add prenorm/postnorm support, refactor using generate.py

example/ck_tile/02_layernorm2d/CMakeLists.txt

+set(LAYERNORM2D_FWD_KNOWN_APIS "fwd;bwd")
+set(LAYERNORM2D_FWD_ENABLE_APIS  "fwd" CACHE STRING
+    "semicolon-separated list of APIs to generate (${LAYERNORM2D_FWD_KNOWN_APIS}) & link, or \"all\".")
+if(LAYERNORM2D_FWD_ENABLE_APIS  STREQUAL "all")
+  set(LAYERNORM2D_FWD_ENABLE_APIS  ${LAYERNORM2D_FWD_KNOWN_APIS})
+endif()
+
+# generate a list of kernels, but not actually emit files at config sta
+execute_process(
+  COMMAND ${Python3_EXECUTABLE} ${CMAKE_CURRENT_LIST_DIR}/generate.py
+  --api ${LAYERNORM2D_FWD_ENABLE_APIS} --working_path ${CMAKE_CURRENT_BINARY_DIR} --list_blobs
+  RESULT_VARIABLE ret
+)
+if(ret AND NOT ret EQUAL 0)
+  message( FATAL_ERROR "Fail to generate kernels via Python. ${ret}")
+endif()
+
+file(STRINGS ${CMAKE_CURRENT_BINARY_DIR}/layernorm2d_fwd_blobs.txt LAYERNORM2D_FWD_GEN_BLOBS)
+
+add_custom_command(
+  OUTPUT ${LAYERNORM2D_FWD_GEN_BLOBS}
+  COMMAND ${Python3_EXECUTABLE} ${CMAKE_CURRENT_LIST_DIR}/generate.py
+  --api ${LAYERNORM2D_FWD_ENABLE_APIS} --working_path ${CMAKE_CURRENT_BINARY_DIR} --gen_blobs
+)
+
 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"
+
 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 ${INSTANCE_SRCS})
+target_sources(${EXAMPLE_LAYERNORM2D_FWD} PRIVATE ${LAYERNORM2D_FWD_GEN_BLOBS})
 
 set(EXAMPLE_LAYERNORM2D_FWD_COMPILE_OPTIONS)
 

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

example/ck_tile/02_layernorm2d/instances/layernorm2d_fwd_fp16_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::fp16_t,  1, 2, 1,  256, 8,  true,  false, true>>(const S&, A);
-template float layernorm2d_fwd_<trait_<ck_tile::fp16_t,  1, 4, 1,  256, 4,  true,  false, true>>(const S&, A);
-template float layernorm2d_fwd_<trait_<ck_tile::fp16_t,  1, 2, 1, 1024, 2,  true,  false, true>>(const S&, A);
-template float layernorm2d_fwd_<trait_<ck_tile::fp16_t,  1, 4, 1, 1024, 1,  true,  false, true>>(const S&, A);
-
-// clang-format on