  4,  64, 1,  true,  false, false>>(s, a);
+    }
+    else if(a.n <= 768) {
+        if (a.n % 4 == 0)
+            r = layernorm2d_fwd_<trait_<data_type,  1, 3,  4,  64, 4,  true,  false, false>>(s, a);
+        else if (a.n % 2 == 0)
+            r = layernorm2d_fwd_<trait_<data_type,  1, 6,  4,  64, 2,  true,  false, false>>(s, a);
+        else
+            r = layernorm2d_fwd_<trait_<data_type,  1,12,  4,  64, 1,  true,  false, false>>(s, a);
+    }
+    else if(a.n <= 1024) {
+        if (a.n % 8 == 0)
+            r = layernorm2d_fwd_<trait_<data_type,  1, 1, 2,  128, 8,  true,  false, false>>(s, a);
+        else if (a.n % 4 == 0)
+            r = layernorm2d_fwd_<trait_<data_type,  1, 2, 2,  128, 4,  true,  false, false>>(s, a);
+        else if (a.n % 2 == 0)
+            r = layernorm2d_fwd_<trait_<data_type,  1, 4, 2,  128, 2,  true,  false, false>>(s, a);
+        else
+            r = layernorm2d_fwd_<trait_<data_type,  1, 4, 1,  256, 1,  true,  false, false>>(s, a);
+    }
+    else if(a.n <= 1536) {
+        if (a.n % 8 == 0)
+            r = layernorm2d_fwd_<trait_<data_type,  1, 3, 4,   64, 8,  true,  false, false>>(s, a);
+        else if (a.n % 4 == 0)
+            r = layernorm2d_fwd_<trait_<data_type,  1, 3, 2,  128, 4,  true,  false, false>>(s, a);
+        else if (a.n % 2 == 0)
+            r = layernorm2d_fwd_<trait_<data_type,  1, 3, 1,  256, 2,  true,  false, false>>(s, a);
+        else
+            r = layernorm2d_fwd_<trait_<data_type,  1, 6, 1,  256, 1,  true,  false, false>>(s, a);
+    }
+    else if(a.n <= 2048) {
+        if (a.n % 8 == 0)
+            r = layernorm2d_fwd_<trait_<data_type,  1, 1, 1,  256, 8,  true,  false, false>>(s, a);
+        else if (a.n % 4 == 0)
+            r = layernorm2d_fwd_<trait_<data_type,  1, 2, 1,  256, 4,  true,  false, false>>(s, a);
+        else if (a.n % 2 == 0)
+            r = layernorm2d_fwd_<trait_<data_type,  1, 4, 1,  256, 2,  true,  false, false>>(s, a);
+        else
+            r = layernorm2d_fwd_<trait_<data_type,  1, 8, 1,  256, 1,  true,  false, false>>(s, a);
+    }
+    else if(a.n <= 3072) {
+        if (a.n % 8 == 0)
+            r = layernorm2d_fwd_<trait_<data_type,  1, 3, 1,  128, 8,  true,  false, false>>(s, a);
+        else if (a.n % 4 == 0)
+            r = layernorm2d_fwd_<trait_<data_type,  1, 3, 1,  256, 4,  true,  false, false>>(s, a);
+        else if (a.n % 2 == 0)
+            r = layernorm2d_fwd_<trait_<data_type,  1, 6, 1,  256, 2,  true,  false, false>>(s, a);
+        else
+            r = layernorm2d_fwd_<trait_<data_type,  1, 3, 1, 1024, 1,  true,  false, false>>(s, a);
+    }
+    else if(a.n <= 4096) {
+        if (a.n % 8 == 0)
+            r = layernorm2d_fwd_<trait_<data_type,  1, 2, 1,  256, 8,  true,  false, false>>(s, a);
+        else if (a.n % 4 == 0)
+            r = layernorm2d_fwd_<trait_<data_type,  1, 4, 1,  256, 4,  true,  false, false>>(s, a);
+        else if (a.n % 2 == 0)
+            r = layernorm2d_fwd_<trait_<data_type,  1, 2, 1, 1024, 2,  true,  false, false>>(s, a);
+        else
+            r = layernorm2d_fwd_<trait_<data_type,  1, 4, 1, 1024, 1,  true,  false, false>>(s, a);
+    }
+    else if(a.n > 4096) {
+        if (a.n % 8 == 0)
+            r = layernorm2d_fwd_<trait_<data_type,  1, 2, 1,  256, 8,  true,  false, true>>(s, a);
+        else if (a.n % 4 == 0)
+            r = layernorm2d_fwd_<trait_<data_type,  1, 4, 1,  256, 4,  true,  false, true>>(s, a);
+        else if (a.n % 2 == 0)
+            r = layernorm2d_fwd_<trait_<data_type,  1, 2, 1, 1024, 2,  true,  false, true>>(s, a);
+        else
+            r = layernorm2d_fwd_<trait_<data_type,  1, 4, 1, 1024, 1,  true,  false, true>>(s, a);
+    }
+    return r;
+#else
+    return layernorm2d_fwd_<trait_<data_type,  1, 1,  1,  256, 4,  true,  false, false>>(s, a);
+#endif
+    // clang-format on
+}
+
+float layernorm2d_fwd(layernorm2d_fwd_traits t,
+                      layernorm2d_fwd_args a,
+                      const ck_tile::stream_config& s)
+{
+
+    float r = -1;
+    if(t.data_type.compare("fp16") == 0)
+    {
+        return layernorm2d_fwd_b16_<ck_tile::fp16_t>(t, a, s);
+    }
+    else if(t.data_type.compare("bf16") == 0)
+    {
+        return layernorm2d_fwd_b16_<ck_tile::bf16_t>(t, a, s);
+    }
+    if(r < 0)
+        throw std::runtime_error("Without supported instances!");
+
+    return r;
+}

example/ck_tile/02_layernorm2d/instances/layernorm2d_fwd_bf16_n1024_instance.cpp

+
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "layernorm2d_fwd_instance_common.hpp"
+
+// clang-format off
+//                                                       rm  rn  tm  tn  vn  pd     mv     2p
+#if 0
+template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1,  2,  4,  64, 8,  true , false, false>>(const S&, A);
+template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1,  4,  4,  64, 4,  true , false, false>>(const S&, A);
+template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1,  8,  4,  64, 2,  true , false, false>>(const S&, A);
+template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 16,  4,  64, 1,  true , false, false>>(const S&, A);
+
+template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1,  1,  1, 256, 4,  true , false, false>>(const S&, A);
+#endif
+
+template float layernorm2d_fwd_<trait_<ck_tile::bf16_t,  1, 1, 2,  128, 8,  true,  false, false>>(const S&, A);
+template float layernorm2d_fwd_<trait_<ck_tile::bf16_t,  1, 2, 2,  128, 4,  true,  false, false>>(const S&, A);
+template float layernorm2d_fwd_<trait_<ck_tile::bf16_t,  1, 4, 2,  128, 2,  true,  false, false>>(const S&, A);
+template float layernorm2d_fwd_<trait_<ck_tile::bf16_t,  1, 4, 1,  256, 1,  true,  false, false>>(const S&, A);
+// clang-format on

example/ck_tile/02_layernorm2d/instances/layernorm2d_fwd_bf16_n1536_instance.cpp

+
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "layernorm2d_fwd_instance_common.hpp"
+
+// clang-format off
+//                                                       rm  rn  tm  tn  vn  pd     mv     2p
+template float layernorm2d_fwd_<trait_<ck_tile::bf16_t,  1, 3, 4,   64, 8,  true,  false, false>>(const S&, A);
+template float layernorm2d_fwd_<trait_<ck_tile::bf16_t,  1, 3, 2,  128, 4,  true,  false, false>>(const S&, A);
+template float layernorm2d_fwd_<trait_<ck_tile::bf16_t,  1, 3, 1,  256, 2,  true,  false, false>>(const S&, A);
+template float layernorm2d_fwd_<trait_<ck_tile::bf16_t,  1, 6, 1,  256, 1,  true,  false, false>>(const S&, A);
+// clang-format on

example/ck_tile/02_layernorm2d/instances/layernorm2d_fwd_bf16_n2048_instance.cpp

+
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "layernorm2d_fwd_instance_common.hpp"
+
+// clang-format off
+//                                                       rm  rn  tm  tn  vn  pd     mv     2p
+template float layernorm2d_fwd_<trait_<ck_tile::bf16_t,  1, 1, 1,  256, 8,  true,  false, false>>(const S&, A);
+template float layernorm2d_fwd_<trait_<ck_tile::bf16_t,  1, 2, 1,  256, 4,  true,  false, false>>(const S&, A);
+template float layernorm2d_fwd_<trait_<ck_tile::bf16_t,  1, 4, 1,  256, 2,  true,  false, false>>(const S&, A);
+template float layernorm2d_fwd_<trait_<ck_tile::bf16_t,  1, 8, 1,  256, 1,  true,  false, false>>(const S&, A);
+
+// clang-format on

example/ck_tile/02_layernorm2d/instances/layernorm2d_fwd_bf16_n256_instance.cpp

+
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "layernorm2d_fwd_instance_common.hpp"
+
+// clang-format off
+//                                                       rm  rn  tm  tn  vn  pd     mv     2p
+template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1,  1,  4,  64, 4,  true , false, false>>(const S&, A);
+template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1,  2,  4,  64, 2,  true , false, false>>(const S&, A);
+template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1,  4,  4,  64, 1,  true , false, false>>(const S&, A);
+// clang-format on

example/ck_tile/02_layernorm2d/instances/layernorm2d_fwd_bf16_n3072_instance.cpp

+
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "layernorm2d_fwd_instance_common.hpp"
+
+// clang-format off
+//                                                       rm  rn  tm  tn  vn  pd     mv     2p
+template float layernorm2d_fwd_<trait_<ck_tile::bf16_t,  1, 3, 1,  128, 8,  true,  false, false>>(const S&, A);
+template float layernorm2d_fwd_<trait_<ck_tile::bf16_t,  1, 3, 1,  256, 4,  true,  false, false>>(const S&, A);
+template float layernorm2d_fwd_<trait_<ck_tile::bf16_t,  1, 6, 1,  256, 2,  true,  false, false>>(const S&, A);
+template float layernorm2d_fwd_<trait_<ck_tile::bf16_t,  1, 3, 1, 1024, 1,  true,  false, false>>(const S&, A);
+
+// clang-format on

example/ck_tile/02_layernorm2d/instances/layernorm2d_fwd_bf16_n4096_instance.cpp

+
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "layernorm2d_fwd_instance_common.hpp"
+
+// clang-format off
+//                                                       rm  rn  tm  tn  vn  pd     mv     2p
+template float layernorm2d_fwd_<trait_<ck_tile::bf16_t,  1, 2, 1,  256, 8,  true,  false, false>>(const S&, A);
+template float layernorm2d_fwd_<trait_<ck_tile::bf16_t,  1, 4, 1,  256, 4,  true,  false, false>>(const S&, A);
+template float layernorm2d_fwd_<trait_<ck_tile::bf16_t,  1, 2, 1, 1024, 2,  true,  false, false>>(const S&, A);
+template float layernorm2d_fwd_<trait_<ck_tile::bf16_t,  1, 4, 1, 1024, 1,  true,  false, false>>(const S&, A);
+
+// clang-format on

example/ck_tile/02_layernorm2d/instances/layernorm2d_fwd_bf16_n4096_tp_instance.cpp

+
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "layernorm2d_fwd_instance_common.hpp"
+
+// clang-format off
+//                                                       rm  rn  tm  tn  vn  pd     mv     2p
+template float layernorm2d_fwd_<trait_<ck_tile::bf16_t,  1, 2, 1,  256, 8,  true,  false, true>>(const S&, A);
+template float layernorm2d_fwd_<trait_<ck_tile::bf16_t,  1, 4, 1,  256, 4,  true,  false, true>>(const S&, A);
+template float layernorm2d_fwd_<trait_<ck_tile::bf16_t,  1, 2, 1, 1024, 2,  true,  false, true>>(const S&, A);
+template float layernorm2d_fwd_<trait_<ck_tile::bf16_t,  1, 4, 1, 1024, 1,  true,  false, true>>(const S&, A);
+
+// clang-format on

example/ck_tile/02_layernorm2d/instances/layernorm2d_fwd_bf16_n512_instance.cpp

+
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "layernorm2d_fwd_instance_common.hpp"
+
+// clang-format off
+//                                                       rm  rn  tm  tn  vn  pd     mv     2p
+template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1,  1,  4,  64, 8,  true , false, false>>(const S&, A);
+template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1,  2,  4,  64, 4,  true , false, false>>(const S&, A);
+template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1,  4,  4,  64, 2,  true , false, false>>(const S&, A);
+template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1,  8,  4,  64, 1,  true , false, false>>(const S&, A);
+// clang-format on

example/ck_tile/02_layernorm2d/instances/layernorm2d_fwd_bf16_n64_n128_instance.cpp

+
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "layernorm2d_fwd_instance_common.hpp"
+
+// clang-format off
+//                                                       rm  rn  tm  tn  vn  pd     mv     2p
+template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1,  1,  4,  64, 1,  true , false, false>>(const S&, A);
+template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1,  1,  4,  64, 2,  true , false, false>>(const S&, A);
+template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1,  2,  4,  64, 1,  true , false, false>>(const S&, A);
+// clang-format on

example/ck_tile/02_layernorm2d/instances/layernorm2d_fwd_bf16_n768_instance.cpp

+
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "layernorm2d_fwd_instance_common.hpp"
+
+// clang-format off
+//                                                       rm  rn  tm  tn  vn  pd     mv     2p
+template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1,  3,  4,  64, 4,  true , false, false>>(const S&, A);
+template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1,  6,  4,  64, 2,  true , false, false>>(const S&, A);
+template float layernorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 12,  4,  64, 1,  true , false, false>>(const S&, A);
+// clang-format on

example/ck_tile/02_layernorm2d/instances/layernorm2d_fwd_fp16_n1024_instance.cpp

+
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "layernorm2d_fwd_instance_common.hpp"
+
+// clang-format off
+//                                                       rm  rn  tm  tn  vn  pd     mv     2p
+#if 0
+template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1,  2,  4,  64, 8,  true , false, false>>(const S&, A);
+template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1,  4,  4,  64, 4,  true , false, false>>(const S&, A);
+template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1,  8,  4,  64, 2,  true , false, false>>(const S&, A);
+template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 16,  4,  64, 1,  true , false, false>>(const S&, A);
+
+template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1,  1,  1, 256, 4,  true , false, false>>(const S&, A);
+#endif
+
+template float layernorm2d_fwd_<trait_<ck_tile::fp16_t,  1, 1, 2,  128, 8,  true,  false, false>>(const S&, A);
+template float layernorm2d_fwd_<trait_<ck_tile::fp16_t,  1, 2, 2,  128, 4,  true,  false, false>>(const S&, A);
+template float layernorm2d_fwd_<trait_<ck_tile::fp16_t,  1, 4, 2,  128, 2,  true,  false, false>>(const S&, A);
+template float layernorm2d_fwd_<trait_<ck_tile::fp16_t,  1, 4, 1,  256, 1,  true,  false, false>>(const S&, A);
+// clang-format on

example/ck_tile/02_layernorm2d/instances/layernorm2d_fwd_fp16_n1536_instance.cpp

+
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "layernorm2d_fwd_instance_common.hpp"
+
+// clang-format off
+//                                                       rm  rn  tm  tn  vn  pd     mv     2p
+template float layernorm2d_fwd_<trait_<ck_tile::fp16_t,  1, 3, 4,   64, 8,  true,  false, false>>(const S&, A);
+template float layernorm2d_fwd_<trait_<ck_tile::fp16_t,  1, 3, 2,  128, 4,  true,  false, false>>(const S&, A);
+template float layernorm2d_fwd_<trait_<ck_tile::fp16_t,  1, 3, 1,  256, 2,  true,  false, false>>(const S&, A);
+template float layernorm2d_fwd_<trait_<ck_tile::fp16_t,  1, 6, 1,  256, 1,  true,  false, false>>(const S&, A);
+// clang-format on

example/ck_tile/02_layernorm2d/instances/layernorm2d_fwd_fp16_n2048_instance.cpp

+
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "layernorm2d_fwd_instance_common.hpp"
+
+// clang-format off
+//                                                       rm  rn  tm  tn  vn  pd     mv     2p
+template float layernorm2d_fwd_<trait_<ck_tile::fp16_t,  1, 1, 1,  256, 8,  true,  false, false>>(const S&, A);
+template float layernorm2d_fwd_<trait_<ck_tile::fp16_t,  1, 2, 1,  256, 4,  true,  false, false>>(const S&, A);
+template float layernorm2d_fwd_<trait_<ck_tile::fp16_t,  1, 4, 1,  256, 2,  true,  false, false>>(const S&, A);
+template float layernorm2d_fwd_<trait_<ck_tile::fp16_t,  1, 8, 1,  256, 1,  true,  false, false>>(const S&, A);
+
+// clang-format on

example/ck_tile/02_layernorm2d/instances/layernorm2d_fwd_fp16_n256_instance.cpp

+
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "layernorm2d_fwd_instance_common.hpp"
+
+// clang-format off
+//                                                       rm  rn  tm  tn  vn  pd     mv     2p
+template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1,  1,  4,  64, 4,  true , false, false>>(const S&, A);
+template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1,  2,  4,  64, 2,  true , false, false>>(const S&, A);
+template float layernorm2d_fwd_<trait_<ck_tile::fp16_t, 1,  4,  4,  64, 1,  true , false, false>>(const S&, A);
+// clang-format on

example/ck_tile/02_layernorm2d/instances/layernorm2d_fwd_fp16_n3072_instance.cpp

+
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "layernorm2d_fwd_instance_common.hpp"
+
+// clang-format off
+//                                                       rm  rn  tm  tn  vn  pd     mv     2p
+template float layernorm2d_fwd_<trait_<ck_tile::fp16_t,  1, 3, 1,  128, 8,  true,  false, false>>(const S&, A);
+template float layernorm2d_fwd_<trait_<ck_tile::fp16_t,  1, 3, 1,  256, 4,  true,  false, false>>(const S&, A);
+template float layernorm2d_fwd_<trait_<ck_tile::fp16_t,  1, 6, 1,  256, 2,  true,  false, false>>(const S&, A);
+template float layernorm2d_fwd_<trait_<ck_tile::fp16_t,  1, 3, 1, 1024, 1,  true,  false, false>>(const S&, A);
+
+// clang-format on

example/ck_tile/02_layernorm2d/instances/layernorm2d_fwd_fp16_n4096_instance.cpp

+
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "layernorm2d_fwd_instance_common.hpp"
+
+// clang-format off
+//                                                       rm  rn  tm  tn  vn  pd     mv     2p
+template float layernorm2d_fwd_<trait_<ck_tile::fp16_t,  1, 2, 1,  256, 8,  true,  false, false>>(const S&, A);
+template float layernorm2d_fwd_<trait_<ck_tile::fp16_t,  1, 4, 1,  256, 4,  true,  false, false>>(const S&, A);
+template float layernorm2d_fwd_<trait_<ck_tile::fp16_t,  1, 2, 1, 1024, 2,  true,  false, false>>(const S&, A);
+template float layernorm2d_fwd_<trait_<ck_tile::fp16_t,  1, 4, 1, 1024, 1,  true,  false, false>>(const S&, A);
+
+// clang-format on