Merge branch 'develop' into gemm_bf16_sk_muozturk

example/ck_tile/10_rmsnorm2d/instances/rmsnorm2d_fwd_bf16_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::bf16_t,  1, 2, 1,  256, 8,  true,  false, true>>(const S&, A);
+template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t,  1, 4, 1,  256, 4,  true,  false, true>>(const S&, A);
+template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t,  1, 2, 1, 1024, 2,  true,  false, true>>(const S&, A);
+template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t,  1, 4, 1, 1024, 1,  true,  false, true>>(const S&, A);
+
+// clang-format on

example/ck_tile/10_rmsnorm2d/instances/rmsnorm2d_fwd_bf16_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::bf16_t, 1,  1,  4,  64, 8,  true , false, false>>(const S&, A);
+template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1,  2,  4,  64, 4,  true , false, false>>(const S&, A);
+template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1,  4,  4,  64, 2,  true , false, false>>(const S&, A);
+template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1,  8,  4,  64, 1,  true , false, false>>(const S&, A);
+// clang-format on

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.cpp

+#include "ck_tile/host.hpp"
+#include "rmsnorm2d_fwd.hpp"
+#include <cstring>
+
+// different threshold for different dtype
+template <typename DataType>
+auto get_elimit()
+{
+    double rtol = 1e-2;
+    double atol = 1e-2;
+    return ck_tile::make_tuple(rtol, atol);
+}
+
+template <>
+auto get_elimit<ck_tile::bf16_t>()
+{
+    double rtol = 1e-2;
+    double atol = 1e-2;
+    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("e", "1e-5", "epsilon")
+        .insert("save_rms", "0", "save rms(invrms) or not. set to 1 in training case")
+        .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 SaveRms>
+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;
+    float epsilon         = arg_parser.get_float("e");
+    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 = RmsnormTypeConfig<DataType>;
+
+    using XDataType     = typename TypeConfig::XDataType;
+    using YDataType     = typename TypeConfig::YDataType;
+    using GammaDataType = typename TypeConfig::GammaDataType;
+
+    using InvRmsDataType =
+        std::conditional_t<SaveRms, typename TypeConfig::InvRmsDataType, ck_tile::null_type>;
+
+    using ComputeDataType = typename TypeConfig::ComputeDataType;
+
+    // host verify
+    ck_tile::HostTensor<XDataType> x_host({m, n}, {stride, 1});
+    ck_tile::HostTensor<GammaDataType> gamma_host({n});
+
+    ck_tile::HostTensor<YDataType> y_host_ref({m, n}, {stride, 1});
+    ck_tile::HostTensor<YDataType> y_host_dev({m, n}, {stride, 1});
+
+    ck_tile::HostTensor<InvRmsDataType> invRms_host_ref({m});
+
+    ck_tile::FillUniformDistribution<XDataType>{-.5f, .5f}(x_host);
+    ck_tile::FillUniformDistribution<GammaDataType>{-.5f, .5f}(gamma_host);
+
+    ck_tile::DeviceMem x_buf(x_host.get_element_space_size_in_bytes());
+    ck_tile::DeviceMem gamma_buf(gamma_host.get_element_space_size_in_bytes());
+    ck_tile::DeviceMem y_buf(y_host_dev.get_element_space_size_in_bytes());
+
+    x_buf.ToDevice(x_host.data());
+    gamma_buf.ToDevice(gamma_host.data());
+
+    std::cout << "[" << data_type << "]"
+              << " m:" << m << ", n:" << n << ", stride:" << stride << std::flush;
+
+    rmsnorm2d_fwd_traits traits{data_type, SaveRms};
+
+    rmsnorm2d_fwd_args args{x_buf.GetDeviceBuffer(),
+                            gamma_buf.GetDeviceBuffer(),
+                            y_buf.GetDeviceBuffer(),
+                            nullptr,
+                            epsilon,
+                            m,
+                            n,
+                            stride};
+
+    float ave_time = rmsnorm2d_fwd(
+        traits, args, ck_tile::stream_config{nullptr, true, kname ? 1 : 0, warmup, repeat});
+
+    std::size_t num_byte =
+        sizeof(XDataType) * m * n + sizeof(GammaDataType) * n + sizeof(YDataType) * 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)
+    {
+        // reference
+        ck_tile::reference_rmsnorm2d_fwd<XDataType,
+                                         GammaDataType,
+                                         ComputeDataType,
+                                         YDataType,
+                                         InvRmsDataType>(
+            x_host, gamma_host, y_host_ref, invRms_host_ref, epsilon);
+
+        y_buf.FromDevice(y_host_dev.data());
+
+        auto [rtol, atol] = get_elimit<DataType>();
+        if(stride == n)
+        {
+            pass = ck_tile::check_err(
+                y_host_dev, y_host_ref, std::string("OUT Error: Incorrect results!"), rtol, atol);
+        }
+        else
+        {
+            for(int i_r = 0; i_r < m; i_r++)
+            {
+                std::vector<YDataType> y_host_dev_row(y_host_dev.begin() + i_r * stride,
+                                                      y_host_dev.begin() + i_r * stride + n);
+                std::vector<YDataType> y_host_ref_row(y_host_ref.begin() + i_r * stride,
+                                                      y_host_ref.begin() + i_r * stride + n);
+                pass &= ck_tile::check_err(y_host_dev_row,
+                                           y_host_ref_row,
+                                           std::string("OUT[") + 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");
+    int save_rms                = arg_parser.get_int("save_rms");
+    if(data_type == "fp16" && save_rms)
+    {
+        return run<ck_tile::half_t, true>(arg_parser) ? 0 : -2;
+    }
+    else if(data_type == "fp16" && !save_rms)
+    {
+        return run<ck_tile::half_t, false>(arg_parser) ? 0 : -2;
+    }
+    else if(data_type == "bf16" && save_rms)
+    {
+        return run<ck_tile::bf16_t, true>(arg_parser) ? 0 : -2;
+    }
+    else if(data_type == "bf16" && !save_rms)
+    {
+        return run<ck_tile::bf16_t, true>(arg_parser) ? 0 : -2;
+    }
+
+    return -3;
+}

example/ck_tile/10_rmsnorm2d/rmsnorm2d_fwd.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/rmsnorm2d.hpp"
+#include <string>
+
+template <typename DataType>
+struct RmsnormTypeConfig;
+
+template <>
+struct RmsnormTypeConfig<ck_tile::half_t>
+{
+    using XDataType       = ck_tile::half_t;
+    using YDataType       = ck_tile::half_t;
+    using GammaDataType   = ck_tile::half_t;
+    using InvRmsDataType  = ck_tile::half_t;
+    using ComputeDataType = float;
+};
+
+template <>
+struct RmsnormTypeConfig<ck_tile::bf16_t>
+{
+    using XDataType       = ck_tile::bf16_t;
+    using YDataType       = ck_tile::bf16_t;
+    using GammaDataType   = ck_tile::bf16_t;
+    using InvRmsDataType  = ck_tile::bf16_t;
+    using ComputeDataType = float;
+};
+
+// runtime args
+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;
+    bool save_rms;
+};
+
+float rmsnorm2d_fwd(rmsnorm2d_fwd_traits, rmsnorm2d_fwd_args, const ck_tile::stream_config&);

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

+#!/bin/sh
+EXE="$(find . -name tile_rmsnorm2d_fwd -type f | head -n 1)"
+
+$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/10_rmsnorm2d/script/smoke_test.sh

+#!/bin/sh
+EXE="$(find . -name tile_rmsnorm2d_fwd -type f | head -n 1)"
+
+for pr_i in "fp16" "bf16" ; do
+$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/11_add_rmsnorm2d_rdquant/CMakeLists.txt

+set(TILE_ADD_RMSNORM2D_RDQUANT_FWD "tile_add_rmsnorm2d_rdquant_fwd")
+# not using add_example_executable() to add this target, since we don't want this to have
+# to be included in "make all/install/check"
+message("adding ${TILE_ADD_RMSNORM2D_RDQUANT_FWD}")
+file(GLOB INSTANCE_SRCS instances/*.cpp)
+add_executable(${TILE_ADD_RMSNORM2D_RDQUANT_FWD} EXCLUDE_FROM_ALL add_rmsnorm2d_rdquant_fwd.cpp)
+target_include_directories(${TILE_ADD_RMSNORM2D_RDQUANT_FWD} PRIVATE ${CMAKE_CURRENT_LIST_DIR})
+target_sources(${TILE_ADD_RMSNORM2D_RDQUANT_FWD} PRIVATE ${INSTANCE_SRCS})
+
+set(TILE_ADD_RMSNORM2D_RDQUANT_FWD_COMPILE_OPTIONS)
+
+# NOTE: we turn off undefined-func-template to let source compile without explicit declare function specializations
+list(APPEND TILE_ADD_RMSNORM2D_RDQUANT_FWD_COMPILE_OPTIONS -Wno-undefined-func-template -Wno-float-equal)
+
+target_compile_options(${TILE_ADD_RMSNORM2D_RDQUANT_FWD} PRIVATE ${TILE_ADD_RMSNORM2D_RDQUANT_FWD_COMPILE_OPTIONS})
+
+set(EXAMPLE_ADD_RMSNORM2D_RDQUANT_FWD "tile_example_add_rmsnorm2d_rdquant_fwd")
+add_executable(${EXAMPLE_ADD_RMSNORM2D_RDQUANT_FWD} EXCLUDE_FROM_ALL example_add_rmsnorm2d_rdquant_fwd.cpp)
+target_compile_options(${EXAMPLE_ADD_RMSNORM2D_RDQUANT_FWD} PRIVATE ${TILE_ADD_RMSNORM2D_RDQUANT_FWD_COMPILE_OPTIONS})
+
+# TODO: we have to turn off this global prop, otherwise the progress bar generated
+# by cmake will print too many files, execvp: /bin/sh: Argument list too long
+# however, this property may affect global
+# TODO: consider codegen a makefile by us
+set_property(GLOBAL PROPERTY RULE_MESSAGES OFF)