Merge branch 'develop' into ck_tile/fa_bwd_v3

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

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

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

example/ck_tile/14_moe_smoothquant/instances/moe_smoothquant_fp16_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::fp16_t, 1,  1,  4, 64, 8,  true , false>>(const S&, A);
+template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1,  2,  4, 64, 4,  true , false>>(const S&, A);
+template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1,  4,  4, 64, 2,  true , false>>(const S&, A);
+template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1,  8,  4, 64, 1,  true , false>>(const S&, A);
+// clang-format on

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

example/ck_tile/14_moe_smoothquant/instances/moe_smoothquant_fp16_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::fp16_t, 1,  3,  4,  64, 4,  true , false>>(const S&, A);
+template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1,  6,  4,  64, 2,  true , false>>(const S&, A);
+template float moe_smoothquant_<trait_<ck_tile::fp16_t, 1, 12,  4,  64, 1,  true , false>>(const S&, A);
+// clang-format on

example/ck_tile/14_moe_smoothquant/instances/moe_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 "moe_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_ = moe_smoothquant_traits_<DataType_,
+                                       Repeat_M_,
+                                       Repeat_N_,
+                                       ThreadPerBlock_M_,
+                                       ThreadPerBlock_N_,
+                                       Vector_N_,
+                                       kPadN_,
+                                       kTwoPass_>;
+
+template <typename data_type>
+float moe_smoothquant_dispatch(moe_smoothquant_traits /*t*/,
+                               moe_smoothquant_args a,
+                               const ck_tile::stream_config& s)
+{
+    float r = -1;
+    // clang-format off
+    //                                         rm  rn  tm  tn  vn   pd    2p
+    if(a.hidden_size <= 64) {
+            r = moe_smoothquant_<trait_<data_type, 1,  1,  4,  64, 1,  true, false>>(s, a);
+    }
+    else if(a.hidden_size <= 128) {
+        if (a.hidden_size % 2 == 0)
+            r = moe_smoothquant_<trait_<data_type, 1,  1,  4,  64, 2,  true, false>>(s, a);
+        else
+            r = moe_smoothquant_<trait_<data_type, 1,  2,  4,  64, 1,  true, false>>(s, a);
+    }
+    else if(a.hidden_size <= 256) {
+        if (a.hidden_size % 4 == 0)
+            r = moe_smoothquant_<trait_<data_type,  1, 1,  4,  64, 4,  true, false>>(s, a);
+        else if (a.hidden_size % 2 == 0)
+            r = moe_smoothquant_<trait_<data_type,  1, 2,  4,  64, 2,  true, false>>(s, a);
+        else
+            r = moe_smoothquant_<trait_<data_type,  1, 4,  4,  64, 1,  true, false>>(s, a);
+    }
+    else if(a.hidden_size <= 512) {
+        if (a.hidden_size % 8 == 0)
+            r = moe_smoothquant_<trait_<data_type,  1, 1,  4,  64, 8,  true, false>>(s, a);
+        else if (a.hidden_size % 4 == 0)
+            r = moe_smoothquant_<trait_<data_type,  1, 2,  4,  64, 4,  true, false>>(s, a);
+        else if (a.hidden_size % 2 == 0)
+            r = moe_smoothquant_<trait_<data_type,  1, 4,  4,  64, 2,  true, false>>(s, a);
+        else
+            r = moe_smoothquant_<trait_<data_type,  1, 8,  4,  64, 1,  true, false>>(s, a);
+    }
+    else if(a.hidden_size <= 768) {
+        if (a.hidden_size % 4 == 0)
+            r = moe_smoothquant_<trait_<data_type,  1, 3,  4,  64, 4,  true, false>>(s, a);
+        else if (a.hidden_size % 2 == 0)
+            r = moe_smoothquant_<trait_<data_type,  1, 6,  4,  64, 2,  true, false>>(s, a);
+        else
+            r = moe_smoothquant_<trait_<data_type,  1,12,  4,  64, 1,  true, false>>(s, a);
+    }
+    else if(a.hidden_size <= 1024) {
+        if (a.hidden_size % 8 == 0)
+            r = moe_smoothquant_<trait_<data_type,  1, 1, 2,  128, 8,  true, false>>(s, a);
+        else if (a.hidden_size % 4 == 0)
+            r = moe_smoothquant_<trait_<data_type,  1, 2, 2,  128, 4,  true, false>>(s, a);
+        else if (a.hidden_size % 2 == 0)
+            r = moe_smoothquant_<trait_<data_type,  1, 4, 2,  128, 2,  true, false>>(s, a);
+        else
+            r = moe_smoothquant_<trait_<data_type,  1, 4, 1,  256, 1,  true, false>>(s, a);
+    }
+    else if(a.hidden_size <= 1536) {
+        if (a.hidden_size % 8 == 0)
+            r = moe_smoothquant_<trait_<data_type,  1, 3, 4,   64, 8,  true, false>>(s, a);
+        else if (a.hidden_size % 4 == 0)
+            r = moe_smoothquant_<trait_<data_type,  1, 3, 2,  128, 4,  true, false>>(s, a);
+        else if (a.hidden_size % 2 == 0)
+            r = moe_smoothquant_<trait_<data_type,  1, 3, 1,  256, 2,  true, false>>(s, a);
+        else
+            r = moe_smoothquant_<trait_<data_type,  1, 6, 1,  256, 1,  true, false>>(s, a);
+    }
+    else if(a.hidden_size <= 2048) {
+        if (a.hidden_size % 8 == 0)
+            r = moe_smoothquant_<trait_<data_type,  1, 1, 1,  256, 8,  true, false>>(s, a);
+        else if (a.hidden_size % 4 == 0)
+            r = moe_smoothquant_<trait_<data_type,  1, 2, 1,  256, 4,  true, false>>(s, a);
+        else if (a.hidden_size % 2 == 0)
+            r = moe_smoothquant_<trait_<data_type,  1, 4, 1,  256, 2,  true, false>>(s, a);
+        else
+            r = moe_smoothquant_<trait_<data_type,  1, 8, 1,  256, 1,  true, false>>(s, a);
+    }
+    else if(a.hidden_size <= 3072) {
+        if (a.hidden_size % 8 == 0)
+            r = moe_smoothquant_<trait_<data_type,  1, 3, 1,  128, 8,  true, false>>(s, a);
+        else if (a.hidden_size % 4 == 0)
+            r = moe_smoothquant_<trait_<data_type,  1, 3, 1,  256, 4,  true, false>>(s, a);
+        else if (a.hidden_size % 2 == 0)
+            r = moe_smoothquant_<trait_<data_type,  1, 6, 1,  256, 2,  true, false>>(s, a);
+        else
+            r = moe_smoothquant_<trait_<data_type,  1, 3, 1, 1024, 1,  true, false>>(s, a);
+    }
+    else if(a.hidden_size <= 4096) {
+        if (a.hidden_size % 8 == 0)
+            r = moe_smoothquant_<trait_<data_type,  1, 2, 1,  256, 8,  true, false>>(s, a);
+        else if (a.hidden_size % 4 == 0)
+            r = moe_smoothquant_<trait_<data_type,  1, 4, 1,  256, 4,  true, false>>(s, a);
+        else if (a.hidden_size % 2 == 0)
+            r = moe_smoothquant_<trait_<data_type,  1, 2, 1, 1024, 2,  true, false>>(s, a);
+        else
+            r = moe_smoothquant_<trait_<data_type,  1, 4, 1, 1024, 1,  true, false>>(s, a);
+    }
+    else if(a.hidden_size > 4096) {
+        if (a.hidden_size % 8 == 0)
+            r = moe_smoothquant_<trait_<data_type,  1, 2, 1,  256, 8,  true, true>>(s, a);
+        else if (a.hidden_size % 4 == 0)
+            r = moe_smoothquant_<trait_<data_type,  1, 4, 1,  256, 4,  true, true>>(s, a);
+        else if (a.hidden_size % 2 == 0)
+            r = moe_smoothquant_<trait_<data_type,  1, 2, 1, 1024, 2,  true, true>>(s, a);
+        else
+            r = moe_smoothquant_<trait_<data_type,  1, 4, 1, 1024, 1,  true, true>>(s, a);
+    }
+    return r;
+    // clang-format on
+}
+
+float moe_smoothquant(moe_smoothquant_traits t,
+                      moe_smoothquant_args a,
+                      const ck_tile::stream_config& s)
+{
+    if(t.data_type.compare("fp16") == 0)
+    {
+        return moe_smoothquant_dispatch<ck_tile::fp16_t>(t, a, s);
+    }
+    else if(t.data_type.compare("bf16") == 0)
+    {
+        return moe_smoothquant_dispatch<ck_tile::bf16_t>(t, a, s);
+    }
+    else
+        throw std::runtime_error("Without supported instances!");
+}

example/ck_tile/14_moe_smoothquant/instances/moe_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 "moe_smoothquant.hpp"
+#include <iostream>
+
+#pragma once
+
+using S = ck_tile::stream_config;
+using A = moe_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_ = moe_smoothquant_traits_<DataType_,
+                                       Repeat_M_,
+                                       Repeat_N_,
+                                       ThreadPerBlock_M_,
+                                       ThreadPerBlock_N_,
+                                       Vector_N_,
+                                       kPadN_,
+                                       kTwoPass_>;
+
+template <typename Traits_>
+float moe_smoothquant_(const S& s, A a)
+{
+    using DataType = typename Traits_::DataType;
+
+    using PipelineProblem = ck_tile::SmoothquantPipelineProblem<
+        typename MoeSmoothquantTypeConfig<DataType>::XDataType,
+        typename MoeSmoothquantTypeConfig<DataType>::XScaleDataType,
+        typename MoeSmoothquantTypeConfig<DataType>::ComputeDataType,
+        typename MoeSmoothquantTypeConfig<DataType>::YScaleDataType,
+        typename MoeSmoothquantTypeConfig<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::MoeSmoothquant<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/14_moe_smoothquant/moe_smoothquant.cpp

+#include "ck_tile/host.hpp"
+#include "moe_smoothquant.hpp"
+#include <cstring>
+#include <set>
+
+// 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);
+}
+
+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;
+    }
+}
+
+auto create_args(int argc, char* argv[])
+{
+    ck_tile::ArgParser arg_parser;
+    arg_parser.insert("t", "3328", "tokens dimension")
+        .insert("h", "4096", "hidden_size dimension")
+        .insert("e", "32", "experts")
+        .insert("k", "5", "topk")
+        .insert("stride", "-1", "stride per row, if -1 then equal to hidden_size")
+        .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 tokens      = arg_parser.get_int("t");
+    ck_tile::index_t hidden_size = arg_parser.get_int("h");
+    ck_tile::index_t stride      = arg_parser.get_int("stride");
+    if(stride < 0)
+        stride = hidden_size;
+    ck_tile::index_t experts = arg_parser.get_int("e");
+    ck_tile::index_t topk    = arg_parser.get_int("k");
+    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 >= hidden_size);
+
+    using TypeConfig = MoeSmoothquantTypeConfig<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({tokens, hidden_size}, {stride, 1});
+    ck_tile::HostTensor<XScaleDataType> xscale_host({experts * hidden_size});
+    ck_tile::HostTensor<ck_tile::index_t> topk_ids_host({tokens, topk});
+
+    ck_tile::HostTensor<YScaleDataType> yscale_host_ref({topk * tokens}, {1});
+    ck_tile::HostTensor<YScaleDataType> yscale_host_dev({topk * tokens}, {1});
+
+    ck_tile::HostTensor<QYDataType> qy_host_ref({topk * tokens, hidden_size}, {stride, 1});
+    ck_tile::HostTensor<QYDataType> qy_host_dev({topk * tokens, hidden_size}, {stride, 1});
+
+    topid_unique_gen<ck_tile::index_t>(topk_ids_host.mData, tokens, topk, experts, 11937);
+    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 topk_ids_buf(topk_ids_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());
+    topk_ids_buf.ToDevice(topk_ids_host.data());
+
+    std::cout << "[" << data_type << "]"
+              << " tokens:" << tokens << ", hidden_size:" << hidden_size << ", stride:" << stride
+              << ", experts:" << experts << ", topk:" << topk << std::flush;
+
+    moe_smoothquant_traits traits{data_type};
+
+    moe_smoothquant_args args{x_buf.GetDeviceBuffer(),
+                              xscale_buf.GetDeviceBuffer(),
+                              topk_ids_buf.GetDeviceBuffer(),
+                              yscale_buf.GetDeviceBuffer(),
+                              qy_buf.GetDeviceBuffer(),
+                              tokens,
+                              hidden_size,
+                              experts,
+                              topk,
+                              stride,
+                              stride};
+
+    float ave_time = moe_smoothquant(
+        traits, args, ck_tile::stream_config{nullptr, true, kname ? 1 : 0, warmup, repeat});
+
+    std::size_t num_byte =
+        sizeof(XDataType) * tokens * hidden_size + sizeof(XScaleDataType) * topk * hidden_size +
+        sizeof(YScaleDataType) * topk * tokens + sizeof(QYDataType) * topk * tokens * hidden_size;
+
+    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({topk * tokens, hidden_size}, {stride, 1});
+        // smooth outlier
+        {
+            auto f = [&](auto i_token) {
+                for(int i_topk = 0; i_topk < topk; i_topk++)
+                {
+                    auto i_expert = topk_ids_host(i_token, i_topk);
+
+                    for(int i_h = 0; i_h < hidden_size; ++i_h)
+                    {
+                        auto v_xscale = ck_tile::type_convert<ComputeDataType>(
+                            xscale_host(i_expert * hidden_size + i_h));
+                        auto v_x = ck_tile::type_convert<ComputeDataType>(x_host(i_token, i_h));
+                        // y_host(i_token * topk + i_topk, i_h) = v_x * v_xscale;
+                        y_host(i_topk * tokens + i_token, i_h) = v_x * v_xscale;
+                    }
+                }
+            };
+
+            ck_tile::make_ParallelTensorFunctor(f, tokens)(std::thread::hardware_concurrency());
+        }
+
+        // yscale
+        {
+            ck_tile::HostTensor<YDataType> y_rowwise_amax_host({topk * tokens});
+
+            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 == hidden_size)
+            {
+                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 < topk * tokens; i_r++)
+                {
+                    std::vector<QYDataType> qy_host_dev_row(qy_host_dev.begin() + i_r * stride,
+                                                            qy_host_dev.begin() + i_r * stride +
+                                                                hidden_size);
+                    std::vector<QYDataType> qy_host_ref_row(qy_host_ref.begin() + i_r * stride,
+                                                            qy_host_ref.begin() + i_r * stride +
+                                                                hidden_size);
+                    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/14_moe_smoothquant/moe_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 MoeSmoothquantTypeConfig;
+
+template <>
+struct MoeSmoothquantTypeConfig<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 MoeSmoothquantTypeConfig<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 moe_smoothquant_args : public ck_tile::MoeSmoothquantHostArgs
+{
+};
+
+// 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 moe_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 moe_smoothquant_(const ck_tile::stream_config& s, moe_smoothquant_args a);
+
+// This is the public API, will be generated by script
+struct moe_smoothquant_traits
+{
+    std::string data_type;
+};
+
+float moe_smoothquant(moe_smoothquant_traits, moe_smoothquant_args, const ck_tile::stream_config&);

example/ck_tile/14_moe_smoothquant/script/perf_test.sh

+
+EXE=build/bin/tile_example_moe_smoothquant
+
+$EXE -t=1 -h=1  -v=1 -prec=bf16 -repeat=1000
+$EXE -t=700 -h=80  -v=1 -prec=bf16 -repeat=1000
+$EXE -t=700 -h=128  -v=1 -prec=bf16 -repeat=1000
+$EXE -t=700 -h=144  -v=1 -prec=bf16 -repeat=1000
+$EXE -t=700 -h=168  -v=1 -prec=bf16 -repeat=1000
+$EXE -t=700 -h=184  -v=1 -prec=bf16 -repeat=1000
+$EXE -t=700 -h=256  -v=1 -prec=bf16 -repeat=1000
+$EXE -t=700 -h=288  -v=1 -prec=bf16 -repeat=1000
+$EXE -t=700 -h=344  -v=1 -prec=bf16 -repeat=1000
+$EXE -t=700 -h=376  -v=1 -prec=bf16 -repeat=1000
+$EXE -t=700 -h=448  -v=1 -prec=bf16 -repeat=1000
+$EXE -t=700 -h=512  -v=1 -prec=bf16 -repeat=1000
+$EXE -t=700 -h=924  -v=1 -prec=bf16 -repeat=1000
+$EXE -t=700 -h=1024  -v=1 -prec=bf16 -repeat=1000
+$EXE -t=700 -h=1078  -v=1 -prec=bf16 -repeat=1000
+$EXE -t=700 -h=1996  -v=1 -prec=bf16 -repeat=1000
+$EXE -t=700 -h=4080  -v=1 -prec=bf16 -repeat=1000
+
+$EXE -t=700 -h=80  -v=1  -prec=fp16 -repeat=1000
+$EXE -t=700 -h=128  -v=1 -prec=fp16 -repeat=1000
+$EXE -t=700 -h=144  -v=1 -prec=fp16 -repeat=1000
+$EXE -t=700 -h=168  -v=1 -prec=fp16 -repeat=1000
+$EXE -t=700 -h=184  -v=1 -prec=fp16 -repeat=1000
+$EXE -t=700 -h=256  -v=1 -prec=fp16 -repeat=1000
+$EXE -t=700 -h=288  -v=1 -prec=fp16 -repeat=1000
+$EXE -t=700 -h=344  -v=1 -prec=fp16 -repeat=1000
+$EXE -t=700 -h=376  -v=1 -prec=fp16 -repeat=1000
+$EXE -t=700 -h=448  -v=1 -prec=fp16 -repeat=1000
+$EXE -t=700 -h=512  -v=1 -prec=fp16 -repeat=1000
+$EXE -t=700 -h=924  -v=1 -prec=fp16 -repeat=1000
+$EXE -t=700 -h=1024  -v=1 -prec=fp16 -repeat=1000
+$EXE -t=700 -h=1078  -v=1 -prec=fp16 -repeat=1000
+$EXE -t=700 -h=1996  -v=1 -prec=fp16 -repeat=1000
+$EXE -t=700 -h=4080  -v=1 -prec=fp16 -repeat=1000
\ No newline at end of file