Merge branch 'develop' into codegen-enable-hiprtc

example/ck_tile/12_smoothquant/instances/smoothquant_fp16_n1024_instance.cpp

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

example/ck_tile/12_smoothquant/instances/smoothquant_fp16_n1536_instance.cpp

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

example/ck_tile/12_smoothquant/instances/smoothquant_fp16_n2048_instance.cpp

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

example/ck_tile/12_smoothquant/instances/smoothquant_fp16_n256_instance.cpp

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

example/ck_tile/12_smoothquant/instances/smoothquant_fp16_n3072_instance.cpp

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

example/ck_tile/12_smoothquant/instances/smoothquant_fp16_n4096_instance.cpp

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

example/ck_tile/12_smoothquant/instances/smoothquant_fp16_n4096_tp_instance.cpp

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

example/ck_tile/12_smoothquant/instances/smoothquant_fp16_n512_instance.cpp

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

example/ck_tile/12_smoothquant/instances/smoothquant_fp16_n64_n128_instance.cpp

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

example/ck_tile/12_smoothquant/instances/smoothquant_fp16_n768_instance.cpp

+
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "smoothquant_instance_common.hpp"
+
+// clang-format off
+//                                                  rm  rn  tm  tn  vn  pd       2p
+template float smoothquant_<trait_<ck_tile::fp16_t, 1,  3,  4,  64, 4,  true , false>>(const S&, A);
+template float smoothquant_<trait_<ck_tile::fp16_t, 1,  6,  4,  64, 2,  true , false>>(const S&, A);
+template float smoothquant_<trait_<ck_tile::fp16_t, 1, 12,  4,  64, 1,  true , false>>(const S&, A);
+// clang-format on

example/ck_tile/12_smoothquant/instances/smoothquant_fwd_api.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <ck_tile/core.hpp>
+#include "smoothquant.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 kTwoPass_>
+using trait_ = smoothquant_traits_<DataType_,
+                                   Repeat_M_,
+                                   Repeat_N_,
+                                   ThreadPerBlock_M_,
+                                   ThreadPerBlock_N_,
+                                   Vector_N_,
+                                   kPadN_,
+                                   kTwoPass_>;
+
+template <typename data_type>
+float smoothquant_dispatch(smoothquant_traits /*t*/,
+                           smoothquant_args a,
+                           const ck_tile::stream_config& s)
+{
+    float r = -1;
+    // clang-format off
+    //                                         rm  rn  tm  tn  vn   pd    2p
+    if(a.n <= 64) {
+            r = smoothquant_<trait_<data_type, 1,  1,  4,  64, 1,  true, false>>(s, a);
+    }
+    else if(a.n <= 128) {
+        if (a.n % 2 == 0)
+            r = smoothquant_<trait_<data_type, 1,  1,  4,  64, 2,  true, false>>(s, a);
+        else
+            r = smoothquant_<trait_<data_type, 1,  2,  4,  64, 1,  true, false>>(s, a);
+    }
+    else if(a.n <= 256) {
+        if (a.n % 4 == 0)
+            r = smoothquant_<trait_<data_type,  1, 1,  4,  64, 4,  true, false>>(s, a);
+        else if (a.n % 2 == 0)
+            r = smoothquant_<trait_<data_type,  1, 2,  4,  64, 2,  true, false>>(s, a);
+        else
+            r = smoothquant_<trait_<data_type,  1, 4,  4,  64, 1,  true, false>>(s, a);
+    }
+    else if(a.n <= 512) {
+        if (a.n % 8 == 0)
+            r = smoothquant_<trait_<data_type,  1, 1,  4,  64, 8,  true, false>>(s, a);
+        else if (a.n % 4 == 0)
+            r = smoothquant_<trait_<data_type,  1, 2,  4,  64, 4,  true, false>>(s, a);
+        else if (a.n % 2 == 0)
+            r = smoothquant_<trait_<data_type,  1, 4,  4,  64, 2,  true, false>>(s, a);
+        else
+            r = smoothquant_<trait_<data_type,  1, 8,  4,  64, 1,  true, false>>(s, a);
+    }
+    else if(a.n <= 768) {
+        if (a.n % 4 == 0)
+            r = smoothquant_<trait_<data_type,  1, 3,  4,  64, 4,  true, false>>(s, a);
+        else if (a.n % 2 == 0)
+            r = smoothquant_<trait_<data_type,  1, 6,  4,  64, 2,  true, false>>(s, a);
+        else
+            r = smoothquant_<trait_<data_type,  1,12,  4,  64, 1,  true, false>>(s, a);
+    }
+    else if(a.n <= 1024) {
+        if (a.n % 8 == 0)
+            r = smoothquant_<trait_<data_type,  1, 1, 2,  128, 8,  true, false>>(s, a);
+        else if (a.n % 4 == 0)
+            r = smoothquant_<trait_<data_type,  1, 2, 2,  128, 4,  true, false>>(s, a);
+        else if (a.n % 2 == 0)
+            r = smoothquant_<trait_<data_type,  1, 4, 2,  128, 2,  true, false>>(s, a);
+        else
+            r = smoothquant_<trait_<data_type,  1, 4, 1,  256, 1,  true, false>>(s, a);
+    }
+    else if(a.n <= 1536) {
+        if (a.n % 8 == 0)
+            r = smoothquant_<trait_<data_type,  1, 3, 4,   64, 8,  true, false>>(s, a);
+        else if (a.n % 4 == 0)
+            r = smoothquant_<trait_<data_type,  1, 3, 2,  128, 4,  true, false>>(s, a);
+        else if (a.n % 2 == 0)
+            r = smoothquant_<trait_<data_type,  1, 3, 1,  256, 2,  true, false>>(s, a);
+        else
+            r = smoothquant_<trait_<data_type,  1, 6, 1,  256, 1,  true, false>>(s, a);
+    }
+    else if(a.n <= 2048) {
+        if (a.n % 8 == 0)
+            r = smoothquant_<trait_<data_type,  1, 1, 1,  256, 8,  true, false>>(s, a);
+        else if (a.n % 4 == 0)
+            r = smoothquant_<trait_<data_type,  1, 2, 1,  256, 4,  true, false>>(s, a);
+        else if (a.n % 2 == 0)
+            r = smoothquant_<trait_<data_type,  1, 4, 1,  256, 2,  true, false>>(s, a);
+        else
+            r = smoothquant_<trait_<data_type,  1, 8, 1,  256, 1,  true, false>>(s, a);
+    }
+    else if(a.n <= 3072) {
+        if (a.n % 8 == 0)
+            r = smoothquant_<trait_<data_type,  1, 3, 1,  128, 8,  true, false>>(s, a);
+        else if (a.n % 4 == 0)
+            r = smoothquant_<trait_<data_type,  1, 3, 1,  256, 4,  true, false>>(s, a);
+        else if (a.n % 2 == 0)
+            r = smoothquant_<trait_<data_type,  1, 6, 1,  256, 2,  true, false>>(s, a);
+        else
+            r = smoothquant_<trait_<data_type,  1, 3, 1, 1024, 1,  true, false>>(s, a);
+    }
+    else if(a.n <= 4096) {
+        if (a.n % 8 == 0)
+            r = smoothquant_<trait_<data_type,  1, 2, 1,  256, 8,  true, false>>(s, a);
+        else if (a.n % 4 == 0)
+            r = smoothquant_<trait_<data_type,  1, 4, 1,  256, 4,  true, false>>(s, a);
+        else if (a.n % 2 == 0)
+            r = smoothquant_<trait_<data_type,  1, 2, 1, 1024, 2,  true, false>>(s, a);
+        else
+            r = smoothquant_<trait_<data_type,  1, 4, 1, 1024, 1,  true, false>>(s, a);
+    }
+    else if(a.n > 4096) {
+        if (a.n % 8 == 0)
+            r = smoothquant_<trait_<data_type,  1, 2, 1,  256, 8,  true, true>>(s, a);
+        else if (a.n % 4 == 0)
+            r = smoothquant_<trait_<data_type,  1, 4, 1,  256, 4,  true, true>>(s, a);
+        else if (a.n % 2 == 0)
+            r = smoothquant_<trait_<data_type,  1, 2, 1, 1024, 2,  true, true>>(s, a);
+        else
+            r = smoothquant_<trait_<data_type,  1, 4, 1, 1024, 1,  true, true>>(s, a);
+    }
+    return r;
+    // clang-format on
+}
+
+float smoothquant(smoothquant_traits t, smoothquant_args a, const ck_tile::stream_config& s)
+{
+    if(t.data_type.compare("fp16") == 0)
+    {
+        return smoothquant_dispatch<ck_tile::fp16_t>(t, a, s);
+    }
+    else if(t.data_type.compare("bf16") == 0)
+    {
+        return smoothquant_dispatch<ck_tile::bf16_t>(t, a, s);
+    }
+    else
+        throw std::runtime_error("Without supported instances!");
+}

example/ck_tile/12_smoothquant/instances/smoothquant_instance_common.hpp

+
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <ck_tile/core.hpp>
+#include "smoothquant.hpp"
+#include <iostream>
+
+#pragma once
+
+using S = ck_tile::stream_config;
+using A = smoothquant_args;
+
+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 kTwoPass_>
+using trait_ = smoothquant_traits_<DataType_,
+                                   Repeat_M_,
+                                   Repeat_N_,
+                                   ThreadPerBlock_M_,
+                                   ThreadPerBlock_N_,
+                                   Vector_N_,
+                                   kPadN_,
+                                   kTwoPass_>;
+
+template <typename Traits_>
+float smoothquant_(const S& s, A a)
+{
+    using DataType = typename Traits_::DataType;
+
+    using PipelineProblem = ck_tile::SmoothquantPipelineProblem<
+        typename SmoothquantTypeConfig<DataType>::XDataType,
+        typename SmoothquantTypeConfig<DataType>::XScaleDataType,
+        typename SmoothquantTypeConfig<DataType>::ComputeDataType,
+        typename SmoothquantTypeConfig<DataType>::YScaleDataType,
+        typename SmoothquantTypeConfig<DataType>::QYDataType,
+        typename Traits_::Shape,
+        Traits_::kPadN,
+        Traits_::kTwoPass>;
+
+    using OnePassPipeline = ck_tile::SmoothquantPipelineOnePass<PipelineProblem>;
+    using TwoPassPipeline = ck_tile::SmoothquantPipelineTwoPass<PipelineProblem>;
+    using Pipeline        = std::conditional_t<Traits_::kTwoPass, TwoPassPipeline, OnePassPipeline>;
+
+    using Kernel = ck_tile::Smoothquant<Pipeline>;
+
+    const dim3 grids                       = Kernel::GridSize(a);
+    constexpr dim3 blocks                  = Kernel::BlockSize();
+    constexpr ck_tile::index_t kBlockPerCu = 1;
+
+    auto kargs = Kernel::MakeKargs(a);
+    if(s.log_level_ > 0)
+        std::cout << ", " << Kernel::GetName() << std::flush;
+
+    return ck_tile::launch_kernel(
+        s, ck_tile::make_kernel<blocks.x, kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
+}

example/ck_tile/12_smoothquant/script/perf_test.sh

+
+EXE="$(find . -name tile_smoothquant -type f | head -n 1)"
+
+$EXE -m=1 -n=1 -e=1e-12 -v=1 -prec=bf16 -repeat=1000
+$EXE -m=700 -n=80 -e=1e-12 -v=1 -prec=bf16 -repeat=1000
+$EXE -m=700 -n=128 -e=1e-12 -v=1 -prec=bf16 -repeat=1000
+$EXE -m=700 -n=144 -e=1e-12 -v=1 -prec=bf16 -repeat=1000
+$EXE -m=700 -n=168 -e=1e-12 -v=1 -prec=bf16 -repeat=1000
+$EXE -m=700 -n=184 -e=1e-12 -v=1 -prec=bf16 -repeat=1000
+$EXE -m=700 -n=256 -e=1e-12 -v=1 -prec=bf16 -repeat=1000
+$EXE -m=700 -n=288 -e=1e-12 -v=1 -prec=bf16 -repeat=1000
+$EXE -m=700 -n=344 -e=1e-12 -v=1 -prec=bf16 -repeat=1000
+$EXE -m=700 -n=376 -e=1e-12 -v=1 -prec=bf16 -repeat=1000
+$EXE -m=700 -n=448 -e=1e-12 -v=1 -prec=bf16 -repeat=1000
+$EXE -m=700 -n=512 -e=1e-12 -v=1 -prec=bf16 -repeat=1000
+$EXE -m=700 -n=924 -e=1e-12 -v=1 -prec=bf16 -repeat=1000
+$EXE -m=700 -n=1024 -e=1e-12 -v=1 -prec=bf16 -repeat=1000
+$EXE -m=700 -n=1078 -e=1e-12 -v=1 -prec=bf16 -repeat=1000
+$EXE -m=700 -n=1996 -e=1e-12 -v=1 -prec=bf16 -repeat=1000
+$EXE -m=700 -n=4080 -e=1e-12 -v=1 -prec=bf16 -repeat=1000
+
+$EXE -m=700 -n=80 -e=1e-12 -v=1  -prec=fp16 -repeat=1000
+$EXE -m=700 -n=128 -e=1e-12 -v=1 -prec=fp16 -repeat=1000
+$EXE -m=700 -n=144 -e=1e-12 -v=1 -prec=fp16 -repeat=1000
+$EXE -m=700 -n=168 -e=1e-12 -v=1 -prec=fp16 -repeat=1000
+$EXE -m=700 -n=184 -e=1e-12 -v=1 -prec=fp16 -repeat=1000
+$EXE -m=700 -n=256 -e=1e-12 -v=1 -prec=fp16 -repeat=1000
+$EXE -m=700 -n=288 -e=1e-12 -v=1 -prec=fp16 -repeat=1000
+$EXE -m=700 -n=344 -e=1e-12 -v=1 -prec=fp16 -repeat=1000
+$EXE -m=700 -n=376 -e=1e-12 -v=1 -prec=fp16 -repeat=1000
+$EXE -m=700 -n=448 -e=1e-12 -v=1 -prec=fp16 -repeat=1000
+$EXE -m=700 -n=512 -e=1e-12 -v=1 -prec=fp16 -repeat=1000
+$EXE -m=700 -n=924 -e=1e-12 -v=1 -prec=fp16 -repeat=1000
+$EXE -m=700 -n=1024 -e=1e-12 -v=1 -prec=fp16 -repeat=1000
+$EXE -m=700 -n=1078 -e=1e-12 -v=1 -prec=fp16 -repeat=1000
+$EXE -m=700 -n=1996 -e=1e-12 -v=1 -prec=fp16 -repeat=1000
+$EXE -m=700 -n=4080 -e=1e-12 -v=1 -prec=fp16 -repeat=1000
\ No newline at end of file

example/ck_tile/12_smoothquant/script/smoke_test.sh

+#!/bin/sh
+EXE="$(find . -name tile_smoothquant -type f | head -n 1)"
+
+for pr_i in "fp16" "bf16" ; do
+$EXE -prec=$pr_i -m=99  -n=13
+$EXE -prec=$pr_i -m=17  -n=16
+$EXE -prec=$pr_i -m=1   -n=100
+$EXE -prec=$pr_i -m=4   -n=128
+$EXE -prec=$pr_i -m=80  -n=127
+$EXE -prec=$pr_i -m=22  -n=255 -stride=256
+$EXE -prec=$pr_i -m=7   -n=599
+$EXE -prec=$pr_i -m=19  -n=512
+$EXE -prec=$pr_i -m=33  -n=313 -stride=1000
+$EXE -prec=$pr_i -m=11  -n=510
+$EXE -prec=$pr_i -m=171 -n=676 -stride=818
+$EXE -prec=$pr_i -m=91  -n=636
+$EXE -prec=$pr_i -m=12  -n=768 -stride=800
+$EXE -prec=$pr_i -m=100 -n=766 -stride=812
+$EXE -prec=$pr_i -m=31  -n=1024
+$EXE -prec=$pr_i -m=64  -n=1000 -stride=1004
+$EXE -prec=$pr_i -m=8   -n=1501
+$EXE -prec=$pr_i -m=3   -n=1826
+$EXE -prec=$pr_i -m=5   -n=2040
+$EXE -prec=$pr_i -m=7   -n=2734
+$EXE -prec=$pr_i -m=1   -n=3182
+$EXE -prec=$pr_i -m=9   -n=4096
+$EXE -prec=$pr_i -m=3   -n=8192
+$EXE -prec=$pr_i -m=1   -n=10547
+$EXE -prec=$pr_i -m=3   -n=17134
+done

example/ck_tile/12_smoothquant/smoothquant.cpp

+#include "ck_tile/host.hpp"
+#include "smoothquant.hpp"
+#include <cstring>
+
+// different threshold for different dtype
+template <typename DataType>
+auto get_elimit()
+{
+    double rtol = 1e-5;
+    double atol = 1e-5;
+    return ck_tile::make_tuple(rtol, atol);
+}
+
+template <>
+auto get_elimit<ck_tile::bf16_t>()
+{
+    double rtol = 1e-5;
+    double atol = 1e-5;
+    return ck_tile::make_tuple(rtol, atol);
+}
+
+template <>
+auto get_elimit<ck_tile::int8_t>()
+{
+    // due to rounding, int8 quantization might have 1 abs error
+    double rtol = 1;
+    double atol = 1;
+    return ck_tile::make_tuple(rtol, atol);
+}
+
+auto create_args(int argc, char* argv[])
+{
+    ck_tile::ArgParser arg_parser;
+    arg_parser.insert("m", "3328", "m dimension")
+        .insert("n", "4096", "n dimension")
+        .insert("stride", "-1", "stride per row, if -1 then equal to n")
+        .insert("v", "1", "cpu validation or not")
+        .insert("kname", "1", "print kernel name or not")
+        .insert("prec", "fp16", "precision")
+        .insert("warmup", "5", "cold iter")
+        .insert("repeat", "20", "hot iter");
+
+    bool result = arg_parser.parse(argc, argv);
+    return std::make_tuple(result, arg_parser);
+}
+
+template <typename DataType>
+bool run(const ck_tile::ArgParser& arg_parser)
+{
+    ck_tile::index_t m      = arg_parser.get_int("m");
+    ck_tile::index_t n      = arg_parser.get_int("n");
+    ck_tile::index_t stride = arg_parser.get_int("stride");
+    if(stride < 0)
+        stride = n;
+    std::string data_type = arg_parser.get_str("prec");
+    int kname             = arg_parser.get_int("kname");
+    int do_validation     = arg_parser.get_int("v");
+    int warmup            = arg_parser.get_int("warmup");
+    int repeat            = arg_parser.get_int("repeat");
+
+    assert(stride >= n);
+
+    using TypeConfig = SmoothquantTypeConfig<DataType>;
+
+    using XDataType       = typename TypeConfig::XDataType;
+    using XScaleDataType  = typename TypeConfig::XScaleDataType;
+    using YScaleDataType  = typename TypeConfig::YScaleDataType;
+    using QYDataType      = typename TypeConfig::QYDataType;
+    using ComputeDataType = typename TypeConfig::ComputeDataType;
+
+    // host verify
+    ck_tile::HostTensor<XDataType> x_host({m, n}, {stride, 1});
+    ck_tile::HostTensor<XScaleDataType> xscale_host({n});
+
+    ck_tile::HostTensor<YScaleDataType> yscale_host_ref({m}, {1});
+    ck_tile::HostTensor<YScaleDataType> yscale_host_dev({m}, {1});
+
+    ck_tile::HostTensor<QYDataType> qy_host_ref({m, n}, {stride, 1});
+    ck_tile::HostTensor<QYDataType> qy_host_dev({m, n}, {stride, 1});
+
+    ck_tile::FillUniformDistribution<XDataType>{-.5f, .5f}(x_host);
+    ck_tile::FillUniformDistribution<XScaleDataType>{1e-3, .5f}(xscale_host);
+
+    ck_tile::DeviceMem x_buf(x_host.get_element_space_size_in_bytes());
+    ck_tile::DeviceMem xscale_buf(xscale_host.get_element_space_size_in_bytes());
+    ck_tile::DeviceMem yscale_buf(yscale_host_dev.get_element_space_size_in_bytes());
+    ck_tile::DeviceMem qy_buf(qy_host_dev.get_element_space_size_in_bytes());
+
+    x_buf.ToDevice(x_host.data());
+    xscale_buf.ToDevice(xscale_host.data());
+
+    std::cout << "[" << data_type << "]"
+              << " m:" << m << ", n:" << n << ", stride:" << stride << std::flush;
+
+    smoothquant_traits traits{data_type};
+
+    smoothquant_args args{x_buf.GetDeviceBuffer(),
+                          xscale_buf.GetDeviceBuffer(),
+                          yscale_buf.GetDeviceBuffer(),
+                          qy_buf.GetDeviceBuffer(),
+                          m,
+                          n,
+                          stride};
+
+    float ave_time = smoothquant(
+        traits, args, ck_tile::stream_config{nullptr, true, kname ? 1 : 0, warmup, repeat});
+
+    std::size_t num_byte = sizeof(XDataType) * m * n + sizeof(XScaleDataType) * n +
+                           sizeof(YScaleDataType) * m + sizeof(QYDataType) * m * n;
+
+    float gb_per_sec = num_byte / 1.E6 / ave_time;
+    std::cout << ", " << ave_time * 1.E3 << " us, " << gb_per_sec << " GB/s" << std::flush;
+
+    bool pass = true;
+
+    if(do_validation)
+    {
+        using YDataType = ComputeDataType;
+        ck_tile::HostTensor<ComputeDataType> y_host({m, n}, {stride, 1});
+        // smooth outlier
+        {
+            auto f = [&](auto n_) {
+                auto v_xscale = ck_tile::type_convert<ComputeDataType>(xscale_host(n_));
+
+                for(int m_ = 0; m_ < m; ++m_)
+                {
+                    auto v_x       = ck_tile::type_convert<ComputeDataType>(x_host(m_, n_));
+                    y_host(m_, n_) = v_x * v_xscale;
+                }
+            };
+
+            ck_tile::make_ParallelTensorFunctor(f, xscale_host.get_element_space_size())(
+                std::thread::hardware_concurrency());
+        }
+
+        // yscale
+        {
+            ck_tile::HostTensor<YDataType> y_rowwise_amax_host({m});
+
+            using ReduceAmax = ck_tile::ReduceOp::AbsMax;
+            ck_tile::reference_reduce<ComputeDataType, ComputeDataType, YDataType>(
+                y_host, y_rowwise_amax_host, ReduceAmax{});
+
+            auto op = [](const auto& v0) {
+                return v0 /
+                       ck_tile::type_convert<ComputeDataType>(ck_tile::numeric<QYDataType>::max());
+            };
+            ck_tile::reference_unary_elementwise<YDataType, YScaleDataType, ComputeDataType>(
+                y_rowwise_amax_host, yscale_host_ref, op);
+
+            yscale_buf.FromDevice(yscale_host_dev.mData.data());
+
+            auto [rtol, atol] = get_elimit<YScaleDataType>();
+            pass &= ck_tile::check_err(yscale_host_dev,
+                                       yscale_host_ref,
+                                       std::string("yscale Error: Incorrect results!"),
+                                       rtol,
+                                       atol);
+        }
+
+        // rowwise quantization
+        {
+            ck_tile::reference_rowwise_quantization2d<YDataType, YScaleDataType, QYDataType>(
+                y_host, yscale_host_ref, qy_host_ref);
+
+            qy_buf.FromDevice(qy_host_dev.data());
+            auto [rtol, atol] = get_elimit<QYDataType>();
+
+            if(stride == n)
+            {
+                pass = ck_tile::check_err(qy_host_dev,
+                                          qy_host_ref,
+                                          std::string("qy Error: Incorrect results!"),
+                                          rtol,
+                                          atol);
+            }
+            else
+            {
+                for(int i_r = 0; i_r < m; i_r++)
+                {
+                    std::vector<QYDataType> qy_host_dev_row(qy_host_dev.begin() + i_r * stride,
+                                                            qy_host_dev.begin() + i_r * stride + n);
+                    std::vector<QYDataType> qy_host_ref_row(qy_host_ref.begin() + i_r * stride,
+                                                            qy_host_ref.begin() + i_r * stride + n);
+                    pass &= ck_tile::check_err(qy_host_dev_row,
+                                               qy_host_ref_row,
+                                               std::string("qy[") + std::to_string(i_r) +
+                                                   std::string("] Error: Incorrect results!"),
+                                               rtol,
+                                               atol);
+                }
+            }
+        }
+
+        std::cout << ", valid:" << (pass ? "y" : "n") << std::flush << std::endl;
+    }
+
+    return pass;
+}
+
+int main(int argc, char* argv[])
+{
+    auto [result, arg_parser] = create_args(argc, argv);
+    if(!result)
+        return -1;
+
+    const std::string data_type = arg_parser.get_str("prec");
+    if(data_type == "fp16")
+    {
+        return run<ck_tile::half_t>(arg_parser) ? 0 : -2;
+    }
+    else if(data_type == "bf16")
+    {
+        return run<ck_tile::bf16_t>(arg_parser) ? 0 : -2;
+    }
+
+    return -3;
+}

example/ck_tile/12_smoothquant/smoothquant.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck_tile/core.hpp"
+#include "ck_tile/host/kernel_launch.hpp"
+#include "ck_tile/ops/smoothquant.hpp"
+#include <string>
+
+template <typename DataType>
+struct SmoothquantTypeConfig;
+
+template <>
+struct SmoothquantTypeConfig<ck_tile::half_t>
+{
+    using XDataType       = ck_tile::half_t;
+    using XScaleDataType  = float;
+    using YScaleDataType  = float;
+    using QYDataType      = ck_tile::int8_t;
+    using ComputeDataType = float;
+};
+
+template <>
+struct SmoothquantTypeConfig<ck_tile::bf16_t>
+{
+    using XDataType       = ck_tile::bf16_t;
+    using XScaleDataType  = float;
+    using YScaleDataType  = float;
+    using QYDataType      = ck_tile::int8_t;
+    using ComputeDataType = float;
+};
+
+// runtime args
+struct smoothquant_args : public ck_tile::SmoothquantHostArgs
+{
+};
+
+// this is used to pattern-match internl kernel implementation, not to instantiate kernel
+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 kTwoPass_>
+struct smoothquant_traits_
+{
+    using DataType = ck_tile::remove_cvref_t<DataType_>;
+
+    static constexpr bool is_warp_per_row = ThreadPerBlock_N_ <= warpSize;
+    static_assert((ThreadPerBlock_M_ * ThreadPerBlock_N_) % warpSize == 0);
+    static constexpr ck_tile::index_t total_warps =
+        (ThreadPerBlock_M_ * ThreadPerBlock_N_) / warpSize;
+
+    // num of warps along m
+    static constexpr ck_tile::index_t BlockWarps_M = []() {
+        if constexpr(is_warp_per_row)
+        {
+            static_assert(warpSize % ThreadPerBlock_N_ == 0);
+            return total_warps * (warpSize / ThreadPerBlock_N_);
+        }
+        else
+        {
+            // static_assert(warpSize % ThreadPerBlock_M_ == 0);
+            return total_warps / (ThreadPerBlock_N_ / warpSize);
+        }
+    }();
+
+    // num of warps along n
+    static constexpr ck_tile::index_t BlockWarps_N = []() {
+        if constexpr(is_warp_per_row)
+        {
+            static_assert(warpSize % ThreadPerBlock_N_ == 0);
+            return 1;
+        }
+        else
+        {
+            static_assert(ThreadPerBlock_N_ % warpSize == 0);
+            return ThreadPerBlock_N_ / warpSize;
+        }
+    }();
+
+    static constexpr ck_tile::index_t Repeat_M = Repeat_M_;
+    static constexpr ck_tile::index_t Repeat_N = Repeat_N_;
+
+    static constexpr ck_tile::index_t Block_M = Repeat_M_ * ThreadPerBlock_M_;
+    static constexpr ck_tile::index_t Block_N = Repeat_N_ * ThreadPerBlock_N_ * Vector_N_;
+
+    static constexpr ck_tile::index_t Warp_M = ThreadPerBlock_M_ / BlockWarps_M;
+    static constexpr ck_tile::index_t Warp_N = ThreadPerBlock_N_ / BlockWarps_N * Vector_N_;
+
+    using BlockTile  = ck_tile::sequence<Block_M, Block_N>;
+    using BlockWarps = ck_tile::sequence<BlockWarps_M, BlockWarps_N>;
+    using WarpTile   = ck_tile::sequence<Warp_M, Warp_N>;
+    using Vector     = ck_tile::sequence<1, Vector_N_>;
+
+    using Shape = ck_tile::Generic2dBlockShape<BlockTile, BlockWarps, WarpTile, Vector>;
+
+    static constexpr bool kPadN    = kPadN_;
+    static constexpr bool kTwoPass = kTwoPass_;
+};
+
+template <typename Traits_>
+float smoothquant_(const ck_tile::stream_config& s, smoothquant_args a);
+
+// This is the public API, will be generated by script
+struct smoothquant_traits
+{
+    std::string data_type;
+};
+
+float smoothquant(smoothquant_traits, smoothquant_args, const ck_tile::stream_config&);

example/ck_tile/13_moe_sorting/CMakeLists.txt

+add_executable(tile_example_moe_sorting EXCLUDE_FROM_ALL moe_sorting.cpp moe_sorting_api.cpp)
+target_include_directories(tile_example_moe_sorting PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/)
+
+set(EXAMPLE_MOE_SORTING_COMPILE_OPTIONS)
+# NOTE: we turn off undefined-func-template to let source compile without explicit declare function specializations
+list(APPEND EXAMPLE_MOE_SORTING_COMPILE_OPTIONS -Wno-undefined-func-template -Wno-float-equal)
+# list(APPEND EXAMPLE_MOE_SORTING_COMPILE_OPTIONS -v --save-temps -Wno-gnu-line-marker)
+target_compile_options(tile_example_moe_sorting PRIVATE ${EXAMPLE_MOE_SORTING_COMPILE_OPTIONS})

example/ck_tile/13_moe_sorting/README.md

+# moe-sorting
+
+This folder contains example for moe-sorting kernel using ck_tile tile-programming implementation. This kernel is often used in Moe model, before launching the fused-moe-gemm block. The input&weight is a `token*topk` 2d matrix. The op rearange the input weight ids into different experts and feed into fuse moe gemm kernel.
+
+## 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_sorting -j
+```
+This will result in an executable `build/bin/tile_example_moe_sorting`
+
+## example
+```
+args:
+          -v    weather do CPU validation or not (default:1)
+       -pr_i    index data type. (currently only fp32 supported now) (default:int32)
+       -pr_w    output weight data type(currently only fp32 supported now) (default:fp32)
+          -t    number of input tokens (default:32)
+          -e    number of experts (default:8)
+          -k    topk (default:2)
+       -st_i    row stride of input, -1 means same as experts (default:-1)
+       -seed    seed to be used, -1 means random every time (default:-1)
+      -kname    when set to 1 it will print kernel name (default:0)
+
+```

example/ck_tile/13_moe_sorting/moe_sorting.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <set>
+#include <vector>
+#include <iostream>
+#include <numeric>
+#include <cassert>
+#include <cstdlib>
+#include <iostream>
+#include <time.h>
+#include <unordered_set>
+
+#include "ck_tile/core.hpp"
+#include "ck_tile/ops/reduce.hpp"
+#include "moe_sorting_api.hpp"
+
+auto create_args(int argc, char* argv[])
+{
+    ck_tile::ArgParser arg_parser;
+    arg_parser.insert("v", "1", "weather do CPU validation or not")
+        .insert("pr_i", "int32", "index data type. (currently only int32 supported now)")
+        .insert("pr_w", "fp32", "output weight data type(currently only fp32 supported now)")
+        .insert("t", "128", "number of input tokens")
+        .insert("e", "8", "number of num_experts")
+        .insert("k", "4", "topk")
+        .insert("unit", "32", "unit_size")
+        .insert("moe_buf_size", "0", "moe_buf_size")
+        .insert("seed", "-1", "seed to be used, -1 means random every time")
+        .insert("kname", "0", "when set to 1 it will print kernel name")
+        .insert("warmup", "5", "number of iterations before benchmark the kernel")
+        .insert("repeat", "20", "number of iterations to benchmark the kernel");
+
+    bool result = arg_parser.parse(argc, argv);
+    return std::make_tuple(result, arg_parser);
+}
+
+template <typename IndexType>
+void topid_unique_gen(
+    std::vector<IndexType>& host_tensor, int tokens, int topk, int num_expert, int seed)
+{
+    size_t total_size = topk * tokens;
+    std::srand(seed);
+    std::set<IndexType> unique_set;
+    IndexType current_v;
+    for(size_t i = 0; i < total_size; i++)
+    {
+        if(i % topk == 0)
+        {
+            unique_set.clear();
+        }
+        current_v = std::rand() % num_expert;
+        while(unique_set.find(current_v) != unique_set.end())
+        {
+            current_v = std::rand() % num_expert;
+        }
+        unique_set.insert(current_v);
+        host_tensor[i] = current_v;
+    }
+}
+
+template <typename WeightType, typename IndexType = ck_tile::index_t>
+bool test_moe_sorting(ck_tile::ArgParser args)
+{
+    int validate            = args.get_int("v");
+    std::string index_prec  = args.get_str("pr_i");
+    std::string weight_prec = args.get_str("pr_w");
+    int tokens              = args.get_int("t");
+    int num_experts         = args.get_int("e");
+    int topk                = args.get_int("k");
+    int seed                = args.get_int("seed");
+    int unit_size           = args.get_int("unit");
+    int moe_buf_size        = args.get_int("moe_buf_size");
+    int kname               = args.get_int("kname");
+    int warmup              = args.get_int("warmup");
+    int repeat              = args.get_int("repeat");
+    int max_output_ids =
+        ck_tile::integer_least_multiple(topk * tokens + num_experts * unit_size - topk, unit_size);
+
+    if(seed < 0)
+    {
+        seed = std::time(nullptr);
+    }
+
+    if(topk > num_experts)
+    {
+        printf("topk:%d value should be smaller than, or equal to number of num_experts:%d\n",
+               topk,
+               num_experts);
+        return false;
+    }
+
+    // tokens already considered batch size
+    ck_tile::HostTensor<IndexType> topk_ids_host({tokens, topk}, {topk, 1});
+    ck_tile::HostTensor<WeightType> weights_host({tokens, topk}, {topk, 1});
+    ck_tile::HostTensor<IndexType> sorted_ids_host({max_output_ids}, {1});
+    ck_tile::HostTensor<WeightType> sorted_weights_host({max_output_ids}, {1});
+    ck_tile::HostTensor<IndexType> sorted_expert_ids_host({max_output_ids / unit_size}, {1});
+    ck_tile::HostTensor<IndexType> sorted_id_cnt_host({1}, {1});
+    ck_tile::HostTensor<float> moe_buf_host({moe_buf_size});
+
+    ck_tile::FillUniformDistribution<WeightType>{-.5f, .5f}(weights_host);
+    ck_tile::FillUniformDistribution<WeightType>{-.5f, .5f}(moe_buf_host);
+    topid_unique_gen<IndexType>(topk_ids_host.mData, tokens, topk, num_experts, seed);
+
+    ck_tile::DeviceMem topk_ids_dev(topk_ids_host.get_element_space_size_in_bytes());
+    ck_tile::DeviceMem weights_dev(weights_host.get_element_space_size_in_bytes());
+    ck_tile::DeviceMem sorted_ids_dev(sorted_ids_host.get_element_space_size_in_bytes());
+    ck_tile::DeviceMem sorted_weights_dev(sorted_weights_host.get_element_space_size_in_bytes());
+    ck_tile::DeviceMem sorted_expert_ids_dev(
+        sorted_expert_ids_host.get_element_space_size_in_bytes());
+    ck_tile::DeviceMem sorted_id_cnt_dev(sorted_id_cnt_host.get_element_space_size_in_bytes());
+    ck_tile::DeviceMem moe_buf_dev(moe_buf_host.get_element_space_size_in_bytes());
+
+    topk_ids_dev.ToDevice(topk_ids_host.data());
+    weights_dev.ToDevice(weights_host.data());
+    if(moe_buf_size > 0)
+    {
+        moe_buf_dev.ToDevice(moe_buf_host.data());
+    }
+
+    moe_sorting_trait trait{index_prec, weight_prec};
+
+    moe_sorting_args karg{topk_ids_dev.GetDeviceBuffer(),
+                          weights_dev.GetDeviceBuffer(),
+                          sorted_ids_dev.GetDeviceBuffer(),
+                          sorted_weights_dev.GetDeviceBuffer(),
+                          sorted_expert_ids_dev.GetDeviceBuffer(),
+                          sorted_id_cnt_dev.GetDeviceBuffer(),
+                          moe_buf_size > 0 ? moe_buf_dev.GetDeviceBuffer() : nullptr,
+                          tokens,
+                          unit_size,
+                          num_experts,
+                          topk,
+                          static_cast<ck_tile::index_t>(moe_buf_size * sizeof(float))};
+
+    ck_tile::stream_config sc{nullptr,
+                              true,
+                              /* log_level = */ (kname ? 1 : 0),
+                              warmup,
+                              repeat};
+    auto ms = moe_sorting(trait, karg, sc);
+    printf("[%s|%s]tokens:%d, num_experts:%d, topk:%d,  ms:%f , ",
+           index_prec.c_str(),
+           weight_prec.c_str(),
+           tokens,
+           num_experts,
+           topk,
+           ms);
+    if(ms < 0)
+        printf("not supported\n");
+    fflush(stdout);
+    if(ms < 0)
+    {
+        return false;
+    }
+
+    sorted_ids_dev.FromDevice(sorted_ids_host.data());
+    sorted_weights_dev.FromDevice(sorted_weights_host.data());
+    sorted_expert_ids_dev.FromDevice(sorted_expert_ids_host.data());
+    sorted_id_cnt_dev.FromDevice(sorted_id_cnt_host.data());
+    if(moe_buf_size > 0)
+    {
+        moe_buf_dev.FromDevice(moe_buf_host.data());
+    }
+
+    bool rtn = true;
+    if(validate)
+    {
+        ck_tile::HostTensor<IndexType> sorted_ids_ref({max_output_ids}, {1});
+        ck_tile::HostTensor<WeightType> sorted_weights_ref({max_output_ids}, {1});
+        ck_tile::HostTensor<IndexType> sorted_expert_ids_ref({max_output_ids / unit_size}, {1});
+
+        int32_t ref_total_tokens_post_pad = 0;
+        ck_tile::reference_moe_sorting<WeightType, IndexType>(topk_ids_host,
+                                                              weights_host,
+                                                              sorted_ids_ref,
+                                                              sorted_weights_ref,
+                                                              sorted_expert_ids_ref,
+                                                              ref_total_tokens_post_pad,
+                                                              num_experts,
+                                                              unit_size);
+        rtn &= ck_tile::check_err(
+            sorted_ids_host, sorted_ids_ref, std::string("OUT Error: Incorrect ids!"), 1e-6, 1e-6);
+        rtn &= ck_tile::check_err(sorted_weights_host,
+                                  sorted_weights_ref,
+                                  std::string("OUT Error: Incorrect w!"),
+                                  1e-6,
+                                  1e-6);
+        rtn &= ck_tile::check_err(sorted_expert_ids_host,
+                                  sorted_expert_ids_ref,
+                                  std::string("OUT Error: Incorrect eid!"),
+                                  1e-6,
+                                  1e-6);
+        if(moe_buf_size)
+        {
+            ck_tile::HostTensor<WeightType> moe_buf_ref({moe_buf_size});
+            rtn &= ck_tile::check_err(
+                moe_buf_host, moe_buf_ref, std::string("OUT Error: Incorrect zero buf!"), 0, 0);
+        }
+        rtn &= ref_total_tokens_post_pad == sorted_id_cnt_host.mData[0];
+    }
+
+    printf("valid:%s\n", rtn ? "y" : "n");
+    fflush(stdout);
+    return rtn;
+}
+
+int main(int argc, char** argv)
+{
+    auto [result, args] = create_args(argc, argv);
+    if(!result)
+        return -1;
+    std::string index_prec  = args.get_str("pr_i");
+    std::string weight_prec = args.get_str("pr_w");
+
+    bool r = true;
+    if(weight_prec.compare("fp32") == 0 && index_prec.compare("int32") == 0)
+    {
+        r &= test_moe_sorting<float, ck_tile::index_t>(args);
+    }
+    return r ? 0 : -1;
+}

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;
+}