Export ROCm/rocm-libraries@2d4a3223cb

example/ck_tile/10_rmsnorm2d/CMakeLists.txt

@@ -0,0 +1,51 @@
+# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
+# SPDX-License-Identifier: MIT
+
+set(RMSNORM2D_FWD_KNOWN_APIS "fwd;bwd")
+set(RMSNORM2D_FWD_ENABLE_APIS  "fwd" CACHE STRING
+    "semicolon-separated list of APIs to generate (${RMSNORM2D_FWD_KNOWN_APIS}) & link, or \"all\".")
+if(RMSNORM2D_FWD_ENABLE_APIS  STREQUAL "all")
+  set(RMSNORM2D_FWD_ENABLE_APIS  ${RMSNORM2D_FWD_KNOWN_APIS})
+endif()
+
+# generate a list of kernels, but not actually emit files at config sta
+execute_process(
+  COMMAND ${Python3_EXECUTABLE} ${CMAKE_CURRENT_LIST_DIR}/generate.py
+  --api ${RMSNORM2D_FWD_ENABLE_APIS} --working_path ${CMAKE_CURRENT_BINARY_DIR} --list_blobs
+  RESULT_VARIABLE ret
+)
+if(ret AND NOT ret EQUAL 0)
+  message( FATAL_ERROR "Fail to generate kernels via Python. ${ret}")
+endif()
+
+file(STRINGS ${CMAKE_CURRENT_BINARY_DIR}/rmsnorm2d_fwd_blobs.txt RMSNORM2D_FWD_GEN_BLOBS)
+
+add_custom_command(
+  OUTPUT ${RMSNORM2D_FWD_GEN_BLOBS}
+  COMMAND ${Python3_EXECUTABLE} ${CMAKE_CURRENT_LIST_DIR}/generate.py
+  --api ${RMSNORM2D_FWD_ENABLE_APIS} --working_path ${CMAKE_CURRENT_BINARY_DIR} --gen_blobs
+)
+
+set(TILE_RMSNORM2D_FWD "tile_rmsnorm2d_fwd")
+
+message(DEBUG "adding ${TILE_RMSNORM2D_FWD}")
+add_executable(${TILE_RMSNORM2D_FWD} rmsnorm2d_fwd.cpp)
+target_include_directories(${TILE_RMSNORM2D_FWD} PRIVATE ${CMAKE_CURRENT_LIST_DIR})
+target_sources(${TILE_RMSNORM2D_FWD} PRIVATE ${RMSNORM2D_FWD_GEN_BLOBS})
+
+set(TILE_RMSNORM2D_FWD_COMPILE_OPTIONS)
+
+# NOTE: we turn off undefined-func-template to let source compile without explicit declare function specializations
+list(APPEND TILE_RMSNORM2D_FWD_COMPILE_OPTIONS -Wno-undefined-func-template -Wno-float-equal --offload-compress)
+
+target_compile_options(${TILE_RMSNORM2D_FWD} PRIVATE ${TILE_RMSNORM2D_FWD_COMPILE_OPTIONS})
+
+set(EXAMPLE_RMSNORM2D_FWD "tile_example_rmsnorm2d_fwd")
+add_executable(${EXAMPLE_RMSNORM2D_FWD} example_rmsnorm2d_fwd.cpp)
+target_compile_options(${EXAMPLE_RMSNORM2D_FWD} PRIVATE ${TILE_RMSNORM2D_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)

example/ck_tile/10_rmsnorm2d/README.md

@@ -0,0 +1,87 @@
+# RMSNorm2D Forward with CK Tile
+
+This example demonstrates 2D Root Mean Square Layer Normalization (RMSNorm) using the CK Tile programming model, a normalization technique widely used in LLMs and transformers.
+
+---
+
+## Algorithm and Math
+
+For each row $x$:
+$$
+\text{rms}(x) = \sqrt{\frac{1}{N} \sum_{i=1}^N x_i^2 + \epsilon}
+$$
+$$
+y_i = \frac{x_i}{\text{rms}(x)} \cdot \gamma_i
+$$
+where $\gamma$ is a learnable scale parameter.
+
+- **Tilewise RMSNorm**: Each thread block processes a tile (row or block), computes the mean square, normalizes, and applies scale.
+
+---
+
+## Tile Programming Model
+
+- **Tiles**: Each thread block processes a tile of the input matrix.
+- **Pipeline**: Modular, can be extended for fused operations.
+
+---
+
+## Build & Run
+
+```bash
+# in the root of ck_tile
+mkdir build && cd build
+sh ../script/cmake-ck-dev.sh  ../ <arch>  # you can replace this <arch> to gfx90a, gfx942...
+make tile_rmsnorm2d_fwd -j`nproc`
+```
+
+This will result in an executable `build/bin/tile_rmsnorm2d_fwd`
+
+### Arguments
+
+```bash
+args:
+           -m    m dimension (default:3328)
+           -n    n dimension (default:4096)
+    -x_stride    x row_stride, if -1 then equal to n (default:-1)
+   -xr_stride    x residule row_stride, if -1 then equal to n (default:-1)
+    -y_stride    y row_stride, if -1 then equal to n (default:-1)
+   -yr_stride    y residule row_stride, if -1 then equal to n (default:-1)
+           -e    epsilon (default:1e-5)
+    -save_rms    save rms(invrms) or not. set to 1 in training case (default:0)
+-save_unquant    save result before quant (default:0)
+           -v    cpu validation or not (default:1)
+       -kname    print kernel name or not (default:1)
+      -prec_i    input precision (default:fp16)
+      -prec_o    output precision, set auto will be the same as input (default:auto)
+     -prec_sm    output quant scale type, set auto will use fp32. used when fquant=1 (default:auto)
+     -prec_sy    output quant scale type, set auto will use fp32. used when fquant=1 or 2 (default:auto)
+        -fadd    fused-add, 0:no fused add, 1:preadd+store, 2:preadd only (default:0)
+      -fquant    fused-quant, 0:no, 1:smooth-dynamic-quant, 2:dynamic-quant (default:0)
+      -warmup    cold iter (default:5)
+      -repeat    hot iter (default:20)
+           -s    sensitive model mode, 0: for no specific model, 1: for T5-like model (default:0)
+        -json    0: No Json, 1: Dump Results in Json format (default:0)
+    -jsonfile    json file name to dump results (default:rmsnorm2d_fwd.json)
+```
+
+---
+
+## Source Structure
+
+- **Kernel**: [`rmsnorm2d_fwd.hpp`](rmsnorm2d_fwd.hpp) (tile-programming kernel template)
+- **Executable**: [`rmsnorm2d_fwd.cpp`](rmsnorm2d_fwd.cpp) (argument parsing, kernel launch)
+- **Build**: `CMakeLists.txt`, `generate.py`, `script/`
+
+---
+
+## Related CK Tile Examples
+
+- [02_layernorm2d](../02_layernorm2d/README.md): LayerNorm2D with tiles
+- [12_smoothquant](../12_smoothquant/README.md): SmoothQuant with tiles
+- [05_reduce](../05_reduce/README.md): Reductions with tiles
+
+For distribution, see  [`include/ck_tile/tile_program/tile_distribution/`](../../../include/ck_tile/tile_program/tile_distribution/).
+
+---
+[Back to CK Tile Examples](../README.md)

example/ck_tile/10_rmsnorm2d/example_rmsnorm2d_fwd.cpp

@@ -0,0 +1,218 @@
+// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
+// SPDX-License-Identifier: MIT
+
+#include "ck_tile/host.hpp"
+#include "ck_tile/core.hpp"
+#include "ck_tile/host/kernel_launch.hpp"
+#include "ck_tile/ops/epilogue.hpp"
+#include "ck_tile/ops/rmsnorm2d.hpp"
+#include <cstring>
+
+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("v", "1", "cpu validation or not")
+        .insert("prec", "fp16", "precision")
+        .insert("warmup", "0", "cold iter")
+        .insert("repeat", "1", "hot iter")
+        .insert("s", "0", "sensitive model mode, 0: for no specific model, 1: for T5-like model");
+
+    bool result = arg_parser.parse(argc, argv);
+    return std::make_tuple(result, arg_parser);
+}
+
+template <typename DataType, int USEModelSensitive>
+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 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 XDataType           = DataType;
+    using YDataType           = DataType;
+    using GammaDataType       = DataType;
+    using InvRmsDataType      = ck_tile::null_type;
+    using UnquantYDataType    = ck_tile::null_type;
+    using SmoothScaleDataType = ck_tile::null_type;
+    using YScaleDataType      = ck_tile::null_type;
+
+    using ComputeDataType = float;
+
+    // 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::HostTensor<UnquantYDataType> unquant_y_host_ref({m, n}, {stride, 1});
+
+    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());
+
+    constexpr bool kTwoPass = true;
+
+    using BlockTile      = ck_tile::sequence<2, 128>;
+    using Vector         = ck_tile::sequence<1, 1>;
+    using ThreadPerBlock = ck_tile::sequence<2, 128>;
+
+    using Shape = ck_tile::Generic2dBlockShape<BlockTile, ThreadPerBlock, Vector>;
+
+    using PipelineTraits =
+        ck_tile::Rmsnorm2dFwdTraits<true,  // kPadN
+                                    false, // kSaveInvRms
+                                    false, // kSaveUnquant
+                                    kTwoPass,
+                                    ck_tile::Rmsnorm2dFusedAddEnum::NO_ADD,     // fuse add
+                                    ck_tile::Rmsnorm2dFusedQuantEnum::NO_SWEEP, // fuse quant
+                                    static_cast<ck_tile::Rmsnorm2dSensitiveEnum>(
+                                        USEModelSensitive)>;
+
+    using Problem = ck_tile::Rmsnorm2dFwdPipelineProblem<XDataType,
+                                                         GammaDataType,
+                                                         ComputeDataType,
+                                                         YDataType,
+                                                         InvRmsDataType,
+                                                         UnquantYDataType,
+                                                         SmoothScaleDataType,
+                                                         YScaleDataType,
+                                                         Shape,
+                                                         PipelineTraits>;
+
+    using OnePassPipeline = ck_tile::Rmsnorm2dFwdPipelineOnePass<Problem>;
+    using TwoPassPipeline = ck_tile::Rmsnorm2dFwdPipelineTwoPass<Problem>;
+    using T5PassPipeline  = ck_tile::Rmsnorm2dFwdPipelineModelSensitiveT5Pass<Problem>;
+
+    using Pipeline =
+        std::conditional_t<(PipelineTraits::kUseModelSensitiveRMSNorm ==
+                                ck_tile::Rmsnorm2dSensitiveEnum::NO_SPECIFIC_MODEL ||
+                            PipelineTraits::kTwoPass), // TODO: consider TwoPass for T5PassPipeline
+                           std::conditional_t<PipelineTraits::kTwoPass,
+                                              TwoPassPipeline,
+                                              OnePassPipeline>, // kUseModelSensitiveRMSNorm
+                                                                // == 0
+                           T5PassPipeline>;
+
+    using Default2DEpilogueProblem = ck_tile::
+        Default2DEpilogueProblem<ComputeDataType, YDataType, false, PipelineTraits::kPadN, false>;
+    using Default2DEpilogue = ck_tile::Default2DEpilogue<Default2DEpilogueProblem>;
+
+    using Kernel = ck_tile::Rmsnorm2dFwd<Pipeline, Default2DEpilogue>;
+
+    ck_tile::Rmsnorm2dFwdHostArgs args{x_buf.GetDeviceBuffer(),
+                                       nullptr,
+                                       nullptr,
+                                       gamma_buf.GetDeviceBuffer(),
+                                       y_buf.GetDeviceBuffer(),
+                                       nullptr,
+                                       nullptr,
+                                       nullptr,
+                                       nullptr,
+                                       epsilon,
+                                       m,
+                                       n,
+                                       stride,
+                                       stride,
+                                       stride,
+                                       stride};
+
+    auto kargs = Kernel::MakeKargs(args);
+
+    const dim3 grids                       = Kernel::GridSize(args);
+    const dim3 blocks                      = Kernel::BlockSize();
+    constexpr ck_tile::index_t kBlockPerCu = 1;
+    auto s = ck_tile::stream_config{nullptr, true, 0, warmup, repeat};
+
+    ck_tile::launch_kernel(s, ck_tile::make_kernel<kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
+
+    bool pass = true;
+
+    if(do_validation)
+    {
+        // reference
+        ck_tile::reference_rmsnorm2d_fwd<XDataType,
+                                         GammaDataType,
+                                         ComputeDataType,
+                                         YDataType,
+                                         InvRmsDataType,
+                                         UnquantYDataType>(
+            x_host, gamma_host, y_host_ref, invRms_host_ref, unquant_y_host_ref, epsilon);
+
+        y_buf.FromDevice(y_host_dev.data());
+
+        auto [rtol, atol] = ck_tile::make_tuple(1e-3, 1e-3);
+        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 << "[" << data_type << "]" << " m:" << m << ", n:" << n << ", stride:" << stride
+                  << ", s:" << USEModelSensitive << ", 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");
+    const int use_model_sensitive_rmsnorm = arg_parser.get_int("s");
+
+    if(data_type == "fp16")
+    {
+        if(use_model_sensitive_rmsnorm == 0) // 0: for no specific RMSNorm
+        {
+            return run<ck_tile::half_t, 0>(arg_parser) ? 0 : -2;
+        }
+        else if(use_model_sensitive_rmsnorm == 1) // 1: for T5-like RMSNorm
+        {
+            return run<ck_tile::half_t, 1>(arg_parser) ? 0 : -2;
+        }
+    }
+
+    return -3;
+}

example/ck_tile/10_rmsnorm2d/rmsnorm2d_fwd.cpp

@@ -0,0 +1,586 @@
+// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
+// SPDX-License-Identifier: MIT
+
+#include "ck_tile/host.hpp"
+#include "rmsnorm2d_fwd.hpp"
+#include <cstring>
+#include "ck_tile/utility/json_dump.hpp"
+
+// 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);
+}
+
+template <>
+auto get_elimit<ck_tile::int8_t>()
+{
+    double rtol = 1e-02;
+    double atol = 1.0;
+    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("x_stride", "-1", "x row_stride, if -1 then equal to n")
+        .insert("xr_stride", "-1", "x residule row_stride, if -1 then equal to n")
+        .insert("y_stride", "-1", "y row_stride, if -1 then equal to n")
+        .insert("yr_stride", "-1", "y residule row_stride, 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("save_unquant", "0", "save result before quant")
+        .insert("v", "1", "cpu validation or not")
+        .insert("kname", "1", "print kernel name or not")
+        .insert("prec_i", "fp16", "input precision")
+        .insert("prec_o", "auto", "output precision, set auto will be the same as input")
+        .insert("prec_sm",
+                "auto",
+                "output quant scale type, set auto will use fp32. used when fquant=1")
+        .insert("prec_sy",
+                "auto",
+                "output quant scale type, set auto will use fp32. used when fquant=1 or 2")
+        .insert("fadd", "0", "fused-add, 0:no fused add, 1:preadd+store, 2:preadd only")
+        .insert("fquant", "0", "fused-quant, 0:no, 1:smooth-dynamic-quant, 2:dynamic-quant")
+        .insert("warmup", "5", "cold iter")
+        .insert("repeat", "20", "hot iter")
+        .insert("s", "0", "sensitive model mode, 0: for no specific model, 1: for T5-like model")
+        .insert("json", "0", "0: No Json, 1: Dump Results in Json format")
+        .insert("jsonfile", "rmsnorm2d_fwd.json", "json file name to dump results");
+
+    bool result = arg_parser.parse(argc, argv);
+    return std::make_tuple(result, arg_parser);
+}
+
+template <typename InDataType,
+          typename OutDataType,
+          typename SmoothScaleDataType,
+          typename YScaleDataType,
+          bool SaveRms,
+          bool SaveUnquant>
+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");
+    float epsilon                   = arg_parser.get_float("e");
+    int kname                       = arg_parser.get_int("kname");
+    int do_validation               = arg_parser.get_int("v");
+    int fused_add                   = arg_parser.get_int("fadd");
+    int fused_quant                 = arg_parser.get_int("fquant");
+    int warmup                      = arg_parser.get_int("warmup");
+    int repeat                      = arg_parser.get_int("repeat");
+    int use_model_sensitive_rmsnorm = arg_parser.get_int("s");
+
+    ck_tile::index_t x_stride = arg_parser.get_int("x_stride");
+    if(x_stride < 0)
+        x_stride = n;
+    ck_tile::index_t xr_stride = arg_parser.get_int("xr_stride");
+    if(xr_stride < 0)
+        xr_stride = n;
+    ck_tile::index_t y_stride = arg_parser.get_int("y_stride");
+    if(y_stride < 0)
+        y_stride = n;
+    ck_tile::index_t yr_stride = arg_parser.get_int("yr_stride");
+    if(yr_stride < 0)
+        yr_stride = n;
+    assert(x_stride >= n);
+
+    std::string prec_i  = arg_parser.get_str("prec_i");
+    std::string prec_o  = arg_parser.get_str("prec_o");
+    std::string prec_sm = arg_parser.get_str("prec_sm");
+    std::string prec_sy = arg_parser.get_str("prec_sy");
+    if(prec_o == "auto")
+    {
+        prec_o = prec_i;
+    }
+    if(prec_sm == "auto")
+    {
+        prec_sm = "fp32";
+    }
+    if(prec_sy == "auto")
+    {
+        prec_sy = "fp32";
+    }
+
+    if((fused_quant == 1 || fused_quant == 2) && prec_o != "int8" && prec_o != "fp8")
+    {
+        std::cout
+            << "if fused_quant is 1 or 2, only support \"-prec_o=int8\" or \"-prec_o=fp8\" cases."
+            << std::endl;
+        return false;
+    }
+
+    if((fused_quant == 0) && SaveUnquant)
+    {
+        std::cout
+            << "save_unquant should be 0 if quant output is not enabled because it is meaningless. "
+            << "Output Y is what wanted." << std::endl;
+        return false;
+    }
+
+    using TypeConfig =
+        RmsnormTypeConfig<InDataType, OutDataType, SmoothScaleDataType, YScaleDataType>;
+
+    using XDataType         = typename TypeConfig::XDataType;
+    using YDataType         = typename TypeConfig::YDataType;
+    using GammaDataType     = typename TypeConfig::GammaDataType;
+    using XResidualDataType = XDataType;
+    using YResidualDataType = XDataType;
+
+    using InvRmsDataType =
+        std::conditional_t<SaveRms, typename TypeConfig::InvRmsDataType, ck_tile::null_type>;
+    using UnquantYDataType =
+        std::conditional_t<SaveUnquant, typename TypeConfig::UnquantYDataType, ck_tile::null_type>;
+
+    using ComputeDataType = typename TypeConfig::ComputeDataType;
+
+    // host verify
+    ck_tile::HostTensor<XDataType> x_host({m, n}, {x_stride, 1});
+    ck_tile::HostTensor<GammaDataType> gamma_host({n});
+    ck_tile::HostTensor<SmoothScaleDataType> sm_scale_host({n});
+    ck_tile::HostTensor<SmoothScaleDataType> sm_scale_host_dev({n});
+
+    ck_tile::HostTensor<XResidualDataType> x_residual_host({m, n}, {xr_stride, 1});
+    ck_tile::HostTensor<YResidualDataType> y_residual_host({m, n}, {yr_stride, 1});
+
+    ck_tile::HostTensor<YDataType> y_host_ref({m, n}, {y_stride, 1});
+    ck_tile::HostTensor<YDataType> y_host_dev({m, n}, {y_stride, 1});
+    ck_tile::HostTensor<YScaleDataType> y_scale_host_ref({m});
+    ck_tile::HostTensor<YScaleDataType> y_scale_host_dev({m});
+
+    ck_tile::HostTensor<InvRmsDataType> invRms_host_ref({m});
+
+    ck_tile::HostTensor<UnquantYDataType> unquant_y_host_ref({m, n}, {y_stride, 1});
+    ck_tile::HostTensor<UnquantYDataType> unquant_y_host_dev({m, n}, {y_stride, 1});
+    ck_tile::HostTensor<ck_tile::null_type> unquant_y_null({1});
+
+    ck_tile::FillUniformDistribution<XDataType>{-.5f, .5f}(x_host);
+    ck_tile::FillUniformDistribution<XResidualDataType>{-.5f, .5f}(x_residual_host);
+    ck_tile::FillUniformDistribution<SmoothScaleDataType>{-1.f, 1.f}(sm_scale_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());
+    ck_tile::DeviceMem y_scale_buf(y_scale_host_dev.get_element_space_size_in_bytes());
+    ck_tile::DeviceMem sm_scale_buf(sm_scale_host_dev.get_element_space_size_in_bytes());
+    ck_tile::DeviceMem x_residual_buf(x_residual_host.get_element_space_size_in_bytes());
+    ck_tile::DeviceMem y_residual_buf(y_residual_host.get_element_space_size_in_bytes());
+    ck_tile::DeviceMem unquant_y_buf(unquant_y_host_dev.get_element_space_size_in_bytes());
+
+    x_buf.ToDevice(x_host.data());
+    gamma_buf.ToDevice(gamma_host.data());
+    x_residual_buf.ToDevice(x_residual_host.data());
+    sm_scale_buf.ToDevice(sm_scale_host.data());
+
+    auto prec_str = [&]() {
+        auto base_str = prec_i;
+        if(prec_i != prec_o)
+        {
+            base_str += "|" + prec_o;
+        }
+        if(fused_quant == 1)
+        {
+            base_str += std::string("(") + prec_sy + ")";
+        }
+        return base_str;
+    }();
+
+    if(n > 8192)
+    {
+        use_model_sensitive_rmsnorm = 0;
+    }
+
+    std::cout << "[" << prec_str << "]" << " m:" << m << ", n:" << n << ", x_stride:" << x_stride
+              << ", xr_stride:" << xr_stride << ", y_stride:" << y_stride
+              << ", yr_stride:" << yr_stride << ", s:" << use_model_sensitive_rmsnorm << std::flush;
+
+    rmsnorm2d_fwd_traits traits{prec_i,
+                                prec_o,
+                                prec_sm,
+                                prec_sy,
+                                SaveRms,
+                                SaveUnquant,
+                                fused_add,
+                                fused_quant,
+                                use_model_sensitive_rmsnorm};
+
+    rmsnorm2d_fwd_args args{x_buf.GetDeviceBuffer(),
+                            fused_add != 0 ? x_residual_buf.GetDeviceBuffer() : nullptr,
+                            fused_quant == 1 ? sm_scale_buf.GetDeviceBuffer() : nullptr,
+                            gamma_buf.GetDeviceBuffer(),
+                            y_buf.GetDeviceBuffer(),
+                            fused_add == 1 ? y_residual_buf.GetDeviceBuffer() : nullptr,
+                            fused_quant != 0 ? y_scale_buf.GetDeviceBuffer() : nullptr,
+                            nullptr, // p_invRms, unsupported yet
+                            SaveUnquant ? unquant_y_buf.GetDeviceBuffer() : nullptr,
+                            epsilon,
+                            m,
+                            n,
+                            x_stride,   // x row_stride
+                            xr_stride,  // x residule row stride
+                            y_stride,   // y row stride
+                            yr_stride}; // y residule row 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;
+    num_byte += SaveRms ? sizeof(InvRmsDataType) * m * n : 0;
+    num_byte += SaveUnquant ? sizeof(UnquantYDataType) * m * n : 0;
+    num_byte += fused_add ? sizeof(XResidualDataType) * m * n : 0;
+    num_byte += ((fused_quant == 1) || (fused_quant == 2)) ? sizeof(YScaleDataType) * m : 0;
+    num_byte += (fused_quant == 1) ? sizeof(SmoothScaleDataType) * n : 0;
+
+    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
+        if(fused_add != 0)
+        {
+            // fused pre_add/pre_add_store
+            // TODO we accumulate directly to x_host for simplcity here...
+            std::transform(x_host.mData.cbegin(),
+                           x_host.mData.cend(),
+                           x_residual_host.mData.cbegin(),
+                           x_host.mData.begin(),
+                           [](auto x_, auto r_) {
+                               auto o_ = ck_tile::type_convert<ComputeDataType>(x_) +
+                                         ck_tile::type_convert<ComputeDataType>(r_);
+                               return ck_tile::type_convert<XDataType>(o_);
+                           });
+        }
+
+        if(fused_quant != 0)
+        {
+            auto dquant_functor = [&](int m_, auto& o_, auto& acc_) {
+                int N_ = acc_.mDesc.get_lengths()[1];
+                if(fused_quant == 1)
+                {
+                    for(int n_ = 0; n_ < N_; n_++)
+                    {
+                        // input smooth outlier
+                        acc_(m_, n_) = acc_(m_, n_) *
+                                       ck_tile::type_convert<ComputeDataType>(sm_scale_host(n_));
+                    }
+                }
+                ComputeDataType absmax = static_cast<ComputeDataType>(0);
+                for(int n_ = 0; n_ < N_; n_++)
+                {
+                    const auto a = ck_tile::abs(acc_(m_, n_));
+                    absmax       = a > absmax ? a : absmax;
+                }
+                // printf("cpu:absmax:%f\n", absmax);
+                constexpr ComputeDataType kMaxY =
+                    std::is_same<YDataType, ck_tile::fp8_t>::value    ? 240.0
+                    : std::is_same<YDataType, ck_tile::int8_t>::value ? 127.0
+                                                                      : 0.0;
+                ComputeDataType y_scale = absmax / kMaxY;
+                y_scale_host_ref(m_)    = ck_tile::type_convert<YScaleDataType>(y_scale);
+                for(int n_ = 0; n_ < N_; n_++)
+                {
+                    o_(m_, n_) = ck_tile::type_convert<YDataType>(acc_(m_, n_) / y_scale);
+                }
+            };
+
+            auto default_and_dquant_functor = [&](int m_, auto& o_unquant_, auto& o_, auto& acc_) {
+                const int N = acc_.mDesc.get_lengths()[1];
+                for(int n_ = 0; n_ < N; ++n_)
+                {
+                    o_unquant_(m_, n_) = ck_tile::type_convert<UnquantYDataType>(acc_(m_, n_));
+                }
+
+                dquant_functor(m_, o_, acc_);
+            };
+
+            if constexpr(SaveUnquant)
+            {
+                ck_tile::reference_rmsnorm2d_fwd<XDataType,
+                                                 GammaDataType,
+                                                 ComputeDataType,
+                                                 YDataType,
+                                                 InvRmsDataType,
+                                                 UnquantYDataType>(x_host,
+                                                                   gamma_host,
+                                                                   y_host_ref,
+                                                                   invRms_host_ref,
+                                                                   unquant_y_host_ref,
+                                                                   epsilon,
+                                                                   default_and_dquant_functor,
+                                                                   use_model_sensitive_rmsnorm);
+            }
+            else
+            {
+                ck_tile::reference_rmsnorm2d_fwd<XDataType,
+                                                 GammaDataType,
+                                                 ComputeDataType,
+                                                 YDataType,
+                                                 InvRmsDataType,
+                                                 UnquantYDataType>(x_host,
+                                                                   gamma_host,
+                                                                   y_host_ref,
+                                                                   invRms_host_ref,
+                                                                   unquant_y_host_ref,
+                                                                   epsilon,
+                                                                   dquant_functor,
+                                                                   use_model_sensitive_rmsnorm);
+            }
+        }
+        else
+        {
+            assert(SaveUnquant == false);
+            ck_tile::reference_rmsnorm2d_fwd<XDataType,
+                                             GammaDataType,
+                                             ComputeDataType,
+                                             YDataType,
+                                             InvRmsDataType,
+                                             ck_tile::null_type>(
+                x_host,
+                gamma_host,
+                y_host_ref,
+                invRms_host_ref,
+                unquant_y_null,
+                epsilon,
+                ck_tile::reference_rmsnorm2d_default_epilogue{},
+                use_model_sensitive_rmsnorm);
+        }
+
+        y_buf.FromDevice(y_host_dev.data());
+
+        ck_tile::HostTensor<YResidualDataType> y_residual_host_dev({m, n}, {yr_stride, 1});
+        if(fused_add == 1)
+        {
+            y_residual_buf.FromDevice(y_residual_host_dev.data());
+        }
+
+        if constexpr(SaveUnquant)
+        {
+            unquant_y_buf.FromDevice(unquant_y_host_dev.data());
+        }
+
+        auto [rtol, atol] = get_elimit<YDataType>();
+        if(x_stride == n)
+        {
+            pass = ck_tile::check_err(
+                y_host_dev, y_host_ref, std::string("\nOUT Error: Incorrect results!"), rtol, atol);
+
+            if constexpr(SaveUnquant)
+            {
+                pass &= ck_tile::check_err(unquant_y_host_dev,
+                                           unquant_y_host_ref,
+                                           std::string("\n OUT ERROR: Incorrect unquant results!"),
+                                           rtol,
+                                           atol);
+            }
+
+            if(fused_add == 1)
+            {
+                pass &= ck_tile::check_err(y_residual_host_dev,
+                                           x_host,
+                                           std::string("\nADD 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 * y_stride,
+                                                      y_host_dev.begin() + i_r * y_stride + n);
+                std::vector<YDataType> y_host_ref_row(y_host_ref.begin() + i_r * y_stride,
+                                                      y_host_ref.begin() + i_r * y_stride + n);
+                pass &= ck_tile::check_err(y_host_dev_row,
+                                           y_host_ref_row,
+                                           std::string("\nOUT[") + std::to_string(i_r) +
+                                               std::string("] Error: Incorrect results!"),
+                                           rtol,
+                                           atol);
+
+                if(fused_add == 1)
+                {
+                    std::vector<YResidualDataType> y_residual_host_dev_row(
+                        y_residual_host_dev.begin() + i_r * yr_stride,
+                        y_residual_host_dev.begin() + i_r * yr_stride + n);
+                    std::vector<YResidualDataType> y_residual_host_ref_row(
+                        x_host.begin() + i_r * yr_stride, x_host.begin() + i_r * yr_stride + n);
+                    pass &= ck_tile::check_err(y_residual_host_dev_row,
+                                               y_residual_host_ref_row,
+                                               std::string("\nADD[") + std::to_string(i_r) +
+                                                   std::string("] Error: Incorrect results!"),
+                                               rtol,
+                                               atol);
+                }
+
+                if constexpr(SaveUnquant)
+                {
+                    std::vector<UnquantYDataType> unquant_y_host_dev_row(
+                        unquant_y_host_dev.begin() + i_r * y_stride,
+                        unquant_y_host_dev.begin() + i_r * y_stride + n);
+                    std::vector<UnquantYDataType> unquant_y_host_ref_row(
+                        unquant_y_host_ref.begin() + i_r * y_stride,
+                        unquant_y_host_ref.begin() + i_r * y_stride + n);
+                    pass &=
+                        ck_tile::check_err(unquant_y_host_dev_row,
+                                           unquant_y_host_ref_row,
+                                           std::string("\nOUT[") + std::to_string(i_r) +
+                                               std::string("] Error: Incorrect unquant y results!"),
+                                           rtol,
+                                           atol);
+                }
+            }
+        }
+
+        if(fused_quant == 1)
+        {
+            y_scale_buf.FromDevice(y_scale_host_dev.data());
+            pass &= ck_tile::check_err(y_scale_host_dev,
+                                       y_scale_host_ref,
+                                       std::string("\nSCALE Error: Incorrect results!"),
+                                       rtol,
+                                       atol);
+        }
+
+        std::cout << ", valid:" << (pass ? "y" : "n") << std::flush << std::endl;
+    }
+
+    if(arg_parser.get_int("json") == 1)
+    {
+        dump_rmsnorm2d_fwd_json(arg_parser.get_str("jsonfile"),
+                                prec_str,
+                                m,
+                                n,
+                                x_stride,
+                                xr_stride,
+                                y_stride,
+                                yr_stride,
+                                use_model_sensitive_rmsnorm,
+                                ave_time,
+                                0,
+                                gb_per_sec,
+                                pass);
+    }
+
+    return pass;
+}
+
+bool is_quant_data_type(const std::string& prec) { return (prec == "int8") || (prec == "fp8"); }
+
+int main(int argc, char* argv[])
+{
+    auto [result, arg_parser] = create_args(argc, argv);
+    if(!result)
+        return -1;
+
+    std::string prec_i  = arg_parser.get_str("prec_i");
+    std::string prec_o  = arg_parser.get_str("prec_o");
+    std::string prec_sm = arg_parser.get_str("prec_sm");
+    std::string prec_sy = arg_parser.get_str("prec_sy");
+    if(prec_o == "auto")
+    {
+        prec_o = prec_i;
+    }
+    if(prec_sm == "auto")
+    {
+        prec_sm = "fp32";
+    }
+    if(prec_sy == "auto")
+    {
+        prec_sy = "fp32";
+    }
+
+    int save_rms    = arg_parser.get_int("save_rms");
+    int fused_quant = arg_parser.get_int("fquant");
+    int save_unquant =
+        arg_parser.get_int("save_unquant") && is_quant_data_type(prec_o) && (fused_quant != 0);
+
+    if(prec_i == "fp16" && prec_o == "fp16" && prec_sm == "fp32" && prec_sy == "fp32" && save_rms)
+    {
+        return run<ck_tile::half_t, ck_tile::half_t, float, float, true, false>(arg_parser) ? 0
+                                                                                            : -2;
+    }
+    else if(prec_i == "fp16" && prec_o == "fp16" && prec_sm == "fp32" && prec_sy == "fp32" &&
+            !save_rms)
+    {
+        return run<ck_tile::half_t, ck_tile::half_t, float, float, false, false>(arg_parser) ? 0
+                                                                                             : -2;
+    }
+    else if(prec_i == "bf16" && prec_o == "bf16" && prec_sm == "fp32" && prec_sy == "fp32" &&
+            save_rms)
+    {
+        return run<ck_tile::bf16_t, ck_tile::bf16_t, float, float, true, false>(arg_parser) ? 0
+                                                                                            : -2;
+    }
+    else if(prec_i == "bf16" && prec_o == "bf16" && prec_sm == "fp32" && prec_sy == "fp32" &&
+            !save_rms)
+    {
+        return run<ck_tile::bf16_t, ck_tile::bf16_t, float, float, false, false>(arg_parser) ? 0
+                                                                                             : -2;
+    }
+
+    // dynamic quant case, only in inference
+    else if(prec_i == "fp16" && prec_o == "int8" && prec_sm == "fp32" && prec_sy == "fp32" &&
+            !save_rms && !save_unquant)
+    {
+        return run<ck_tile::half_t, ck_tile::int8_t, float, float, true, false>(arg_parser) ? 0
+                                                                                            : -2;
+    }
+    else if(prec_i == "bf16" && prec_o == "int8" && prec_sm == "fp32" && prec_sy == "fp32" &&
+            !save_rms && !save_unquant)
+    {
+        return run<ck_tile::bf16_t, ck_tile::int8_t, float, float, true, false>(arg_parser) ? 0
+                                                                                            : -2;
+    }
+    else if(prec_i == "fp16" && prec_o == "fp8" && prec_sm == "fp32" && prec_sy == "fp32" &&
+            !save_rms && !save_unquant)
+    {
+        return run<ck_tile::half_t, ck_tile::fp8_t, float, float, false, false>(arg_parser) ? 0
+                                                                                            : -2;
+    }
+    else if(prec_i == "bf16" && prec_o == "fp8" && prec_sm == "fp32" && prec_sy == "fp32" &&
+            !save_rms && !save_unquant)
+    {
+        return run<ck_tile::bf16_t, ck_tile::fp8_t, float, float, false, false>(arg_parser) ? 0
+                                                                                            : -2;
+    }
+    else if(prec_i == "fp16" && prec_o == "int8" && prec_sm == "fp32" && prec_sy == "fp32" &&
+            !save_rms && save_unquant)
+    {
+        return run<ck_tile::half_t, ck_tile::int8_t, float, float, true, true>(arg_parser) ? 0 : -2;
+    }
+    else if(prec_i == "bf16" && prec_o == "int8" && prec_sm == "fp32" && prec_sy == "fp32" &&
+            !save_rms && save_unquant)
+    {
+        return run<ck_tile::bf16_t, ck_tile::int8_t, float, float, true, true>(arg_parser) ? 0 : -2;
+    }
+    else if(prec_i == "fp16" && prec_o == "fp8" && prec_sm == "fp32" && prec_sy == "fp32" &&
+            !save_rms && save_unquant)
+    {
+        return run<ck_tile::half_t, ck_tile::fp8_t, float, float, false, true>(arg_parser) ? 0 : -2;
+    }
+    else if(prec_i == "bf16" && prec_o == "fp8" && prec_sm == "fp32" && prec_sy == "fp32" &&
+            !save_rms && save_unquant)
+    {
+        return run<ck_tile::bf16_t, ck_tile::fp8_t, float, float, false, true>(arg_parser) ? 0 : -2;
+    }
+
+    return -3;
+}

example/ck_tile/10_rmsnorm2d/rmsnorm2d_fwd.hpp

@@ -0,0 +1,71 @@
+// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
+// SPDX-License-Identifier: MIT
+
+#pragma once
+
+#include "ck_tile/core.hpp"
+#include "ck_tile/host/kernel_launch.hpp"
+#include "ck_tile/ops/rmsnorm2d.hpp"
+#include <string>
+
+template <typename InType,
+          typename OutType,
+          typename SmoothScaleDataType_,
+          typename YScaleDataType_>
+struct RmsnormTypeConfig;
+
+template <typename OutType, typename SmoothScaleDataType_, typename YScaleDataType_>
+struct RmsnormTypeConfig<ck_tile::half_t, OutType, SmoothScaleDataType_, YScaleDataType_>
+{
+    using XDataType           = ck_tile::half_t;
+    using YDataType           = OutType;
+    using GammaDataType       = ck_tile::half_t;
+    using InvRmsDataType      = ck_tile::half_t;
+    using UnquantYDataType    = ck_tile::half_t;
+    using ComputeDataType     = float;
+    using SmoothScaleDataType = SmoothScaleDataType_;
+    using YScaleDataType      = YScaleDataType_;
+};
+
+template <typename OutType, typename SmoothScaleDataType_, typename YScaleDataType_>
+struct RmsnormTypeConfig<ck_tile::bf16_t, OutType, SmoothScaleDataType_, YScaleDataType_>
+{
+    using XDataType           = ck_tile::bf16_t;
+    using YDataType           = OutType;
+    using GammaDataType       = ck_tile::bf16_t;
+    using InvRmsDataType      = ck_tile::bf16_t;
+    using UnquantYDataType    = ck_tile::bf16_t;
+    using ComputeDataType     = float;
+    using SmoothScaleDataType = SmoothScaleDataType_;
+    using YScaleDataType      = YScaleDataType_;
+};
+
+// runtime args
+struct rmsnorm2d_fwd_args : public ck_tile::Rmsnorm2dFwdHostArgs
+{
+};
+
+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 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;
+    bool save_unquant;
+    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
+
+    int use_model_sensitive_rmsnorm = 0; // 0: Use default RMSNorm; 1: Use T5-like implementation
+};
+
+float rmsnorm2d_fwd(rmsnorm2d_fwd_traits, rmsnorm2d_fwd_args, const ck_tile::stream_config&);

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

@@ -0,0 +1,77 @@
+#!/bin/sh
+# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
+# SPDX-License-Identifier: MIT
+
+EXE="$(find . -name tile_rmsnorm2d_fwd -type f | head -n 1)"
+
+# 0: for no specific RMSNorm
+$EXE -m=1 -n=1 -e=1e-12 -v=1 -prec_i=bf16 -repeat=1000 -s=0
+$EXE -m=700 -n=80 -e=1e-12 -v=1 -prec_i=bf16 -repeat=1000 -s=0
+$EXE -m=700 -n=128 -e=1e-12 -v=1 -prec_i=bf16 -repeat=1000 -s=0
+$EXE -m=700 -n=144 -e=1e-12 -v=1 -prec_i=bf16 -repeat=1000 -s=0
+$EXE -m=700 -n=168 -e=1e-12 -v=1 -prec_i=bf16 -repeat=1000 -s=0
+$EXE -m=700 -n=184 -e=1e-12 -v=1 -prec_i=bf16 -repeat=1000 -s=0
+$EXE -m=700 -n=256 -e=1e-12 -v=1 -prec_i=bf16 -repeat=1000 -s=0
+$EXE -m=700 -n=288 -e=1e-12 -v=1 -prec_i=bf16 -repeat=1000 -s=0
+$EXE -m=700 -n=344 -e=1e-12 -v=1 -prec_i=bf16 -repeat=1000 -s=0
+$EXE -m=700 -n=376 -e=1e-12 -v=1 -prec_i=bf16 -repeat=1000 -s=0
+$EXE -m=700 -n=448 -e=1e-12 -v=1 -prec_i=bf16 -repeat=1000 -s=0
+$EXE -m=700 -n=512 -e=1e-12 -v=1 -prec_i=bf16 -repeat=1000 -s=0
+$EXE -m=700 -n=924 -e=1e-12 -v=1 -prec_i=bf16 -repeat=1000 -s=0
+$EXE -m=700 -n=1024 -e=1e-12 -v=1 -prec_i=bf16 -repeat=1000 -s=0
+$EXE -m=700 -n=1078 -e=1e-12 -v=1 -prec_i=bf16 -repeat=1000 -s=0
+$EXE -m=700 -n=1996 -e=1e-12 -v=1 -prec_i=bf16 -repeat=1000 -s=0
+$EXE -m=700 -n=4080 -e=1e-12 -v=1 -prec_i=bf16 -repeat=1000 -s=0
+
+$EXE -m=700 -n=80 -e=1e-12 -v=1  -prec_i=fp16 -repeat=1000 -s=0
+$EXE -m=700 -n=128 -e=1e-12 -v=1 -prec_i=fp16 -repeat=1000 -s=0
+$EXE -m=700 -n=144 -e=1e-12 -v=1 -prec_i=fp16 -repeat=1000 -s=0
+$EXE -m=700 -n=168 -e=1e-12 -v=1 -prec_i=fp16 -repeat=1000 -s=0
+$EXE -m=700 -n=184 -e=1e-12 -v=1 -prec_i=fp16 -repeat=1000 -s=0
+$EXE -m=700 -n=256 -e=1e-12 -v=1 -prec_i=fp16 -repeat=1000 -s=0
+$EXE -m=700 -n=288 -e=1e-12 -v=1 -prec_i=fp16 -repeat=1000 -s=0
+$EXE -m=700 -n=344 -e=1e-12 -v=1 -prec_i=fp16 -repeat=1000 -s=0
+$EXE -m=700 -n=376 -e=1e-12 -v=1 -prec_i=fp16 -repeat=1000 -s=0
+$EXE -m=700 -n=448 -e=1e-12 -v=1 -prec_i=fp16 -repeat=1000 -s=0
+$EXE -m=700 -n=512 -e=1e-12 -v=1 -prec_i=fp16 -repeat=1000 -s=0
+$EXE -m=700 -n=924 -e=1e-12 -v=1 -prec_i=fp16 -repeat=1000 -s=0
+$EXE -m=700 -n=1024 -e=1e-12 -v=1 -prec_i=fp16 -repeat=1000 -s=0
+$EXE -m=700 -n=1078 -e=1e-12 -v=1 -prec_i=fp16 -repeat=1000 -s=0
+$EXE -m=700 -n=1996 -e=1e-12 -v=1 -prec_i=fp16 -repeat=1000 -s=0
+$EXE -m=700 -n=4080 -e=1e-12 -v=1 -prec_i=fp16 -repeat=1000 -s=0
+
+# 1: for T5-like RMSNorm
+$EXE -m=1 -n=1 -e=1e-12 -v=1 -prec_i=bf16 -repeat=1000 -s=1
+$EXE -m=700 -n=80 -e=1e-12 -v=1 -prec_i=bf16 -repeat=1000 -s=1
+$EXE -m=700 -n=128 -e=1e-12 -v=1 -prec_i=bf16 -repeat=1000 -s=1
+$EXE -m=700 -n=144 -e=1e-12 -v=1 -prec_i=bf16 -repeat=1000 -s=1
+$EXE -m=700 -n=168 -e=1e-12 -v=1 -prec_i=bf16 -repeat=1000 -s=1
+$EXE -m=700 -n=184 -e=1e-12 -v=1 -prec_i=bf16 -repeat=1000 -s=1
+$EXE -m=700 -n=256 -e=1e-12 -v=1 -prec_i=bf16 -repeat=1000 -s=1
+$EXE -m=700 -n=288 -e=1e-12 -v=1 -prec_i=bf16 -repeat=1000 -s=1
+$EXE -m=700 -n=344 -e=1e-12 -v=1 -prec_i=bf16 -repeat=1000 -s=1
+$EXE -m=700 -n=376 -e=1e-12 -v=1 -prec_i=bf16 -repeat=1000 -s=1
+$EXE -m=700 -n=448 -e=1e-12 -v=1 -prec_i=bf16 -repeat=1000 -s=1
+$EXE -m=700 -n=512 -e=1e-12 -v=1 -prec_i=bf16 -repeat=1000 -s=1
+$EXE -m=700 -n=924 -e=1e-12 -v=1 -prec_i=bf16 -repeat=1000 -s=1
+$EXE -m=700 -n=1024 -e=1e-12 -v=1 -prec_i=bf16 -repeat=1000 -s=1
+$EXE -m=700 -n=1078 -e=1e-12 -v=1 -prec_i=bf16 -repeat=1000 -s=1
+$EXE -m=700 -n=1996 -e=1e-12 -v=1 -prec_i=bf16 -repeat=1000 -s=1
+$EXE -m=700 -n=4080 -e=1e-12 -v=1 -prec_i=bf16 -repeat=1000 -s=1
+
+$EXE -m=700 -n=80 -e=1e-12 -v=1  -prec_i=fp16 -repeat=1000 -s=1
+$EXE -m=700 -n=128 -e=1e-12 -v=1 -prec_i=fp16 -repeat=1000 -s=1
+$EXE -m=700 -n=144 -e=1e-12 -v=1 -prec_i=fp16 -repeat=1000 -s=1
+$EXE -m=700 -n=168 -e=1e-12 -v=1 -prec_i=fp16 -repeat=1000 -s=1
+$EXE -m=700 -n=184 -e=1e-12 -v=1 -prec_i=fp16 -repeat=1000 -s=1
+$EXE -m=700 -n=256 -e=1e-12 -v=1 -prec_i=fp16 -repeat=1000 -s=1
+$EXE -m=700 -n=288 -e=1e-12 -v=1 -prec_i=fp16 -repeat=1000 -s=1
+$EXE -m=700 -n=344 -e=1e-12 -v=1 -prec_i=fp16 -repeat=1000 -s=1
+$EXE -m=700 -n=376 -e=1e-12 -v=1 -prec_i=fp16 -repeat=1000 -s=1
+$EXE -m=700 -n=448 -e=1e-12 -v=1 -prec_i=fp16 -repeat=1000 -s=1
+$EXE -m=700 -n=512 -e=1e-12 -v=1 -prec_i=fp16 -repeat=1000 -s=1
+$EXE -m=700 -n=924 -e=1e-12 -v=1 -prec_i=fp16 -repeat=1000 -s=1
+$EXE -m=700 -n=1024 -e=1e-12 -v=1 -prec_i=fp16 -repeat=1000 -s=1
+$EXE -m=700 -n=1078 -e=1e-12 -v=1 -prec_i=fp16 -repeat=1000 -s=1
+$EXE -m=700 -n=1996 -e=1e-12 -v=1 -prec_i=fp16 -repeat=1000 -s=1
+$EXE -m=700 -n=4080 -e=1e-12 -v=1 -prec_i=fp16 -repeat=1000 -s=1

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

@@ -0,0 +1,88 @@
+#!/bin/bash
+# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
+# SPDX-License-Identifier: MIT
+
+
+EXE="$(find . -name tile_rmsnorm2d_fwd -type f | head -n 1)"
+
+total=0
+valid=0
+
+run_case() {
+  cmd="$EXE -prec_i=$1 -fadd=$2 -s=$3 $4 -m=$5 -n=$6 $7"
+  echo "[CMD] $cmd"
+  output=$($cmd 2>&1)
+  echo "$output"
+  if echo "$output" | grep -q "valid:y"; then
+    valid=$((valid + 1))
+  fi
+  total=$((total + 1))
+}
+
+fquant_list=(
+  ""
+  "-fquant=1 -prec_o=int8"
+  "-fquant=2 -prec_o=int8"
+  "-fquant=1 -prec_o=fp8"
+  "-fquant=2 -prec_o=fp8"
+  "-fquant=1 -prec_o=int8 -save_unquant=1"
+  "-fquant=2 -prec_o=int8 -save_unquant=1"
+  "-fquant=1 -prec_o=fp8 -save_unquant=1"
+  "-fquant=2 -prec_o=fp8 -save_unquant=1"
+)
+
+m_n_list=(
+  "99 13" "17 16" "1 100" "4 128" "80 127"
+  "7 599" "19 512" "11 510" "91 636"
+  "31 1024" "8 1501" "3 1826" "5 2040"
+  "7 2734" "1 3182" "9 4096" "3 8192"
+)
+
+### Add special stride test ###
+m_n_stride_list=(
+  "22 255 -x_stride=256 -xr_stride=256 -y_stride=256 -yr_stride=256"
+  "33 313 -x_stride=1000 -xr_stride=1000 -y_stride=1000 -yr_stride=1000"
+  "171 676 -x_stride=818 -xr_stride=818 -y_stride=818 -yr_stride=818"
+  "12 768 -x_stride=800 -xr_stride=800 -y_stride=800 -yr_stride=800"
+  "100 766 -x_stride=812 -xr_stride=812 -y_stride=812 -yr_stride=812"
+  "64 1000 -x_stride=1004 -xr_stride=1004 -y_stride=1004 -yr_stride=1004"
+)
+
+for fquant in "${fquant_list[@]}"; do
+  for pr_i in "fp16" "bf16"; do
+    for fadd in "0" "1"; do
+      for s in "0" "1"; do
+        for pair in "${m_n_list[@]}"; do
+          m=$(echo $pair | cut -d ' ' -f1)
+          n=$(echo $pair | cut -d ' ' -f2)
+          run_case "$pr_i" "$fadd" "$s" "$fquant" "$m" "$n" ""
+        done
+
+        ### Running tests with stride ###
+        for triple in "${m_n_stride_list[@]}"; do
+          m=$(echo $triple | cut -d ' ' -f1)
+          n=$(echo $triple | cut -d ' ' -f2)
+          stride_args=$(echo $triple | cut -d ' ' -f3-)
+          run_case "$pr_i" "$fadd" "$s" "$fquant" "$m" "$n" "$stride_args"
+        done
+      done
+    done
+  done
+done
+
+# Special two-pass only
+for pr_i in "fp16" "bf16"; do
+  for fadd in "0" "1"; do
+    for s in "0" "1"; do
+      run_case "$pr_i" "$fadd" "$s" "" "1" "10547" ""
+    done
+  done
+done
+
+# Summary
+echo "=============================="
+echo "Total cases: $total"
+echo "Valid cases: $valid"
+accuracy=$(awk "BEGIN {printf \"%.2f\", ($valid / $total) * 100}")
+echo "Accuracy: $accuracy%"
+echo "=============================="