Merge branch 'develop' into gemm_bf16_sk_muozturk

example/ck_tile/13_moe_sorting/moe_sorting_api.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "moe_sorting_api.hpp"
+
+#define MOE_SORTING_DISPATCH(unroll_num_)                                                   \
+    constexpr ck_tile::index_t unroll_num = unroll_num_;                                    \
+    using ms_problem     = ck_tile::MoeSortingProblem<index_t, ms_weight_type, unroll_num>; \
+    using kernel         = ck_tile::MoeSortingKernel<ms_problem>;                           \
+    auto kargs           = kernel::MakeKargs(a);                                            \
+    const dim3 grids     = kernel::GridSize(a);                                             \
+    const dim3 blocks    = kernel::BlockSize(a);                                            \
+    const auto lds_bytes = kernel::GetSmemSize(a);                                          \
+    float ave_time       = ck_tile::launch_kernel(                                          \
+        s, ck_tile::make_kernel(kernel{}, grids, blocks, lds_bytes, kargs));          \
+    return ave_time;
+
+float moe_sorting(moe_sorting_trait t, moe_sorting_args a, ck_tile::stream_config s)
+{
+    if(t.weight_type == "fp32" && t.index_type == "int32")
+    {
+        if(a.num_experts > 127)
+        {
+            printf("lds size exceed, only support experts <127 \n");
+            return -1;
+        }
+        if(a.moe_buf_bytes % 16)
+        {
+            printf("buf set size %d unaligned, must be multiple of 16\n", a.moe_buf_bytes);
+            return -1;
+        }
+        using index_t              = ck_tile::index_t;
+        using ms_weight_type       = float;
+        index_t smem_io_unroll_num = ck_tile::integer_divide_ceil(a.tokens * a.topk, 64);
+        switch(smem_io_unroll_num)
+        {
+        case(1): {
+            MOE_SORTING_DISPATCH(1);
+        }
+        case(2): {
+            MOE_SORTING_DISPATCH(2);
+        }
+        case(3): {
+            MOE_SORTING_DISPATCH(3);
+        }
+        case(5): {
+            MOE_SORTING_DISPATCH(5);
+        }
+        case(6): {
+            MOE_SORTING_DISPATCH(6);
+        }
+        case(7): {
+            MOE_SORTING_DISPATCH(7);
+        }
+        case(8): {
+            MOE_SORTING_DISPATCH(8);
+        }
+        case(9): {
+            MOE_SORTING_DISPATCH(9);
+        }
+        case(10): {
+            MOE_SORTING_DISPATCH(10);
+        }
+        case(11): {
+            MOE_SORTING_DISPATCH(11);
+        }
+        default: {
+            MOE_SORTING_DISPATCH(4);
+        }
+        }
+    }
+    return -1;
+}

example/ck_tile/13_moe_sorting/moe_sorting_api.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+#include <string>
+#include "ck_tile/core.hpp"
+#include "ck_tile/host.hpp"
+#include "ck_tile/ops/fused_moe.hpp"
+
+struct moe_sorting_trait
+{
+    std::string index_type;
+    std::string weight_type; // currently always float
+};
+
+struct moe_sorting_args : public ck_tile::MoeSortingHostArgs
+{
+};
+
+float moe_sorting(moe_sorting_trait t, moe_sorting_args a, ck_tile::stream_config s);

example/ck_tile/13_moe_sorting/script/smoke_test.sh

+# #!/bin/sh
+
+EXE=./build/bin/tile_example_moe_sorting
+
+$EXE -t=80 -e=17 -moe_buf_size=16
+$EXE -t=111 -e=117 -moe_buf_size=4
+$EXE -t=1000 -e=55 -moe_buf_size=1024
+$EXE -t=99 -e=120  -moe_buf_size=10244
+$EXE -t=175 -e=64 -k=8
+$EXE -t=65 -e=8 -k=2
+$EXE -t=1 -e=25
+$EXE -t=31 -e=19 -k=15
+$EXE -t=81 -e=37 -k=7
+$EXE -t=23 -e=1 -k=1
+$EXE -t=127 -e=99 -k=19
+$EXE -t=71 -e=11 -k=11
+$EXE -t=1 -e=1 -k=1
+$EXE -t=99 -e=2 -k=1
+$EXE -t=333 -e=99 -k=13
\ No newline at end of file

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

example/ck_tile/14_moe_smoothquant/instances/moe_smoothquant_fp16_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::fp16_t, 1, 1, 1,  256, 8, true, false>>(const S&, A);
+template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 2, 1,  256, 4, true, false>>(const S&, A);
+template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 4, 1,  256, 2, true, false>>(const S&, A);
+template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 8, 1,  256, 1, true, false>>(const S&, A);
+
+// clang-format on

example/ck_tile/14_moe_smoothquant/instances/moe_smoothquant_fp16_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::fp16_t, 1,  1,  4, 64, 4, true, false>>(const S&, A);
+template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1,  2,  4, 64, 2, true, false>>(const S&, A);
+template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1,  4,  4, 64, 1, true, false>>(const S&, A);
+// clang-format on

example/ck_tile/14_moe_smoothquant/instances/moe_smoothquant_fp16_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::fp16_t,  1, 3, 1,  128, 8,true, false>>(const S&, A);
+template float moe_smoothquant_<trait_<ck_tile::fp16_t,  1, 3, 1,  256, 4,true, false>>(const S&, A);
+template float moe_smoothquant_<trait_<ck_tile::fp16_t,  1, 6, 1,  256, 2,true, false>>(const S&, A);
+template float moe_smoothquant_<trait_<ck_tile::fp16_t,  1, 3, 1, 1024, 1,true, false>>(const S&, A);
+
+// clang-format on