merge from public repo

example/ck_tile/10_rmsnorm2d/instances/rmsnorm2d_fwd_bf16_n64_n128_instance.cpp

-
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-#include "rmsnorm2d_fwd_instance_common.hpp"
-
-// clang-format off
-//                                                       rm  rn  tm  tn  vn  pd    rms     2p
-template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1,  1,  4,  64, 1,  true , false, false>>(const S&, A);
-template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1,  1,  4,  64, 2,  true , false, false>>(const S&, A);
-template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1,  2,  4,  64, 1,  true , false, false>>(const S&, A);
-// clang-format on

example/ck_tile/10_rmsnorm2d/instances/rmsnorm2d_fwd_bf16_n768_instance.cpp

-
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-#include "rmsnorm2d_fwd_instance_common.hpp"
-
-// clang-format off
-//                                                       rm  rn  tm  tn  vn  pd    rms     2p
-template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1,  3,  4,  64, 4,  true , false, false>>(const S&, A);
-template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1,  6,  4,  64, 2,  true , false, false>>(const S&, A);
-template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 12,  4,  64, 1,  true , false, false>>(const S&, A);
-// clang-format on

example/ck_tile/10_rmsnorm2d/instances/rmsnorm2d_fwd_fp16_n1024_instance.cpp

-
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-#include "rmsnorm2d_fwd_instance_common.hpp"
-
-// clang-format off
-//                                                       rm  rn  tm  tn  vn  pd    rms     2p
-#if 0
-template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t, 1,  2,  4,  64, 8,  true , false, false>>(const S&, A);
-template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t, 1,  4,  4,  64, 4,  true , false, false>>(const S&, A);
-template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t, 1,  8,  4,  64, 2,  true , false, false>>(const S&, A);
-template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 16,  4,  64, 1,  true , false, false>>(const S&, A);
-
-template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t, 1,  1,  1, 256, 4,  true , false, false>>(const S&, A);
-#endif
-
-template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t,  1, 1, 2,  128, 8,  true,  false, false>>(const S&, A);
-template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t,  1, 2, 2,  128, 4,  true,  false, false>>(const S&, A);
-template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t,  1, 4, 2,  128, 2,  true,  false, false>>(const S&, A);
-template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t,  1, 4, 1,  256, 1,  true,  false, false>>(const S&, A);
-// clang-format on

example/ck_tile/10_rmsnorm2d/instances/rmsnorm2d_fwd_fp16_n1536_instance.cpp

-
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-#include "rmsnorm2d_fwd_instance_common.hpp"
-
-// clang-format off
-//                                                       rm  rn  tm  tn  vn  pd    rms     2p
-template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t,  1, 3, 4,   64, 8,  true,  false, false>>(const S&, A);
-template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t,  1, 3, 2,  128, 4,  true,  false, false>>(const S&, A);
-template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t,  1, 3, 1,  256, 2,  true,  false, false>>(const S&, A);
-template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t,  1, 6, 1,  256, 1,  true,  false, false>>(const S&, A);
-// clang-format on

example/ck_tile/10_rmsnorm2d/instances/rmsnorm2d_fwd_fp16_n2048_instance.cpp

-
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-#include "rmsnorm2d_fwd_instance_common.hpp"
-
-// clang-format off
-//                                                       rm  rn  tm  tn  vn  pd    rms     2p
-template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t,  1, 1, 1,  256, 8,  true,  false, false>>(const S&, A);
-template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t,  1, 2, 1,  256, 4,  true,  false, false>>(const S&, A);
-template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t,  1, 4, 1,  256, 2,  true,  false, false>>(const S&, A);
-template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t,  1, 8, 1,  256, 1,  true,  false, false>>(const S&, A);
-
-// clang-format on

example/ck_tile/10_rmsnorm2d/instances/rmsnorm2d_fwd_fp16_n256_instance.cpp

-
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-#include "rmsnorm2d_fwd_instance_common.hpp"
-
-// clang-format off
-//                                                       rm  rn  tm  tn  vn  pd    rms     2p
-template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t, 1,  1,  4,  64, 4,  true , false, false>>(const S&, A);
-template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t, 1,  2,  4,  64, 2,  true , false, false>>(const S&, A);
-template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t, 1,  4,  4,  64, 1,  true , false, false>>(const S&, A);
-// clang-format on

example/ck_tile/10_rmsnorm2d/instances/rmsnorm2d_fwd_fp16_n3072_instance.cpp

-
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-#include "rmsnorm2d_fwd_instance_common.hpp"
-
-// clang-format off
-//                                                       rm  rn  tm  tn  vn  pd    rms     2p
-template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t,  1, 3, 1,  128, 8,  true,  false, false>>(const S&, A);
-template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t,  1, 3, 1,  256, 4,  true,  false, false>>(const S&, A);
-template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t,  1, 6, 1,  256, 2,  true,  false, false>>(const S&, A);
-template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t,  1, 3, 1, 1024, 1,  true,  false, false>>(const S&, A);
-
-// clang-format on

example/ck_tile/10_rmsnorm2d/instances/rmsnorm2d_fwd_fp16_n4096_instance.cpp

-
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-#include "rmsnorm2d_fwd_instance_common.hpp"
-
-// clang-format off
-//                                                       rm  rn  tm  tn  vn  pd    rms     2p
-template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t,  1, 2, 1,  256, 8,  true,  false, false>>(const S&, A);
-template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t,  1, 4, 1,  256, 4,  true,  false, false>>(const S&, A);
-template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t,  1, 2, 1, 1024, 2,  true,  false, false>>(const S&, A);
-template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t,  1, 4, 1, 1024, 1,  true,  false, false>>(const S&, A);
-
-// clang-format on

example/ck_tile/10_rmsnorm2d/instances/rmsnorm2d_fwd_fp16_n4096_tp_instance.cpp

-
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-#include "rmsnorm2d_fwd_instance_common.hpp"
-
-// clang-format off
-//                                                       rm  rn  tm  tn  vn  pd    rms     2p
-template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t,  1, 2, 1,  256, 8,  true,  false, true>>(const S&, A);
-template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t,  1, 4, 1,  256, 4,  true,  false, true>>(const S&, A);
-template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t,  1, 2, 1, 1024, 2,  true,  false, true>>(const S&, A);
-template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t,  1, 4, 1, 1024, 1,  true,  false, true>>(const S&, A);
-
-// clang-format on

example/ck_tile/10_rmsnorm2d/instances/rmsnorm2d_fwd_fp16_n512_instance.cpp

-
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-#include "rmsnorm2d_fwd_instance_common.hpp"
-
-// clang-format off
-//                                                       rm  rn  tm  tn  vn  pd    rms     2p
-template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t, 1,  1,  4,  64, 8,  true , false, false>>(const S&, A);
-template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t, 1,  2,  4,  64, 4,  true , false, false>>(const S&, A);
-template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t, 1,  4,  4,  64, 2,  true , false, false>>(const S&, A);
-template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t, 1,  8,  4,  64, 1,  true , false, false>>(const S&, A);
-// clang-format on

example/ck_tile/10_rmsnorm2d/instances/rmsnorm2d_fwd_fp16_n64_n128_instance.cpp

-
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-#include "rmsnorm2d_fwd_instance_common.hpp"
-
-// clang-format off
-//                                                       rm  rn  tm  tn  vn  pd    rms     2p
-template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t, 1,  1,  4,  64, 1,  true , false, false>>(const S&, A);
-template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t, 1,  1,  4,  64, 2,  true , false, false>>(const S&, A);
-template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t, 1,  2,  4,  64, 1,  true , false, false>>(const S&, A);
-// clang-format on

example/ck_tile/10_rmsnorm2d/instances/rmsnorm2d_fwd_fp16_n768_instance.cpp

-
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-#include "rmsnorm2d_fwd_instance_common.hpp"
-
-// clang-format off
-//                                                       rm  rn  tm  tn  vn  pd    rms     2p
-template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t, 1,  3,  4,  64, 4,  true , false, false>>(const S&, A);
-template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t, 1,  6,  4,  64, 2,  true , false, false>>(const S&, A);
-template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 12,  4,  64, 1,  true , false, false>>(const S&, A);
-// clang-format on

example/ck_tile/10_rmsnorm2d/instances/rmsnorm2d_fwd_instance_common.hpp

-
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
-
-#include <ck_tile/core.hpp>
-#include "rmsnorm2d_fwd.hpp"
-#include <iostream>
-
-#pragma once
-
-using S = ck_tile::stream_config;
-using A = rmsnorm2d_fwd_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 kSaveInvRms_,
-          bool kTwoPass_>
-using trait_ = rmsnorm2d_fwd_traits_<DataType_,
-                                     Repeat_M_,
-                                     Repeat_N_,
-                                     ThreadPerBlock_M_,
-                                     ThreadPerBlock_N_,
-                                     Vector_N_,
-                                     kPadN_,
-                                     kSaveInvRms_,
-                                     kTwoPass_>;
-
-template <typename Traits_>
-float rmsnorm2d_fwd_(const S& s, A a)
-{
-    using DataType = typename Traits_::DataType;
-
-    using PipelineProblem =
-        ck_tile::Rmsnorm2dFwdPipelineProblem<typename RmsnormTypeConfig<DataType>::XDataType,
-                                             typename RmsnormTypeConfig<DataType>::GammaDataType,
-                                             typename RmsnormTypeConfig<DataType>::ComputeDataType,
-                                             typename RmsnormTypeConfig<DataType>::YDataType,
-                                             typename RmsnormTypeConfig<DataType>::InvRmsDataType,
-                                             typename Traits_::Shape,
-                                             Traits_::kPadN,
-                                             Traits_::kSaveInvRms,
-                                             Traits_::kTwoPass>;
-
-    using OnePassPipeline = ck_tile::Rmsnorm2dFwdPipelineOnePass<PipelineProblem>;
-    using TwoPassPipeline = ck_tile::Rmsnorm2dFwdPipelineTwoPass<PipelineProblem>;
-    using Pipeline        = std::conditional_t<Traits_::kTwoPass, TwoPassPipeline, OnePassPipeline>;
-
-    using Kernel = ck_tile::Rmsnorm2dFwd<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/10_rmsnorm2d/rmsnorm2d_fwd.hpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
 
 #pragma once
 
@@ -8,27 +8,34 @@
 #include "ck_tile/ops/rmsnorm2d.hpp"
 #include <string>
 
-template <typename DataType>
+template <typename InType,
+          typename OutType,
+          typename SmoothScaleDataType_,
+          typename YScaleDataType_>
 struct RmsnormTypeConfig;
 
-template <>
-struct RmsnormTypeConfig<ck_tile::half_t>
+template <typename OutType, typename SmoothScaleDataType_, typename YScaleDataType_>
+struct RmsnormTypeConfig<ck_tile::half_t, OutType, SmoothScaleDataType_, YScaleDataType_>
 {
     using XDataType           = ck_tile::half_t;
-    using YDataType       = ck_tile::half_t;
+    using YDataType           = OutType;
     using GammaDataType       = ck_tile::half_t;
     using InvRmsDataType      = ck_tile::half_t;
     using ComputeDataType     = float;
+    using SmoothScaleDataType = SmoothScaleDataType_;
+    using YScaleDataType      = YScaleDataType_;
 };
 
-template <>
-struct RmsnormTypeConfig<ck_tile::bf16_t>
+template <typename OutType, typename SmoothScaleDataType_, typename YScaleDataType_>
+struct RmsnormTypeConfig<ck_tile::bf16_t, OutType, SmoothScaleDataType_, YScaleDataType_>
 {
     using XDataType           = ck_tile::bf16_t;
-    using YDataType       = ck_tile::bf16_t;
+    using YDataType           = OutType;
     using GammaDataType       = ck_tile::bf16_t;
     using InvRmsDataType      = ck_tile::bf16_t;
     using ComputeDataType     = float;
+    using SmoothScaleDataType = SmoothScaleDataType_;
+    using YScaleDataType      = YScaleDataType_;
 };
 
 // runtime args
@@ -36,82 +43,24 @@ struct rmsnorm2d_fwd_args : public ck_tile::Rmsnorm2dFwdHostArgs
 {
 };
 
-// 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 kSaveInvRms_,
-          bool kTwoPass_>
-struct rmsnorm2d_fwd_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 kSaveInvRms = kSaveInvRms_;
-    static constexpr bool kTwoPass    = kTwoPass_;
-};
-
 template <typename Traits_>
 float rmsnorm2d_fwd_(const ck_tile::stream_config& s, rmsnorm2d_fwd_args a);
 
 // This is the public API, will be generated by script
 struct rmsnorm2d_fwd_traits
 {
-    std::string data_type;
+    std::string prec_i; // input precision
+    std::string prec_o; // output precision
+
+    // if fused_quant == 1, need set prec_sm/prec_sy to proper string, otherwise can set
+    // arbitrary(will skip check) if fused_quant == 2, need set prec_sy to proper string, otherwise
+    // can set arbitrary(will skip check)
+    std::string prec_sm; // x-scale, used for [1*N] input smooth quant
+    std::string prec_sy; // y-scale, used for [M*1] output for next layer
+
     bool save_rms;
+    int fused_add;   // 0:no-add, 1:pre-add-store, 2:pre-add
+    int fused_quant; // 0:no-sweep, 1:smooth-dynamic-quant, 2:dynamic-quant
 };
 
 float rmsnorm2d_fwd(rmsnorm2d_fwd_traits, rmsnorm2d_fwd_args, const ck_tile::stream_config&);

example/ck_tile/10_rmsnorm2d/script/smoke_test.sh

 #!/bin/sh
 EXE="$(find . -name tile_rmsnorm2d_fwd -type f | head -n 1)"
 
+for fquant in "" "-fquant=1 -prec_o=int8" "-fquant=2 -prec_o=int8"; do
 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
+for fadd in "0" "1"; do
+$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=99  -n=13
+$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=17  -n=16
+$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=1   -n=100
+$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=4   -n=128
+$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=80  -n=127
+$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=22  -n=255 -stride=256
+$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=7   -n=599
+$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=19  -n=512
+$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=33  -n=313 -stride=1000
+$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=11  -n=510
+$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=171 -n=676 -stride=818
+$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=91  -n=636
+$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=12  -n=768 -stride=800
+$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=100 -n=766 -stride=812
+$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=31  -n=1024
+$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=64  -n=1000 -stride=1004
+$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=8   -n=1501
+$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=3   -n=1826
+$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=5   -n=2040
+$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=7   -n=2734
+$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=1   -n=3182
+$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=9   -n=4096
+$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=3   -n=8192
+#$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=1   -n=10547
+#$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=3   -n=17134
+done
+done
 done

example/ck_tile/12_smoothquant/example_smoothquant.cpp

@@ -63,17 +63,17 @@ bool run(const ck_tile::ArgParser& arg_parser)
     int warmup            = arg_parser.get_int("warmup");
     int repeat            = arg_parser.get_int("repeat");
 
-    assert(stride >= n);
+    assert(x_stride >= n);
 
     using XDataType           = DataType;
-    using XScaleDataType  = float;
+    using SmoothScaleDataType = float;
     using YScaleDataType      = float;
     using QYDataType          = ck_tile::int8_t;
     using ComputeDataType     = float;
 
     // host verify
     ck_tile::HostTensor<XDataType> x_host({m, n}, {x_stride, 1});
-    ck_tile::HostTensor<XScaleDataType> xscale_host({n});
+    ck_tile::HostTensor<SmoothScaleDataType> smscale_host({n});
 
     ck_tile::HostTensor<YScaleDataType> yscale_host_ref({m}, {1});
     ck_tile::HostTensor<YScaleDataType> yscale_host_dev({m}, {1});
@@ -82,15 +82,15 @@ bool run(const ck_tile::ArgParser& arg_parser)
     ck_tile::HostTensor<QYDataType> qy_host_dev({m, n}, {y_stride, 1});
 
     ck_tile::FillUniformDistribution<XDataType>{-.5f, .5f}(x_host);
-    ck_tile::FillUniformDistribution<XScaleDataType>{1e-3, .5f}(xscale_host);
+    ck_tile::FillUniformDistribution<SmoothScaleDataType>{1e-3, .5f}(smscale_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 smscale_buf(smscale_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());
+    smscale_buf.ToDevice(smscale_host.data());
 
     constexpr bool kTwoPass = true;
 
@@ -101,7 +101,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
 
     using Shape   = ck_tile::Generic2dBlockShape<BlockTile, BlockWarps, WarpTile, Vector>;
     using Problem = ck_tile::SmoothquantPipelineProblem<XDataType,
-                                                        XScaleDataType,
+                                                        SmoothScaleDataType,
                                                         ComputeDataType,
                                                         YScaleDataType,
                                                         QYDataType,
@@ -115,7 +115,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
     using Kernel          = ck_tile::Smoothquant<Pipeline>;
 
     ck_tile::SmoothquantHostArgs args{x_buf.GetDeviceBuffer(),
-                                      xscale_buf.GetDeviceBuffer(),
+                                      smscale_buf.GetDeviceBuffer(),
                                       yscale_buf.GetDeviceBuffer(),
                                       qy_buf.GetDeviceBuffer(),
                                       m,
@@ -142,16 +142,16 @@ bool run(const ck_tile::ArgParser& arg_parser)
         // smooth outlier
         {
             auto f = [&](auto n_) {
-                auto v_xscale = ck_tile::type_convert<ComputeDataType>(xscale_host(n_));
+                auto v_smscale = ck_tile::type_convert<ComputeDataType>(smscale_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;
+                    y_host(m_, n_) = v_x * v_smscale;
                 }
             };
 
-            ck_tile::make_ParallelTensorFunctor(f, xscale_host.get_element_space_size())(
+            ck_tile::make_ParallelTensorFunctor(f, smscale_host.get_element_space_size())(
                 std::thread::hardware_concurrency());
         }
 

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.
+// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
 
 #include <ck_tile/core.hpp>
 #include "smoothquant.hpp"
@@ -35,7 +35,7 @@ float smoothquant_(const S& s, A a)
 
     using PipelineProblem = ck_tile::SmoothquantPipelineProblem<
         typename SmoothquantTypeConfig<DataType>::XDataType,
-        typename SmoothquantTypeConfig<DataType>::XScaleDataType,
+        typename SmoothquantTypeConfig<DataType>::SmoothScaleDataType,
         typename SmoothquantTypeConfig<DataType>::ComputeDataType,
         typename SmoothquantTypeConfig<DataType>::YScaleDataType,
         typename SmoothquantTypeConfig<DataType>::QYDataType,

example/ck_tile/12_smoothquant/smoothquant.cpp

@@ -67,14 +67,14 @@ bool run(const ck_tile::ArgParser& arg_parser)
     using TypeConfig = SmoothquantTypeConfig<DataType>;
 
     using XDataType           = typename TypeConfig::XDataType;
-    using XScaleDataType  = typename TypeConfig::XScaleDataType;
+    using SmoothScaleDataType = typename TypeConfig::SmoothScaleDataType;
     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}, {x_stride, 1});
-    ck_tile::HostTensor<XScaleDataType> xscale_host({n});
+    ck_tile::HostTensor<SmoothScaleDataType> smscale_host({n});
 
     ck_tile::HostTensor<YScaleDataType> yscale_host_ref({m}, {1});
     ck_tile::HostTensor<YScaleDataType> yscale_host_dev({m}, {1});
@@ -83,15 +83,15 @@ bool run(const ck_tile::ArgParser& arg_parser)
     ck_tile::HostTensor<QYDataType> qy_host_dev({m, n}, {y_stride, 1});
 
     ck_tile::FillUniformDistribution<XDataType>{-.5f, .5f}(x_host);
-    ck_tile::FillUniformDistribution<XScaleDataType>{1e-3, .5f}(xscale_host);
+    ck_tile::FillUniformDistribution<SmoothScaleDataType>{1e-3, .5f}(smscale_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 smscale_buf(smscale_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());
+    smscale_buf.ToDevice(smscale_host.data());
 
     std::cout << "[" << data_type << "]"
               << " m:" << m << ", n:" << n << ", x_stride:" << x_stride << ", y_stride:" << y_stride
@@ -100,7 +100,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
     smoothquant_traits traits{data_type};
 
     smoothquant_args args{x_buf.GetDeviceBuffer(),
-                          xscale_buf.GetDeviceBuffer(),
+                          smscale_buf.GetDeviceBuffer(),
                           yscale_buf.GetDeviceBuffer(),
                           qy_buf.GetDeviceBuffer(),
                           m,
@@ -111,7 +111,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
     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 +
+    std::size_t num_byte = sizeof(XDataType) * m * n + sizeof(SmoothScaleDataType) * n +
                            sizeof(YScaleDataType) * m + sizeof(QYDataType) * m * n;
 
     float gb_per_sec = num_byte / 1.E6 / ave_time;
@@ -126,16 +126,16 @@ bool run(const ck_tile::ArgParser& arg_parser)
         // smooth outlier
         {
             auto f = [&](auto n_) {
-                auto v_xscale = ck_tile::type_convert<ComputeDataType>(xscale_host(n_));
+                auto v_smscale = ck_tile::type_convert<ComputeDataType>(smscale_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;
+                    y_host(m_, n_) = v_x * v_smscale;
                 }
             };
 
-            ck_tile::make_ParallelTensorFunctor(f, xscale_host.get_element_space_size())(
+            ck_tile::make_ParallelTensorFunctor(f, smscale_host.get_element_space_size())(
                 std::thread::hardware_concurrency());
         }