Merge branch 'develop' into gfx950

example/ck_tile/03_gemm/gemm_mem_pipeline.cpp → example/ck_tile/03_gemm/universal_gemm.cpp

@@ -14,12 +14,34 @@
 #include "ck_tile/host.hpp"
 #include "gemm_basic.hpp"
 
+#define CK_TILE_PIPELINE_COMPUTE 1
+#define CK_TILE_PIPELINE_MEMORY 2
+
+#ifndef CK_TILE_PIPELINE_DEFAULT
+#define CK_TILE_PIPELINE_DEFAULT CK_TILE_PIPELINE_COMPUTE
+#endif
+
 template <typename ALayout, typename BLayout, typename CLayout>
 float gemm_calc(const gemm_basic_args& args, const ck_tile::stream_config& s)
 {
-    // ToDo: This will be modified by the codegen code later.
+#if(CK_TILE_PIPELINE_DEFAULT == CK_TILE_PIPELINE_MEMORY)
+    // Memory friendly for Interwave scheduler
     constexpr ck_tile::index_t M_Tile = 128;
-    constexpr ck_tile::index_t N_Tile = 128;
+    constexpr ck_tile::index_t N_Tile = 32;
+    constexpr ck_tile::index_t K_Tile = 64;
+
+    constexpr ck_tile::index_t M_Warp = 4;
+    constexpr ck_tile::index_t N_Warp = 1;
+    constexpr ck_tile::index_t K_Warp = 1;
+
+    constexpr ck_tile::index_t M_Warp_Tile = 32;
+    constexpr ck_tile::index_t N_Warp_Tile = 32;
+    constexpr ck_tile::index_t K_Warp_Tile = 8;
+
+#elif(CK_TILE_PIPELINE_DEFAULT == CK_TILE_PIPELINE_COMPUTE)
+    // Compute friendly for Intrawave scheduler
+    constexpr ck_tile::index_t M_Tile = 256;
+    constexpr ck_tile::index_t N_Tile = 256;
     constexpr ck_tile::index_t K_Tile = 32;
 
     constexpr ck_tile::index_t M_Warp = 2;
@@ -28,12 +50,12 @@ float gemm_calc(const gemm_basic_args& args, const ck_tile::stream_config& s)
 
     constexpr ck_tile::index_t M_Warp_Tile = 32;
     constexpr ck_tile::index_t N_Warp_Tile = 32;
-    constexpr ck_tile::index_t K_Warp_Tile = 8;
+    constexpr ck_tile::index_t K_Warp_Tile = 16;
+#endif
 
-    // The kPadA, kPadB, kPadC & kBlockPerCu should also come from the Codegen part.
-    constexpr bool kPadM = true;
-    constexpr bool kPadN = true;
-    constexpr bool kPadK = true;
+    constexpr bool kPadM = false;
+    constexpr bool kPadN = false;
+    constexpr bool kPadK = false;
 
     constexpr int kBlockPerCu = 1;
 
@@ -49,8 +71,11 @@ float gemm_calc(const gemm_basic_args& args, const ck_tile::stream_config& s)
         ck_tile::Default2DEpilogueProblem<AccDataType, CDataType, kPadM, kPadN>>;
 
     using Traits = ck_tile::TileGemmTraits<kPadM, kPadN, kPadK, ALayout, BLayout, CLayout>;
-
+#if(CK_TILE_PIPELINE_DEFAULT == CK_TILE_PIPELINE_MEMORY)
     using BaseGemmPipeline = ck_tile::BaseGemmPipelineAgBgCrMem<
+#elif(CK_TILE_PIPELINE_DEFAULT == CK_TILE_PIPELINE_COMPUTE)
+    using BaseGemmPipeline                 = ck_tile::BaseGemmPipelineAgBgCrCompV3<
+#endif
         ck_tile::GemmPipelineProblem<ADataType, BDataType, AccDataType, GemmShape, Traits>>;
 
     const ck_tile::index_t num_loop    = TilePartitioner::GetLoopNum(args.K);
@@ -63,13 +88,21 @@ float gemm_calc(const gemm_basic_args& args, const ck_tile::stream_config& s)
         constexpr bool has_hot_loop_v = has_hot_loop_.value;
         constexpr auto tail_number_v  = tail_number_.value;
 
+#if(CK_TILE_PIPELINE_DEFAULT == CK_TILE_PIPELINE_MEMORY)
         using GemmPipeline = ck_tile::GemmPipelineAgBgCrMem<
+#elif(CK_TILE_PIPELINE_DEFAULT == CK_TILE_PIPELINE_COMPUTE)
+        using GemmPipeline = ck_tile::GemmPipelineAgBgCrCompV3<
+#endif
             ck_tile::UniversalGemmPipelineProblem<ADataType,
                                                   BDataType,
                                                   AccDataType,
                                                   GemmShape,
                                                   Traits,
+#if(CK_TILE_PIPELINE_DEFAULT == CK_TILE_PIPELINE_MEMORY)
+                                                  ck_tile::GemmPipelineScheduler::Interwave,
+#elif(CK_TILE_PIPELINE_DEFAULT == CK_TILE_PIPELINE_COMPUTE)
                                                   ck_tile::GemmPipelineScheduler::Intrawave,
+#endif
                                                   has_hot_loop_v,
                                                   tail_number_v>>;
         using Kernel = ck_tile::GemmKernel<TilePartitioner, GemmPipeline, GemmEpilogue>;
@@ -86,6 +119,11 @@ float gemm_calc(const gemm_basic_args& args, const ck_tile::stream_config& s)
         const dim3 grids      = Kernel::GridSize(args.M, args.N, args.kbatch);
         constexpr dim3 blocks = Kernel::BlockSize();
 
+        if(!Kernel::IsSupportedArgument(kargs))
+        {
+            throw std::runtime_error("Wrong! Arguments not supported! Skipping gemm!\n");
+        }
+
         if(s.log_level_ > 0)
         {
             std::cout << "Launching kernel with args:"
@@ -174,8 +212,8 @@ float gemm_calc(const gemm_basic_args& args, const ck_tile::stream_config& s)
         {
             std::ostringstream err;
             err << "When there's no hot loop, this tail number \"" << tail_num
-                << "\" is not supported! " << __FILE__ << ":" << __LINE__
-                << ", in function: " << __func__;
+                << "\" is not supported! PrefetchStages: " << BaseGemmPipeline::PrefetchStages
+                << "\n File: " << __FILE__ << ":" << __LINE__ << ", in function: " << __func__;
             throw std::runtime_error(err.str());
         }
     }

example/ck_tile/06_permute/alternative_impl/matrix_core_swizzle.cpp

@@ -40,7 +40,7 @@ float matrix_core_swizzle(matrix_core_swizzle_traits t,
             else if(t.permute.compare("0,1,3,4,2,5") == 0)
             {
                 constexpr matrix_core_permute_style pstyle =
-                    matrix_core_permute_style::permute_b_nr_kr_kw_nw_kv;
+                    matrix_core_permute_style::b_nr_kr_kw_nw_kv;
                 using Kernel =
                     matrix_core_swizzle_kernel<BLOCK_SIZE, NPerBlock, KPerBlock, pstyle, Inst>;
 
@@ -83,7 +83,7 @@ float matrix_core_swizzle(matrix_core_swizzle_traits t,
             else if(t.permute.compare("0,1,3,4,2,5") == 0)
             {
                 constexpr matrix_core_permute_style pstyle =
-                    matrix_core_permute_style::permute_b_nr_kr_kw_nw_kv;
+                    matrix_core_permute_style::b_nr_kr_kw_nw_kv;
                 using Kernel =
                     matrix_core_swizzle_kernel<BLOCK_SIZE, NPerBlock, KPerBlock, pstyle, Inst>;
 

example/ck_tile/06_permute/alternative_impl/matrix_core_swizzle_kernel.hpp

@@ -42,8 +42,8 @@ enum class matrix_core_permute_style
 {
     permute_b_n0_k0_n1_k1_n2_k2 = 0, // 0,1,4,2,5,3,6
     permute_b_n0_n1_k0_k1_n2_k2 = 1, // 0,1,2,4,5,3,6
-    permute_b_nr_kr_kw_nw_kv    = 2, // 0,1,3,4,2,5
-    permute_b_nr_kr_waveflatten = permute_b_nr_kr_kw_nw_kv,
+    b_nr_kr_kw_nw_kv            = 2, // 0,1,3,4,2,5
+    b_nr_kr_waveflatten         = b_nr_kr_kw_nw_kv,
 };
 
 // assume this is B matrix, originally we have batch*n*k
@@ -203,7 +203,7 @@ struct matrix_core_swizzle_kernel
             else
             {
                 // clang-format off
-                // permute_b_nr_kr_kw_nw_kv or permute_b_nr_kr_waveflatten
+                // b_nr_kr_kw_nw_kv or b_nr_kr_waveflatten
                 constexpr index_t Kv = Alignment;
                 constexpr index_t Nw = WarpGemm::WarpGemmAttribute::Impl::kAMLane;
                 constexpr index_t Kw = WarpGemm::WarpGemmAttribute::Impl::kABKLane;
@@ -332,7 +332,7 @@ struct matrix_core_swizzle_kernel
                         make_tuple(sequence<0>{}, sequence<1>{}));
                     return tmp_1;
 #else
-                    // permute_b_nr_kr_waveflatten = permute_b_nr_kr_kw_nw_kv,
+                    // b_nr_kr_waveflatten = b_nr_kr_kw_nw_kv,
                     constexpr index_t kv = Alignment;
                     constexpr index_t nw = WarpGemm::WarpGemmAttribute::Impl::kAMLane;
                     constexpr index_t kw = WarpGemm::WarpGemmAttribute::Impl::kABKLane;
@@ -376,13 +376,13 @@ struct matrix_core_swizzle_kernel
                 else
                 {
 #if MERGE_2D_013425
-                    // permute_b_nr_kr_waveflatten = permute_b_nr_kr_kw_nw_kv
+                    // b_nr_kr_waveflatten = b_nr_kr_kw_nw_kv
                     return make_tile_window(dst_view,
                                             make_tuple(number<NPerBlock>{}, number<KPerBlock>{}),
                                             {i_n * NPerBlock, i_k * KPerBlock},
                                             get_dst_dist());
 #else
-                    // permute_b_nr_kr_waveflatten = permute_b_nr_kr_kw_nw_kv
+                    // b_nr_kr_waveflatten = b_nr_kr_kw_nw_kv
                     constexpr index_t kv = Alignment;
                     constexpr index_t nw = WarpGemm::WarpGemmAttribute::Impl::kAMLane;
                     constexpr index_t kw = WarpGemm::WarpGemmAttribute::Impl::kABKLane;

example/ck_tile/06_permute/permute.cpp

@@ -264,7 +264,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
     {
         if(arg_parser.get_str("perm") == std::string("0,1,3,4,2,5"))
         {
-            // permute_b_nr_kr_kw_nw_kv = 2,   // 0,1,3,4,2,5
+            // b_nr_kr_kw_nw_kv = 2,   // 0,1,3,4,2,5
             matrix_core_swizzle_traits t;
             t.data_type = data_type;
             t.permute   = arg_parser.get_str("perm");

example/ck_tile/12_smoothquant/CMakeLists.txt

@@ -18,7 +18,7 @@ function (add_smoothquant_example TARGET_NAME MAIN_SRC)
     target_compile_options(${TARGET_NAME} PRIVATE ${COMPILE_OPTIONS})
 endfunction(add_smoothquant_example TARGET_NAME MAIN_SRC)
 
-file(GLOB INSTANCE_SRCS instances/*.cpp)
 
-add_smoothquant_example(tile_smoothquant smoothquant.cpp ${INSTANCE_SRCS})
 add_smoothquant_example(tile_example_smoothquant example_smoothquant.cpp)
+file(GLOB INSTANCE_SRCS instances/*.cpp)
+add_smoothquant_example(tile_smoothquant smoothquant.cpp ${INSTANCE_SRCS})

example/ck_tile/13_moe_sorting/moe_sorting_api.hpp

@@ -5,7 +5,7 @@
 #include <string>
 #include "ck_tile/core.hpp"
 #include "ck_tile/host.hpp"
-#include "ck_tile/ops/moe_sorting.hpp"
+#include "ck_tile/ops/fused_moe.hpp"
 
 struct moe_sorting_trait
 {

example/ck_tile/14_moe_smoothquant/CMakeLists.txt

+function (add_moe_smoothquant_example TARGET_NAME MAIN_SRC)
+    message("adding ${TARGET_NAME}")
+    # not using add_example_executable() to add target, since we don't want this to have
+    # to be included in "make all/install/check"
+    add_executable(${TARGET_NAME} EXCLUDE_FROM_ALL ${MAIN_SRC})
+    target_include_directories(${TARGET_NAME} PRIVATE ${CMAKE_CURRENT_LIST_DIR})
+
+    foreach(source IN LISTS ARGN)
+        list(APPEND INSTANCE_SRCS ${source})
+    endforeach()
+
+    target_sources(${TARGET_NAME} PRIVATE ${INSTANCE_SRCS})
+
+    set(COMPILE_OPTIONS)
+    # NOTE: we turn off undefined-func-template to let source compile without explicit declare function specializations
+    list(APPEND COMPILE_OPTIONS -Wno-undefined-func-template -Wno-float-equal)
+    # list(APPEND COMPILE_OPTIONS -v --save-temps -Wno-gnu-line-marker)
+
+    target_compile_options(${TARGET_NAME} PRIVATE ${COMPILE_OPTIONS})
+endfunction(add_moe_smoothquant_example TARGET_NAME MAIN_SRC)
+
+file(GLOB INSTANCE_SRCS instances/*.cpp)
+
+add_moe_smoothquant_example(tile_example_moe_smoothquant moe_smoothquant.cpp ${INSTANCE_SRCS})
+

example/ck_tile/14_moe_smoothquant/README.md

+# moe-smoothquant
+
+This folder contains example for moe-smoothquant using ck_tile tile-programming implementation.
+![](misc/moe-sm.png)
+
+Unlike standard smoothquant op, the input scale is from different expert `[expert, hidden]`, we need reuse the `topk-id` from previous `topk-softmax` and select the corresponding `expert` from current topk, and expand the output/per-token-scale by `topk`
+
+## build
+```
+# in the root of ck_tile
+mkdir build && cd build
+sh ../script/cmake-ck-dev.sh  ../ <arch>  # you can replace this <arch> to gfx90a, gfx942...
+make tile_example_moe_smoothquant -j
+```
+This will result in an executable `build/bin/tile_example_moe_smoothquant`

example/ck_tile/14_moe_smoothquant/instances/moe_smoothquant_bf16_n1024_instance.cpp

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

example/ck_tile/14_moe_smoothquant/instances/moe_smoothquant_bf16_n1536_instance.cpp

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

example/ck_tile/14_moe_smoothquant/instances/moe_smoothquant_bf16_n2048_instance.cpp

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

example/ck_tile/14_moe_smoothquant/instances/moe_smoothquant_bf16_n256_instance.cpp

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

example/ck_tile/14_moe_smoothquant/instances/moe_smoothquant_bf16_n3072_instance.cpp

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

example/ck_tile/14_moe_smoothquant/instances/moe_smoothquant_bf16_n4096_instance.cpp

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

example/ck_tile/14_moe_smoothquant/instances/moe_smoothquant_bf16_n4096_tp_instance.cpp

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

example/ck_tile/14_moe_smoothquant/instances/moe_smoothquant_bf16_n512_instance.cpp

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

example/ck_tile/14_moe_smoothquant/instances/moe_smoothquant_bf16_n64_n128_instance.cpp

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

example/ck_tile/14_moe_smoothquant/instances/moe_smoothquant_bf16_n768_instance.cpp

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

example/ck_tile/14_moe_smoothquant/instances/moe_smoothquant_fp16_n1024_instance.cpp

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

example/ck_tile/14_moe_smoothquant/instances/moe_smoothquant_fp16_n1536_instance.cpp

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