Merge remote-tracking branch 'origin/feiw/mxfp4_moe_2Stages' into wip-f4

clang_format_git.py

@@ -0,0 +1,140 @@
+#!/usr/bin/env python3
+
+import subprocess
+import sys
+import re
+import os
+from pathlib import Path
+from typing import List
+
+# 定义颜色代码
+class Color:
+    GREEN = '\033[0;32m'
+    YELLOW = '\033[0;33m'
+    RED = '\033[0;31m'
+    NC = '\033[0m'  # 无颜色
+
+def print_color(color: str, message: str):
+    """打印带颜色的消息"""
+    print(f"{color}{message}{Color.NC}")
+
+def check_command_exists(command: str) -> bool:
+    """检查命令是否存在"""
+    try:
+        subprocess.run(['which', command], stdout=subprocess.PIPE, stderr=subprocess.PIPE, check=True)
+        return True
+    except subprocess.CalledProcessError:
+        return False
+
+def is_git_repo() -> bool:
+    """检查当前目录是否是Git仓库"""
+    try:
+        subprocess.run(['git', 'rev-parse', '--is-inside-work-tree'], 
+                     stdout=subprocess.PIPE, stderr=subprocess.PIPE, check=True)
+        return True
+    except subprocess.CalledProcessError:
+        return False
+
+def get_git_modified_files() -> List[str]:
+    """获取Git中已修改但未提交的文件及未跟踪的新文件"""
+    try:
+        # 获取已修改的文件
+        modified_result = subprocess.run(['git', 'status', '--porcelain'], 
+                                      stdout=subprocess.PIPE, text=True, check=True)
+        
+        files = []
+        for line in modified_result.stdout.splitlines():
+            if not line.strip():
+                continue
+            
+            # 检查文件状态（M=修改, A=添加, ??=未跟踪）
+            if re.match(r'^\s*[AM]', line) or line.startswith('??'):
+                # 提取文件名
+                parts = line.strip().split(maxsplit=1)
+                if len(parts) > 1:
+                    files.append(parts[1])
+                else:
+                    # 处理未跟踪文件的情况
+                    files.append(parts[0][2:].strip())
+                    
+        return files
+    except subprocess.CalledProcessError as e:
+        print_color(Color.RED, f"获取Git修改文件失败: {e}")
+        return []
+
+def filter_cpp_files(files: List[str]) -> List[str]:
+    """筛选C++相关文件"""
+    cpp_extensions = ['.cpp', '.hpp', '.h', '.cc', '.c', '.cxx']
+    return [file for file in files if Path(file).suffix.lower() in cpp_extensions]
+
+def format_files(files: List[str]) -> tuple:
+    """使用clang-format格式化文件"""
+    success_count = 0
+    fail_count = 0
+    
+    for file in files:
+        print(f"格式化: {file} ... ", end="")
+        try:
+            # 先运行dos2unix确保文件使用Unix换行符
+            subprocess.run(['dos2unix', file], 
+                         stdout=subprocess.PIPE, stderr=subprocess.PIPE, check=True)
+            
+            # 使用clang-format格式化文件
+            subprocess.run(['clang-format-12', '-style=file', '-i', file], 
+                         stdout=subprocess.PIPE, stderr=subprocess.PIPE, check=True)
+            
+            print_color(Color.GREEN, "成功")
+            success_count += 1
+        except subprocess.CalledProcessError as e:
+            print_color(Color.RED, f"失败: {e}")
+            fail_count += 1
+    
+    return success_count, fail_count
+
+def main():
+    # 检查clang-format-12是否安装
+    if not check_command_exists('clang-format-12'):
+        print_color(Color.RED, "错误: clang-format-12 未安装或不在PATH中")
+        sys.exit(1)
+        
+    # 检查dos2unix是否安装
+    if not check_command_exists('dos2unix'):
+        print_color(Color.YELLOW, "警告: dos2unix 未安装，将跳过行尾符转换")
+        
+    # 检查是否在Git仓库中
+    if not is_git_repo():
+        print_color(Color.RED, "错误: 当前目录不是Git仓库")
+        sys.exit(1)
+    
+    print_color(Color.YELLOW, "获取Git修改的文件列表...")
+    
+    # 获取修改的文件
+    all_files = get_git_modified_files()
+    
+    # 筛选C++文件
+    cpp_files = filter_cpp_files(all_files)
+    
+    if not cpp_files:
+        print_color(Color.YELLOW, "没有找到需要格式化的C++文件")
+        sys.exit(0)
+    
+    print_color(Color.GREEN, "找到以下文件需要格式化:")
+    for file in cpp_files:
+        print(f"  - {file}")
+    
+    print_color(Color.YELLOW, "开始格式化文件...")
+    
+    # 格式化文件
+    success_count, fail_count = format_files(cpp_files)
+    
+    print()
+    print_color(Color.GREEN, "格式化完成!")
+    print_color(Color.GREEN, f"  - 成功: {success_count} 个文件")
+    if fail_count > 0:
+        print_color(Color.RED, f"  - 失败: {fail_count} 个文件")
+    
+    print()
+    print_color(Color.YELLOW, "提示: 您可以使用 'git diff' 查看格式化后的变更")
+
+if __name__ == "__main__":
+    main()

example/01_gemm/CMakeLists.txt

@@ -38,6 +38,12 @@ add_example_dependencies(example_gemm_xdl example_gemm_xdl_fp16_fp8_streamk_v3)
 
 add_example_executable(example_gemm_xdl_bf16_v3 gemm_xdl_bf16_v3.cpp)
 add_example_dependencies(example_gemm_xdl example_gemm_xdl_bf16_v3)
+set(GEMM_OPTIONS)
+list(APPEND GEMM_OPTIONS "SHELL: -mllvm -greedy-reverse-local-assignment=1 -mllvm --slp-threshold=-16")
+list(APPEND GEMM_OPTIONS -v --save-temps -Wno-gnu-line-marker)
+target_compile_options(example_gemm_xdl_bf16_v3 PRIVATE ${GEMM_OPTIONS})
+target_compile_options(example_gemm_xdl_fp8_v3 PRIVATE ${GEMM_OPTIONS})
+
 
 list(APPEND gpu_list gfx942 gfx950)
 set(target 0)

example/01_gemm/gemm_xdl_fp8_v3.cpp

@@ -28,10 +28,10 @@ using DeviceGemmV2Instance =
         ADataType,   BDataType,  CDataType,  AccDataType,  CShuffleDataType, 
         PassThrough, PassThrough, PassThrough, GemmDefault, 
         256,
-        224, 256, 
+        256, 256, 
         128, 16, 16,
         16,   16,
-        7,    8, 
+        8,    8, 
         S<8, 32, 1>,  S<1, 0, 2>,  S<1, 0, 2>, 
         2, 16, 16, 0,
         S<8, 32, 1>,  S<1, 0, 2>,  S<1, 0, 2>, 

example/65_gemm_multiply_multiply/CMakeLists.txt

@@ -30,3 +30,10 @@ foreach(gpu IN LISTS GPU_TARGETS)
         set(target 1)
     endif()
 endforeach()
+
+set(GEMM_OPTIONS)
+list(APPEND GEMM_OPTIONS "SHELL: -mllvm -greedy-reverse-local-assignment=1 -mllvm --slp-threshold=-32")
+list(APPEND GEMM_OPTIONS -v --save-temps -Wno-gnu-line-marker)
+target_compile_options(example_gemm_multiply_multiply_xdl_fp8_bpreshuffle PRIVATE ${GEMM_OPTIONS})
+target_compile_options(example_moe_gemm1_xdl_fp8 PRIVATE ${GEMM_OPTIONS})
+target_compile_options(example_moe_gemm2_xdl_fp8 PRIVATE ${GEMM_OPTIONS})

example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8_bpreshuffle.cpp

@@ -1,5 +1,5 @@
 // SPDX-License-Identifier: MIT
-// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
 
 #include <iostream>
 #include <numeric>
@@ -139,7 +139,7 @@ using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultiD_Xdl_CShu
     // clang-format off
     <   Row, Col, DsLayout, ELayout, A0DataType, B0DataType, DsDataType, EDataType, AccDataType, CShuffleDataType,
         AElementOp,  BElementOp, CDEElementOp, GemmSpec, 256,
-        128,   128,    128,
+        256,   256,    128,
         16,   16,
         32,   32,
         4,    1,

example/65_gemm_multiply_multiply/moe_gemm1_xdl_fp8.cpp

@@ -158,11 +158,14 @@ using BElementOp = PassThrough;
 
 static constexpr auto GemmSpec         = ck::tensor_operation::device::GemmSpecialization::Default;
 static constexpr ck::index_t MPerBlock = 128;
-static constexpr ck::index_t MXDLPerWave = 4;
-static constexpr ck::index_t NXDLPerWave = 2;
-static constexpr ck::index_t BLOCKSIZE   = 256;
-static constexpr ck::index_t NPerBlock   = 64;
+static constexpr ck::index_t NPerBlock   = 256;
 static constexpr ck::index_t MNPerXDL    = 16;
+static constexpr ck::index_t MXDLPerWave = MPerBlock / (MNPerXDL * 1);
+static constexpr ck::index_t NXDLPerWave = NPerBlock / (MNPerXDL * 4);
+static constexpr ck::index_t CShuffleMXDLPerWave = MXDLPerWave;
+static constexpr ck::index_t CShuffleNXDLPerWave = NXDLPerWave;
+static constexpr ck::index_t BLOCKSIZE   = 256;
+
 static constexpr ck::index_t KPerBlock   = 128 / sizeof(A0DataType);
 static constexpr ck::index_t Nswizzle    = false;
 static constexpr ck::index_t AK1         = 16 / sizeof(A0DataType);
@@ -183,15 +186,15 @@ using DeviceOpInstance                   = ck::tensor_operation::device::DeviceM
                // mn_perxdl
                MNPerXDL,   MNPerXDL,
                // mn_xdlperwave 
-               MXDLPerWave,    NXDLPerWave,
+               MXDLPerWave,  NXDLPerWave,
                // a,b: loadtranfer cluster, cluster order, srcorder,VECDIM, srcpervec, dstpervec, lds_extra
                S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, AK1, AK1, 0,
                S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, BK1, BK1, 0,
                //    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
                //    MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
                 //  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
-                2,    2,   S<1, 32, 1, 8>, S<EVec, D0Vec, D1Vec>,
-               ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v1, ActOP, Nswizzle, true, MulRoutedWeight, true, int32_t, A0DataType>;
+                CShuffleMXDLPerWave,    CShuffleNXDLPerWave,   S<1, 32, 1, 8>, S<EVec, D0Vec, D1Vec, 1>,
+               ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v3, ActOP, Nswizzle, true, MulRoutedWeight, true, int32_t, A0DataType>;
 
 // clang-format on
 
@@ -205,9 +208,9 @@ int main(int argc, char* argv[])
     ck::index_t N               = 4096;
     ck::index_t K               = 6144;
     ck::index_t experts         = 8;
-    ck::index_t sorted_tile_num = 16;
-    ck::index_t valid_tile_num  = 13;
-    ck::index_t tokens          = 64;
+    ck::index_t sorted_tile_num = 133;
+    ck::index_t valid_tile_num  = 128;
+    ck::index_t tokens          = 8192;
     ck::index_t topk            = 2;
 
     if(argc == 1)
@@ -263,11 +266,12 @@ int main(int argc, char* argv[])
     Tensor<ck::index_t> sorted_token_ids(HostTensorDescriptor({sorted_size}, {1}));
     Tensor<ck::index_t> max_token_id(HostTensorDescriptor({1 + sorted_tile_num}));
     max_token_id.mData = {valid_size};
-    int eids[]         = {0, 0, 1, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 3, 3, 3};
+    // int eids[]         = {0, 0, 1, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 3, 3, 3};
     for(int i = 0; i < sorted_tile_num; i++)
     {
-        expert_ids.mData[i] = eids[i];
+        expert_ids.mData[i] = i / (valid_tile_num / experts);
     }
+
     int token_per_tile = (tokens * topk + valid_tile_num - 1) / valid_tile_num;
     int tokenid        = 0;
 
@@ -307,7 +311,7 @@ int main(int argc, char* argv[])
     case 0: break;
     case 1:
         a0_t_k.GenerateTensorValue(GeneratorTensor_3<A0DataType>{0.0, 1.0});
-        b0_e_n_k.GenerateTensorValue(GeneratorTensor_3<B0DataType>{-0.5, 0.5});
+        b0_e_n_k.GenerateTensorValue(GeneratorTensor_3<B0DataType>{-0.1, 0.1});
         d0_t_n.GenerateTensorValue(GeneratorTensor_3<D0DataType>{0.0, 1.0});
         d1_e_n.GenerateTensorValue(GeneratorTensor_3<D1DataType>{0.0, 1.0});
         d2_e_n.GenerateTensorValue(GeneratorTensor_3<D2DataType>{0.0, 1.0});

example/65_gemm_multiply_multiply/moe_gemm2_xdl_fp8.cpp

@@ -123,11 +123,11 @@ using BElementOp   = PassThrough;
 using CDEElementOp = MulABScaleExpertWeight;
 
 static constexpr auto GemmSpec         = ck::tensor_operation::device::GemmSpecialization::Default;
-static constexpr ck::index_t MPerBlock = 128;
+static constexpr ck::index_t MPerBlock = 256;
 static constexpr ck::index_t BLOCKSIZE = 256;
-static constexpr ck::index_t MXDLPerWave = 4;
+static constexpr ck::index_t MXDLPerWave = 16;
 static constexpr ck::index_t NXDLPerWave = 4;
-static constexpr ck::index_t NPerBlock   = 128;
+static constexpr ck::index_t NPerBlock   = 256;
 static constexpr ck::index_t MNPerXDL    = 16;
 static constexpr ck::index_t KPerBlock   = 128 / sizeof(A0DataType);
 
@@ -136,11 +136,12 @@ static constexpr ck::index_t CShuffleMLane = BLOCKSIZE / CShuffleNLane;
 static constexpr ck::index_t AK1           = 16 / sizeof(A0DataType);
 static constexpr ck::index_t BK1           = 16 / sizeof(B0DataType);
 static constexpr ck::index_t EVec          = 2;
-static constexpr ck::index_t D0Vec         = 1;
-static constexpr ck::index_t D1Vec         = 1;
-static constexpr ck::index_t D2Vec         = 1;
-static constexpr bool MulRoutedWeight      = true;
-using DeviceOpInstance                     = ck::tensor_operation::device::DeviceMoeGemm
+// TODO: Epilogue performance issue. AtomicAdd lose 15~20% performance compare with Set.
+static constexpr ck::index_t D0Vec    = 1;
+static constexpr ck::index_t D1Vec    = 1;
+static constexpr ck::index_t D2Vec    = 1;
+static constexpr bool MulRoutedWeight = true;
+using DeviceOpInstance                = ck::tensor_operation::device::DeviceMoeGemm
     // clang-format off
 ///######|  ALayout|  BLayout| DsLayout| ELayout|      AData|      BData|     DsData|     EData|     AccData|         CShuffle|           A|           B|          CDE|           GEMM| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
 ///######|         |         |         |        |       Type|       Type|       Type|      Type|        Type|         DataType| Elementwise| Elementwise|  Elementwise| Spacialization|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
@@ -164,7 +165,7 @@ using DeviceOpInstance                     = ck::tensor_operation::device::Devic
             //    S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0,
             //    S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0,
                S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, AK1, AK1, 0,
-               S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, AK1, AK1, 0,
+               S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, BK1, BK1, 0,
                //    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
                //    MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
                 //  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
@@ -186,8 +187,8 @@ int main(int argc, char* argv[])
     ck::index_t N               = 4096;
     ck::index_t K               = 4096;
     ck::index_t experts         = 8;
-    ck::index_t sorted_tile_num = 16;
-    ck::index_t valid_tile_num  = 13;
+    ck::index_t sorted_tile_num = 133;
+    ck::index_t valid_tile_num  = 128;
     ck::index_t sorted_size     = sorted_tile_num * MPerBlock;
     ck::index_t valid_size      = valid_tile_num * MPerBlock;
     ck::index_t tokens          = 128;
@@ -247,11 +248,11 @@ int main(int argc, char* argv[])
     Tensor<ck::index_t> max_token_id(HostTensorDescriptor({1}));
 
     max_token_id.mData = {valid_size, 0, 2, 3, 4, 6, 8, 10, 12, 13};
-    int eids[]         = {0, 0, 1, 2, 3, 3, 4, 4, 5, 5, 6, 7, 7, 3, 3, 3};
+    // int eids[]         = {0, 0, 1, 2, 3, 3, 4, 4, 5, 5, 6, 7, 7, 3, 3, 3};
 
     for(int i = 0; i < sorted_tile_num; i++)
     {
-        expert_ids.mData[i] = eids[i];
+        expert_ids.mData[i] = i / ((valid_tile_num + experts - 1) / experts);
     }
     if(tokens * topk > valid_size)
     {

example/67_gemm_microscaling/CMakeLists.txt

@@ -7,7 +7,10 @@ add_example_executable(example_gemm_mx_bf8 gemm_mx_bf8.cpp)
 add_example_dependencies(example_gemm_mx example_gemm_mx_bf8)
 
 add_example_executable(example_gemm_mx_fp8_bf8 gemm_mx_fp8_bf8.cpp)
-# add_example_dependencies(example_gemm_mx example_gemm_mx_fp8_bf8)
+# add_example_dependencies(example_gemm_mx example_gemm_mx_fp8_bf8) TOFO: Fix RRR
+
+add_example_executable(example_gemm_mx_fp8_bpreshuffle gemm_mx_fp8_bpreshuffle.cpp)
+# add_example_dependencies(example_gemm_mx example_gemm_mx_fp8_bpreshuffle) TODO: Fix
 
 add_example_executable(example_gemm_mx_fp4 gemm_mx_fp4.cpp)
 add_example_dependencies(example_gemm_mx example_gemm_mx_fp4)
@@ -15,6 +18,18 @@ add_example_dependencies(example_gemm_mx example_gemm_mx_fp4)
 add_example_executable(example_gemm_mx_fp4_bpreshuffle gemm_mx_fp4_bpreshuffle.cpp)
 add_example_dependencies(example_gemm_mx example_gemm_mx_fp4_bpreshuffle)
 
+add_example_executable(example_moe_gemm1_xdl_mx_fp4 moe_gemm1_xdl_mx_fp4.cpp)
+# add_example_dependencies(example_gemm_mx example_moe_gemm1_xdl_mx_fp4) TODO: Fix
+
+add_example_executable(example_moe_gemm1_xdl_mx_fp4_bns moe_gemm1_xdl_mx_fp4_bns.cpp)
+add_example_dependencies(example_gemm_mx example_moe_gemm1_xdl_mx_fp4_bns)
+
+add_example_executable(example_moe_gemm2_xdl_mx_fp4 moe_gemm2_xdl_mx_fp4.cpp)
+# add_example_dependencies(example_gemm_mx example_moe_gemm2_xdl_mx_fp4) TODO: Fix
+
+add_example_executable(example_moe_gemm2_xdl_mx_fp4_bns moe_gemm2_xdl_mx_fp4_bns.cpp)
+add_example_dependencies(example_gemm_mx example_moe_gemm2_xdl_mx_fp4_bns)
+
 set(FP4_MXGEMM_OPTIONS)
 list(APPEND FP4_MXGEMM_OPTIONS "SHELL: -mllvm -greedy-reverse-local-assignment=1 -mllvm --amdgpu-use-amdgpu-trackers=1")
 list(APPEND FP4_MXGEMM_OPTIONS -v --save-temps -Wno-gnu-line-marker -ftemplate-backtrace-limit=0)
@@ -24,4 +39,10 @@ target_compile_options(example_gemm_mx_fp4_bpreshuffle PRIVATE ${FP4_MXGEMM_OPTI
 set(FP8_MXGEMM_OPTIONS)
 list(APPEND FP8_MXGEMM_OPTIONS "SHELL: -mllvm -greedy-reverse-local-assignment=1 -mllvm --slp-threshold=-32")
 list(APPEND FP8_MXGEMM_OPTIONS -v --save-temps -Wno-gnu-line-marker -ftemplate-backtrace-limit=0)
-target_compile_options(example_gemm_mx_fp8 PRIVATE ${FP8_MXGEMM_OPTIONS})
+target_compile_options(example_moe_gemm1_xdl_mx_fp4 PRIVATE ${FP4_MXGEMM_OPTIONS})
+target_compile_options(example_moe_gemm2_xdl_mx_fp4 PRIVATE ${FP4_MXGEMM_OPTIONS})
+target_compile_options(example_moe_gemm1_xdl_mx_fp4_bns PRIVATE ${FP4_MXGEMM_OPTIONS})
+target_compile_options(example_moe_gemm2_xdl_mx_fp4_bns PRIVATE ${FP4_MXGEMM_OPTIONS})
+
+target_compile_options(example_gemm_mx_fp8 PRIVATE ${FP8_MXGEMM_OPTIONS})
+target_compile_options(example_gemm_mx_fp8_bpreshuffle PRIVATE ${FP8_MXGEMM_OPTIONS})

example/67_gemm_microscaling/gemm_mx_fp8_bpreshuffle.cpp

@@ -0,0 +1,359 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_v3_mx_b_preshuffle.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/utility/blkgemmpipe_scheduler.hpp"
+#include "ck/utility/data_type.hpp"
+#include "ck/utility/sequence.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_mx_gemm.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/fill.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F8        = ck::f8_t;
+using F16       = ck::half_t;
+using BF16      = ck::bhalf_t;
+using F32       = float;
+using XDataType = ck::e8m0_bexp_t;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using A0DataType       = F8;
+using A1DataType       = XDataType;
+using B0DataType       = F8;
+using B1DataType       = XDataType;
+using AccDataType      = F32;
+using DsDataType       = ck::Tuple<>;
+using CDataType        = BF16;
+using CShuffleDataType = CDataType;
+
+using A0Layout = Row;
+using B0Layout = Col;
+using CLayout  = Row;
+
+void preShuffleBuffer(const F8* src, F8* dst, int N, int K, int NXdl)
+{
+    int KPack = 16;
+    int NLane = NXdl;
+    int KLane = 64 / NLane;
+
+    int K0 = K / (KLane * KPack);
+    // K -> K0 KLane KPack
+    // N -> N0 NLane
+    // N, K -> N0 K0 KLane NLane KPack
+    int tempk;
+    for(int n = 0; n < N; ++n)
+    {
+        for(int k = 0; k < K; ++k)
+        {
+            int n0 = n / NLane;
+            int n1 = n % NLane;
+
+            int k0 = k / (KLane * KPack);
+            tempk  = k % (KLane * KPack);
+            int k1 = tempk / KPack;
+            int k2 = tempk % KPack;
+
+            int outputIndex = n0 * KPack * NLane * KLane * K0 + k0 * KPack * NLane * KLane +
+                              k1 * KPack * NLane + n1 * KPack + k2;
+
+            dst[outputIndex] = src[n * K + k];
+        }
+    }
+}
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using AElementOp = PassThrough; // elementwise transformation for A matrix
+using BElementOp = PassThrough; // elementwise transformation for B matrix
+using CElementOp = PassThrough; // elementwise transformation for C matrix
+
+constexpr ck::index_t ScaleBlockSize = 32; // scaling block size
+
+constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMX_Xdl_CShuffleV3_BPreShuffle<
+    A0Layout,    B0Layout,    CLayout,          
+    A0DataType,  A1DataType,  B0DataType,   B1DataType,   CDataType,    AccDataType,  CShuffleDataType, 
+    AElementOp, BElementOp, CElementOp,  GemmSpec,         
+    ScaleBlockSize,   256,   
+    128,  128,   128,        
+    16,    16,               
+    16,    16,               
+    8,     2,                
+    S<8, 32, 1>,   S<1, 0, 2>,   S<1, 0, 2>,   2,   16,   16,   0,            
+    S<8, 32, 1>,   S<1, 0, 2>,   S<1, 0, 2>,   2,   16,   16,   0,            
+    2,   1,   S<1, 32, 1, 8>,  8,                
+    ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v1, A0DataType, B0DataType>;
+// clang-format on
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+    bool flush_cache     = true;
+
+    // GEMM shape
+    ck::index_t M = 3840;
+    ck::index_t N = 4096;
+    ck::index_t K = 4096;
+
+    ck::index_t StrideA = K;
+    ck::index_t StrideB = K;
+    ck::index_t StrideC = N;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 8)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        M = std::stoi(argv[4]);
+        N = std::stoi(argv[5]);
+        K = std::stoi(argv[6]);
+
+        flush_cache = std::stoi(argv[7]);
+
+        StrideA = K;
+        StrideB = K;
+        StrideC = N;
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=no, 1=yes)\n");
+        printf("arg4 to 6: M, N, K\n");
+        printf("arg7: flush both I$ and L2$ (0=no, 1=yes)\n");
+        exit(0);
+    }
+
+    ck::index_t Scale_Stride_AM = (K + ScaleBlockSize - 1) / ScaleBlockSize;
+    ck::index_t Scale_Stride_BN = (K + ScaleBlockSize - 1) / ScaleBlockSize;
+
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
+            if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    Tensor<A0DataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, A0Layout{}));
+    Tensor<A1DataType> a_m_k_scale(f_host_tensor_descriptor(
+        M, (K + ScaleBlockSize - 1) / ScaleBlockSize, Scale_Stride_AM, A0Layout{}));
+    Tensor<B0DataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, B0Layout{}));
+    Tensor<B0DataType> b_preshuffled(f_host_tensor_descriptor(K, N, StrideB, B0Layout{}));
+    Tensor<B1DataType> b_k_n_scale(f_host_tensor_descriptor(
+        (K + ScaleBlockSize - 1) / ScaleBlockSize, N, Scale_Stride_BN, B0Layout{}));
+    Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+    Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+
+    std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
+    std::cout << "a_m_k_scale: " << a_m_k_scale.mDesc << std::endl;
+    std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
+    std::cout << "b_k_n_scale: " << b_k_n_scale.mDesc << std::endl;
+    std::cout << "e_m_n: " << c_m_n_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-2, 2});
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
+        a_m_k_scale.GenerateTensorValue(GeneratorTensor_3<A1DataType>{0, 1.0});
+        b_k_n_scale.GenerateTensorValue(GeneratorTensor_3<B1DataType>{0, 1.0});
+        break;
+    case 2:
+        a_m_k.GenerateTensorValue(GeneratorTensor_1<A0DataType>{});
+        b_k_n.GenerateTensorValue(GeneratorTensor_1<B0DataType>{});
+        a_m_k_scale.GenerateTensorValue(GeneratorTensor_1<A1DataType>{});
+        b_k_n_scale.GenerateTensorValue(GeneratorTensor_1<B1DataType>{});
+        break;
+    case 3:
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-2, 2});
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
+        a_m_k_scale.GenerateTensorValue(GeneratorTensor_1<A1DataType>{});
+        b_k_n_scale.GenerateTensorValue(GeneratorTensor_1<B1DataType>{});
+        break;
+    case 4:
+        a_m_k.GenerateTensorValue(GeneratorTensor_1<A0DataType>{});
+        b_k_n.GenerateTensorValue(GeneratorTensor_1<B0DataType>{});
+        a_m_k_scale.GenerateTensorValue(GeneratorTensor_3<A1DataType>{0, 1.0});
+        b_k_n_scale.GenerateTensorValue(GeneratorTensor_3<B1DataType>{0, 1.0});
+        break;
+    case 5:
+        a_m_k.GenerateTensorValue(GeneratorTensor_1<A0DataType>{});
+        b_k_n.GenerateTensorValue(GeneratorTensor_1<B0DataType>{});
+        a_m_k_scale.GenerateTensorValue(GeneratorTensor_1<A1DataType>{});
+        b_k_n_scale.GenerateTensorValue(GeneratorTensor_3<B1DataType>{0, 1.0});
+        break;
+    case 6:
+        a_m_k.GenerateTensorValue(GeneratorTensor_1<A0DataType>{});
+        b_k_n.GenerateTensorValue(GeneratorTensor_1<B0DataType>{});
+        a_m_k_scale.GenerateTensorValue(GeneratorTensor_3<A1DataType>{0, 1.0});
+        b_k_n_scale.GenerateTensorValue(GeneratorTensor_1<B1DataType>{});
+        break;
+    default:
+        a_m_k.GenerateTensorValue(GeneratorTensor_3<A0DataType>{-0.5, 0.5});
+        b_k_n.GenerateTensorValue(GeneratorTensor_3<B0DataType>{-0.5, 0.5});
+        a_m_k_scale.GenerateTensorValue(GeneratorTensor_3<A1DataType>{0, 1.0});
+        b_k_n_scale.GenerateTensorValue(GeneratorTensor_3<B1DataType>{0, 1.0});
+    }
+
+    DeviceMem a_device_buf(sizeof(A0DataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem a_scale_device_buf(sizeof(A1DataType) * a_m_k_scale.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(B0DataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem b_scale_device_buf(sizeof(B1DataType) * b_k_n_scale.mDesc.GetElementSpaceSize());
+    DeviceMem c_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_m_k.mData.data());
+    a_scale_device_buf.ToDevice(a_m_k_scale.mData.data());
+    b_scale_device_buf.ToDevice(b_k_n_scale.mData.data());
+
+#if 1
+    printf("print a_m_k_scale:\n");
+    for(int m = 0; m < M; ++m)
+    {
+        for(int k = 0; k < (K + ScaleBlockSize - 1) / ScaleBlockSize; ++k)
+        {
+            printf("%f ", ck::type_convert<float>(a_m_k_scale(m, k)));
+        }
+        printf("\n");
+    }
+#endif
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CElementOp{};
+
+    // do GEMM
+    auto device_op = DeviceOpInstance{};
+    int NPerXdl    = device_op.GetPreShuffleParameters();
+
+    preShuffleBuffer(b_k_n.mData.data(), b_preshuffled.mData.data(), N, K, NPerXdl);
+    b_device_buf.ToDevice(b_preshuffled.mData.data());
+
+    auto invoker = device_op.MakeInvoker();
+    auto argument =
+        device_op.MakeArgument(static_cast<A0DataType*>(a_device_buf.GetDeviceBuffer()),
+                               static_cast<XDataType*>(a_scale_device_buf.GetDeviceBuffer()),
+                               static_cast<B0DataType*>(b_device_buf.GetDeviceBuffer()),
+                               static_cast<XDataType*>(b_scale_device_buf.GetDeviceBuffer()),
+                               static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
+                               M,
+                               N,
+                               K,
+                               StrideA,
+                               Scale_Stride_AM,
+                               StrideB,
+                               Scale_Stride_BN,
+                               StrideC,
+                               1, // KBatch
+                               a_element_op,
+                               b_element_op,
+                               cde_element_op);
+
+    if(!device_op.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_gemm with the specified compilation parameters does "
+            "not support this GEMM problem");
+    }
+
+    std::size_t flop = std::size_t(2) * M * N * K + std::size_t(2) * M * N * K / ScaleBlockSize;
+    std::size_t num_btype = sizeof(A0DataType) * M * K + sizeof(B0DataType) * K * N +
+                            sizeof(CDataType) * M * N +
+                            sizeof(XDataType) * (M * K + K * N) / ScaleBlockSize;
+
+    float ave_time = .0;
+
+    if(flush_cache)
+    {
+        int rotating_buf = (512 * 1024 * 1024 + num_btype - 1) / num_btype;
+
+        ave_time = invoker.Run(argument,
+                               StreamConfig{nullptr, time_kernel, 0, 50, 100, true, rotating_buf});
+    }
+    else
+    {
+        ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel, 0, 50, 100});
+    }
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << device_op.GetTypeString() << std::endl;
+
+    if(do_verification)
+    {
+        using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceMXGemm<A0DataType,
+                                                                                  B0DataType,
+                                                                                  CDataType,
+                                                                                  AccDataType,
+                                                                                  XDataType,
+                                                                                  PassThrough,
+                                                                                  PassThrough,
+                                                                                  PassThrough,
+                                                                                  float,
+                                                                                  float>;
+        auto ref_gemm               = ReferenceGemmInstance{};
+        auto ref_invoker            = ref_gemm.MakeInvoker();
+
+        auto ref_argument = ref_gemm.MakeArgument(a_m_k,
+                                                  a_m_k_scale,
+                                                  b_k_n,
+                                                  b_k_n_scale,
+                                                  c_m_n_host_result,
+                                                  PassThrough{},
+                                                  PassThrough{},
+                                                  PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        c_device_buf.FromDevice(c_m_n_device_result.mData.data());
+
+        return ck::utils::check_err(
+                   c_m_n_device_result, c_m_n_host_result, "Error: Incorrect results!", 5e-2, 5e-2)
+                   ? 0
+                   : 1;
+    }
+
+    return 0;
+}

example/67_gemm_microscaling/moe_gemm1_xdl_mx_fp4.cpp

@@ -0,0 +1,528 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_moe_mx_gemm.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp"
+
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_moe_mx_gemm1.hpp"
+#include "ck/library/utility/check_err.hpp"
+
+#include "ck/utility/blkgemmpipe_scheduler.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F4        = ck::f4x2_pk_t;
+using F16       = ck::half_t;
+using BF16      = ck::bhalf_t;
+using F32       = float;
+using XDataType = ck::e8m0_bexp_t;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using A0DataType       = F4;
+using A1DataType       = XDataType;
+using B0DataType       = F4;
+using B1DataType       = XDataType;
+using EDataType        = F16;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using D0DataType       = F32;
+using D1DataType       = F32;
+using D2DataType       = F32;
+using DsDataType       = ck::Tuple<D0DataType, D1DataType, D2DataType>;
+
+using A0Layout = Row;
+using B0Layout = Col;
+using ELayout  = Row;
+using D0Layout = Row;
+using D1Layout = Col;
+using D2Layout = ELayout;
+using DsLayout = ck::Tuple<D0Layout, D1Layout, D2Layout>;
+
+// d0: ascale, d1: bscale, d2:expert weight
+struct MulABScaleExpertWeight
+{
+    template <typename E, typename C, typename D0, typename D1, typename D2>
+    __host__ __device__ constexpr void
+    operator()(E& e, const C& c, const D0& d0, const D1& d1, const D2& d2) const;
+    // for real kernel use
+    template <>
+    __host__ __device__ constexpr void operator()<EDataType, float, float, float, float>(
+        EDataType& e, const float& c, const float& d0, const float& d1, const float& d2) const
+    {
+        (void)d0;
+        (void)d1;
+        (void)d2;
+
+        e = ck::type_convert<EDataType>(c);
+    }
+    // for reference cpu
+    template <>
+    __host__ __device__ constexpr void operator()<float, float, float, float, float>(
+        float& e, const float& c, const float& d0, const float& d1, const float& d2) const
+    {
+        // for reference cpu
+        (void)d0;
+        (void)d1;
+        (void)d2;
+        e = ck::type_convert<EDataType>(c);
+    }
+};
+
+using CDEElementOp = MulABScaleExpertWeight; // combine MulRoutedWeight = true
+
+// using CDEElementOp = MulABScale; // combine MulRoutedWeight = true
+
+void preShuffleBuffer(const F4* src, F4* dst, int N, int K, int NXdl)
+{
+    int KPack = 32;
+    int NLane = NXdl;
+    int KLane = 64 / NLane;
+
+    int K0 = K / (KLane * KPack);
+    // K -> K0 KLane KPack
+    // N -> N0 NLane
+    // N, K -> N0 K0 KLane NLane KPack
+    int tempk;
+    for(int n = 0; n < N; ++n)
+    {
+        for(int k = 0; k < K; ++k)
+        {
+            int n0 = n / NLane;
+            int n1 = n % NLane;
+
+            int k0 = k / (KLane * KPack);
+            tempk  = k % (KLane * KPack);
+            int k1 = tempk / KPack;
+            int k2 = tempk % KPack;
+
+            int outputIndex = n0 * KPack * NLane * KLane * K0 + k0 * KPack * NLane * KLane +
+                              k1 * KPack * NLane + n1 * KPack + k2;
+
+            dst[outputIndex / 2] = src[(n * K + k) / 2];
+        }
+    }
+}
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = MulABScaleExpertWeight;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::Default;
+
+constexpr ck::index_t ScaleBlockSize  = 32; // scaling block size
+static constexpr ck::index_t Nswizzle = false;
+static constexpr ck::index_t ActOP    = 0; // 0: gelu_and_mul, 1: silu_and_mul
+static constexpr bool MulRoutedWeight = false;
+
+#if 0
+static constexpr ck::index_t MPerBlock = 128;
+static constexpr ck::index_t BLOCKSIZE = 256;
+static constexpr ck::index_t MXDLPerWave   = 8;
+static constexpr ck::index_t NXDLPerWave   = 2;
+static constexpr ck::index_t NPerBlock     = 128;
+static constexpr ck::index_t MNPerXDL      = 16;
+static constexpr ck::index_t KPerBlock     = 128 / sizeof(A0DataType);
+static constexpr ck::index_t CShuffleNLane = 32;
+static constexpr ck::index_t CShuffleMLane = BLOCKSIZE / CShuffleNLane;
+static constexpr ck::index_t AK1           = 16 / sizeof(A0DataType);
+static constexpr ck::index_t BK1           = 32 / sizeof(B0DataType);
+static constexpr ck::index_t EVec          = 2;
+static constexpr ck::index_t D0Vec         = 1;
+static constexpr ck::index_t D1Vec         = 1;
+static constexpr ck::index_t D2Vec         = 1;
+using DeviceOpInstance                     = ck::tensor_operation::device::DeviceMoeGemmMX
+    // clang-format off
+        <      Row, Col, DsLayout, ELayout, A0DataType, B0DataType, DsDataType, EDataType, AccDataType, CShuffleDataType,
+               AElementOp,  BElementOp, CDEElementOp,       GemmSpec,   
+               ScaleBlockSize, BLOCKSIZE,   
+               MPerBlock,   NPerBlock,    KPerBlock,
+               AK1,   BK1,
+               MNPerXDL,   MNPerXDL,
+               MXDLPerWave,    NXDLPerWave,
+               S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, AK1, AK1, 0,
+               S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, BK1, BK1, 0,
+               2,    2,   S<1, CShuffleMLane, 1, CShuffleNLane>, S<EVec, D0Vec, D1Vec, D2Vec>,
+               ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v1, ActOP, Nswizzle, false, MulRoutedWeight, ck::index_t, A0DataType>;
+// clang-format on
+
+#else
+static constexpr ck::index_t MPerBlock = 128;
+
+// clang-format off
+using DeviceOpInstance = ck::tensor_operation::device::DeviceMoeGemmMX<
+            A0Layout,    B0Layout,    DsLayout,    ELayout, 
+            A0DataType,  A1DataType,  B0DataType,  B1DataType,  DsDataType, EDataType, AccDataType, CShuffleDataType,
+            AElementOp,  BElementOp, CDEElementOp, GemmSpec,   
+            ScaleBlockSize,  256, MPerBlock,  128,  128,
+            32,   32,
+            16,   16,
+            8,    2,
+            S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 32, 32, 0,
+            S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 32, 32, 0,
+            2,    2,   S<1, 32, 1, 8>, S<2, 1, 1, 1>,
+            ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v3, ActOP, Nswizzle, true, MulRoutedWeight, ck::index_t, A0DataType>;
+#endif
+// clang-format on
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = true;
+
+    // per expert:
+    // GEMM shape
+    ck::index_t N               = 6144;
+    ck::index_t K               = 4096;
+    ck::index_t experts         = 8;
+    ck::index_t sorted_tile_num = 16;
+    ck::index_t valid_tile_num  = 13;
+    ck::index_t sorted_size     = sorted_tile_num * MPerBlock;
+    ck::index_t valid_size      = valid_tile_num * MPerBlock;
+    ck::index_t tokens          = 832;
+    ck::index_t topk            = 2;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 4)
+    {
+        // use default case
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 7)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+        N               = std::stoi(argv[4]);
+        K               = std::stoi(argv[5]);
+        tokens          = std::stoi(argv[6]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=no, 1=yes)\n");
+        printf("arg4 to 6: N, K, tokens\n");
+        exit(0);
+    }
+
+    if(tokens * topk > valid_size)
+    {
+        printf("err config, tokens * topk > valid_size\n");
+        exit(-1);
+    }
+    ck::index_t StrideA              = K;
+    ck::index_t StrideB              = K;
+    ck::index_t StrideE              = N;
+    ck::index_t Scale_Stride_AM      = (K + ScaleBlockSize - 1) / ScaleBlockSize;
+    ck::index_t Scale_Stride_BN      = (K + ScaleBlockSize - 1) / ScaleBlockSize;
+    constexpr ck::index_t NumDTensor = DsDataType::Size();
+    constexpr auto StrideDs          = std::array<ck::index_t, NumDTensor>{0, 0, 0};
+
+    ck::index_t KBatch = 1;
+
+    Tensor<ck::index_t> expert_ids(HostTensorDescriptor({sorted_tile_num}, {1}));
+    Tensor<ck::index_t> sorted_token_ids(HostTensorDescriptor({sorted_size}, {1}));
+    Tensor<ck::index_t> max_token_id(HostTensorDescriptor({1 + sorted_tile_num}));
+    max_token_id.mData = {valid_size};
+    // int eids[]         = {0, 0, 1, 2, 3, 3, 4, 4, 5, 5, 6, 7, 7, 3, 3, 3};
+    for(int i = 0; i < sorted_tile_num; i++)
+    {
+        expert_ids.mData[i] = i / ck::math::integer_divide_ceil(valid_tile_num, experts);
+    }
+    int token_per_tile = (tokens * topk + valid_tile_num - 1) / valid_tile_num;
+    int tokenid        = 0;
+    for(int i = 0; i < sorted_size; i++)
+    {
+        int tile_off = i % MPerBlock;
+        if(tile_off < token_per_tile)
+        {
+            sorted_token_ids.mData[i] = (tokenid % tokens) | ((tokenid / tokens) << 24);
+            tokenid++;
+        }
+        else
+        {
+            sorted_token_ids.mData[i] = tokens;
+        }
+    }
+
+    Tensor<A0DataType> a0_t_k(HostTensorDescriptor({tokens, K}, {K, 1}));
+    Tensor<A1DataType> a1_t_k(HostTensorDescriptor(
+        {tokens, (K + ScaleBlockSize - 1) / ScaleBlockSize}, {Scale_Stride_AM, 1}));
+    Tensor<B0DataType> b0_e_n_k(HostTensorDescriptor({experts, K, N * 2}, {N * 2 * K, 1, K}));
+    Tensor<B1DataType> b1_e_n_k(
+        HostTensorDescriptor({experts, (K + ScaleBlockSize - 1) / ScaleBlockSize, N * 2},
+                             {(N * 2 * Scale_Stride_BN), 1, Scale_Stride_BN}));
+    Tensor<B0DataType> b0_preshuffled(HostTensorDescriptor({experts, K, N * 2}, {N * 2 * K, 1, K}));
+    Tensor<D0DataType> d0_t_n(HostTensorDescriptor({tokens, N}, {StrideDs[0], 0}));
+    Tensor<D1DataType> d1_e_n(
+        HostTensorDescriptor({experts, N * 2}, {StrideDs[1] * N * 2, StrideDs[1]}));
+    Tensor<D2DataType> d2_e_n(HostTensorDescriptor({sorted_size, N}, {1, 0}));
+    Tensor<EDataType> e_t_n_host_result(HostTensorDescriptor({tokens, topk, N}, {topk * N, N, 1}));
+    Tensor<EDataType> e_t_n_device_result(
+        HostTensorDescriptor({tokens, topk, N}, {topk * N, N, 1}));
+
+    std::cout << "a0_t_k: " << a0_t_k.mDesc << std::endl;
+    std::cout << "a1_t_k: " << a1_t_k.mDesc << std::endl;
+    std::cout << "b0_e_n_k: " << b0_e_n_k.mDesc << std::endl;
+    std::cout << "b1_e_n_k: " << b1_e_n_k.mDesc << std::endl;
+    std::cout << "d2_e_n: " << d2_e_n.mDesc << std::endl;
+    std::cout << "d1_e_n: " << d1_e_n.mDesc << std::endl;
+    std::cout << "d0_t_n: " << d0_t_n.mDesc << std::endl;
+    std::cout << "e_t_n: " << e_t_n_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a0_t_k.GenerateTensorValue(GeneratorTensor_3<A0DataType>{-0.5, 0.5});
+        b0_e_n_k.GenerateTensorValue(GeneratorTensor_3<B0DataType>{-0.5, 0.5});
+        a1_t_k.GenerateTensorValue(GeneratorTensor_3<A1DataType>{1.0, 1.0});
+        b1_e_n_k.GenerateTensorValue(GeneratorTensor_3<B1DataType>{1.0, 1.0});
+        d2_e_n.GenerateTensorValue(GeneratorTensor_3<D2DataType>{1.0, 1.0});
+        break;
+    case 2:
+        a0_t_k.GenerateTensorValue(GeneratorTensor_1<A0DataType>{});
+        b0_e_n_k.GenerateTensorValue(GeneratorTensor_1<B0DataType>{});
+        a1_t_k.GenerateTensorValue(GeneratorTensor_1<A1DataType>{});
+        b1_e_n_k.GenerateTensorValue(GeneratorTensor_1<B1DataType>{});
+        d2_e_n.GenerateTensorValue(GeneratorTensor_3<D2DataType>{});
+        break;
+    default:
+        a0_t_k.GenerateTensorValue(GeneratorTensor_3<A0DataType>{0.0, 1.0});
+        b0_e_n_k.GenerateTensorValue(GeneratorTensor_3<B0DataType>{-0.5, 0.5});
+        a1_t_k.GenerateTensorValue(GeneratorTensor_3<A1DataType>{0.0, 1.0});
+        b1_e_n_k.GenerateTensorValue(GeneratorTensor_3<B1DataType>{0.0, 1.0});
+        d2_e_n.GenerateTensorValue(GeneratorTensor_3<D2DataType>{0.0, 1.0});
+    }
+
+    DeviceMem sorted_token_ids_dev(sizeof(ck::index_t) *
+                                   sorted_token_ids.mDesc.GetElementSpaceSize());
+    DeviceMem expert_ids_dev(sizeof(ck::index_t) * expert_ids.mDesc.GetElementSpaceSize());
+    DeviceMem max_token_id_dev(sizeof(ck::index_t) * max_token_id.mDesc.GetElementSpaceSize());
+    DeviceMem a0_device_buf(sizeof(A0DataType) * a0_t_k.mDesc.GetElementSpaceSize() / 2);
+    DeviceMem a1_device_buf(sizeof(A1DataType) * a1_t_k.mDesc.GetElementSpaceSize());
+    DeviceMem b0_device_buf(sizeof(B0DataType) * b0_e_n_k.mDesc.GetElementSpaceSize() / 2);
+    DeviceMem b1_device_buf(sizeof(B1DataType) * b1_e_n_k.mDesc.GetElementSpaceSize());
+    DeviceMem d0_device_buf(sizeof(D0DataType) * d0_t_n.mDesc.GetElementSpaceSize());
+    DeviceMem d1_device_buf(sizeof(D1DataType) * d1_e_n.mDesc.GetElementSpaceSize());
+    DeviceMem d2_device_buf(sizeof(D2DataType) * d2_e_n.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) * e_t_n_device_result.mDesc.GetElementSpaceSize());
+
+    sorted_token_ids_dev.ToDevice(sorted_token_ids.mData.data());
+    expert_ids_dev.ToDevice(expert_ids.mData.data());
+    max_token_id_dev.ToDevice(max_token_id.mData.data());
+    a0_device_buf.ToDevice(a0_t_k.mData.data());
+    a1_device_buf.ToDevice(a1_t_k.mData.data());
+    b1_device_buf.ToDevice(b1_e_n_k.mData.data());
+    d0_device_buf.ToDevice(d0_t_n.mData.data());
+    d1_device_buf.ToDevice(d1_e_n.mData.data());
+    d2_device_buf.ToDevice(d2_e_n.mData.data());
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{};
+
+    // do GEMM
+    auto device_op = DeviceOpInstance{};
+
+#if 1
+    preShuffleBuffer(b0_e_n_k.mData.data(),
+                     b0_preshuffled.mData.data(),
+                     N * 2 * experts,
+                     K,
+                     device_op.GetPreShuffleParameters());
+#else
+    // weight pre-shuffle
+    int KPack = 32; // int4 -> 32, fp8 -> 16, fp16 -> 8
+    int NLane = device_op.GetPreShuffleParameters();
+    int KLane = 64 / NLane;
+
+    int K0 = K / (KLane * KPack);
+    // K -> K0 KLane KPack
+    // N -> N0 NLane
+    // N, K -> N0 K0 KLane NLane KPack
+    int tempk;
+    for(int e = 0; e < experts; ++e)
+    {
+        for(int n = 0; n < N; ++n)
+        {
+            for(int k = 0; k < K; ++k)
+            {
+                int n0 = n / NLane;
+                int n1 = n % NLane;
+
+                int k0 = k / (KLane * KPack);
+                tempk  = k % (KLane * KPack);
+                int k1 = tempk / KPack;
+                int k2 = tempk % KPack;
+
+                int outputIndex = n0 * KPack * NLane * KLane * K0 + k0 * KPack * NLane * KLane +
+                                  k1 * KPack * NLane + n1 * KPack + k2;
+
+                b0_preshuffled(e, outputIndex % K, outputIndex / K) = b0_e_n_k(e, k, n);
+            }
+        }
+    }
+#endif
+
+    b0_device_buf.ToDevice(b0_preshuffled.mData.data());
+
+    auto invoker  = device_op.MakeInvoker();
+    auto argument = device_op.MakeArgument(
+        sorted_token_ids_dev.GetDeviceBuffer(),
+        expert_ids_dev.GetDeviceBuffer(),
+        max_token_id_dev.GetDeviceBuffer(),
+        a0_device_buf.GetDeviceBuffer(),
+        a1_device_buf.GetDeviceBuffer(),
+        b0_device_buf.GetDeviceBuffer(),
+        b1_device_buf.GetDeviceBuffer(),
+        std::array<const void*, NumDTensor>{nullptr, nullptr, d2_device_buf.GetDeviceBuffer()},
+        e_device_buf.GetDeviceBuffer(),
+        tokens,
+        topk,
+        sorted_size,
+        N,
+        K,
+        StrideA,
+        Scale_Stride_AM,
+        StrideB,
+        Scale_Stride_BN,
+        StrideDs,
+        StrideE,
+        KBatch,
+        a_element_op,
+        b_element_op,
+        cde_element_op);
+
+    if(!device_op.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_gemm with the specified compilation parameters does "
+            "not support this GEMM problem");
+    }
+
+    if(!(ck::get_device_name() == "gfx942" || ck::get_device_name() == "gfx950"))
+    {
+        std::cout << "This kernel support gfx942 and gfx950 only" << std::endl;
+    }
+
+    if(time_kernel)
+    {
+        float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+        std::size_t flop =
+            // FMA * tokens * N * (Gate+Up) * topk * K +
+            // FMA * tokens * N * (Gate+Up) * topk * (K/BlockScale)
+            std::size_t(2) * tokens * N * 2 * topk * K +
+            std::size_t(2) * tokens * N * 2 * topk * K / ScaleBlockSize;
+
+        std::size_t num_btype = sizeof(A0DataType) / 2 * tokens * topk * K +
+                                sizeof(B0DataType) / 2 * K * N * 2 * experts +
+                                sizeof(XDataType) * tokens * topk * K / ScaleBlockSize +
+                                sizeof(XDataType) * K / ScaleBlockSize * N * 2 * experts +
+                                sizeof(EDataType) * tokens * topk * N;
+
+        float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+        float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+        std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
+                  << " GB/s\n"
+                  << device_op.GetTypeString() << std::endl;
+    }
+
+    if(do_verification)
+    {
+        invoker.Run(argument, StreamConfig{nullptr, false, 0, 0, 1});
+
+        e_device_buf.FromDevice(e_t_n_device_result.mData.data());
+
+        Tensor<CShuffleDataType> c_t_k_n({tokens, topk, N}, {topk * N, N, 1});
+
+        using ReferenceGemmInstance =
+            ck::tensor_operation::host::ReferenceMoeMXGemm1<A0DataType,
+                                                            A1DataType,
+                                                            B0DataType,
+                                                            B1DataType,
+                                                            CShuffleDataType,
+                                                            D2DataType,
+                                                            AccDataType,
+                                                            PassThrough,
+                                                            PassThrough,
+                                                            PassThrough,
+                                                            ActOP,
+                                                            MulRoutedWeight>;
+        auto ref_moe_gemm = ReferenceGemmInstance{};
+        auto ref_invoker  = ref_moe_gemm.MakeInvoker();
+
+        auto ref_argument = ref_moe_gemm.MakeArgument(sorted_token_ids,
+                                                      expert_ids,
+                                                      max_token_id,
+                                                      MPerBlock,
+                                                      a0_t_k,
+                                                      a1_t_k,
+                                                      b0_e_n_k,
+                                                      b1_e_n_k,
+                                                      d2_e_n,
+                                                      c_t_k_n,
+                                                      PassThrough{},
+                                                      PassThrough{},
+                                                      PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+        for(int m = 0; m < valid_size; ++m)
+        {
+
+            const int fuse_t  = sorted_token_ids.mData[m];
+            const int t       = fuse_t & 0xffffff;
+            const int topk_id = (fuse_t & 0xff000000) >> 24;
+
+            if(t >= tokens)
+            {
+                continue;
+            }
+            for(int n = 0; n < N; ++n)
+            {
+                e_t_n_host_result(t, topk_id, n) =
+                    ck::type_convert<EDataType>(c_t_k_n(t, topk_id, n));
+            }
+        }
+
+        e_device_buf.FromDevice(e_t_n_device_result.mData.data());
+        auto status =
+            ck::utils::check_err(
+                e_t_n_device_result, e_t_n_host_result, "Error: Incorrect results!", 1e-3, 5e-1)
+                ? 0
+                : 1;
+        if(status == 0)
+        {
+            printf("Validation Pass.\n");
+        }
+        return status;
+    }
+
+    return 0;
+}

example/67_gemm_microscaling/moe_gemm1_xdl_mx_fp4_bns.cpp

@@ -0,0 +1,691 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_moe_mx_gemm_bns.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp"
+
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_moe_mx_gemm1.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/fill.hpp"
+#include "ck/utility/blkgemmpipe_scheduler.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F4              = ck::f4x2_pk_t;
+using F16             = ck::half_t;
+using BF16            = ck::bhalf_t;
+using F32             = float;
+using XDataType       = ck::e8m0_bexp_t;
+using XPackedDataType = int32_t; // 4 packed e8m0_bexp_t
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using A0DataType       = F4;
+using A1DataType       = XPackedDataType;
+using B0DataType       = F4;
+using B1DataType       = XPackedDataType;
+using EDataType        = F16;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using D0DataType       = F32;
+using D1DataType       = F32;
+using D2DataType       = F32;
+using DsDataType       = ck::Tuple<D0DataType, D1DataType, D2DataType>;
+
+using A0Layout = Row;
+using B0Layout = Col;
+using ELayout  = Row;
+using D0Layout = Row;
+using D1Layout = Col;
+using D2Layout = ELayout;
+using DsLayout = ck::Tuple<D0Layout, D1Layout, D2Layout>;
+
+// d0: ascale, d1: bscale, d2:expert weight
+struct MulABScaleExpertWeight
+{
+    template <typename E, typename C, typename D0, typename D1, typename D2>
+    __host__ __device__ constexpr void
+    operator()(E& e, const C& c, const D0& d0, const D1& d1, const D2& d2) const;
+    // for real kernel use
+    template <>
+    __host__ __device__ constexpr void operator()<EDataType, float, float, float, float>(
+        EDataType& e, const float& c, const float& d0, const float& d1, const float& d2) const
+    {
+        (void)d0;
+        (void)d1;
+        (void)d2;
+
+        e = ck::type_convert<EDataType>(c);
+    }
+    // for reference cpu
+    template <>
+    __host__ __device__ constexpr void operator()<float, float, float, float, float>(
+        float& e, const float& c, const float& d0, const float& d1, const float& d2) const
+    {
+        // for reference cpu
+        (void)d0;
+        (void)d1;
+        (void)d2;
+        e = ck::type_convert<EDataType>(c);
+    }
+};
+
+using CDEElementOp = MulABScaleExpertWeight;
+
+// A, B Scale preshuffle
+template <bool KLast>
+void preShuffleScaleBuffer(ck::e8m0_bexp_t* src, ck::e8m0_bexp_t* dst, int MN, int K)
+{
+    int MNXdlPack = 2;
+    int KXdlPack  = 2;
+
+    int XdlMNThread = 16;
+    int XdlKThread  = 64 / XdlMNThread;
+
+    int K0 = K / KXdlPack / XdlKThread; // KRepeat
+
+    // The 4 16x128 building blocks will be packed into 1 32x256 for F4
+    // The 8 16x16x128 mfma will be packed into 1 32x32x256 for F4
+
+    // unfold the MN32xK(256/32) scale buffer
+    //    4            16             2           2
+    // To XdlKThread-> XdlMNThread -> KXdlPack -> MNXdlPack
+    // Then, MNRepeat->KRepeat
+
+    for(int n = 0; n < MN; ++n)
+    {
+        for(int k = 0; k < K; ++k)
+        {
+            int n0    = n / (XdlMNThread * MNXdlPack); // i MNRepeat
+            int tempn = n % (XdlMNThread * MNXdlPack);
+            int n1    = tempn % XdlMNThread; // i XdlMNThread
+            int n2    = tempn / XdlMNThread; // i MNXdlPack
+
+            int k0    = k / (XdlKThread * KXdlPack); // i KRepeat
+            int tempk = k % (XdlKThread * KXdlPack);
+            int k1    = tempk % XdlKThread; // i XdlKThread
+            int k2    = tempk / XdlKThread; // i KXdlPack
+
+            int outputIndex = n0 * MNXdlPack * KXdlPack * XdlMNThread * XdlKThread * K0 +
+                              k0 * MNXdlPack * KXdlPack * XdlMNThread * XdlKThread +
+                              k1 * MNXdlPack * KXdlPack * XdlMNThread + n1 * MNXdlPack * KXdlPack +
+                              k2 * MNXdlPack + n2;
+            // src[n * K + k] = ck::type_convert<ck::e8m0_bexp_t>(static_cast<float>(powf(2.0f, n2 +
+            // k2 * MNXdlPack)));
+            if constexpr(KLast)
+                dst[outputIndex] = src[n * K + k];
+            else
+                dst[outputIndex] = src[k * MN + n];
+        }
+    }
+}
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = MulABScaleExpertWeight;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::Default;
+
+constexpr ck::index_t DataPackedSize   = 2;                    // Packed representation of data
+constexpr ck::index_t ScaleBlockSize   = 32;                   // scaling block size
+constexpr ck::index_t KPerBlock        = 256 / DataPackedSize; // 256 f4 = 128 fp4x2
+static constexpr ck::index_t Nswizzle  = false;
+static constexpr ck::index_t ActOP     = 0; // 0: gelu_and_mul, 1: silu_and_mul
+static constexpr ck::index_t MPerBlock = 128;
+static constexpr ck::index_t NPerBlock = 64;
+static constexpr ck::index_t BlockSize = 256;
+static constexpr bool MulRoutedWeight  = true;
+
+// clang-format off
+using DeviceOpInstance                     = ck::tensor_operation::device::DeviceMoeGemmMXBNS<      
+    A0Layout,    B0Layout,    DsLayout,    ELayout, 
+    A0DataType,  A1DataType,  B0DataType,  B1DataType,  DsDataType, EDataType, AccDataType, CShuffleDataType,
+    AElementOp,  BElementOp, CDEElementOp, GemmSpec,   
+    ScaleBlockSize, BlockSize,   
+    MPerBlock,      NPerBlock,    KPerBlock,
+    16,   16, 
+    16,   16,
+    4,     2,
+    S<8, 32, 1>, S<1, 0, 2>,     S<1, 0, 2>,    2, 16, 16, 0,
+    S<8, 32, 1>, S<1, 0, 2>,     S<1, 0, 2>,    2, 16, 16, 0,
+    2,    2,     S<1, 32, 1, 8>, S<8, 1, 1, 1>,
+    ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v3, 
+    ActOP, Nswizzle, true, MulRoutedWeight, ck::index_t, A0DataType>;
+// clang-format on
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = true;
+
+    // per expert:
+    // GEMM shape
+    constexpr ck::index_t sorted_tile_num = 13;
+    constexpr ck::index_t valid_tile_num  = sorted_tile_num;
+    ck::index_t sorted_size               = sorted_tile_num * MPerBlock;
+    ck::index_t valid_size                = valid_tile_num * MPerBlock;
+
+    ck::index_t N       = 4096;
+    ck::index_t K       = 6144;
+    ck::index_t experts = 8;
+    ck::index_t tokens  = 832;
+    ck::index_t topk    = 2;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 3)
+    {
+        // use default case
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 7)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+        N               = std::stoi(argv[4]);
+        K               = std::stoi(argv[5]);
+        tokens          = std::stoi(argv[6]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=no, 1=yes)\n");
+        printf("arg4 to 6: N, K, tokens\n");
+        exit(0);
+    }
+
+    if(K % ScaleBlockSize != 0)
+    {
+        throw std::runtime_error("wrong! K must be multiple of ScaleBlockSize.");
+    };
+
+    ck::index_t StrideA              = K;
+    ck::index_t StrideB              = K;
+    ck::index_t StrideE              = N;
+    ck::index_t Scale_Stride_AM      = (K + ScaleBlockSize - 1) / ScaleBlockSize;
+    ck::index_t Scale_Stride_BN      = (K + ScaleBlockSize - 1) / ScaleBlockSize;
+    constexpr ck::index_t NumDTensor = DsDataType::Size();
+    constexpr auto StrideDs          = std::array<ck::index_t, NumDTensor>{0, 0, 0};
+
+    ck::index_t KBatch = 1;
+
+    Tensor<ck::index_t> expert_ids(HostTensorDescriptor({sorted_tile_num}, {1}));
+    Tensor<ck::index_t> sorted_token_ids(HostTensorDescriptor({sorted_size}, {1}));
+    Tensor<ck::index_t> max_token_id(HostTensorDescriptor({sorted_tile_num + 1}));
+    max_token_id.mData[0] = valid_size;
+
+    if(tokens * topk > valid_size)
+    {
+        printf("err config, tokens * topk > valid_size\n");
+        exit(-1);
+    }
+
+    for(int i = 0; i < sorted_tile_num; i++)
+    {
+        expert_ids.mData[i] = i / ck::math::integer_divide_ceil(valid_tile_num, experts);
+    }
+    int token_per_tile = (tokens * topk + valid_tile_num - 1) / valid_tile_num;
+    int tokenid        = 0;
+    for(int i = 0; i < sorted_size; i++)
+    {
+        int tile_off = i % MPerBlock;
+        if(tile_off < token_per_tile)
+        {
+            sorted_token_ids.mData[i] = (tokenid % tokens) | ((tokenid / tokens) << 24);
+            tokenid++;
+        }
+        else
+        {
+            sorted_token_ids.mData[i] = tokens;
+        }
+    }
+
+    expert_ids.savetxt("expert_ids.txt", "int");
+    sorted_token_ids.savetxt("sorted_token_ids.txt", "int");
+
+    Tensor<A0DataType> a0_t_k(HostTensorDescriptor({tokens, K}, {K, 1}));
+    Tensor<XDataType> a1_t_k(HostTensorDescriptor(
+        {tokens, (K + ScaleBlockSize - 1) / ScaleBlockSize}, {Scale_Stride_AM, 1}));
+    Tensor<B0DataType> b0_e_n_k(HostTensorDescriptor({experts, K, N * 2}, {N * 2 * K, 1, K}));
+    Tensor<XDataType> b1_e_n_k(
+        HostTensorDescriptor({experts, (K + ScaleBlockSize - 1) / ScaleBlockSize, N * 2},
+                             {(N * 2 * Scale_Stride_BN), 1, Scale_Stride_BN}));
+
+    // A, B Scale preshuffle
+    Tensor<XDataType> a_scale_sorted(HostTensorDescriptor(
+        {sorted_size, (K + ScaleBlockSize - 1) / ScaleBlockSize}, {Scale_Stride_AM, 1}));
+    Tensor<XDataType> a_scale_preshuffled(HostTensorDescriptor(
+        {sorted_size, (K + ScaleBlockSize - 1) / ScaleBlockSize}, {Scale_Stride_AM, 1}));
+    Tensor<XDataType> b_scale_preshuffled(
+        HostTensorDescriptor({experts, (K + ScaleBlockSize - 1) / ScaleBlockSize, N * 2},
+                             {N * 2 * Scale_Stride_BN, 1, Scale_Stride_BN}));
+    Tensor<D2DataType> d2_e_n(HostTensorDescriptor({sorted_size, N}, {1, 0}));
+    Tensor<EDataType> e_t_k_n_host_result(
+        HostTensorDescriptor({tokens, topk, N}, {topk * N, N, 1}));
+    Tensor<EDataType> e_t_k_n_device_result(
+        HostTensorDescriptor({tokens, topk, N}, {topk * N, N, 1}));
+
+    e_t_k_n_device_result.SetZero();
+    std::cout << "a0_t_k:   " << a0_t_k.mDesc << std::endl;
+    std::cout << "a1_t_k:   " << a1_t_k.mDesc << std::endl;
+    std::cout << "b0_e_n_k: " << b0_e_n_k.mDesc << std::endl;
+    std::cout << "b1_e_n_k: " << b1_e_n_k.mDesc << std::endl;
+    std::cout << "d2_e_n:   " << d2_e_n.mDesc << std::endl;
+    std::cout << "e_t_k_n:  " << e_t_k_n_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a0_t_k.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-1, 1});
+        b0_e_n_k.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-1, 1});
+        a1_t_k.GenerateTensorValue(GeneratorTensor_3<XDataType>{0, 1.0});
+        b1_e_n_k.GenerateTensorValue(GeneratorTensor_3<XDataType>{0, 1.0});
+        d2_e_n.GenerateTensorValue(GeneratorTensor_3<D2DataType>{0, 1.0});
+        break;
+    case 2:
+        a0_t_k.GenerateTensorValue(GeneratorTensor_1<A0DataType>{});
+        b0_e_n_k.GenerateTensorValue(GeneratorTensor_1<B0DataType>{});
+        a1_t_k.GenerateTensorValue(GeneratorTensor_1<XDataType>{});
+        b1_e_n_k.GenerateTensorValue(GeneratorTensor_1<XDataType>{});
+        d2_e_n.GenerateTensorValue(GeneratorTensor_1<D2DataType>{0.1f});
+
+        // a0_t_k.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-1, 1});
+        // b0_e_n_k.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-5, 5});
+        // a1_t_k.GenerateTensorValue(GeneratorTensor_3<XDataType>{0, 1.0});
+        // b1_e_n_k.GenerateTensorValue(GeneratorTensor_3<XDataType>{0, 1.0});
+        break;
+    case 3:
+        a0_t_k.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-1, 1});
+        b0_e_n_k.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-1, 1});
+        a1_t_k.GenerateTensorValue(GeneratorTensor_3<XDataType>{0, 1.0});
+        b1_e_n_k.GenerateTensorValue(GeneratorTensor_3<XDataType>{0, 1.0});
+        d2_e_n.GenerateTensorValue(GeneratorTensor_1<D2DataType>{});
+        break;
+    case 4:
+        a0_t_k.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-2, 2});
+        b0_e_n_k.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
+        a1_t_k.GenerateTensorValue(GeneratorTensor_1<XDataType>{});
+        b1_e_n_k.GenerateTensorValue(GeneratorTensor_3<XDataType>{0, 5.0});
+        d2_e_n.GenerateTensorValue(GeneratorTensor_1<D2DataType>{});
+        break;
+    case 5:
+        a0_t_k.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-2, 2});
+        b0_e_n_k.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
+        a1_t_k.GenerateTensorValue(GeneratorTensor_3<XDataType>{0, 1.0});
+        b1_e_n_k.GenerateTensorValue(GeneratorTensor_1<XDataType>{});
+        d2_e_n.GenerateTensorValue(GeneratorTensor_1<D2DataType>{1});
+        break;
+    case 6:
+        a0_t_k.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-2, 2});
+        b0_e_n_k.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
+        a1_t_k.GenerateTensorValue(GeneratorTensor_3<XDataType>{0, 1.0});
+        b1_e_n_k.GenerateTensorValue(GeneratorTensor_1<XDataType>{});
+        d2_e_n.GenerateTensorValue(GeneratorTensor_1<D2DataType>{});
+        break;
+    case 7:
+        a0_t_k.GenerateTensorValue(GeneratorTensor_1<A0DataType>{0.5f});
+        b0_e_n_k.GenerateTensorValue(GeneratorTensor_1<B0DataType>{1.5f});
+        a1_t_k.GenerateTensorValue(GeneratorTensor_1<XDataType>{1.0f});
+        b1_e_n_k.GenerateTensorValue(GeneratorTensor_1<XDataType>{1.0f});
+        d2_e_n.GenerateTensorValue(GeneratorTensor_1<D2DataType>{0.1f});
+        break;
+    default:
+        a0_t_k.GenerateTensorValue(GeneratorTensor_3<A0DataType>{0.0, 1.0});
+        b0_e_n_k.GenerateTensorValue(GeneratorTensor_3<B0DataType>{-0.5, 0.5});
+        a1_t_k.GenerateTensorValue(GeneratorTensor_3<XDataType>{0.0, 1.0});
+        b1_e_n_k.GenerateTensorValue(GeneratorTensor_3<XDataType>{0.0, 1.0});
+        d2_e_n.GenerateTensorValue(GeneratorTensor_3<D2DataType>{0.0, 1.0});
+    }
+    DeviceMem sorted_token_ids_dev(sizeof(ck::index_t) * sorted_token_ids.GetElementSpaceSize());
+    DeviceMem expert_ids_dev(sizeof(ck::index_t) * expert_ids.GetElementSpaceSize());
+    DeviceMem max_token_id_dev(sizeof(ck::index_t) * max_token_id.GetElementSpaceSize());
+    DeviceMem a0_device_buf(sizeof(A0DataType) * a0_t_k.GetElementSpaceSize());
+    DeviceMem a1_device_buf(sizeof(XDataType) * a_scale_sorted.GetElementSpaceSize());
+    DeviceMem b0_device_buf(sizeof(B0DataType) * b0_e_n_k.GetElementSpaceSize());
+    DeviceMem b1_device_buf(sizeof(XDataType) * b1_e_n_k.GetElementSpaceSize());
+    DeviceMem d2_device_buf(sizeof(D2DataType) * d2_e_n.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) * e_t_k_n_device_result.GetElementSpaceSize());
+
+    // A scale sorted
+    for(int i = 0; i < sorted_size; i++)
+    {
+        int token_id = sorted_token_ids.mData[i] & 0x00FFFFFF;
+
+        for(int k = 0; k < (K + ScaleBlockSize - 1) / ScaleBlockSize; k++)
+        {
+            if(token_id == tokens)
+            {
+                a_scale_sorted(i, k) = ck::type_convert<XDataType>(0);
+            }
+            else
+            {
+                a_scale_sorted(i, k) = a1_t_k(token_id, k);
+            }
+        }
+    }
+
+    // A/B scale shuffle
+    preShuffleScaleBuffer<ck::is_same_v<A0Layout, Row>>(a_scale_sorted.mData.data(),
+                                                        a_scale_preshuffled.mData.data(),
+                                                        sorted_size,
+                                                        K / ScaleBlockSize);
+    preShuffleScaleBuffer<ck::is_same_v<B0Layout, Col>>(b1_e_n_k.mData.data(),
+                                                        b_scale_preshuffled.mData.data(),
+                                                        N * 2 * experts,
+                                                        K / ScaleBlockSize);
+
+    sorted_token_ids_dev.ToDevice(sorted_token_ids.mData.data());
+    expert_ids_dev.ToDevice(expert_ids.mData.data());
+    max_token_id_dev.ToDevice(max_token_id.mData.data());
+    a0_device_buf.ToDevice(a0_t_k.mData.data());
+    b0_device_buf.ToDevice(b0_e_n_k.mData.data());
+    a1_device_buf.ToDevice(a_scale_preshuffled.mData.data());
+    b1_device_buf.ToDevice(b_scale_preshuffled.mData.data());
+    d2_device_buf.ToDevice(d2_e_n.mData.data());
+    e_device_buf.ToDevice(e_t_k_n_device_result.mData.data());
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{};
+
+#if 0
+    printf("a0_t_k_k:\n");
+    for(int t = 0; t < tokens; ++t)
+    {
+        //for(int tk = 0; tk < topk; ++tk)
+        {
+            for(int k = 0; k < K; ++k)
+            {
+                auto f4x2 = a0_t_k(t, k).data;
+                if(k % 2 == 0)
+                {
+                    ck::f4_t f4 = (f4x2 >> 4) & 0xf;
+                    printf("%.2f ", ck::type_convert<float>(f4));
+                }
+                else
+                {
+                    ck::f4_t f4 = (f4x2 >> 0) & 0xf;
+                    printf("%.2f ", ck::type_convert<float>(f4));
+                }
+            }
+            printf("\n");
+        }
+        printf("\n");
+    }
+
+    printf("a1_t_k_k:\n");
+    for(int t = 0; t < tokens; ++t)
+    {
+        for(int tk = 0; tk < topk; ++tk)
+        {
+            for(int k = 0; k < (K + ScaleBlockSize - 1) / ScaleBlockSize; ++k)
+            {
+                printf("%.2f ", ck::type_convert<float>(a1_t_k_k(t, tk, k)));
+            }
+            printf("\n");
+        }
+        printf("\n");
+    }
+
+    printf("a_scale_sorted: K/scale: %d\n", (K + ScaleBlockSize - 1) / ScaleBlockSize);
+    for(int i = 0; i < sorted_size; ++i)
+    {
+        for(int k = 0; k < (K + ScaleBlockSize - 1) / ScaleBlockSize; ++k)
+        {
+            printf("%.2f ", ck::type_convert<float>(a_scale_sorted(i, k)));
+        }
+        printf("\n");
+    }
+
+    printf("a_scale_preshuffled:\n");
+    for(int i = 0; i < sorted_size; ++i)
+    {
+        for(int k = 0; k < (K + ScaleBlockSize - 1) / ScaleBlockSize; ++k)
+        {
+            printf("%.2f ", ck::type_convert<float>(a_scale_preshuffled(i, k)));
+        }
+        printf("\n");
+    }
+
+    printf("b0_e_n_k:\n");
+    for(int e = 0; e < experts; ++e)
+    {
+        for(int n = 0; n < N; ++n)
+        {
+            for(int k = 0; k < K; ++k)
+            {
+                auto f4x2 = b0_e_n_k(e, k, n).data;
+                if(k % 2 == 0)
+                {
+                    ck::f4_t f4 = f4x2 >> 4 & 0xf;
+                    printf("%.2f ", ck::type_convert<float>(f4));
+                }
+                else
+                {
+                    ck::f4_t f4 = f4x2 >> 0 & 0xf;
+                    printf("%.2f ", ck::type_convert<float>(f4));
+                }
+            }
+            printf("\n");
+        }
+        printf("\n");
+    }
+
+    printf("b1_e_n_k:\n");
+    for(int e = 0; e < experts; ++e)
+    {
+        for(int k = 0; k < (K + ScaleBlockSize - 1) / ScaleBlockSize; ++k)
+        {
+            for(int n = 0; n < N; ++n)
+            {
+                printf("%.2f ", ck::type_convert<float>(b1_e_n_k(e, k, n)));
+            }
+            printf("\n");
+        }
+        printf("\n");
+    }
+
+    printf("d2_e_n:\n");
+    for(int i = 0; i < sorted_size; ++i)
+    {
+        for(int n = 0; n < 1; ++n)
+        {
+            printf("%.2f ", ck::type_convert<float>(d2_e_n(i, n)));
+        }
+    }
+#endif
+
+    // do GEMM
+    auto device_op = DeviceOpInstance{};
+
+    auto invoker  = device_op.MakeInvoker();
+    auto argument = device_op.MakeArgument(
+        sorted_token_ids_dev.GetDeviceBuffer(),
+        expert_ids_dev.GetDeviceBuffer(),
+        max_token_id_dev.GetDeviceBuffer(),
+        a0_device_buf.GetDeviceBuffer(),
+        a1_device_buf.GetDeviceBuffer(),
+        b0_device_buf.GetDeviceBuffer(),
+        b1_device_buf.GetDeviceBuffer(),
+        std::array<const void*, NumDTensor>{nullptr, nullptr, d2_device_buf.GetDeviceBuffer()},
+        e_device_buf.GetDeviceBuffer(),
+        tokens,
+        topk,
+        sorted_size,
+        N,
+        K,
+        StrideA,
+        Scale_Stride_AM,
+        StrideB,
+        Scale_Stride_BN,
+        StrideDs,
+        StrideE,
+        KBatch,
+        a_element_op,
+        b_element_op,
+        cde_element_op);
+
+    if(!device_op.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_gemm with the specified compilation parameters does "
+            "not support this GEMM problem");
+    }
+
+    if(!(ck::get_device_name() == "gfx942" || ck::get_device_name() == "gfx950"))
+    {
+        std::cout << "This kernel support gfx942 and gfx950 only" << std::endl;
+    }
+
+    if(time_kernel)
+    {
+        // not result correct here because output buf not setzero
+        float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+        std::size_t flop =
+            // FMA * tokens * N * (Gate+Up) * topk * K +
+            // FMA * tokens * N * (Gate+Up) * topk * (K/BlockScale)
+            std::size_t(2) * tokens * N * 2 * topk * K +
+            std::size_t(2) * tokens * N * 2 * topk * K / ScaleBlockSize;
+
+        std::size_t num_btype = sizeof(A0DataType) / 2 * tokens * topk * K +
+                                sizeof(B0DataType) / 2 * K * N * 2 * experts +
+                                sizeof(XDataType) * tokens * topk * K / ScaleBlockSize +
+                                sizeof(XDataType) * K / ScaleBlockSize * N * 2 * experts +
+                                sizeof(EDataType) * tokens * topk * N;
+
+        float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+        float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+        std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
+                  << " GB/s" << device_op.GetTypeString() << std::endl;
+    }
+
+    if(do_verification)
+    {
+        // gemm2 use atomic, so need to reinit outputs
+        e_device_buf.ToDevice(e_t_k_n_device_result.mData.data());
+        invoker.Run(argument, StreamConfig{nullptr, false, 0, 0, 1});
+
+        Tensor<CShuffleDataType> c_t_k_n({tokens, topk, N}, {topk * N, N, 1});
+
+        using ReferenceGemmInstance =
+            ck::tensor_operation::host::ReferenceMoeMXGemm1<A0DataType,
+                                                            XDataType,
+                                                            B0DataType,
+                                                            XDataType,
+                                                            CShuffleDataType,
+                                                            D2DataType,
+                                                            AccDataType,
+                                                            PassThrough,
+                                                            PassThrough,
+                                                            PassThrough,
+                                                            ActOP,
+                                                            MulRoutedWeight>;
+        auto ref_moe_gemm = ReferenceGemmInstance{};
+        auto ref_invoker  = ref_moe_gemm.MakeInvoker();
+
+        auto ref_argument = ref_moe_gemm.MakeArgument(sorted_token_ids,
+                                                      expert_ids,
+                                                      max_token_id,
+                                                      MPerBlock,
+                                                      a0_t_k,
+                                                      a1_t_k,
+                                                      b0_e_n_k,
+                                                      b1_e_n_k,
+                                                      d2_e_n,
+                                                      c_t_k_n,
+                                                      PassThrough{},
+                                                      PassThrough{},
+                                                      PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+        for(int m = 0; m < valid_size; ++m)
+        {
+            const int fuse_t  = sorted_token_ids.mData[m];
+            const int t       = fuse_t & 0xffffff;
+            const int topk_id = (fuse_t & 0xff000000) >> 24;
+
+            if(t >= tokens)
+            {
+                continue;
+            }
+            for(int n = 0; n < N; ++n)
+            {
+                e_t_k_n_host_result(t, topk_id, n) =
+                    ck::type_convert<EDataType>(c_t_k_n(t, topk_id, n));
+            }
+        }
+
+        e_device_buf.FromDevice(e_t_k_n_device_result.mData.data());
+
+#if 0
+        e_t_k_n_device_result.savetxt("e_t_k_n_device_result.txt", "float");
+        printf("e_t_k_n_device_result:\n");
+        for(int t = 0; t < tokens; ++t)
+        {
+            for(int k = 0; k < topk; k++)
+            {
+                printf("[%d,%d]: ", t, k);
+                for(int n = 0; n < N; ++n)
+                {
+                    printf("%.2f ", ck::type_convert<float>(e_t_k_n_device_result(t, k, n)));
+                }
+                printf("\n");
+            }
+        }
+
+        printf("e_t_k_n_host_result:\n");
+        for(int t = 0; t < tokens; ++t)
+        {
+            for(int k = 0; k < topk; k++)
+            {
+                printf("[%d,%d]: ", t, k);
+                for(int n = 0; n < N; ++n)
+                {
+                    printf("%.2f ", ck::type_convert<float>(e_t_k_n_host_result(t, k, n)));
+                }
+                printf("\n");
+            }
+        }
+#endif
+
+        auto status =
+            ck::utils::check_err(
+                e_t_k_n_device_result, e_t_k_n_host_result, "Error: Incorrect results!", 1e-3, 5e-1)
+                ? 0
+                : 1;
+        if(status == 0)
+        {
+            printf("Validation Pass.\n");
+        }
+        return status;
+    }
+
+    return 0;
+}

example/67_gemm_microscaling/moe_gemm2_xdl_mx_fp4.cpp

@@ -0,0 +1,721 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_moe_mx_gemm.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp"
+
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_moe_mx_gemm2.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/fill.hpp"
+#include "ck/utility/blkgemmpipe_scheduler.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F4              = ck::f4x2_pk_t;
+using F16             = ck::half_t;
+using BF16            = ck::bhalf_t;
+using F32             = float;
+using XDataType       = ck::e8m0_bexp_t;
+using XPackedDataType = int32_t; // 4 packed e8m0_bexp_t
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using A0DataType       = F4;
+using A1DataType       = XPackedDataType;
+using B0DataType       = F4;
+using B1DataType       = XPackedDataType;
+using EDataType        = F16;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using D0DataType       = F32;
+using D1DataType       = F32;
+using D2DataType       = F32;
+using DsDataType       = ck::Tuple<D0DataType, D1DataType, D2DataType>;
+
+using A0Layout = Row;
+using B0Layout = Col;
+using ELayout  = Row;
+using D0Layout = Row;
+using D1Layout = Col;
+using D2Layout = ELayout;
+using DsLayout = ck::Tuple<D0Layout, D1Layout, D2Layout>;
+
+// d0: ascale, d1: bscale, d2:expert weight
+struct MulABScaleExpertWeight
+{
+    template <typename E, typename C, typename D0, typename D1, typename D2>
+    __host__ __device__ constexpr void
+    operator()(E& e, const C& c, const D0& d0, const D1& d1, const D2& d2) const;
+    // for real kernel use
+    template <>
+    __host__ __device__ constexpr void operator()<EDataType, float, float, float, float>(
+        EDataType& e, const float& c, const float& d0, const float& d1, const float& d2) const
+    {
+        (void)d0;
+        (void)d1;
+        (void)d2;
+
+        e = ck::type_convert<EDataType>(c);
+    }
+    // for reference cpu
+    template <>
+    __host__ __device__ constexpr void operator()<float, float, float, float, float>(
+        float& e, const float& c, const float& d0, const float& d1, const float& d2) const
+    {
+        // for reference cpu
+        e = ck::type_convert<EDataType>(c * d0 * d1 * d2);
+    }
+};
+
+using CDEElementOp = MulABScaleExpertWeight;
+
+// B preshuffle
+void preShuffleBuffer(const F4* src, F4* dst, int N, int K, int NXdl)
+{
+    int KPack = 32;
+    int NLane = NXdl;
+    int KLane = 64 / NLane;
+
+    int K0 = K / (KLane * KPack);
+    // K -> K0 KLane KPack
+    // N -> N0 NLane
+    // N, K -> N0 K0 KLane NLane KPack
+    int tempk;
+    for(int n = 0; n < N; ++n)
+    {
+        for(int k = 0; k < K; ++k)
+        {
+            int n0 = n / NLane;
+            int n1 = n % NLane;
+
+            int k0 = k / (KLane * KPack);
+            tempk  = k % (KLane * KPack);
+            int k1 = tempk / KPack;
+            int k2 = tempk % KPack;
+
+            int outputIndex = n0 * KPack * NLane * KLane * K0 + k0 * KPack * NLane * KLane +
+                              k1 * KPack * NLane + n1 * KPack + k2;
+
+            dst[outputIndex / 2] = src[(n * K + k) / 2];
+        }
+    }
+}
+
+// A, B Scale preshuffle
+template <bool KLast>
+void preShuffleScaleBuffer(ck::e8m0_bexp_t* src, ck::e8m0_bexp_t* dst, int MN, int K)
+{
+    int MNXdlPack = 2;
+    int KXdlPack  = 2;
+
+    int XdlMNThread = 16;
+    int XdlKThread  = 64 / XdlMNThread;
+
+    int K0 = K / KXdlPack / XdlKThread; // KRepeat
+
+    // The 4 16x128 building blocks will be packed into 1 32x256 for F4
+    // The 8 16x16x128 mfma will be packed into 1 32x32x256 for F4
+
+    // unfold the MN32xK(256/32) scale buffer
+    //    4            16             2           2
+    // To XdlKThread-> XdlMNThread -> KXdlPack -> MNXdlPack
+    // Then, MNRepeat->KRepeat
+
+    for(int n = 0; n < MN; ++n)
+    {
+        for(int k = 0; k < K; ++k)
+        {
+            int n0    = n / (XdlMNThread * MNXdlPack); // i MNRepeat
+            int tempn = n % (XdlMNThread * MNXdlPack);
+            int n1    = tempn % XdlMNThread; // i XdlMNThread
+            int n2    = tempn / XdlMNThread; // i MNXdlPack
+
+            int k0    = k / (XdlKThread * KXdlPack); // i KRepeat
+            int tempk = k % (XdlKThread * KXdlPack);
+            int k1    = tempk % XdlKThread; // i XdlKThread
+            int k2    = tempk / XdlKThread; // i KXdlPack
+
+            int outputIndex = n0 * MNXdlPack * KXdlPack * XdlMNThread * XdlKThread * K0 +
+                              k0 * MNXdlPack * KXdlPack * XdlMNThread * XdlKThread +
+                              k1 * MNXdlPack * KXdlPack * XdlMNThread + n1 * MNXdlPack * KXdlPack +
+                              k2 * MNXdlPack + n2;
+            // src[n * K + k] = ck::type_convert<ck::e8m0_bexp_t>(static_cast<float>(powf(2.0f, n2 +
+            // k2 * MNXdlPack)));
+            if constexpr(KLast)
+                dst[outputIndex] = src[n * K + k];
+            else
+                dst[outputIndex] = src[k * MN + n];
+        }
+    }
+}
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = MulABScaleExpertWeight;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::Default;
+
+constexpr ck::index_t DataPackedSize = 2;                    // Packed representation of data
+constexpr ck::index_t ScaleBlockSize = 32;                   // scaling block size
+constexpr ck::index_t KPerBlock      = 256 / DataPackedSize; // 256 f4 = 128 fp4x2
+
+#if 0
+static constexpr ck::index_t MPerBlock = 128;
+static constexpr ck::index_t BLOCKSIZE = 256;
+static constexpr ck::index_t MXDLPerWave   = 8;
+static constexpr ck::index_t NXDLPerWave   = 2;
+static constexpr ck::index_t NPerBlock     = 128;
+static constexpr ck::index_t MNPerXDL      = 16;
+static constexpr ck::index_t KPerBlock     = 128 / sizeof(A0DataType);
+static constexpr ck::index_t CShuffleNLane = 32;
+static constexpr ck::index_t CShuffleMLane = BLOCKSIZE / CShuffleNLane;
+static constexpr ck::index_t AK1           = 16 / sizeof(A0DataType);
+static constexpr ck::index_t BK1           = 32 / sizeof(B0DataType);
+static constexpr ck::index_t EVec          = 2;
+static constexpr ck::index_t D0Vec         = 1;
+static constexpr ck::index_t D1Vec         = 1;
+static constexpr ck::index_t D2Vec         = 1;
+static constexpr bool MulRoutedWeight      = true;
+using DeviceOpInstance                     = ck::tensor_operation::device::DeviceMoeGemm
+    // clang-format off
+        <      Row, Col, DsLayout, ELayout, A0DataType, B0DataType, DsDataType, EDataType, AccDataType, CShuffleDataType,
+               AElementOp,  BElementOp, CDEElementOp,       GemmSpec,   
+               BLOCKSIZE,   MPerBlock,   NPerBlock,    KPerBlock,
+               AK1,   BK1,
+               MNPerXDL,   MNPerXDL,
+               MXDLPerWave,    NXDLPerWave,
+               S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, AK1, AK1, 0,
+               S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, BK1, BK1, 0,
+               2,    2,   S<1, CShuffleMLane, 1, CShuffleNLane>, S<EVec, D0Vec, D1Vec, D2Vec>,
+               ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v1, 0, false, false, MulRoutedWeight, false, ck::index_t, A0DataType>;
+// clang-format on
+
+#else
+static constexpr ck::index_t MPerBlock = 32;
+static constexpr bool MulRoutedWeight  = true;
+
+// clang-format off
+using DeviceOpInstance                     = ck::tensor_operation::device::DeviceMoeGemmMX<      
+    A0Layout,    B0Layout,    DsLayout,    ELayout, 
+    A0DataType,  A1DataType,  B0DataType,  B1DataType,  DsDataType, EDataType, AccDataType, CShuffleDataType,
+    AElementOp,  BElementOp, CDEElementOp, GemmSpec,   
+    ScaleBlockSize,      64,   
+    MPerBlock,   32,    KPerBlock,
+    16,   16,
+    16,   16,
+    2,    2,
+    S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0,
+    S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0,
+    1,    1,   S<1, 8, 1, 8>, S<2, 1, 1, 1>,
+    ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v3, 0, false, false, MulRoutedWeight, ck::index_t, A0DataType>;
+// clang-format on
+#endif
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = true;
+
+    // per expert:
+    // GEMM shape
+    constexpr ck::index_t sorted_tile_num = 2;
+    constexpr ck::index_t valid_tile_num  = 2;
+    ck::index_t sorted_size               = sorted_tile_num * MPerBlock;
+    ck::index_t valid_size                = valid_tile_num * MPerBlock;
+
+    ck::index_t N       = 6144;
+    ck::index_t K       = 4096;
+    ck::index_t experts = 2;
+    ck::index_t tokens  = 832;
+    ck::index_t topk    = 2;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 3)
+    {
+        // use default case
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 7)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+        N               = std::stoi(argv[4]);
+        K               = std::stoi(argv[5]);
+        tokens          = std::stoi(argv[6]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=no, 1=yes)\n");
+        printf("arg4 to 6: N, K, tokens\n");
+        exit(0);
+    }
+
+    if(K % ScaleBlockSize != 0)
+    {
+        throw std::runtime_error("wrong! K must be multiple of ScaleBlockSize.");
+    };
+
+    ck::index_t StrideA              = K;
+    ck::index_t StrideB              = K;
+    ck::index_t StrideE              = N;
+    ck::index_t Scale_Stride_AM      = (K + ScaleBlockSize - 1) / ScaleBlockSize;
+    ck::index_t Scale_Stride_BN      = (K + ScaleBlockSize - 1) / ScaleBlockSize;
+    constexpr ck::index_t NumDTensor = DsDataType::Size();
+    constexpr auto StrideDs          = std::array<ck::index_t, NumDTensor>{0, 0, 0};
+
+    ck::index_t KBatch = 1;
+
+    Tensor<ck::index_t> expert_ids(HostTensorDescriptor({sorted_tile_num}, {1}));
+    Tensor<ck::index_t> sorted_token_ids(HostTensorDescriptor({sorted_size}, {1}));
+    Tensor<ck::index_t> max_token_id(HostTensorDescriptor({1}));
+    max_token_id.mData[0] = valid_size;
+    // int eids[]            = {0, 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 3, 3, 3};
+    int eids[sorted_tile_num]{};
+    for(int i = 0; i < sorted_tile_num; i++)
+    {
+        if(i < valid_tile_num)
+        {
+            eids[i] = (i * experts) / valid_tile_num;
+        }
+        else
+        {
+            eids[i] = 3;
+        }
+    }
+
+    for(int i = 0; i < sorted_tile_num; i++)
+    {
+        expert_ids.mData[i] = eids[i];
+    }
+    if(tokens * topk > valid_size)
+    {
+        printf("err config, tokens * topk > valid_size\n");
+        exit(-1);
+    }
+    int token_per_tile = tokens * topk / valid_tile_num;
+    int tokenid        = 0;
+    for(int i = 0; i < sorted_size; i++)
+    {
+        int tile_off = i % MPerBlock;
+        if(tile_off < token_per_tile)
+        {
+            sorted_token_ids.mData[i] = (tokenid % tokens) | ((tokenid / tokens) << 24);
+            tokenid++;
+        }
+        else
+        {
+            sorted_token_ids.mData[i] = tokens;
+        }
+    }
+
+    expert_ids.savetxt("expert_ids.txt", "int");
+    sorted_token_ids.savetxt("sorted_token_ids.txt", "int");
+    Tensor<A0DataType> a0_t_k_k(HostTensorDescriptor({tokens, topk, K}, {topk * K, K, 1}));
+    Tensor<XDataType> a1_t_k_k(
+        HostTensorDescriptor({tokens, topk, (K + ScaleBlockSize - 1) / ScaleBlockSize},
+                             {(topk * Scale_Stride_AM), Scale_Stride_AM, 1}));
+    Tensor<B0DataType> b0_e_n_k(HostTensorDescriptor({experts, K, N}, {N * K, 1, K}));
+    Tensor<XDataType> b1_e_n_k(
+        HostTensorDescriptor({experts, (K + ScaleBlockSize - 1) / ScaleBlockSize, N},
+                             {(N * Scale_Stride_BN), 1, Scale_Stride_BN}));
+    // B preshuffle
+    Tensor<B0DataType> b0_preshuffled(HostTensorDescriptor({experts, K, N}, {N * K, 1, K}));
+
+    // A, B Scale preshuffle
+    Tensor<XDataType> a_scale_sorted(HostTensorDescriptor(
+        {sorted_size, (K + ScaleBlockSize - 1) / ScaleBlockSize}, {Scale_Stride_AM, 1}));
+    Tensor<XDataType> a_scale_preshuffled(HostTensorDescriptor(
+        {sorted_size, (K + ScaleBlockSize - 1) / ScaleBlockSize}, {Scale_Stride_AM, 1}));
+    Tensor<XDataType> b_scale_preshuffled(
+        HostTensorDescriptor({experts, (K + ScaleBlockSize - 1) / ScaleBlockSize, N},
+                             {N * Scale_Stride_BN, 1, Scale_Stride_BN}));
+    Tensor<D2DataType> d2_e_n(HostTensorDescriptor({sorted_size, N}, {1, 0}));
+    Tensor<EDataType> e_t_n_host_result(HostTensorDescriptor({tokens, N}, {N, 1}));
+    Tensor<EDataType> e_t_n_device_result(HostTensorDescriptor({tokens, N}, {N, 1}));
+
+    e_t_n_device_result.SetZero();
+    std::cout << "a0_t_k_k: " << a0_t_k_k.mDesc << std::endl;
+    std::cout << "a1_t_k_k: " << a1_t_k_k.mDesc << std::endl;
+    std::cout << "b0_e_n_k: " << b0_e_n_k.mDesc << std::endl;
+    std::cout << "b1_e_n_k: " << b1_e_n_k.mDesc << std::endl;
+    std::cout << "d2_e_n: " << d2_e_n.mDesc << std::endl;
+    std::cout << "e_t_n: " << e_t_n_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a0_t_k_k.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-1, 1});
+        b0_e_n_k.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-1, 1});
+        a1_t_k_k.GenerateTensorValue(GeneratorTensor_3<XDataType>{0, 1.0});
+        b1_e_n_k.GenerateTensorValue(GeneratorTensor_3<XDataType>{0, 1.0});
+        d2_e_n.GenerateTensorValue(GeneratorTensor_2<D2DataType>{-1, 1});
+        break;
+    case 2:
+        a0_t_k_k.GenerateTensorValue(GeneratorTensor_1<A0DataType>{});
+        b0_e_n_k.GenerateTensorValue(GeneratorTensor_1<B0DataType>{});
+        a1_t_k_k.GenerateTensorValue(GeneratorTensor_1<XDataType>{});
+        b1_e_n_k.GenerateTensorValue(GeneratorTensor_1<XDataType>{});
+        d2_e_n.GenerateTensorValue(GeneratorTensor_1<D2DataType>{});
+        break;
+    case 3:
+        a0_t_k_k.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-2, 2});
+        b0_e_n_k.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
+        a1_t_k_k.GenerateTensorValue(GeneratorTensor_1<XDataType>{});
+        b1_e_n_k.GenerateTensorValue(GeneratorTensor_3<XDataType>{0, 1.0});
+        d2_e_n.GenerateTensorValue(GeneratorTensor_2<D2DataType>{-2, 2});
+        break;
+    case 4:
+        a0_t_k_k.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-2, 2});
+        b0_e_n_k.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
+        a1_t_k_k.GenerateTensorValue(GeneratorTensor_3<XDataType>{0, 1.0});
+        b1_e_n_k.GenerateTensorValue(GeneratorTensor_1<XDataType>{});
+        d2_e_n.GenerateTensorValue(GeneratorTensor_2<D2DataType>{-2, 2});
+        break;
+    case 5:
+        a0_t_k_k.GenerateTensorValue(GeneratorTensor_1<A0DataType>{});
+        b0_e_n_k.GenerateTensorValue(GeneratorTensor_1<B0DataType>{});
+        a1_t_k_k.GenerateTensorValue(GeneratorTensor_1<XDataType>{});
+        b1_e_n_k.GenerateTensorValue(GeneratorTensor_3<XDataType>{0, 1.0});
+        d2_e_n.GenerateTensorValue(GeneratorTensor_1<D2DataType>{});
+        break;
+    case 6:
+        a0_t_k_k.GenerateTensorValue(GeneratorTensor_1<A0DataType>{});
+        b0_e_n_k.GenerateTensorValue(GeneratorTensor_1<B0DataType>{});
+        a1_t_k_k.GenerateTensorValue(GeneratorTensor_3<XDataType>{0, 1.0});
+        b1_e_n_k.GenerateTensorValue(GeneratorTensor_1<XDataType>{});
+        d2_e_n.GenerateTensorValue(GeneratorTensor_1<D2DataType>{});
+        break;
+    default:
+        a0_t_k_k.GenerateTensorValue(GeneratorTensor_3<A0DataType>{0.0, 1.0});
+        b0_e_n_k.GenerateTensorValue(GeneratorTensor_3<B0DataType>{-0.5, 0.5});
+        a1_t_k_k.GenerateTensorValue(GeneratorTensor_3<XDataType>{0.0, 1.0});
+        b1_e_n_k.GenerateTensorValue(GeneratorTensor_3<XDataType>{0.0, 1.0});
+        d2_e_n.GenerateTensorValue(GeneratorTensor_3<D2DataType>{0.0, 1.0});
+    }
+    DeviceMem sorted_token_ids_dev(sizeof(ck::index_t) *
+                                   sorted_token_ids.mDesc.GetElementSpaceSize());
+    DeviceMem expert_ids_dev(sizeof(ck::index_t) * expert_ids.mDesc.GetElementSpaceSize());
+    DeviceMem max_token_id_dev(sizeof(ck::index_t) * max_token_id.mDesc.GetElementSpaceSize());
+    DeviceMem a0_device_buf(sizeof(A0DataType) * a0_t_k_k.mDesc.GetElementSpaceSize() / 2);
+    DeviceMem a1_device_buf(sizeof(XDataType) * a_scale_sorted.mDesc.GetElementSpaceSize());
+    DeviceMem b0_device_buf(sizeof(B0DataType) * b0_e_n_k.mDesc.GetElementSpaceSize() / 2);
+    DeviceMem b1_device_buf(sizeof(XDataType) * b1_e_n_k.mDesc.GetElementSpaceSize());
+    DeviceMem d2_device_buf(sizeof(D2DataType) * d2_e_n.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) * e_t_n_device_result.mDesc.GetElementSpaceSize());
+
+    // A scale sorted
+    for(int i = 0; i < sorted_size; i++)
+    {
+        int token_id = sorted_token_ids.mData[i] & 0x00FFFFFF;
+        int topk_id  = (sorted_token_ids.mData[i] >> 24) & 0x000000FF;
+
+        for(int k = 0; k < (K + ScaleBlockSize - 1) / ScaleBlockSize; k++)
+        {
+            if(token_id == tokens)
+            {
+                a_scale_sorted(i, k) = ck::type_convert<XDataType>(0);
+            }
+            else
+            {
+                a_scale_sorted(i, k) = a1_t_k_k(token_id, topk_id, k);
+            }
+        }
+    }
+
+    preShuffleScaleBuffer<ck::is_same_v<A0Layout, Row>>(a_scale_sorted.mData.data(),
+                                                        a_scale_preshuffled.mData.data(),
+                                                        sorted_size,
+                                                        K / ScaleBlockSize);
+    preShuffleScaleBuffer<ck::is_same_v<B0Layout, Col>>(
+        b1_e_n_k.mData.data(), b_scale_preshuffled.mData.data(), N * experts, K / ScaleBlockSize);
+
+    sorted_token_ids_dev.ToDevice(sorted_token_ids.mData.data());
+    expert_ids_dev.ToDevice(expert_ids.mData.data());
+    max_token_id_dev.ToDevice(max_token_id.mData.data());
+    a0_device_buf.ToDevice(a0_t_k_k.mData.data());
+    a1_device_buf.ToDevice(a_scale_preshuffled.mData.data());
+    b1_device_buf.ToDevice(b_scale_preshuffled.mData.data());
+    d2_device_buf.ToDevice(d2_e_n.mData.data());
+    e_device_buf.ToDevice(e_t_n_device_result.mData.data());
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{};
+
+#if 1
+    printf("a0_t_k_k:\n");
+    for(int t = 0; t < tokens; ++t)
+    {
+        for(int tk = 0; tk < topk; ++tk)
+        {
+            for(int k = 0; k < K; ++k)
+            {
+                auto f4x2 = a0_t_k_k(t, tk, k).data;
+                if(k % 2 == 0)
+                {
+                    ck::f4_t f4 = (f4x2 >> 4) & 0xf;
+                    printf("%.2f ", ck::type_convert<float>(f4));
+                }
+                else
+                {
+                    ck::f4_t f4 = (f4x2 >> 0) & 0xf;
+                    printf("%.2f ", ck::type_convert<float>(f4));
+                }
+            }
+            printf("\n");
+        }
+        printf("\n");
+    }
+
+    printf("a1_t_k_k:\n");
+    for(int t = 0; t < tokens; ++t)
+    {
+        for(int tk = 0; tk < topk; ++tk)
+        {
+            for(int k = 0; k < (K + ScaleBlockSize - 1) / ScaleBlockSize; ++k)
+            {
+                printf("%.2f ", ck::type_convert<float>(a1_t_k_k(t, tk, k)));
+            }
+            printf("\n");
+        }
+        printf("\n");
+    }
+
+    printf("a_scale_sorted: K/scale: %d\n", (K + ScaleBlockSize - 1) / ScaleBlockSize);
+    for(int i = 0; i < sorted_size; ++i)
+    {
+        for(int k = 0; k < (K + ScaleBlockSize - 1) / ScaleBlockSize; ++k)
+        {
+            printf("%.2f ", ck::type_convert<float>(a_scale_sorted(i, k)));
+        }
+        printf("\n");
+    }
+
+    printf("a_scale_preshuffled:\n");
+    for(int i = 0; i < sorted_size; ++i)
+    {
+        for(int k = 0; k < (K + ScaleBlockSize - 1) / ScaleBlockSize; ++k)
+        {
+            printf("%.2f ", ck::type_convert<float>(a_scale_preshuffled(i, k)));
+        }
+        printf("\n");
+    }
+
+    printf("b0_e_n_k:\n");
+    for(int e = 0; e < experts; ++e)
+    {
+        for(int n = 0; n < N; ++n)
+        {
+            for(int k = 0; k < K; ++k)
+            {
+                auto f4x2 = b0_e_n_k(e, k, n).data;
+                if(k % 2 == 0)
+                {
+                    ck::f4_t f4 = f4x2 >> 4 & 0xf;
+                    printf("%.2f ", ck::type_convert<float>(f4));
+                }
+                else
+                {
+                    ck::f4_t f4 = f4x2 >> 0 & 0xf;
+                    printf("%.2f ", ck::type_convert<float>(f4));
+                }
+            }
+            printf("\n");
+        }
+        printf("\n");
+    }
+
+    printf("b1_e_n_k:\n");
+    for(int e = 0; e < experts; ++e)
+    {
+        for(int k = 0; k < (K + ScaleBlockSize - 1) / ScaleBlockSize; ++k)
+        {
+            for(int n = 0; n < N; ++n)
+            {
+                printf("%.2f ", ck::type_convert<float>(b1_e_n_k(e, k, n)));
+            }
+            printf("\n");
+        }
+        printf("\n");
+    }
+#endif
+
+    // do GEMM
+    auto device_op = DeviceOpInstance{};
+
+    preShuffleBuffer(b0_e_n_k.mData.data(),
+                     b0_preshuffled.mData.data(),
+                     N * experts,
+                     K,
+                     device_op.GetPreShuffleParameters());
+
+    b0_device_buf.ToDevice(b0_preshuffled.mData.data());
+
+    auto invoker  = device_op.MakeInvoker();
+    auto argument = device_op.MakeArgument(
+        sorted_token_ids_dev.GetDeviceBuffer(),
+        expert_ids_dev.GetDeviceBuffer(),
+        max_token_id_dev.GetDeviceBuffer(),
+        a0_device_buf.GetDeviceBuffer(),
+        a1_device_buf.GetDeviceBuffer(),
+        b0_device_buf.GetDeviceBuffer(),
+        b1_device_buf.GetDeviceBuffer(),
+        std::array<const void*, NumDTensor>{nullptr, nullptr, d2_device_buf.GetDeviceBuffer()},
+        e_device_buf.GetDeviceBuffer(),
+        tokens,
+        topk,
+        sorted_size,
+        N,
+        K,
+        StrideA,
+        Scale_Stride_AM,
+        StrideB,
+        Scale_Stride_BN,
+        StrideDs,
+        StrideE,
+        KBatch,
+        a_element_op,
+        b_element_op,
+        cde_element_op);
+
+    if(!device_op.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_gemm with the specified compilation parameters does "
+            "not support this GEMM problem");
+    }
+
+    if(!(ck::get_device_name() == "gfx942" || ck::get_device_name() == "gfx950"))
+    {
+        std::cout << "This kernel support gfx942 and gfx950 only" << std::endl;
+    }
+
+    if(time_kernel)
+    {
+        // not result correct here because output buf not setzero
+        float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+        // FMA * tokens * N * topk * K +
+        // FMA * tokens * N * topk * (K/BlockScale)
+        std::size_t flop = std::size_t(2) * tokens * topk * N * K +
+                           std::size_t(2) * tokens * topk * N * K / ScaleBlockSize;
+
+        std::size_t num_btype =
+            sizeof(A0DataType) / 2 * tokens * K * topk + sizeof(B0DataType) / 2 * K * N * experts +
+            sizeof(XDataType) * tokens * topk * K / ScaleBlockSize +
+            sizeof(XDataType) * K / ScaleBlockSize * N * experts + sizeof(EDataType) * tokens * N;
+
+        float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+        float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+        std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
+                  << " GB/s" << device_op.GetTypeString() << std::endl;
+    }
+
+    if(do_verification)
+    {
+        // gemm2 use atomic, so need to reinit outputs
+        e_device_buf.ToDevice(e_t_n_device_result.mData.data());
+        invoker.Run(argument, StreamConfig{nullptr, false, 0, 0, 1});
+
+        Tensor<CShuffleDataType> c_t_n({tokens, N});
+
+        using ReferenceGemmInstance =
+            ck::tensor_operation::host::ReferenceMoeMXGemm2<A0DataType,
+                                                            XDataType,
+                                                            B0DataType,
+                                                            XDataType,
+                                                            D2DataType,
+                                                            CShuffleDataType,
+                                                            AccDataType,
+                                                            PassThrough,
+                                                            PassThrough,
+                                                            CDEElementOp,
+                                                            MulRoutedWeight,
+                                                            float,
+                                                            float>;
+
+        auto ref_moe_gemm = ReferenceGemmInstance{};
+        auto ref_invoker  = ref_moe_gemm.MakeInvoker();
+        auto ref_argument = ref_moe_gemm.MakeArgument(sorted_token_ids,
+                                                      expert_ids,
+                                                      max_token_id,
+                                                      MPerBlock,
+                                                      a0_t_k_k,
+                                                      a1_t_k_k,
+                                                      b0_e_n_k,
+                                                      b1_e_n_k,
+                                                      d2_e_n, // topk weights
+                                                      c_t_n,
+                                                      PassThrough{},
+                                                      PassThrough{},
+                                                      cde_element_op);
+
+        ref_invoker.Run(ref_argument);
+        for(int t = 0; t < tokens; ++t)
+        {
+            for(int n = 0; n < N; ++n)
+            {
+                e_t_n_host_result(t, n) = ck::type_convert<EDataType>(c_t_n(t, n));
+            }
+        }
+
+        e_device_buf.FromDevice(e_t_n_device_result.mData.data());
+
+#if 1
+        printf("e_t_n_device_result:\n");
+        for(int t = 0; t < tokens; ++t)
+        {
+            for(int n = 0; n < N; ++n)
+            {
+                printf("%.2f ", ck::type_convert<float>(e_t_n_device_result(t, n)));
+            }
+            printf("\n");
+        }
+
+        printf("e_t_n_host_result:\n");
+        for(int t = 0; t < tokens; ++t)
+        {
+            for(int n = 0; n < N; ++n)
+            {
+                printf("%.2f ", ck::type_convert<float>(e_t_n_host_result(t, n)));
+            }
+            printf("\n");
+        }
+#endif
+
+        return ck::utils::check_err(
+                   e_t_n_device_result, e_t_n_host_result, "Error: Incorrect results!", 1e-3, 5e-2)
+                   ? 0
+                   : 1;
+    }
+
+    return 0;
+}

example/67_gemm_microscaling/moe_gemm2_xdl_mx_fp4_bns.cpp

@@ -0,0 +1,658 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_moe_mx_gemm_bns.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp"
+
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_moe_mx_gemm2.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/fill.hpp"
+#include "ck/utility/blkgemmpipe_scheduler.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F4              = ck::f4x2_pk_t;
+using F16             = ck::half_t;
+using BF16            = ck::bhalf_t;
+using F32             = float;
+using XDataType       = ck::e8m0_bexp_t;
+using XPackedDataType = int32_t; // 4 packed e8m0_bexp_t
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using A0DataType       = F4;
+using A1DataType       = XPackedDataType;
+using B0DataType       = F4;
+using B1DataType       = XPackedDataType;
+using EDataType        = F16;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using D0DataType       = F32;
+using D1DataType       = F32;
+using D2DataType       = F32;
+using DsDataType       = ck::Tuple<D0DataType, D1DataType, D2DataType>;
+
+using A0Layout = Row;
+using B0Layout = Col;
+using ELayout  = Row;
+using D0Layout = Row;
+using D1Layout = Col;
+using D2Layout = ELayout;
+using DsLayout = ck::Tuple<D0Layout, D1Layout, D2Layout>;
+
+// d0: ascale, d1: bscale, d2:expert weight
+struct MulABScaleExpertWeight
+{
+    template <typename E, typename C, typename D0, typename D1, typename D2>
+    __host__ __device__ constexpr void
+    operator()(E& e, const C& c, const D0& d0, const D1& d1, const D2& d2) const;
+    // for real kernel use
+    template <>
+    __host__ __device__ constexpr void operator()<EDataType, float, float, float, float>(
+        EDataType& e, const float& c, const float& d0, const float& d1, const float& d2) const
+    {
+        (void)d0;
+        (void)d1;
+        (void)d2;
+
+        e = ck::type_convert<EDataType>(c);
+    }
+    // for reference cpu
+    template <>
+    __host__ __device__ constexpr void operator()<float, float, float, float, float>(
+        float& e, const float& c, const float& d0, const float& d1, const float& d2) const
+    {
+        // for reference cpu
+        e = ck::type_convert<EDataType>(c * d0 * d1 * d2);
+    }
+};
+
+using CDEElementOp = MulABScaleExpertWeight;
+
+// A, B Scale preshuffle
+template <bool KLast>
+void preShuffleScaleBuffer(ck::e8m0_bexp_t* src, ck::e8m0_bexp_t* dst, int MN, int K)
+{
+    int MNXdlPack = 2;
+    int KXdlPack  = 2;
+
+    int XdlMNThread = 16;
+    int XdlKThread  = 64 / XdlMNThread;
+
+    int K0 = K / KXdlPack / XdlKThread; // KRepeat
+
+    // The 4 16x128 building blocks will be packed into 1 32x256 for F4
+    // The 8 16x16x128 mfma will be packed into 1 32x32x256 for F4
+
+    // unfold the MN32xK(256/32) scale buffer
+    //    4            16             2           2
+    // To XdlKThread-> XdlMNThread -> KXdlPack -> MNXdlPack
+    // Then, MNRepeat->KRepeat
+
+    for(int n = 0; n < MN; ++n)
+    {
+        for(int k = 0; k < K; ++k)
+        {
+            int n0    = n / (XdlMNThread * MNXdlPack); // i MNRepeat
+            int tempn = n % (XdlMNThread * MNXdlPack);
+            int n1    = tempn % XdlMNThread; // i XdlMNThread
+            int n2    = tempn / XdlMNThread; // i MNXdlPack
+
+            int k0    = k / (XdlKThread * KXdlPack); // i KRepeat
+            int tempk = k % (XdlKThread * KXdlPack);
+            int k1    = tempk % XdlKThread; // i XdlKThread
+            int k2    = tempk / XdlKThread; // i KXdlPack
+
+            int outputIndex = n0 * MNXdlPack * KXdlPack * XdlMNThread * XdlKThread * K0 +
+                              k0 * MNXdlPack * KXdlPack * XdlMNThread * XdlKThread +
+                              k1 * MNXdlPack * KXdlPack * XdlMNThread + n1 * MNXdlPack * KXdlPack +
+                              k2 * MNXdlPack + n2;
+            // src[n * K + k] = ck::type_convert<ck::e8m0_bexp_t>(static_cast<float>(powf(2.0f, n2 +
+            // k2 * MNXdlPack)));
+            if constexpr(KLast)
+                dst[outputIndex] = src[n * K + k];
+            else
+                dst[outputIndex] = src[k * MN + n];
+        }
+    }
+}
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = MulABScaleExpertWeight;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::Default;
+
+constexpr ck::index_t DataPackedSize = 2;                    // Packed representation of data
+constexpr ck::index_t ScaleBlockSize = 32;                   // scaling block size
+constexpr ck::index_t KPerBlock      = 256 / DataPackedSize; // 256 f4 = 128 fp4x2
+
+static constexpr ck::index_t MPerBlock = 128;
+static constexpr bool MulRoutedWeight  = true;
+
+// clang-format off
+using DeviceOpInstance                     = ck::tensor_operation::device::DeviceMoeGemmMXBNS<      
+    A0Layout,    B0Layout,    DsLayout,    ELayout, 
+    A0DataType,  A1DataType,  B0DataType,  B1DataType,  DsDataType, EDataType, AccDataType, CShuffleDataType,
+    AElementOp,  BElementOp, CDEElementOp, GemmSpec,   
+    ScaleBlockSize,      256,   
+    MPerBlock,  128,    KPerBlock,
+    16,   16,
+    16,   16,
+    4,    4,
+    S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0,
+    S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0,
+    2,    2,   S<1, 32, 1, 8>, S<2, 1, 1, 1>,
+    ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v3, 0, false, false, MulRoutedWeight, ck::index_t, A0DataType>;
+// clang-format on
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = true;
+
+    // per expert:
+    // GEMM shape
+    constexpr ck::index_t sorted_tile_num = 13;
+    constexpr ck::index_t valid_tile_num  = sorted_tile_num;
+    ck::index_t sorted_size               = sorted_tile_num * MPerBlock;
+    ck::index_t valid_size                = valid_tile_num * MPerBlock;
+
+    ck::index_t N       = 6144;
+    ck::index_t K       = 4096;
+    ck::index_t experts = 8;
+    ck::index_t tokens  = 832;
+    ck::index_t topk    = 2;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 3)
+    {
+        // use default case
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 7)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+        N               = std::stoi(argv[4]);
+        K               = std::stoi(argv[5]);
+        tokens          = std::stoi(argv[6]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=no, 1=yes)\n");
+        printf("arg4 to 6: N, K, tokens\n");
+        exit(0);
+    }
+
+    if(K % ScaleBlockSize != 0)
+    {
+        throw std::runtime_error("wrong! K must be multiple of ScaleBlockSize.");
+    };
+
+    ck::index_t StrideA              = K;
+    ck::index_t StrideB              = K;
+    ck::index_t StrideE              = N;
+    ck::index_t Scale_Stride_AM      = (K + ScaleBlockSize - 1) / ScaleBlockSize;
+    ck::index_t Scale_Stride_BN      = (K + ScaleBlockSize - 1) / ScaleBlockSize;
+    constexpr ck::index_t NumDTensor = DsDataType::Size();
+    constexpr auto StrideDs          = std::array<ck::index_t, NumDTensor>{0, 0, 0};
+
+    ck::index_t KBatch = 1;
+
+    Tensor<ck::index_t> expert_ids(HostTensorDescriptor({sorted_tile_num}, {1}));
+    Tensor<ck::index_t> sorted_token_ids(HostTensorDescriptor({sorted_size}, {1}));
+    Tensor<ck::index_t> max_token_id(HostTensorDescriptor({1}));
+    max_token_id.mData[0] = valid_size;
+    // int eids[]            = {0, 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 3, 3, 3};
+    int eids[sorted_tile_num]{};
+    for(int i = 0; i < sorted_tile_num; i++)
+    {
+        if(i < valid_tile_num)
+        {
+            eids[i] = (i * experts) / valid_tile_num;
+        }
+        else
+        {
+            eids[i] = 3;
+        }
+    }
+
+    for(int i = 0; i < sorted_tile_num; i++)
+    {
+        expert_ids.mData[i] = eids[i];
+    }
+    if(tokens * topk > valid_size)
+    {
+        printf("err config, tokens * topk > valid_size\n");
+        exit(-1);
+    }
+    int token_per_tile = tokens * topk / valid_tile_num;
+    int tokenid        = 0;
+    for(int i = 0; i < sorted_size; i++)
+    {
+        int tile_off = i % MPerBlock;
+        if(tile_off < token_per_tile)
+        {
+            sorted_token_ids.mData[i] = (tokenid % tokens) | ((tokenid / tokens) << 24);
+            tokenid++;
+        }
+        else
+        {
+            sorted_token_ids.mData[i] = tokens;
+        }
+    }
+
+    expert_ids.savetxt("expert_ids.txt", "int");
+    sorted_token_ids.savetxt("sorted_token_ids.txt", "int");
+    Tensor<A0DataType> a0_t_k_k(HostTensorDescriptor({tokens, topk, K}, {topk * K, K, 1}));
+    Tensor<XDataType> a1_t_k_k(
+        HostTensorDescriptor({tokens, topk, (K + ScaleBlockSize - 1) / ScaleBlockSize},
+                             {(topk * Scale_Stride_AM), Scale_Stride_AM, 1}));
+    Tensor<B0DataType> b0_e_n_k(HostTensorDescriptor({experts, K, N}, {N * K, 1, K}));
+    Tensor<XDataType> b1_e_n_k(
+        HostTensorDescriptor({experts, (K + ScaleBlockSize - 1) / ScaleBlockSize, N},
+                             {(N * Scale_Stride_BN), 1, Scale_Stride_BN}));
+    // B preshuffle
+    Tensor<B0DataType> b0_preshuffled(HostTensorDescriptor({experts, K, N}, {N * K, 1, K}));
+
+    // A, B Scale preshuffle
+    Tensor<XDataType> a_scale_sorted(HostTensorDescriptor(
+        {sorted_size, (K + ScaleBlockSize - 1) / ScaleBlockSize}, {Scale_Stride_AM, 1}));
+    Tensor<XDataType> a_scale_preshuffled(HostTensorDescriptor(
+        {sorted_size, (K + ScaleBlockSize - 1) / ScaleBlockSize}, {Scale_Stride_AM, 1}));
+    Tensor<XDataType> b_scale_preshuffled(
+        HostTensorDescriptor({experts, (K + ScaleBlockSize - 1) / ScaleBlockSize, N},
+                             {N * Scale_Stride_BN, 1, Scale_Stride_BN}));
+    Tensor<D2DataType> d2_e_n(HostTensorDescriptor({sorted_size, N}, {1, 0}));
+    Tensor<EDataType> e_t_n_host_result(HostTensorDescriptor({tokens, N}, {N, 1}));
+    Tensor<EDataType> e_t_n_device_result(HostTensorDescriptor({tokens, N}, {N, 1}));
+
+    e_t_n_device_result.SetZero();
+    std::cout << "a0_t_k_k: " << a0_t_k_k.mDesc << std::endl;
+    std::cout << "a1_t_k_k: " << a1_t_k_k.mDesc << std::endl;
+    std::cout << "b0_e_n_k: " << b0_e_n_k.mDesc << std::endl;
+    std::cout << "b1_e_n_k: " << b1_e_n_k.mDesc << std::endl;
+    std::cout << "d2_e_n: " << d2_e_n.mDesc << std::endl;
+    std::cout << "e_t_n: " << e_t_n_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a0_t_k_k.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-1, 1});
+        b0_e_n_k.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-1, 1});
+        a1_t_k_k.GenerateTensorValue(GeneratorTensor_3<XDataType>{0, 1.0});
+        b1_e_n_k.GenerateTensorValue(GeneratorTensor_3<XDataType>{0, 1.0});
+        d2_e_n.GenerateTensorValue(GeneratorTensor_3<D2DataType>{0, 1.0});
+        break;
+    case 2:
+        a0_t_k_k.GenerateTensorValue(GeneratorTensor_1<A0DataType>{});
+        b0_e_n_k.GenerateTensorValue(GeneratorTensor_1<B0DataType>{});
+        a1_t_k_k.GenerateTensorValue(GeneratorTensor_1<XDataType>{});
+        b1_e_n_k.GenerateTensorValue(GeneratorTensor_1<XDataType>{});
+        d2_e_n.GenerateTensorValue(GeneratorTensor_1<D2DataType>{});
+        break;
+    case 3:
+        a0_t_k_k.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-1, 1});
+        b0_e_n_k.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-1, 1});
+        a1_t_k_k.GenerateTensorValue(GeneratorTensor_3<XDataType>{0, 1.0});
+        b1_e_n_k.GenerateTensorValue(GeneratorTensor_3<XDataType>{0, 1.0});
+        d2_e_n.GenerateTensorValue(GeneratorTensor_1<D2DataType>{});
+        break;
+    case 4:
+        a0_t_k_k.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-2, 2});
+        b0_e_n_k.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
+        a1_t_k_k.GenerateTensorValue(GeneratorTensor_1<XDataType>{});
+        b1_e_n_k.GenerateTensorValue(GeneratorTensor_3<XDataType>{0, 5.0});
+        d2_e_n.GenerateTensorValue(GeneratorTensor_1<D2DataType>{});
+        break;
+    case 5:
+        a0_t_k_k.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-2, 2});
+        b0_e_n_k.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
+        a1_t_k_k.GenerateTensorValue(GeneratorTensor_3<XDataType>{0, 1.0});
+        b1_e_n_k.GenerateTensorValue(GeneratorTensor_1<XDataType>{});
+        d2_e_n.GenerateTensorValue(GeneratorTensor_1<D2DataType>{1});
+        break;
+    case 6:
+        a0_t_k_k.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-2, 2});
+        b0_e_n_k.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
+        a1_t_k_k.GenerateTensorValue(GeneratorTensor_3<XDataType>{0, 1.0});
+        b1_e_n_k.GenerateTensorValue(GeneratorTensor_1<XDataType>{});
+        d2_e_n.GenerateTensorValue(GeneratorTensor_1<D2DataType>{});
+        break;
+    default:
+        a0_t_k_k.GenerateTensorValue(GeneratorTensor_3<A0DataType>{0.0, 1.0});
+        b0_e_n_k.GenerateTensorValue(GeneratorTensor_3<B0DataType>{-0.5, 0.5});
+        a1_t_k_k.GenerateTensorValue(GeneratorTensor_3<XDataType>{0.0, 1.0});
+        b1_e_n_k.GenerateTensorValue(GeneratorTensor_3<XDataType>{0.0, 1.0});
+        d2_e_n.GenerateTensorValue(GeneratorTensor_3<D2DataType>{0.0, 1.0});
+    }
+    DeviceMem sorted_token_ids_dev(sizeof(ck::index_t) * sorted_token_ids.GetElementSpaceSize());
+    DeviceMem expert_ids_dev(sizeof(ck::index_t) * expert_ids.GetElementSpaceSize());
+    DeviceMem max_token_id_dev(sizeof(ck::index_t) * max_token_id.GetElementSpaceSize());
+    DeviceMem a0_device_buf(sizeof(A0DataType) * a0_t_k_k.GetElementSpaceSize());
+    DeviceMem a1_device_buf(sizeof(XDataType) * a_scale_sorted.GetElementSpaceSize());
+    DeviceMem b0_device_buf(sizeof(B0DataType) * b0_e_n_k.GetElementSpaceSize());
+    DeviceMem b1_device_buf(sizeof(XDataType) * b1_e_n_k.GetElementSpaceSize());
+    DeviceMem d2_device_buf(sizeof(D2DataType) * d2_e_n.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) * e_t_n_device_result.GetElementSpaceSize());
+    // d2_e_n.savetxt("weight.txt", "int");
+
+    // A scale sorted
+    for(int i = 0; i < sorted_size; i++)
+    {
+        int token_id = sorted_token_ids.mData[i] & 0x00FFFFFF;
+        int topk_id  = (sorted_token_ids.mData[i] >> 24) & 0x000000FF;
+
+        for(int k = 0; k < (K + ScaleBlockSize - 1) / ScaleBlockSize; k++)
+        {
+            if(token_id == tokens)
+            {
+                a_scale_sorted(i, k) = ck::type_convert<XDataType>(0);
+            }
+            else
+            {
+                a_scale_sorted(i, k) = a1_t_k_k(token_id, topk_id, k);
+            }
+        }
+    }
+
+    preShuffleScaleBuffer<ck::is_same_v<A0Layout, Row>>(a_scale_sorted.mData.data(),
+                                                        a_scale_preshuffled.mData.data(),
+                                                        sorted_size,
+                                                        K / ScaleBlockSize);
+    preShuffleScaleBuffer<ck::is_same_v<B0Layout, Col>>(
+        b1_e_n_k.mData.data(), b_scale_preshuffled.mData.data(), N * experts, K / ScaleBlockSize);
+
+    sorted_token_ids_dev.ToDevice(sorted_token_ids.mData.data());
+    expert_ids_dev.ToDevice(expert_ids.mData.data());
+    max_token_id_dev.ToDevice(max_token_id.mData.data());
+    a0_device_buf.ToDevice(a0_t_k_k.mData.data());
+    b0_device_buf.ToDevice(b0_e_n_k.mData.data());
+    a1_device_buf.ToDevice(a_scale_preshuffled.mData.data());
+    b1_device_buf.ToDevice(b_scale_preshuffled.mData.data());
+    d2_device_buf.ToDevice(d2_e_n.mData.data());
+    e_device_buf.ToDevice(e_t_n_device_result.mData.data());
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{};
+
+#if 0
+    printf("a0_t_k_k:\n");
+    for(int t = 0; t < tokens; ++t)
+    {
+        for(int tk = 0; tk < topk; ++tk)
+        {
+            for(int k = 0; k < K; ++k)
+            {
+                auto f4x2 = a0_t_k_k(t, tk, k).data;
+                if(k % 2 == 0)
+                {
+                    ck::f4_t f4 = (f4x2 >> 4) & 0xf;
+                    printf("%.2f ", ck::type_convert<float>(f4));
+                }
+                else
+                {
+                    ck::f4_t f4 = (f4x2 >> 0) & 0xf;
+                    printf("%.2f ", ck::type_convert<float>(f4));
+                }
+            }
+            printf("\n");
+        }
+        printf("\n");
+    }
+
+    printf("a1_t_k_k:\n");
+    for(int t = 0; t < tokens; ++t)
+    {
+        for(int tk = 0; tk < topk; ++tk)
+        {
+            for(int k = 0; k < (K + ScaleBlockSize - 1) / ScaleBlockSize; ++k)
+            {
+                printf("%.2f ", ck::type_convert<float>(a1_t_k_k(t, tk, k)));
+            }
+            printf("\n");
+        }
+        printf("\n");
+    }
+
+    printf("a_scale_sorted: K/scale: %d\n", (K + ScaleBlockSize - 1) / ScaleBlockSize);
+    for(int i = 0; i < sorted_size; ++i)
+    {
+        for(int k = 0; k < (K + ScaleBlockSize - 1) / ScaleBlockSize; ++k)
+        {
+            printf("%.2f ", ck::type_convert<float>(a_scale_sorted(i, k)));
+        }
+        printf("\n");
+    }
+
+    printf("a_scale_preshuffled:\n");
+    for(int i = 0; i < sorted_size; ++i)
+    {
+        for(int k = 0; k < (K + ScaleBlockSize - 1) / ScaleBlockSize; ++k)
+        {
+            printf("%.2f ", ck::type_convert<float>(a_scale_preshuffled(i, k)));
+        }
+        printf("\n");
+    }
+
+    printf("b0_e_n_k:\n");
+    for(int e = 0; e < experts; ++e)
+    {
+        for(int n = 0; n < N; ++n)
+        {
+            for(int k = 0; k < K; ++k)
+            {
+                auto f4x2 = b0_e_n_k(e, k, n).data;
+                if(k % 2 == 0)
+                {
+                    ck::f4_t f4 = f4x2 >> 4 & 0xf;
+                    printf("%.2f ", ck::type_convert<float>(f4));
+                }
+                else
+                {
+                    ck::f4_t f4 = f4x2 >> 0 & 0xf;
+                    printf("%.2f ", ck::type_convert<float>(f4));
+                }
+            }
+            printf("\n");
+        }
+        printf("\n");
+    }
+
+    printf("b1_e_n_k:\n");
+    for(int e = 0; e < experts; ++e)
+    {
+        for(int k = 0; k < (K + ScaleBlockSize - 1) / ScaleBlockSize; ++k)
+        {
+            for(int n = 0; n < N; ++n)
+            {
+                printf("%.2f ", ck::type_convert<float>(b1_e_n_k(e, k, n)));
+            }
+            printf("\n");
+        }
+        printf("\n");
+    }
+
+    printf("d2_e_n:\n");
+    for(int i = 0; i < sorted_size; ++i)
+    {
+        for(int n = 0; n < 1; ++n)
+        {
+            printf("%.2f ", ck::type_convert<float>(d2_e_n(i, n)));
+        }
+    }
+#endif
+
+    // do GEMM
+    auto device_op = DeviceOpInstance{};
+
+    auto invoker  = device_op.MakeInvoker();
+    auto argument = device_op.MakeArgument(
+        sorted_token_ids_dev.GetDeviceBuffer(),
+        expert_ids_dev.GetDeviceBuffer(),
+        max_token_id_dev.GetDeviceBuffer(),
+        a0_device_buf.GetDeviceBuffer(),
+        a1_device_buf.GetDeviceBuffer(),
+        b0_device_buf.GetDeviceBuffer(),
+        b1_device_buf.GetDeviceBuffer(),
+        std::array<const void*, NumDTensor>{nullptr, nullptr, d2_device_buf.GetDeviceBuffer()},
+        e_device_buf.GetDeviceBuffer(),
+        tokens,
+        topk,
+        sorted_size,
+        N,
+        K,
+        StrideA,
+        Scale_Stride_AM,
+        StrideB,
+        Scale_Stride_BN,
+        StrideDs,
+        StrideE,
+        KBatch,
+        a_element_op,
+        b_element_op,
+        cde_element_op);
+
+    if(!device_op.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_gemm with the specified compilation parameters does "
+            "not support this GEMM problem");
+    }
+
+    if(!(ck::get_device_name() == "gfx942" || ck::get_device_name() == "gfx950"))
+    {
+        std::cout << "This kernel support gfx942 and gfx950 only" << std::endl;
+    }
+
+    if(time_kernel)
+    {
+        // not result correct here because output buf not setzero
+        float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+        // FMA * tokens * N * topk * K +
+        // FMA * tokens * N * topk * (K/BlockScale)
+        std::size_t flop = std::size_t(2) * tokens * topk * N * K +
+                           std::size_t(2) * tokens * topk * N * K / ScaleBlockSize;
+
+        std::size_t num_btype =
+            sizeof(A0DataType) / 2 * tokens * K * topk + sizeof(B0DataType) / 2 * K * N * experts +
+            sizeof(XDataType) * tokens * topk * K / ScaleBlockSize +
+            sizeof(XDataType) * K / ScaleBlockSize * N * experts + sizeof(EDataType) * tokens * N;
+
+        float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+        float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+        std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
+                  << " GB/s" << device_op.GetTypeString() << std::endl;
+    }
+
+    if(do_verification)
+    {
+        // gemm2 use atomic, so need to reinit outputs
+        e_device_buf.ToDevice(e_t_n_device_result.mData.data());
+        invoker.Run(argument, StreamConfig{nullptr, false, 0, 0, 1});
+
+        Tensor<CShuffleDataType> c_t_n({tokens, N});
+
+        using ReferenceGemmInstance =
+            ck::tensor_operation::host::ReferenceMoeMXGemm2<A0DataType,
+                                                            XDataType,
+                                                            B0DataType,
+                                                            XDataType,
+                                                            D2DataType,
+                                                            CShuffleDataType,
+                                                            AccDataType,
+                                                            PassThrough,
+                                                            PassThrough,
+                                                            CDEElementOp,
+                                                            MulRoutedWeight,
+                                                            float,
+                                                            float>;
+
+        auto ref_moe_gemm = ReferenceGemmInstance{};
+        auto ref_invoker  = ref_moe_gemm.MakeInvoker();
+        auto ref_argument = ref_moe_gemm.MakeArgument(sorted_token_ids,
+                                                      expert_ids,
+                                                      max_token_id,
+                                                      MPerBlock,
+                                                      a0_t_k_k,
+                                                      a1_t_k_k,
+                                                      b0_e_n_k,
+                                                      b1_e_n_k,
+                                                      d2_e_n, // topk weights
+                                                      c_t_n,
+                                                      PassThrough{},
+                                                      PassThrough{},
+                                                      cde_element_op);
+
+        ref_invoker.Run(ref_argument);
+        for(int t = 0; t < tokens; ++t)
+        {
+            for(int n = 0; n < N; ++n)
+            {
+                e_t_n_host_result(t, n) = ck::type_convert<EDataType>(c_t_n(t, n));
+            }
+        }
+
+        e_device_buf.FromDevice(e_t_n_device_result.mData.data());
+
+#if 0
+        printf("e_t_n_device_result:\n");
+        for(int t = 0; t < tokens; ++t)
+        {
+            for(int n = 0; n < N; ++n)
+            {
+                printf("%.2f ", ck::type_convert<float>(e_t_n_device_result(t, n)));
+            }
+            printf("\n");
+        }
+
+        printf("e_t_n_host_result:\n");
+        for(int t = 0; t < tokens; ++t)
+        {
+            for(int n = 0; n < N; ++n)
+            {
+                printf("%.2f ", ck::type_convert<float>(e_t_n_host_result(t, n)));
+            }
+            printf("\n");
+        }
+#endif
+
+        return ck::utils::check_err(
+                   e_t_n_device_result, e_t_n_host_result, "Error: Incorrect results!", 1e-3, 5e-2)
+                   ? 0
+                   : 1;
+    }
+
+    return 0;
+}

include/ck/library/utility/check_err.hpp

@@ -181,7 +181,7 @@ check_err(const Range& out,
         {
             max_err = err > max_err ? err : max_err;
             err_count++;
-            if(err_count < 500)
+            if(err_count < 5)
             {
                 std::cerr << msg << std::setw(12) << std::setprecision(7) << " out[" << i
                           << "] != ref[" << i << "]: " << o << " != " << r << std::endl;
@@ -232,7 +232,7 @@ check_err(const Range& out,
         {
             max_err = err > max_err ? err : max_err;
             err_count++;
-            if(err_count < 500)
+            if(err_count < 5)
             {
                 std::cerr << msg << std::setw(12) << std::setprecision(7) << " out[" << i
                           << "] != ref[" << i << "]: " << o << " != " << r << std::endl;
@@ -282,7 +282,7 @@ check_err(const Range& out,
         {
             max_err = err > max_err ? err : max_err;
             err_count++;
-            if(err_count < 500)
+            if(err_count < 5)
             {
                 std::cerr << msg << std::setw(12) << std::setprecision(7) << " out[" << i
                           << "] != ref[" << i << "]: " << o << " != " << r << std::endl;
@@ -339,7 +339,7 @@ check_err(const Range& out,
         {
             max_err = err > max_err ? err : max_err;
             err_count++;
-            if(err_count < 500)
+            if(err_count < 5)
             {
                 std::cerr << msg << " out[" << i << "] != ref[" << i << "]: " << o << " != " << r
                           << std::endl;
@@ -390,7 +390,7 @@ check_err(const Range& out,
         {
             max_err = err > max_err ? err : max_err;
             err_count++;
-            if(err_count < 500)
+            if(err_count < 5)
             {
                 std::cerr << msg << std::setw(12) << std::setprecision(7) << " out[" << i
                           << "] != ref[" << i << "]: " << o << " != " << r << std::endl;
@@ -437,7 +437,7 @@ check_err(const Range& out,
         {
             max_err = err > max_err ? err : max_err;
             err_count++;
-            if(err_count < 500)
+            if(err_count < 5)
             {
                 std::cerr << msg << std::setw(12) << std::setprecision(7) << " out[" << i
                           << "] != ref[" << i << "]: " << o << " != " << r << std::endl;
@@ -484,7 +484,7 @@ check_err(const Range& out,
         {
             max_err = err > max_err ? err : max_err;
             err_count++;
-            if(err_count < 500)
+            if(err_count < 5)
             {
                 std::cerr << msg << std::setw(12) << std::setprecision(7) << " out[" << i
                           << "] != ref[" << i << "]: " << o << " != " << r << std::endl;

include/ck/library/utility/host_tensor_generator.hpp

@@ -163,6 +163,18 @@ struct GeneratorTensor_1<ck::pk_i4_t>
     }
 };
 
+template <>
+struct GeneratorTensor_1<ck::e8m0_bexp_t>
+{
+    float value = 1;
+
+    template <typename... Is>
+    ck::e8m0_bexp_t operator()(Is...)
+    {
+        return ck::type_convert<ck::e8m0_bexp_t>(value);
+    }
+};
+
 template <typename T>
 struct GeneratorTensor_2
 {

include/ck/tensor_operation/gpu/block/blockwise_gemm_mx_pipeline_xdlops_base.hpp

@@ -53,7 +53,9 @@ struct BlockwiseGemmXdlops_mx_pipeline_base
     static constexpr index_t A_K0 = ATileDesc{}.GetLength(I0);
     static constexpr index_t B_K0 = BTileDesc{}.GetLength(I0);
     static constexpr index_t A_K1 = ATileDesc{}.GetLength(I2);
-    static constexpr index_t B_K1 = BTileDesc{}.GetLength(I2);
+    // static constexpr index_t B_K1 = BTileDesc{}.GetLength(I2);
+    static constexpr index_t B_K1 =
+        BTileDesc{}.GetLength(Number < BTileDesc{}.GetNumOfDimension() == 4 ? 3 : 2 > {});
 
     static constexpr auto xdlops_gemm = XdlopsGemm<ComputeTypeA,
                                                    MPerXDL,
@@ -373,6 +375,8 @@ struct BlockwiseGemmXdlops_mx_pipeline_base
             c_grid_desc_g_m0_n0_m1_n1_m2_n2);
     }
 
+    __host__ __device__ static constexpr auto GetCThreadDesc() { return c_thread_desc_; }
+
     static constexpr AMmaTileDesc a_block_desc_m0_m1_m2_m3_k;
     static constexpr BMmaTileDesc b_block_desc_n0_n1_n2_n3_k;
 

include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuffle_gufusion_v3.hpp

@@ -0,0 +1,973 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_base.hpp"
+
+namespace ck {
+
+// Compute optimized pipeline
+// GlobalPrefetchStages: 2
+// LocalPreFillStages: 1
+// LocalPreFetchStages: 1
+// LocalSharedMemoryBuffer: 1
+
+template <BlockGemmPipelineScheduler BlkGemmPipelineVer,
+          index_t BlockSize,
+          typename ADataType,
+          typename BDataType,
+          typename ComputeDataType,
+          typename AccDataType,
+          typename ATileDesc,
+          typename BTileDesc,
+          typename AMmaTileDesc,
+          typename BMmaTileDesc,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MRepeat,
+          index_t NRepeat,
+          index_t KPacks>
+struct BlockwiseGemmXdlops_pipeline_bpreshuffle_gufusion_v3
+{
+};
+
+template <index_t BlockSize,
+          typename ADataType,
+          typename BDataType,
+          typename ComputeDataType,
+          typename AccDataType,
+          typename ATileDesc,
+          typename BTileDesc,
+          typename AMmaTileDesc,
+          typename BMmaTileDesc,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MRepeat,
+          index_t NRepeat,
+          index_t KPack
+          // ,bool TransposeC //disable transposec right now...
+          >
+struct BlockwiseGemmXdlops_pipeline_bpreshuffle_gufusion_v3<BlockGemmPipelineScheduler::Intrawave,
+                                                            BlockSize,
+                                                            ADataType,
+                                                            BDataType,
+                                                            ComputeDataType,
+                                                            AccDataType,
+                                                            ATileDesc,
+                                                            BTileDesc,
+                                                            AMmaTileDesc,
+                                                            BMmaTileDesc,
+                                                            ABlockTransferSrcScalarPerVector,
+                                                            BBlockTransferSrcScalarPerVector,
+                                                            MPerBlock,
+                                                            NPerBlock,
+                                                            KPerBlock,
+                                                            MPerXDL,
+                                                            NPerXDL,
+                                                            MRepeat,
+                                                            NRepeat,
+                                                            KPack>
+    : BlockwiseGemmXdlops_pipeline_base<BlockSize,
+                                        ADataType,
+                                        BDataType,
+                                        ComputeDataType,
+                                        AccDataType,
+                                        ATileDesc,
+                                        BTileDesc,
+                                        AMmaTileDesc,
+                                        BMmaTileDesc,
+                                        ABlockTransferSrcScalarPerVector,
+                                        BBlockTransferSrcScalarPerVector,
+                                        MPerBlock,
+                                        NPerBlock,
+                                        KPerBlock,
+                                        MPerXDL,
+                                        NPerXDL,
+                                        MRepeat,
+                                        NRepeat,
+                                        KPack>
+
+{
+    using Base = BlockwiseGemmXdlops_pipeline_base<BlockSize,
+                                                   ADataType,
+                                                   BDataType,
+                                                   ComputeDataType,
+                                                   AccDataType,
+                                                   ATileDesc,
+                                                   BTileDesc,
+                                                   AMmaTileDesc,
+                                                   BMmaTileDesc,
+                                                   ABlockTransferSrcScalarPerVector,
+                                                   BBlockTransferSrcScalarPerVector,
+                                                   MPerBlock,
+                                                   NPerBlock,
+                                                   KPerBlock,
+                                                   MPerXDL,
+                                                   NPerXDL,
+                                                   MRepeat,
+                                                   NRepeat,
+                                                   KPack>;
+    using Base::A_K1;
+    using Base::B_K1;
+    using Base::I0;
+    using Base::I1;
+    using Base::I2;
+    using Base::KGroup;
+    using Base::KRepeat;
+    using Base::xdlops_gemm;
+    using typename Base::HotLoopInstList;
+
+    using Base::a_block_desc_m0_m1_m2_k;
+    using Base::CalculateCThreadOriginDataIndex;
+    using Base::CalculateCThreadOriginDataIndex8D;
+    using Base::GetCBlockDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2;
+    using Base::GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2;
+    using Base::GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4;
+    using Base::GetCThreadBuffer;
+    using Base::GetCThreadDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2;
+    using Base::GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2;
+    using Base::GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4;
+    using Base::MakeCGridDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2;
+    using Base::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2;
+
+    using Base::AMmaKStride;
+    using Base::BMmaKStride;
+
+    using Base::MWaves;
+
+    static constexpr index_t PrefetchStages        = 2;
+    static constexpr index_t PrefillStages         = 1;
+    static constexpr index_t GlobalBufferNum       = 1;
+    static constexpr index_t HotloopLocalBufSwitch = MRepeat % 2 == 0 ? 0 : 1;
+
+    template <typename TileDesc_M0_M1_M2_K>
+    __host__ __device__ static constexpr auto MakeAGemmMmaTileDescriptor(const TileDesc_M0_M1_M2_K&)
+    {
+        constexpr index_t M0 = TileDesc_M0_M1_M2_K{}.GetLength(Number<0>{});
+        constexpr index_t M1 = TileDesc_M0_M1_M2_K{}.GetLength(Number<1>{});
+        constexpr index_t M2 = TileDesc_M0_M1_M2_K{}.GetLength(Number<2>{});
+        constexpr index_t K2 = KPack / KGroup;
+        constexpr index_t K1 = 64 / NPerXDL;
+        constexpr index_t K0 = KRepeat * KGroup;
+
+        return transform_tensor_descriptor(
+            TileDesc_M0_M1_M2_K{},
+            make_tuple(
+                make_pass_through_transform(Number<M0>{}),
+                make_pass_through_transform(Number<M1>{}),
+                make_pass_through_transform(Number<M2>{}),
+                make_unmerge_transform(make_tuple(Number<K0>{}, Number<K1>{}, Number<K2>{}))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3, 4, 5>{}));
+    }
+
+    static constexpr auto a_block_desc_m0_m1_m2_k0_k1_k2 =
+        MakeAGemmMmaTileDescriptor(a_block_desc_m0_m1_m2_k);
+
+    __host__ __device__ static constexpr bool BlockHasHotloop(index_t num_loop)
+    {
+        return num_loop > PrefetchStages;
+    }
+
+    __host__ __device__ static constexpr TailNumber BlockLoopTailNum(index_t num_loop)
+    {
+        return num_loop % 2 == 0 ? TailNumber::Even : TailNumber::Odd;
+    }
+
+    __device__ static constexpr auto HotLoopScheduler()
+    {
+        // A/B split schedule
+        // compiler is likely to use ds_read2 when instruction width smaller than 16bytes
+        constexpr auto num_ds_read_inst_a =
+            HotLoopInstList::A_LDS_Read_Width * sizeof(ADataType) == 16
+                ? HotLoopInstList::A_LDS_Read_Inst_Num
+                : HotLoopInstList::A_LDS_Read_Inst_Num / 2;
+
+        constexpr auto num_ds_write_inst_a = HotLoopInstList::A_LDS_Write_Inst_Num;
+
+        constexpr auto num_buffer_load_inst_a = HotLoopInstList::A_Buffer_Load_Inst_Num;
+        constexpr auto num_buffer_load_inst_b = HotLoopInstList::B_Buffer_Load_Inst_Num * 2;
+
+        static_assert(num_buffer_load_inst_a == num_ds_write_inst_a);
+
+        constexpr auto num_mfma_inst = HotLoopInstList::C_MFMA_Inst_Num * 2;
+        constexpr auto mfma_cycle    = HotLoopInstList::C_MFMA_Inst_Cycle;
+
+        constexpr auto ds_read_a_issue_cycle =
+            HotLoopInstList::A_LDS_Read_Width * sizeof(ADataType) == 16 ? 8 : 4;
+        constexpr auto ds_read_a_mfma_rate =
+            math::integer_divide_ceil(mfma_cycle - 4, 2 * ds_read_a_issue_cycle);
+
+        // constexpr auto num_dsread_a_mfma =
+        //     (num_ds_read_inst_a + ds_read_a_mfma_rate - 1) / ds_read_a_mfma_rate;
+
+        constexpr auto num_total_stages = MRepeat;
+
+        // Group num_mfma_perstage num_ds_read_a_perstage
+        // since we want to reuse a local register buffer
+        constexpr auto num_mfma_perstage      = num_mfma_inst / num_total_stages;
+        constexpr auto num_ds_read_a_perstage = num_ds_read_inst_a / num_total_stages;
+
+        constexpr auto num_ds_read_a_mfma_perstage =
+            math::integer_divide_ceil(num_ds_read_a_perstage, ds_read_a_mfma_rate);
+
+        constexpr auto num_ds_read_a_prefetch_stages = 2;
+
+        constexpr auto buffer_load_perstage_more = math::integer_divide_ceil(
+            (num_buffer_load_inst_a + num_buffer_load_inst_b), (num_total_stages - 2));
+        constexpr auto buffer_load_perstage_less = math::integer_divide_floor(
+            (num_buffer_load_inst_a + num_buffer_load_inst_b), (num_total_stages - 2));
+
+        constexpr auto buffer_load_stages_more =
+            (num_buffer_load_inst_a + num_buffer_load_inst_b) -
+            math::integer_divide_floor((num_buffer_load_inst_a + num_buffer_load_inst_b),
+                                       (num_total_stages - 2)) *
+                ((num_total_stages - 2));
+
+        constexpr auto buffer_load_b_stages =
+            buffer_load_perstage_more * buffer_load_stages_more > num_buffer_load_inst_b
+                ? num_buffer_load_inst_b / buffer_load_perstage_more
+                : (buffer_load_stages_more +
+                   (num_buffer_load_inst_b - buffer_load_perstage_more * buffer_load_stages_more) /
+                       buffer_load_perstage_less);
+
+        constexpr auto buffer_load_a_stages =
+            num_total_stages - num_ds_read_a_prefetch_stages - buffer_load_b_stages;
+
+        constexpr auto buffer_load_issue_point_b = 0;
+        constexpr auto buffer_load_issue_point_interval_more =
+            num_mfma_perstage / buffer_load_perstage_more;
+        constexpr auto buffer_load_issue_point_interval_less =
+            num_mfma_perstage / buffer_load_perstage_less;
+        constexpr auto ds_write_issue_point      = 0;
+        constexpr auto buffer_load_issue_point_a = num_mfma_perstage >= 3 ? 1 : 0;
+
+        // B global read
+        static_for<0, buffer_load_b_stages, 1>{}([&](auto i) {
+            static_for<0, num_mfma_perstage, 1>{}([&](auto imfma) {
+                __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+
+                if constexpr(((i < buffer_load_stages_more) &&
+                              (imfma % buffer_load_issue_point_interval_more ==
+                               buffer_load_issue_point_b)) ||
+                             ((i >= buffer_load_stages_more) &&
+                              (imfma % buffer_load_issue_point_interval_less ==
+                               buffer_load_issue_point_b)))
+                {
+                    __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read
+                }
+
+                if constexpr(imfma >= (num_mfma_perstage - num_ds_read_a_mfma_perstage))
+                {
+                    __builtin_amdgcn_sched_group_barrier(0x100, ds_read_a_mfma_rate, 0); // DS read
+                }
+            });
+        });
+
+        // A global read + A local write
+        static_for<0, buffer_load_a_stages, 1>{}([&](auto i) {
+            static_for<0, num_mfma_perstage, 1>{}([&](auto imfma) {
+                __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                if constexpr((((i + buffer_load_b_stages) < buffer_load_stages_more) &&
+                              (imfma % buffer_load_issue_point_interval_more ==
+                               ds_write_issue_point)) ||
+                             (((i + buffer_load_b_stages) >= buffer_load_stages_more) &&
+                              (imfma % buffer_load_issue_point_interval_less ==
+                               ds_write_issue_point)))
+                {
+                    __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS write
+                }
+                if constexpr((((i + buffer_load_b_stages) < buffer_load_stages_more) &&
+                              (imfma % buffer_load_issue_point_interval_more ==
+                               buffer_load_issue_point_a)) ||
+                             (((i + buffer_load_b_stages) >= buffer_load_stages_more) &&
+                              (imfma % buffer_load_issue_point_interval_less ==
+                               buffer_load_issue_point_a)))
+                {
+                    __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read
+                }
+                if constexpr(imfma >= (num_mfma_perstage - num_ds_read_a_mfma_perstage))
+                {
+                    __builtin_amdgcn_sched_group_barrier(0x100, ds_read_a_mfma_rate, 0); // DS read
+                }
+            });
+        });
+
+        // lds synchronization, prefetch next loop local A
+        static_for<0, num_ds_read_a_prefetch_stages, 1>{}([&](auto i) {
+            ignore = i;
+            static_for<0, num_mfma_perstage, 1>{}([&](auto imfma) {
+                __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                if constexpr(imfma >= (num_mfma_perstage - num_ds_read_a_mfma_perstage))
+                {
+                    __builtin_amdgcn_sched_group_barrier(0x100, ds_read_a_mfma_rate, 0); // DS read
+                }
+            });
+        });
+    }
+
+    template <typename Stage>
+    __device__ static constexpr auto EpilogueScheduler_1(Stage stage)
+    {
+        constexpr auto num_ds_read_inst_a  = HotLoopInstList::A_LDS_Read_Inst_Num;
+        constexpr auto num_ds_write_inst_a = HotLoopInstList::A_LDS_Write_Inst_Num;
+        constexpr auto num_buffer_load_inst_b =
+            MWaves * HotLoopInstList::B_Buffer_Load_Inst_Num * 2;
+
+        constexpr auto num_mfma = HotLoopInstList::C_MFMA_Inst_Num * 2;
+
+        constexpr auto staged_num_ds_read_inst_a = num_ds_read_inst_a / MRepeat;
+        constexpr auto staged_num_mfma           = num_mfma / MRepeat;
+
+        constexpr auto staged_num_mfma_per_ds_read_a = staged_num_mfma / staged_num_ds_read_inst_a;
+
+        if constexpr(stage.value == 0)
+        {
+            constexpr auto staged_num_buffer_load_b_per_ds_read_a =
+                num_buffer_load_inst_b / staged_num_ds_read_inst_a;
+            constexpr auto staged_num_mfma_per_buffer_load_b =
+                staged_num_mfma / num_buffer_load_inst_b;
+            // B global
+            static_for<0, staged_num_ds_read_inst_a, 1>{}([&](auto i_inst) {
+                ignore = i_inst;
+
+                static_for<0, staged_num_buffer_load_b_per_ds_read_a, 1>{}([&](auto ibuf_inst) {
+                    ignore = ibuf_inst;
+                    __builtin_amdgcn_sched_group_barrier(
+                        0x008, staged_num_mfma_per_buffer_load_b, 0);  // MFMA
+                    __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read
+                });
+
+                __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+                __builtin_amdgcn_sched_group_barrier(
+                    0x008, staged_num_mfma_per_buffer_load_b - 1, 0); // MFMA
+                __builtin_amdgcn_sched_group_barrier(0x020, 1, 0);    // VMEM read
+            });
+
+            __builtin_amdgcn_sched_barrier(0);
+        }
+        else if constexpr(stage.value == 1)
+        {
+#if 0
+            constexpr auto staged_num_ds_write_a_per_ds_read_a =
+                num_ds_write_inst_a / staged_num_ds_read_inst_a;
+            constexpr auto staged_num_mfma_per_ds_write_a = staged_num_mfma / num_ds_write_inst_a;
+            // A local write
+            static_for<0, staged_num_ds_read_inst_a, 1>{}([&](auto i_inst) {
+                ignore = i_inst;
+
+                static_for<0, staged_num_ds_write_a_per_ds_read_a, 1>{}([&](auto idswrite_inst) {
+                    ignore = idswrite_inst;
+                    __builtin_amdgcn_sched_group_barrier(
+                        0x008, staged_num_mfma_per_ds_write_a - 1, 0); // MFMA
+                    __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS Write
+                });
+
+                __builtin_amdgcn_sched_group_barrier(
+                    0x008, staged_num_ds_write_a_per_ds_read_a, 0); // MFMA
+                __builtin_amdgcn_sched_group_barrier(0x100, 1, 0);  // DS read
+            });
+#elif 1
+            constexpr auto staged_num_mfma_per_ds_write_a =
+                math::integer_divide_ceil(staged_num_mfma, num_ds_write_inst_a);
+
+            constexpr auto stage_more_mfma =
+                staged_num_mfma - (staged_num_mfma_per_ds_write_a - 1) * num_ds_write_inst_a;
+
+            // A local write
+            static_for<0, num_ds_write_inst_a, 1>{}([&](auto i_inst) {
+                if constexpr(i_inst.value < stage_more_mfma)
+                {
+                    if(i_inst.value < staged_num_ds_read_inst_a)
+                    {
+                        __builtin_amdgcn_sched_group_barrier(
+                            0x008, staged_num_mfma_per_ds_write_a - 1, 0); // MFMA
+                        __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS Write
+                        __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                        __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+                    }
+                    else
+                    {
+                        __builtin_amdgcn_sched_group_barrier(
+                            0x008, staged_num_mfma_per_ds_write_a, 0);     // MFMA
+                        __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS Write
+                    }
+                }
+                else
+                {
+                    if(i_inst.value < staged_num_ds_read_inst_a)
+                    {
+                        __builtin_amdgcn_sched_group_barrier(
+                            0x008, staged_num_mfma_per_ds_write_a - 2, 0); // MFMA
+                        __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS Write
+                        __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                        __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+                    }
+                    else
+                    {
+                        __builtin_amdgcn_sched_group_barrier(
+                            0x008, staged_num_mfma_per_ds_write_a - 1, 0); // MFMA
+                        __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS Write
+                    }
+                }
+            });
+#endif
+            __builtin_amdgcn_sched_barrier(0);
+        }
+        else
+        {
+            // A local Read
+            static_for<0, staged_num_ds_read_inst_a, 1>{}([&](auto i_inst) {
+                ignore = i_inst;
+                __builtin_amdgcn_sched_group_barrier(
+                    0x008, staged_num_mfma_per_ds_read_a, 0);      // MFMA
+                __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+            });
+
+            __builtin_amdgcn_sched_barrier(0);
+        }
+    }
+
+    __device__ static constexpr auto EpilogueScheduler_2()
+    {
+        constexpr auto num_ds_read_inst_a = HotLoopInstList::A_LDS_Read_Inst_Num;
+
+        constexpr auto num_mfma = HotLoopInstList::C_MFMA_Inst_Num * 2;
+
+        constexpr auto staged_num_ds_read_inst_a = num_ds_read_inst_a / MRepeat;
+        constexpr auto staged_num_mfma           = num_mfma / MRepeat;
+
+        constexpr auto staged_num_mfma_per_ds_read_a = staged_num_mfma / staged_num_ds_read_inst_a;
+
+        // A local Read
+        static_for<0, staged_num_ds_read_inst_a, 1>{}([&](auto i_inst) {
+            ignore = i_inst;
+            __builtin_amdgcn_sched_group_barrier(0x008, staged_num_mfma_per_ds_read_a, 0); // MFMA
+            __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+        });
+
+        __builtin_amdgcn_sched_barrier(0);
+    }
+
+    template <bool HasMainLoop,
+              TailNumber TailNum,
+              typename AGridDesc,
+              typename ABlockDesc,
+              typename ABlockTransfer,
+              typename AGridBuffer,
+              typename ABlockBuffer,
+              typename ABlockTransferStep,
+              typename BGridDesc,
+              typename BBlockTransfer,
+              typename BGridBuffer,
+              typename BBlockBuffer,
+              typename BBlockTransferStep,
+              typename CThreadBuffer>
+    __device__ void Run(const AGridDesc& a_grid_desc,
+                        const ABlockDesc& a_block_desc,
+                        ABlockTransfer& a_blockwise_copy,
+                        const AGridBuffer& a_grid_buf,
+                        ABlockBuffer& a_block_buf,
+                        const ABlockTransferStep& a_block_copy_step,
+                        const BGridDesc& b_grid_desc,
+                        BBlockTransfer& b_blockwise_copy,
+                        BBlockTransfer& b_blockwise_copy_up,
+                        const BGridBuffer& b_grid_buf,
+                        const BGridBuffer& b_grid_buf_up,
+                        BBlockBuffer& b_block_buf,
+                        const BBlockTransferStep& b_block_copy_step,
+                        CThreadBuffer& c_thread_buf,
+                        CThreadBuffer& c_thread_buf_up,
+                        index_t num_loop) const
+    {
+        ignore = b_block_buf;
+        __builtin_amdgcn_sched_barrier(0);
+        auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
+            a_thread_desc_.GetElementSpaceSize());
+        auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
+            b_thread_desc_.GetElementSpaceSize());
+
+        StaticallyIndexedArray<decltype(b_thread_buf), Number<2>{}> b_thread_bufs;
+        StaticallyIndexedArray<decltype(b_thread_buf), Number<2>{}> b_thread_bufs_up;
+        constexpr auto b_block_origin_idx = make_tuple(I0, I0, I0, I0);
+
+        // Global prefetch A1 B1
+        b_blockwise_copy.Run(b_grid_desc,
+                             b_grid_buf,
+                             b_block_desc_n0_n1_k0_k1,
+                             b_block_origin_idx,
+                             b_thread_bufs(I0));
+
+        b_blockwise_copy_up.Run(b_grid_desc,
+                                b_grid_buf_up,
+                                b_block_desc_n0_n1_k0_k1,
+                                b_block_origin_idx,
+                                b_thread_bufs_up(I0));
+        b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+        b_blockwise_copy_up.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+        a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
+        a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+        __builtin_amdgcn_sched_barrier(0);
+
+        // // Local prefill A1
+        a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(I0));
+
+        // // Global prefetch A2
+        a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
+        a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+
+        // Local prefetch A1
+        block_sync_lds();
+        static_for<0, 2, 1>{}([&](auto m0) {
+            static_for<0, KRepeat, 1>{}([&](auto k0) {
+                static_for<0, KGroup, 1>{}([&](auto kg0) {
+                    a_thread_copy_.Run(a_block_desc_m0_m1_m2_k0_k1_k2,
+                                       make_tuple(m0, I0, I0, Number<k0 * KGroup + kg0>{}, I0, I0),
+                                       a_block_buf.At(I0),
+                                       a_thread_desc_,
+                                       make_tuple(m0, I0, I0, k0, I0, Number<kg0 * A_K1>{}),
+                                       a_thread_buf);
+                });
+            });
+        });
+
+        // Initialize C
+        c_thread_buf.Clear();
+        c_thread_buf_up.Clear();
+
+        __builtin_amdgcn_sched_barrier(0);
+
+        // main body
+        if constexpr(HasMainLoop)
+        {
+            index_t i = 0;
+            do
+            {
+                auto LoopFunc = [&](auto mfma_reg_buf, auto local_read_buf) {
+                    b_blockwise_copy.Run(b_grid_desc,
+                                         b_grid_buf,
+                                         b_block_desc_n0_n1_k0_k1,
+                                         b_block_origin_idx,
+                                         b_thread_bufs(local_read_buf));
+                    b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+                    b_blockwise_copy_up.Run(b_grid_desc,
+                                            b_grid_buf_up,
+                                            b_block_desc_n0_n1_k0_k1,
+                                            b_block_origin_idx,
+                                            b_thread_bufs_up(local_read_buf));
+                    b_blockwise_copy_up.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+                    a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(local_read_buf));
+                    a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
+                    a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+                    static_for<0, MRepeat, 1>{}([&](auto m0) {
+                        static_for<0, KRepeat, 1>{}([&](auto k0) {
+                            static_for<0, NRepeat, 1>{}([&](auto n0) {
+                                vector_type<ComputeDataType, KPack> a_thread_vec;
+                                vector_type<ComputeDataType, KPack> b_thread_vec;
+                                vector_type<ComputeDataType, KPack> b_thread_vec_up;
+
+                                static_for<0, KPack, 1>{}([&](auto ik) {
+                                    a_thread_vec.template AsType<ComputeDataType>()(ik) =
+                                        a_thread_buf[Number<a_thread_desc_.CalculateOffset(
+                                            make_tuple((m0 + HotloopLocalBufSwitch * mfma_reg_buf) %
+                                                           2,
+                                                       I0,
+                                                       I0,
+                                                       k0,
+                                                       I0,
+                                                       ik))>{}];
+                                    b_thread_vec.template AsType<ComputeDataType>()(ik) =
+                                        b_thread_bufs[mfma_reg_buf]
+                                                     [Number<b_thread_desc_.CalculateOffset(
+                                                         make_tuple(n0, I0, k0, ik))>{}];
+
+                                    b_thread_vec_up.template AsType<ComputeDataType>()(ik) =
+                                        b_thread_bufs_up[mfma_reg_buf]
+                                                        [Number<b_thread_desc_.CalculateOffset(
+                                                            make_tuple(n0, I0, k0, ik))>{}];
+                                });
+
+                                using mfma_input_type =
+                                    typename vector_type<ComputeDataType,
+                                                         xdlops_gemm.K1PerXdlops>::type;
+
+                                constexpr index_t c_offset =
+                                    c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                                xdlops_gemm.Run(
+                                    a_thread_vec.template AsType<mfma_input_type>(),
+                                    b_thread_vec.template AsType<mfma_input_type>(),
+                                    c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+
+                                xdlops_gemm.Run(
+                                    a_thread_vec.template AsType<mfma_input_type>(),
+                                    b_thread_vec_up.template AsType<mfma_input_type>(),
+                                    c_thread_buf_up.GetVectorTypeReference(Number<c_offset>{}));
+                            });
+                        });
+
+                        if constexpr(m0.value == MRepeat - 2)
+                        {
+                            block_sync_lds();
+
+                            static_for<0, KRepeat, 1>{}([&](auto k0) {
+                                static_for<0, KGroup, 1>{}([&](auto kg0) {
+                                    a_thread_copy_.Run(
+                                        a_block_desc_m0_m1_m2_k0_k1_k2,
+                                        make_tuple(Number<(m0 + 2) % MRepeat>{},
+                                                   I0,
+                                                   I0,
+                                                   Number<k0 * KGroup + kg0>{},
+                                                   I0,
+                                                   I0),
+                                        a_block_buf.At(local_read_buf),
+                                        a_thread_desc_,
+                                        make_tuple(
+                                            Number<(m0 + 2 + HotloopLocalBufSwitch * mfma_reg_buf) %
+                                                   2>{},
+                                            I0,
+                                            I0,
+                                            k0,
+                                            I0,
+                                            Number<kg0 * A_K1>{}),
+                                        a_thread_buf);
+                                });
+                            });
+                        }
+                        else if constexpr(m0.value == (MRepeat - 1))
+                        {
+                            static_for<0, KRepeat, 1>{}([&](auto k0) {
+                                static_for<0, KGroup, 1>{}([&](auto kg0) {
+                                    a_thread_copy_.Run(
+                                        a_block_desc_m0_m1_m2_k0_k1_k2,
+                                        make_tuple(Number<(m0 + 2) % MRepeat>{},
+                                                   I0,
+                                                   I0,
+                                                   Number<k0 * KGroup + kg0>{},
+                                                   I0,
+                                                   I0),
+                                        a_block_buf.At(local_read_buf),
+                                        a_thread_desc_,
+                                        make_tuple(
+                                            Number<(m0 + 2 + HotloopLocalBufSwitch * mfma_reg_buf) %
+                                                   2>{},
+                                            I0,
+                                            I0,
+                                            k0,
+                                            I0,
+                                            Number<kg0 * A_K1>{}),
+                                        a_thread_buf);
+                                });
+                            });
+                        }
+                        else
+                        {
+                            static_for<0, KRepeat, 1>{}([&](auto k0) {
+                                static_for<0, KGroup, 1>{}([&](auto kg0) {
+                                    a_thread_copy_.Run(
+                                        a_block_desc_m0_m1_m2_k0_k1_k2,
+                                        make_tuple(Number<(m0 + 2) % MRepeat>{},
+                                                   I0,
+                                                   I0,
+                                                   Number<k0 * KGroup + kg0>{},
+                                                   I0,
+                                                   I0),
+                                        a_block_buf.At(mfma_reg_buf),
+                                        a_thread_desc_,
+                                        make_tuple(
+                                            Number<(m0 + 2 + HotloopLocalBufSwitch * mfma_reg_buf) %
+                                                   2>{},
+                                            I0,
+                                            I0,
+                                            k0,
+                                            I0,
+                                            Number<kg0 * A_K1>{}),
+                                        a_thread_buf);
+                                });
+                            });
+                        }
+                    });
+                    HotLoopScheduler();
+                };
+
+                LoopFunc(I0, I1);
+                LoopFunc(I1, I0);
+
+                i += 2;
+            } while(i < (num_loop - 2));
+        }
+        // tail
+        if constexpr(TailNum == TailNumber::Even)
+        {
+            b_blockwise_copy.Run(b_grid_desc,
+                                 b_grid_buf,
+                                 b_block_desc_n0_n1_k0_k1,
+                                 b_block_origin_idx,
+                                 b_thread_bufs(I1));
+
+            b_blockwise_copy_up.Run(b_grid_desc,
+                                    b_grid_buf_up,
+                                    b_block_desc_n0_n1_k0_k1,
+                                    b_block_origin_idx,
+                                    b_thread_bufs_up(I1));
+            a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(I1));
+            static_for<0, MRepeat, 1>{}([&](auto m0) {
+                static_for<0, KRepeat, 1>{}([&](auto k0) {
+                    static_for<0, NRepeat, 1>{}([&](auto n0) {
+                        vector_type<ComputeDataType, KPack> a_thread_vec;
+                        vector_type<ComputeDataType, KPack> b_thread_vec;
+                        vector_type<ComputeDataType, KPack> b_thread_vec_up;
+
+                        static_for<0, KPack, 1>{}([&](auto ik) {
+                            a_thread_vec.template AsType<ComputeDataType>()(ik) =
+                                a_thread_buf[Number<a_thread_desc_.CalculateOffset(
+                                    make_tuple(m0 % 2, I0, I0, k0, I0, ik))>{}];
+                            b_thread_vec.template AsType<ComputeDataType>()(ik) =
+                                b_thread_bufs[I0][Number<b_thread_desc_.CalculateOffset(
+                                    make_tuple(n0, I0, k0, ik))>{}];
+
+                            b_thread_vec_up.template AsType<ComputeDataType>()(ik) =
+                                b_thread_bufs_up[I0][Number<b_thread_desc_.CalculateOffset(
+                                    make_tuple(n0, I0, k0, ik))>{}];
+                        });
+
+                        using mfma_input_type =
+                            typename vector_type<ComputeDataType, xdlops_gemm.K1PerXdlops>::type;
+
+                        constexpr index_t c_offset =
+                            c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                        xdlops_gemm.Run(a_thread_vec.template AsType<mfma_input_type>(),
+                                        b_thread_vec.template AsType<mfma_input_type>(),
+                                        c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+
+                        xdlops_gemm.Run(a_thread_vec.template AsType<mfma_input_type>(),
+                                        b_thread_vec_up.template AsType<mfma_input_type>(),
+                                        c_thread_buf_up.GetVectorTypeReference(Number<c_offset>{}));
+                    });
+                });
+                if constexpr(m0.value == (MRepeat - 2))
+                {
+                    block_sync_lds();
+
+                    static_for<0, KRepeat, 1>{}([&](auto k0) {
+                        static_for<0, KGroup, 1>{}([&](auto kg0) {
+                            a_thread_copy_.Run(
+                                a_block_desc_m0_m1_m2_k0_k1_k2,
+                                make_tuple(Number<(m0 + 2) % MRepeat>{},
+                                           I0,
+                                           I0,
+                                           Number<k0 * KGroup + kg0>{},
+                                           I0,
+                                           I0),
+                                a_block_buf.At(I1),
+                                a_thread_desc_,
+                                make_tuple(
+                                    Number<(m0 + 2) % 2>{}, I0, I0, k0, I0, Number<kg0 * A_K1>{}),
+                                a_thread_buf);
+                        });
+                    });
+                }
+                else if constexpr(m0.value == MRepeat - 1)
+                {
+                    static_for<0, KRepeat, 1>{}([&](auto k0) {
+                        static_for<0, KGroup, 1>{}([&](auto kg0) {
+                            a_thread_copy_.Run(
+                                a_block_desc_m0_m1_m2_k0_k1_k2,
+                                make_tuple(Number<(m0 + 2) % MRepeat>{},
+                                           I0,
+                                           I0,
+                                           Number<k0 * KGroup + kg0>{},
+                                           I0,
+                                           I0),
+                                a_block_buf.At(I1),
+                                a_thread_desc_,
+                                make_tuple(
+                                    Number<(m0 + 2) % 2>{}, I0, I0, k0, I0, Number<kg0 * A_K1>{}),
+                                a_thread_buf);
+                        });
+                    });
+                }
+                else
+                {
+                    static_for<0, KRepeat, 1>{}([&](auto k0) {
+                        static_for<0, KGroup, 1>{}([&](auto kg0) {
+                            a_thread_copy_.Run(
+                                a_block_desc_m0_m1_m2_k0_k1_k2,
+                                make_tuple(Number<(m0 + 2) % MRepeat>{},
+                                           I0,
+                                           I0,
+                                           Number<k0 * KGroup + kg0>{},
+                                           I0,
+                                           I0),
+                                a_block_buf.At(I0),
+                                a_thread_desc_,
+                                make_tuple(
+                                    Number<(m0 + 2) % 2>{}, I0, I0, k0, I0, Number<kg0 * A_K1>{}),
+                                a_thread_buf);
+                        });
+                    });
+                }
+            });
+
+            HotLoopScheduler();
+
+            static_for<0, MRepeat, 1>{}([&](auto m0) {
+                static_for<0, KRepeat, 1>{}([&](auto k0) {
+                    static_for<0, NRepeat, 1>{}([&](auto n0) {
+                        vector_type<ComputeDataType, KPack> a_thread_vec;
+                        vector_type<ComputeDataType, KPack> b_thread_vec;
+                        vector_type<ComputeDataType, KPack> b_thread_vec_up;
+
+                        static_for<0, KPack, 1>{}([&](auto ik) {
+                            a_thread_vec.template AsType<ComputeDataType>()(ik) =
+                                a_thread_buf[Number<a_thread_desc_.CalculateOffset(make_tuple(
+                                    (m0 + HotloopLocalBufSwitch) % 2, I0, I0, k0, I0, ik))>{}];
+                            b_thread_vec.template AsType<ComputeDataType>()(ik) =
+                                b_thread_bufs[I1][Number<b_thread_desc_.CalculateOffset(
+                                    make_tuple(n0, I0, k0, ik))>{}];
+                            b_thread_vec_up.template AsType<ComputeDataType>()(ik) =
+                                b_thread_bufs_up[I1][Number<b_thread_desc_.CalculateOffset(
+                                    make_tuple(n0, I0, k0, ik))>{}];
+                        });
+
+                        using mfma_input_type =
+                            typename vector_type<ComputeDataType, xdlops_gemm.K1PerXdlops>::type;
+
+                        constexpr index_t c_offset =
+                            c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                        xdlops_gemm.Run(a_thread_vec.template AsType<mfma_input_type>(),
+                                        b_thread_vec.template AsType<mfma_input_type>(),
+                                        c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+
+                        xdlops_gemm.Run(a_thread_vec.template AsType<mfma_input_type>(),
+                                        b_thread_vec_up.template AsType<mfma_input_type>(),
+                                        c_thread_buf_up.GetVectorTypeReference(Number<c_offset>{}));
+                    });
+                });
+
+                if constexpr(m0.value < (MRepeat - 2))
+                {
+                    static_for<0, KRepeat, 1>{}([&](auto k0) {
+                        static_for<0, KGroup, 1>{}([&](auto kg0) {
+                            a_thread_copy_.Run(
+                                a_block_desc_m0_m1_m2_k0_k1_k2,
+                                make_tuple(
+                                    Number<m0 + 2>{}, I0, I0, Number<k0 * KGroup + kg0>{}, I0, I0),
+                                a_block_buf.At(I1),
+                                a_thread_desc_,
+                                make_tuple(Number<(m0 + 2 + HotloopLocalBufSwitch) % 2>{},
+                                           I0,
+                                           I0,
+                                           k0,
+                                           I0,
+                                           Number<kg0 * A_K1>{}),
+                                a_thread_buf);
+                        });
+                    });
+                }
+            });
+
+            HotLoopScheduler();
+            // Let's leak last MFMA block to epilogue region, cover the potential lds-shuffle
+            // latency
+        }
+        else if constexpr(TailNum == TailNumber::Odd)
+        {
+            static_for<0, MRepeat, 1>{}([&](auto m0) {
+                static_for<0, KRepeat, 1>{}([&](auto k0) {
+                    static_for<0, NRepeat, 1>{}([&](auto n0) {
+                        vector_type<ComputeDataType, KPack> a_thread_vec;
+                        vector_type<ComputeDataType, KPack> b_thread_vec;
+                        vector_type<ComputeDataType, KPack> b_thread_vec_up;
+
+                        static_for<0, KPack, 1>{}([&](auto ik) {
+                            a_thread_vec.template AsType<ComputeDataType>()(ik) =
+                                a_thread_buf[Number<a_thread_desc_.CalculateOffset(
+                                    make_tuple(m0 % 2, I0, I0, k0, I0, ik))>{}];
+                            b_thread_vec.template AsType<ComputeDataType>()(ik) =
+                                b_thread_bufs[I0][Number<b_thread_desc_.CalculateOffset(
+                                    make_tuple(n0, I0, k0, ik))>{}];
+                            b_thread_vec_up.template AsType<ComputeDataType>()(ik) =
+                                b_thread_bufs_up[I0][Number<b_thread_desc_.CalculateOffset(
+                                    make_tuple(n0, I0, k0, ik))>{}];
+                        });
+
+                        using mfma_input_type =
+                            typename vector_type<ComputeDataType, xdlops_gemm.K1PerXdlops>::type;
+
+                        constexpr index_t c_offset =
+                            c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                        xdlops_gemm.Run(a_thread_vec.template AsType<mfma_input_type>(),
+                                        b_thread_vec.template AsType<mfma_input_type>(),
+                                        c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                        xdlops_gemm.Run(a_thread_vec.template AsType<mfma_input_type>(),
+                                        b_thread_vec_up.template AsType<mfma_input_type>(),
+                                        c_thread_buf_up.GetVectorTypeReference(Number<c_offset>{}));
+                    });
+                });
+
+                if constexpr(m0.value < (MRepeat - 2))
+                {
+                    static_for<0, KRepeat, 1>{}([&](auto k0) {
+                        static_for<0, KGroup, 1>{}([&](auto kg0) {
+                            a_thread_copy_.Run(
+                                a_block_desc_m0_m1_m2_k0_k1_k2,
+                                make_tuple(
+                                    Number<m0 + 2>{}, I0, I0, Number<k0 * KGroup + kg0>{}, I0, I0),
+                                a_block_buf.At(I0),
+                                a_thread_desc_,
+                                make_tuple(
+                                    Number<(m0 + 2) % 2>{}, I0, I0, k0, I0, Number<kg0 * A_K1>{}),
+                                a_thread_buf);
+                        });
+                    });
+                }
+            });
+        }
+    }
+
+    protected:
+    // MRepeat MWave MLane KRepeat KLane KPack
+    // KRepeat -> MRepeat-> Mwave->KLane->MLane->KPack
+    // Reduce the vgpr usage here.
+    static constexpr auto a_thread_desc_ = make_naive_tensor_descriptor_packed(
+        make_tuple(I2, I1, I1, Number<KRepeat>{}, I1, Number<KPack>{}));
+
+    using AThreadCopy = ThreadwiseTensorSliceTransfer_v4<ADataType,
+                                                         ComputeDataType,
+                                                         decltype(a_block_desc_m0_m1_m2_k0_k1_k2),
+                                                         decltype(a_thread_desc_),
+                                                         Sequence<1, 1, 1, 1, 1, KPack / KGroup>,
+                                                         Sequence<0, 1, 2, 3, 4, 5>,
+                                                         5,
+                                                         A_K1,
+                                                         A_K1>;
+
+    AThreadCopy a_thread_copy_{Base::CalculateAThreadOriginDataIndex6D()};
+
+    static constexpr auto b_thread_desc_ = make_naive_tensor_descriptor_packed(
+        make_tuple(Number<NRepeat>{}, I1, Number<KRepeat>{}, Number<KPack>{}));
+
+    static constexpr BTileDesc b_block_desc_n0_n1_k0_k1;
+
+    using Base::c_thread_desc_;
+};
+
+} // namespace ck

include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuffle_mx_moe_gufusion_v1.hpp

@@ -0,0 +1,968 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_mx_pipeline_xdlops_base.hpp"
+
+namespace ck {
+
+// Naive pipeline with lowest resource request per WGP
+// GlobalPrefetchStages: 2
+// LocalPreFillStages: 1
+// LocalPreFetchStages: 1
+// LocalSharedMemoryBuffer: 1
+
+template <BlockGemmPipelineScheduler BlkGemmPipelineVer,
+          index_t ThreadBlockSize,
+          index_t ScaleBlockSize,
+          typename ADataType,
+          typename AScaleDataType,
+          typename BDataType,
+          typename BScaleDataType,
+          typename ATileDesc,
+          typename BTileDesc,
+          typename AMmaTileDesc,
+          typename BMmaTileDesc,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MRepeat, // MXdlPerWave
+          index_t NRepeat, // NXdlPerWave
+          index_t KPack>
+struct BlockwiseGemmXdlops_pipeline_bpreshuffle_mx_moe_gufusion_v1
+{
+};
+
+template <index_t ThreadBlockSize,
+          index_t ScaleBlockSize,
+          typename ADataType,
+          typename AScaleDataType,
+          typename BDataType,
+          typename BScaleDataType,
+          typename ATileDesc,
+          typename BTileDesc,
+          typename AMmaTileDesc,
+          typename BMmaTileDesc,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MRepeat, // MXdlPerWave
+          index_t NRepeat, // NXdlPerWave
+          index_t KPack>
+struct BlockwiseGemmXdlops_pipeline_bpreshuffle_mx_moe_gufusion_v1<
+    BlockGemmPipelineScheduler::Intrawave,
+    ThreadBlockSize,
+    ScaleBlockSize,
+    ADataType,
+    AScaleDataType,
+    BDataType,
+    BScaleDataType,
+    ATileDesc,
+    BTileDesc,
+    AMmaTileDesc,
+    BMmaTileDesc,
+    ABlockTransferSrcScalarPerVector,
+    BBlockTransferSrcScalarPerVector,
+    MPerBlock,
+    NPerBlock,
+    KPerBlock,
+    MPerXDL,
+    NPerXDL,
+    MRepeat,
+    NRepeat,
+    KPack> : BlockwiseGemmXdlops_mx_pipeline_base<ThreadBlockSize,
+                                                  ADataType,
+                                                  BDataType,
+                                                  ATileDesc,
+                                                  BTileDesc,
+                                                  AMmaTileDesc,
+                                                  BMmaTileDesc,
+                                                  ABlockTransferSrcScalarPerVector,
+                                                  BBlockTransferSrcScalarPerVector,
+                                                  MPerBlock,
+                                                  NPerBlock,
+                                                  KPerBlock,
+                                                  MPerXDL,
+                                                  NPerXDL,
+                                                  MRepeat,
+                                                  NRepeat,
+                                                  KPack>
+
+{
+
+    using Base = BlockwiseGemmXdlops_mx_pipeline_base<ThreadBlockSize,
+                                                      ADataType,
+                                                      BDataType,
+                                                      ATileDesc,
+                                                      BTileDesc,
+                                                      AMmaTileDesc,
+                                                      BMmaTileDesc,
+                                                      ABlockTransferSrcScalarPerVector,
+                                                      BBlockTransferSrcScalarPerVector,
+                                                      MPerBlock,
+                                                      NPerBlock,
+                                                      KPerBlock,
+                                                      MPerXDL,
+                                                      NPerXDL,
+                                                      MRepeat,
+                                                      NRepeat,
+                                                      KPack>;
+    using Base::I0;
+    using Base::I1;
+    using Base::KRepeat;
+    using Base::MWaves;
+    using Base::NWaves;
+    using Base::WaveSize;
+    using Base::xdlops_gemm;
+
+    using Base::CalculateCThreadOriginDataIndex;
+    using Base::GetCBlockDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2;
+    using Base::GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2;
+    using Base::GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4;
+    using Base::GetCThreadBuffer;
+    using Base::GetCThreadDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2;
+    using Base::GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2;
+    using Base::GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4;
+    using Base::GetWaveIdx;
+    using Base::MakeCGridDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2;
+    using Base::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2;
+
+    using Base::a_block_desc_m0_m1_m2_k;
+    using Base::b_block_desc_n0_n1_n2_k;
+
+    using Base::AMmaKStride;
+    using Base::BMmaKStride;
+    using Base::KThreadChunk;
+
+    using Base::APackedSize;
+    using Base::BPackedSize;
+    using Base::ComputePackedSize;
+
+    using AccType      = typename Base::AccType;
+    using Tuple4       = typename Base::Tuple4;
+    using ComputeTypeA = typename Base::ComputeTypeA;
+    using ComputeTypeB = typename Base::ComputeTypeB;
+
+    static constexpr index_t PrefetchStages  = 2;
+    static constexpr index_t PrefillStages   = 1;
+    static constexpr index_t GlobalBufferNum = 2;
+
+    template <typename TileDesc_M0_M1_M2_K>
+    __host__ __device__ static constexpr auto MakeAGemmMmaTileDescriptor(const TileDesc_M0_M1_M2_K&)
+    {
+        constexpr index_t M0 = TileDesc_M0_M1_M2_K{}.GetLength(Number<0>{});
+        constexpr index_t M1 = TileDesc_M0_M1_M2_K{}.GetLength(Number<1>{});
+        constexpr index_t M2 = TileDesc_M0_M1_M2_K{}.GetLength(Number<2>{});
+        constexpr index_t K2 = KPack;
+        constexpr index_t K1 = 64 / NPerXDL;
+        constexpr index_t K0 = KRepeat;
+
+        return transform_tensor_descriptor(
+            TileDesc_M0_M1_M2_K{},
+            make_tuple(
+                make_pass_through_transform(Number<M0>{}),
+                make_pass_through_transform(Number<M1>{}),
+                make_pass_through_transform(Number<M2>{}),
+                make_unmerge_transform(make_tuple(Number<K0>{}, Number<K1>{}, Number<K2>{}))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3, 4, 5>{}));
+    }
+
+    static constexpr auto a_block_desc_m0_m1_m2_k0_k1_k2 =
+        MakeAGemmMmaTileDescriptor(a_block_desc_m0_m1_m2_k);
+
+    static constexpr auto ScalesPerKBlockSize =
+        KPerBlock / ScaleBlockSize; // How many mx-vectors per K block
+
+    //> How many mx-vectors in each row/col is processed in one call to xdlops_gemm.Run()
+    static constexpr auto ScalesPerXdlopsRun = (KPack * xdlops_gemm.K0PerXdlops) / ScaleBlockSize;
+
+    //> How many scales a thread must read to accommodate one call to xdlops_gemm.Run()
+    static constexpr auto ScalesPerXdlopsRunPerThread =
+        ScalesPerXdlopsRun / xdlops_gemm.mfma_instr.num_input_blks;
+
+    __host__ static constexpr bool BlockHasHotloop(index_t num_loop)
+    {
+        return num_loop > PrefetchStages;
+    }
+
+    __host__ static constexpr TailNumber BlockLoopTailNum(index_t num_loop)
+    {
+        return num_loop % 2 == 0 ? TailNumber::Even : TailNumber::Odd;
+    }
+
+    template <bool HasMainLoop,
+              TailNumber TailNum,
+              typename AGridDesc,
+              typename ABlockDesc,
+              typename ABlockTransfer,
+              typename AGridBuffer,
+              typename ABlockBuffer,
+              typename ABlockTransferStep,
+              typename BGridDesc,
+              typename BBlockDesc,
+              typename BBlockTransfer,
+              typename BGridBuffer,
+              typename BBlockBuffer,
+              typename BBlockTransferStep,
+              typename CThreadBuffer,
+              typename AScaleGridBuffer,
+              typename AScaleGridDesc,
+              typename AScaleThreadTransfer,
+              typename BScaleGridBuffer,
+              typename BScaleGridDesc,
+              typename BScaleThreadTransfer>
+    __device__ void Run(
+        // ABlockCopy
+        const AGridDesc& a_grid_desc,
+        const ABlockDesc& a_block_desc,
+        ABlockTransfer& a_blockwise_copy,
+        const AGridBuffer& a_grid_buf,
+        ABlockBuffer& a_block_buf,
+        const ABlockTransferStep& a_block_copy_step,
+        // BBlockCopy
+        const BGridDesc& b_grid_desc,
+        const BBlockDesc& b_block_desc,
+        BBlockTransfer& b_blockwise_copy,
+        BBlockTransfer& b_blockwise_copy_up,
+        const BGridBuffer& b_grid_buf,
+        const BGridBuffer& b_grid_buf_up,
+        BBlockBuffer& b_block_buf,
+        const BBlockTransferStep& b_block_copy_step,
+        // CThread
+        CThreadBuffer& c_thread_buf,
+        CThreadBuffer& c_thread_buf_up,
+        // A and B scales
+        const AScaleGridDesc& a_scale_grid_desc,
+        AScaleThreadTransfer& a_scale_thread_copy,
+        const AScaleGridBuffer& a_scale_grid_buf,
+        const BScaleGridDesc& b_scale_grid_desc,
+        BScaleThreadTransfer& b_scale_thread_copy,
+        BScaleThreadTransfer& b_scale_thread_copy_up,
+        const BScaleGridBuffer& b_scale_grid_buf,
+        const BScaleGridBuffer& b_scale_grid_buf_up,
+        index_t num_loop) const
+    {
+        ignore            = b_block_desc;
+        ignore            = b_block_buf;
+        ignore            = a_scale_grid_buf;
+        ignore            = b_scale_grid_buf;
+        ignore            = b_scale_grid_buf_up;
+        auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeTypeA>(
+            a_thread_desc_.GetElementSpaceSize());
+        auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeTypeB>(
+            b_thread_desc_.GetElementSpaceSize());
+
+        StaticallyIndexedArray<decltype(b_thread_buf), Number<2>{}> b_thread_bufs;
+        StaticallyIndexedArray<decltype(b_thread_buf), Number<2>{}> b_thread_bufs_up;
+        constexpr auto b_block_origin_idx = make_tuple(I0, I0, I0, I0);
+
+        auto a_scale_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, AScaleDataType>(
+            a_scale_thread_desc.GetElementSpaceSize());
+        auto b_scale_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, BScaleDataType>(
+            b_scale_thread_desc.GetElementSpaceSize());
+
+        StaticallyIndexedArray<decltype(a_scale_thread_buf), Number<2>{}> a_scale_thread_bufs;
+        StaticallyIndexedArray<decltype(b_scale_thread_buf), Number<2>{}> b_scale_thread_bufs;
+        StaticallyIndexedArray<decltype(b_scale_thread_buf), Number<2>{}> b_scale_thread_bufs_up;
+
+        // Global prefetch A1 B1
+        a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, I0);
+        b_blockwise_copy.Run(b_grid_desc,
+                             b_grid_buf,
+                             b_block_desc_n0_n1_k0_k1,
+                             b_block_origin_idx,
+                             b_thread_bufs(I0));
+        b_blockwise_copy_up.Run(b_grid_desc,
+                                b_grid_buf_up,
+                                b_block_desc_n0_n1_k0_k1,
+                                b_block_origin_idx,
+                                b_thread_bufs_up(I0));
+
+        a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+        b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+        b_blockwise_copy_up.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+        // Prefetch a_scales to buf 0
+        a_scale_thread_copy.Run(a_scale_grid_desc,
+                                a_scale_grid_buf,
+                                a_scale_thread_desc,
+                                make_tuple(I0, I0, I0),
+                                a_scale_thread_bufs(I0));
+
+        // restore row id and advance to the next set of scales
+        a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc,
+                                               make_multi_index(0, ScalesPerKBlockSize, 0));
+
+        // Prefetch b_scales to buf 0
+        static_for<0, NRepeat, 1>{}([&](auto n0) {
+            static_for<0, KRepeat, 1>{}([&](auto k0) {
+                static_for<0, ScalesPerXdlopsRunPerThread, 1>{}([&](auto s) {
+                    constexpr auto b_scale_offset =
+                        b_scale_thread_desc.CalculateOffset(make_tuple(n0, k0, s));
+                    auto b_scale_thread_buf_copy =
+                        make_static_buffer<AddressSpaceEnum::Vgpr, BScaleDataType>(
+                            b_scale_thread_desc_copy.GetElementSpaceSize());
+                    b_scale_thread_copy.Run(b_scale_grid_desc,
+                                            b_scale_grid_buf,
+                                            b_scale_thread_desc_copy,
+                                            make_tuple(I0, I0),
+                                            b_scale_thread_buf_copy);
+
+                    b_scale_thread_bufs(I0)(Number<b_scale_offset>{}) =
+                        b_scale_thread_buf_copy[Number<0>{}];
+                    b_scale_thread_copy.MoveSrcSliceWindow(
+                        b_scale_grid_desc,
+                        make_multi_index(0, xdlops_gemm.KPerXdlops / ScaleBlockSize));
+
+                    auto b_scale_thread_buf_copy_up =
+                        make_static_buffer<AddressSpaceEnum::Vgpr, BScaleDataType>(
+                            b_scale_thread_desc_copy.GetElementSpaceSize());
+                    b_scale_thread_copy_up.Run(b_scale_grid_desc,
+                                               b_scale_grid_buf_up,
+                                               b_scale_thread_desc_copy,
+                                               make_tuple(I0, I0),
+                                               b_scale_thread_buf_copy_up);
+
+                    b_scale_thread_bufs_up(I0)(Number<b_scale_offset>{}) =
+                        b_scale_thread_buf_copy_up[Number<0>{}];
+                    b_scale_thread_copy_up.MoveSrcSliceWindow(
+                        b_scale_grid_desc,
+                        make_multi_index(0, xdlops_gemm.KPerXdlops / ScaleBlockSize));
+                });
+            });
+            b_scale_thread_copy.MoveSrcSliceWindow(
+                b_scale_grid_desc, make_multi_index(NWaves * NPerXDL, -ScalesPerKBlockSize));
+            b_scale_thread_copy_up.MoveSrcSliceWindow(
+                b_scale_grid_desc, make_multi_index(NWaves * NPerXDL, -ScalesPerKBlockSize));
+        });
+
+        // restore col id and advance to the next set of scales
+        // NWaves * NPerXDL * NRepeat == NPerBlock
+        b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc,
+                                               make_multi_index(-NPerBlock, ScalesPerKBlockSize));
+        b_scale_thread_copy_up.MoveSrcSliceWindow(
+            b_scale_grid_desc, make_multi_index(-NPerBlock, ScalesPerKBlockSize));
+
+        __builtin_amdgcn_sched_barrier(0);
+
+        // Local prefill A1
+        a_blockwise_copy.RunWrite(a_block_desc, a_block_buf, I0);
+
+        // Global prefetch A2
+        a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, I0);
+        a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+
+        // Prefetch a_scales to buf 1
+        a_scale_thread_copy.Run(a_scale_grid_desc,
+                                a_scale_grid_buf,
+                                a_scale_thread_desc,
+                                make_tuple(I0, I0, I0),
+                                a_scale_thread_bufs(I1));
+
+        // restore row id and advance to the next set of scales
+        a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc,
+                                               make_multi_index(0, ScalesPerKBlockSize, 0));
+
+        // Prefetch b_scales to buf 1
+        static_for<0, NRepeat, 1>{}([&](auto n0) {
+            static_for<0, KRepeat, 1>{}([&](auto k0) {
+                static_for<0, ScalesPerXdlopsRunPerThread, 1>{}([&](auto s) {
+                    constexpr auto b_scale_offset =
+                        b_scale_thread_desc.CalculateOffset(make_tuple(n0, k0, s));
+                    auto b_scale_thread_buf_copy =
+                        make_static_buffer<AddressSpaceEnum::Vgpr, BScaleDataType>(
+                            b_scale_thread_desc_copy.GetElementSpaceSize());
+                    b_scale_thread_copy.Run(b_scale_grid_desc,
+                                            b_scale_grid_buf,
+                                            b_scale_thread_desc_copy,
+                                            make_tuple(I0, I0),
+                                            b_scale_thread_buf_copy);
+
+                    b_scale_thread_bufs(I1)(Number<b_scale_offset>{}) =
+                        b_scale_thread_buf_copy[Number<0>{}];
+                    b_scale_thread_copy.MoveSrcSliceWindow(
+                        b_scale_grid_desc,
+                        make_multi_index(0, xdlops_gemm.KPerXdlops / ScaleBlockSize));
+
+                    auto b_scale_thread_buf_copy_up =
+                        make_static_buffer<AddressSpaceEnum::Vgpr, BScaleDataType>(
+                            b_scale_thread_desc_copy.GetElementSpaceSize());
+                    b_scale_thread_copy_up.Run(b_scale_grid_desc,
+                                               b_scale_grid_buf_up,
+                                               b_scale_thread_desc_copy,
+                                               make_tuple(I0, I0),
+                                               b_scale_thread_buf_copy_up);
+
+                    b_scale_thread_bufs_up(I1)(Number<b_scale_offset>{}) =
+                        b_scale_thread_buf_copy_up[Number<0>{}];
+                    b_scale_thread_copy_up.MoveSrcSliceWindow(
+                        b_scale_grid_desc,
+                        make_multi_index(0, xdlops_gemm.KPerXdlops / ScaleBlockSize));
+                });
+            });
+            b_scale_thread_copy.MoveSrcSliceWindow(
+                b_scale_grid_desc, make_multi_index(NWaves * NPerXDL, -ScalesPerKBlockSize));
+            b_scale_thread_copy_up.MoveSrcSliceWindow(
+                b_scale_grid_desc, make_multi_index(NWaves * NPerXDL, -ScalesPerKBlockSize));
+        });
+
+        b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc,
+                                               make_multi_index(-NPerBlock, ScalesPerKBlockSize));
+        b_scale_thread_copy_up.MoveSrcSliceWindow(
+            b_scale_grid_desc, make_multi_index(-NPerBlock, ScalesPerKBlockSize));
+
+        // Local prefetch A1
+        block_sync_lds();
+        static_for<0, KRepeat, 1>{}([&](auto k) {
+            constexpr auto k_step = k * xdlops_gemm.KPerXdlops * (KPack / xdlops_gemm.K1PerXdlops);
+
+            static_for<0, MRepeat, 1>{}([&](auto m0) {
+                static_for<0, xdlops_gemm.K1PerXdlops / KThreadChunk, 1>{}([&](auto chunk) {
+                    constexpr auto a_k_step_chunk =
+                        k_step + chunk * KThreadChunk * xdlops_gemm.mfma_instr.num_input_blks;
+                    a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
+                                       make_tuple(m0, I0, I0, Number<a_k_step_chunk>{}),
+                                       a_block_buf,
+                                       a_thread_desc_,
+                                       make_tuple(m0, I0, k, Number<chunk * KThreadChunk>{}),
+                                       a_thread_buf);
+                });
+            });
+        });
+
+        // Initialize C
+        c_thread_buf.Clear();
+        c_thread_buf_up.Clear();
+
+        // main body
+        if constexpr(HasMainLoop)
+        {
+            // loop over k with the step KPerBlock
+            index_t i = 0;
+            do
+            {
+                auto LoopFunc = [&](auto mfma_reg_buf, auto local_read_buf) {
+                    b_blockwise_copy.Run(b_grid_desc,
+                                         b_grid_buf,
+                                         b_block_desc_n0_n1_k0_k1,
+                                         b_block_origin_idx,
+                                         b_thread_bufs(local_read_buf));
+                    b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+                    b_blockwise_copy_up.Run(b_grid_desc,
+                                            b_grid_buf_up,
+                                            b_block_desc_n0_n1_k0_k1,
+                                            b_block_origin_idx,
+                                            b_thread_bufs_up(local_read_buf));
+                    b_blockwise_copy_up.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+                    block_sync_lds();
+                    a_blockwise_copy.RunWrite(a_block_desc, a_block_buf, mfma_reg_buf);
+
+                    a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, local_read_buf);
+                    a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+
+                    static_for<0, MRepeat, 1>{}([&](auto m0) {
+                        static_for<0, NRepeat, 1>{}([&](auto n0) {
+                            static_for<0, KRepeat, 1>{}([&](auto k0) {
+                                vector_type<ComputeTypeA, KPack> a_thread_vec;
+                                vector_type<ComputeTypeB, KPack> b_thread_vec;
+                                vector_type<ComputeTypeB, KPack> b_thread_vec_up;
+
+                                static_for<0, KPack / ComputePackedSize, 1>{}([&](auto ik) {
+                                    a_thread_vec.template AsType<ComputeTypeA>()(ik) =
+                                        a_thread_buf[Number<a_thread_desc_.CalculateOffset(
+                                            make_tuple(m0, I0, k0, ik))>{}];
+                                    b_thread_vec.template AsType<ComputeTypeB>()(ik) =
+                                        b_thread_bufs[mfma_reg_buf]
+                                                     [Number<b_thread_desc_.CalculateOffset(
+                                                         make_tuple(n0, I0, k0, ik))>{}];
+                                    b_thread_vec_up.template AsType<ComputeTypeB>()(ik) =
+                                        b_thread_bufs_up[mfma_reg_buf]
+                                                        [Number<b_thread_desc_.CalculateOffset(
+                                                            make_tuple(n0, I0, k0, ik))>{}];
+                                });
+
+                                constexpr index_t a_scale_offset =
+                                    a_scale_thread_desc.CalculateOffset(make_tuple(m0, k0, I0));
+                                constexpr index_t b_scale_offset =
+                                    b_scale_thread_desc.CalculateOffset(make_tuple(n0, k0, I0));
+
+                                static_assert(
+                                    0 < ScalesPerXdlopsRunPerThread,
+                                    "Must have at least one scale per Xdlops per Thread.");
+
+                                vector_type<AScaleDataType, ScalesPerXdlopsRunPerThread>
+                                    a_scale_thread_vec;
+                                vector_type<BScaleDataType, ScalesPerXdlopsRunPerThread>
+                                    b_scale_thread_vec;
+                                vector_type<BScaleDataType, ScalesPerXdlopsRunPerThread>
+                                    b_scale_thread_vec_up;
+
+                                // Pack scale_thread_buf into scale_thread_vec
+                                static_for<0, ScalesPerXdlopsRunPerThread, 1>{}([&](auto s) {
+                                    a_scale_thread_vec.template AsType<AScaleDataType>()(s) =
+                                        a_scale_thread_bufs[mfma_reg_buf]
+                                                           [Number<a_scale_offset + s>{}];
+                                    b_scale_thread_vec.template AsType<BScaleDataType>()(s) =
+                                        b_scale_thread_bufs[mfma_reg_buf]
+                                                           [Number<b_scale_offset + s>{}];
+                                    b_scale_thread_vec_up.template AsType<BScaleDataType>()(s) =
+                                        b_scale_thread_bufs_up[mfma_reg_buf]
+                                                              [Number<b_scale_offset + s>{}];
+                                });
+
+                                using mfma_input_type_a =
+                                    typename vector_type<ComputeTypeA,
+                                                         xdlops_gemm.K1PerXdlops /
+                                                             APackedSize>::type;
+                                using mfma_input_type_b =
+                                    typename vector_type<ComputeTypeB,
+                                                         xdlops_gemm.K1PerXdlops /
+                                                             BPackedSize>::type;
+
+                                constexpr index_t c_offset =
+                                    c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                                // MFMA accumulation
+                                xdlops_gemm.template Run<>(
+                                    a_thread_vec.template AsType<mfma_input_type_a>(),
+                                    a_scale_thread_vec.template AsType<AScaleDataType>(),
+                                    b_thread_vec.template AsType<mfma_input_type_b>(),
+                                    b_scale_thread_vec.template AsType<BScaleDataType>(),
+                                    c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                                xdlops_gemm.template Run<>(
+                                    a_thread_vec.template AsType<mfma_input_type_a>(),
+                                    a_scale_thread_vec.template AsType<AScaleDataType>(),
+                                    b_thread_vec_up.template AsType<mfma_input_type_b>(),
+                                    b_scale_thread_vec_up.template AsType<BScaleDataType>(),
+                                    c_thread_buf_up.GetVectorTypeReference(Number<c_offset>{}));
+                            });
+                        });
+                    });
+
+                    block_sync_lds();
+
+                    // a thread copy
+                    static_for<0, KRepeat, 1>{}([&](auto k) {
+                        constexpr auto k_step =
+                            k * xdlops_gemm.KPerXdlops * (KPack / xdlops_gemm.K1PerXdlops);
+
+                        static_for<0, MRepeat, 1>{}([&](auto m0) {
+                            static_for<0, xdlops_gemm.K1PerXdlops / KThreadChunk, 1>{}(
+                                [&](auto chunk) {
+                                    constexpr auto a_k_step_chunk =
+                                        k_step + chunk * KThreadChunk *
+                                                     xdlops_gemm.mfma_instr.num_input_blks;
+                                    a_thread_copy_.Run(
+                                        a_block_desc_m0_m1_m2_k,
+                                        make_tuple(m0, I0, I0, Number<a_k_step_chunk>{}),
+                                        a_block_buf,
+                                        a_thread_desc_,
+                                        make_tuple(m0, I0, k, Number<chunk * KThreadChunk>{}),
+                                        a_thread_buf);
+                                });
+                        });
+                    });
+
+                    // Prefetch a_scales
+                    a_scale_thread_copy.Run(a_scale_grid_desc,
+                                            a_scale_grid_buf,
+                                            a_scale_thread_desc,
+                                            make_tuple(I0, I0, I0),
+                                            a_scale_thread_bufs(mfma_reg_buf));
+
+                    // restore row id and advance to the next set of scales
+                    a_scale_thread_copy.MoveSrcSliceWindow(
+                        a_scale_grid_desc, make_multi_index(0, ScalesPerKBlockSize, 0));
+
+                    // Prefetch b_scales
+                    static_for<0, NRepeat, 1>{}([&](auto n0) {
+                        static_for<0, KRepeat, 1>{}([&](auto k0) {
+                            static_for<0, ScalesPerXdlopsRunPerThread, 1>{}([&](auto s) {
+                                constexpr auto b_scale_offset =
+                                    b_scale_thread_desc.CalculateOffset(make_tuple(n0, k0, s));
+                                auto b_scale_thread_buf_copy =
+                                    make_static_buffer<AddressSpaceEnum::Vgpr, BScaleDataType>(
+                                        b_scale_thread_desc_copy.GetElementSpaceSize());
+                                b_scale_thread_copy.Run(b_scale_grid_desc,
+                                                        b_scale_grid_buf,
+                                                        b_scale_thread_desc_copy,
+                                                        make_tuple(I0, I0),
+                                                        b_scale_thread_buf_copy);
+
+                                b_scale_thread_bufs(mfma_reg_buf)(Number<b_scale_offset>{}) =
+                                    b_scale_thread_buf_copy[Number<0>{}];
+                                b_scale_thread_copy.MoveSrcSliceWindow(
+                                    b_scale_grid_desc,
+                                    make_multi_index(0, xdlops_gemm.KPerXdlops / ScaleBlockSize));
+
+                                auto b_scale_thread_buf_copy_up =
+                                    make_static_buffer<AddressSpaceEnum::Vgpr, BScaleDataType>(
+                                        b_scale_thread_desc_copy.GetElementSpaceSize());
+                                b_scale_thread_copy_up.Run(b_scale_grid_desc,
+                                                           b_scale_grid_buf_up,
+                                                           b_scale_thread_desc_copy,
+                                                           make_tuple(I0, I0),
+                                                           b_scale_thread_buf_copy_up);
+
+                                b_scale_thread_bufs_up(mfma_reg_buf)(Number<b_scale_offset>{}) =
+                                    b_scale_thread_buf_copy_up[Number<0>{}];
+                                b_scale_thread_copy_up.MoveSrcSliceWindow(
+                                    b_scale_grid_desc,
+                                    make_multi_index(0, xdlops_gemm.KPerXdlops / ScaleBlockSize));
+                            });
+                        });
+                        b_scale_thread_copy.MoveSrcSliceWindow(
+                            b_scale_grid_desc,
+                            make_multi_index(NWaves * NPerXDL, -ScalesPerKBlockSize));
+                        b_scale_thread_copy_up.MoveSrcSliceWindow(
+                            b_scale_grid_desc,
+                            make_multi_index(NWaves * NPerXDL, -ScalesPerKBlockSize));
+                    });
+
+                    b_scale_thread_copy.MoveSrcSliceWindow(
+                        b_scale_grid_desc, make_multi_index(-NPerBlock, ScalesPerKBlockSize));
+                    b_scale_thread_copy_up.MoveSrcSliceWindow(
+                        b_scale_grid_desc, make_multi_index(-NPerBlock, ScalesPerKBlockSize));
+                };
+
+                LoopFunc(I0, I1);
+                LoopFunc(I1, I0);
+
+                i += 2;
+            } while(i < (num_loop - 2));
+        }
+
+        // tail
+        if constexpr(TailNum == TailNumber::Even)
+        {
+            b_blockwise_copy.Run(b_grid_desc,
+                                 b_grid_buf,
+                                 b_block_desc_n0_n1_k0_k1,
+                                 b_block_origin_idx,
+                                 b_thread_bufs(I1));
+
+            b_blockwise_copy_up.Run(b_grid_desc,
+                                    b_grid_buf_up,
+                                    b_block_desc_n0_n1_k0_k1,
+                                    b_block_origin_idx,
+                                    b_thread_bufs_up(I1));
+            block_sync_lds();
+            a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
+
+            static_for<0, MRepeat, 1>{}([&](auto m0) {
+                static_for<0, NRepeat, 1>{}([&](auto n0) {
+                    static_for<0, KRepeat, 1>{}([&](auto k0) {
+                        vector_type<ComputeTypeA, KPack> a_thread_vec;
+                        vector_type<ComputeTypeB, KPack> b_thread_vec;
+                        vector_type<ComputeTypeB, KPack> b_thread_vec_up;
+
+                        static_for<0, KPack / ComputePackedSize, 1>{}([&](auto ik) {
+                            a_thread_vec.template AsType<ComputeTypeA>()(ik) =
+                                a_thread_buf[Number<a_thread_desc_.CalculateOffset(
+                                    make_tuple(m0, I0, k0, ik))>{}];
+                            b_thread_vec.template AsType<ComputeTypeB>()(ik) =
+                                b_thread_bufs[I0][Number<b_thread_desc_.CalculateOffset(
+                                    make_tuple(n0, I0, k0, ik))>{}];
+                            b_thread_vec_up.template AsType<ComputeTypeB>()(ik) =
+                                b_thread_bufs_up[I0][Number<b_thread_desc_.CalculateOffset(
+                                    make_tuple(n0, I0, k0, ik))>{}];
+                        });
+
+                        constexpr index_t a_scale_offset =
+                            a_scale_thread_desc.CalculateOffset(make_tuple(m0, k0, I0));
+
+                        constexpr index_t b_scale_offset =
+                            b_scale_thread_desc.CalculateOffset(make_tuple(n0, k0, I0));
+
+                        vector_type<AScaleDataType, ScalesPerXdlopsRunPerThread> a_scale_thread_vec;
+                        vector_type<BScaleDataType, ScalesPerXdlopsRunPerThread> b_scale_thread_vec;
+                        vector_type<BScaleDataType, ScalesPerXdlopsRunPerThread>
+                            b_scale_thread_vec_up;
+
+                        // Pack b_scale_thread_buf into b_scale_thread_vec
+                        static_for<0, ScalesPerXdlopsRunPerThread, 1>{}([&](auto s) {
+                            a_scale_thread_vec.template AsType<AScaleDataType>()(s) =
+                                a_scale_thread_bufs[I0][Number<a_scale_offset + s>{}];
+                            b_scale_thread_vec.template AsType<BScaleDataType>()(s) =
+                                b_scale_thread_bufs[I0][Number<b_scale_offset + s>{}];
+                            b_scale_thread_vec_up.template AsType<BScaleDataType>()(s) =
+                                b_scale_thread_bufs_up[I0][Number<b_scale_offset + s>{}];
+                        });
+
+                        using mfma_input_type_a =
+                            typename vector_type<ComputeTypeA,
+                                                 xdlops_gemm.K1PerXdlops / APackedSize>::type;
+                        using mfma_input_type_b =
+                            typename vector_type<ComputeTypeB,
+                                                 xdlops_gemm.K1PerXdlops / BPackedSize>::type;
+
+                        constexpr index_t c_offset =
+                            c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                        // MFMA accumulation
+                        xdlops_gemm.template Run<>(
+                            a_thread_vec.template AsType<mfma_input_type_a>(),
+                            a_scale_thread_vec.template AsType<AScaleDataType>(),
+                            b_thread_vec.template AsType<mfma_input_type_b>(),
+                            b_scale_thread_vec.template AsType<BScaleDataType>(),
+                            c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                        xdlops_gemm.template Run<>(
+                            a_thread_vec.template AsType<mfma_input_type_a>(),
+                            a_scale_thread_vec.template AsType<AScaleDataType>(),
+                            b_thread_vec_up.template AsType<mfma_input_type_b>(),
+                            b_scale_thread_vec_up.template AsType<BScaleDataType>(),
+                            c_thread_buf_up.GetVectorTypeReference(Number<c_offset>{}));
+                    });
+                });
+            });
+
+            block_sync_lds();
+
+            // a thread copy
+            static_for<0, KRepeat, 1>{}([&](auto k) {
+                constexpr auto k_step =
+                    k * xdlops_gemm.KPerXdlops * (KPack / xdlops_gemm.K1PerXdlops);
+
+                static_for<0, MRepeat, 1>{}([&](auto m0) {
+                    static_for<0, xdlops_gemm.K1PerXdlops / KThreadChunk, 1>{}([&](auto chunk) {
+                        constexpr auto a_k_step_chunk =
+                            k_step + chunk * KThreadChunk * xdlops_gemm.mfma_instr.num_input_blks;
+                        a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
+                                           make_tuple(m0, I0, I0, Number<a_k_step_chunk>{}),
+                                           a_block_buf,
+                                           a_thread_desc_,
+                                           make_tuple(m0, I0, k, Number<chunk * KThreadChunk>{}),
+                                           a_thread_buf);
+                    });
+                });
+            });
+
+            static_for<0, MRepeat, 1>{}([&](auto m0) {
+                static_for<0, NRepeat, 1>{}([&](auto n0) {
+                    static_for<0, KRepeat, 1>{}([&](auto k0) {
+                        vector_type<ComputeTypeA, KPack> a_thread_vec;
+                        vector_type<ComputeTypeB, KPack> b_thread_vec;
+                        vector_type<ComputeTypeB, KPack> b_thread_vec_up;
+
+                        static_for<0, KPack / ComputePackedSize, 1>{}([&](auto ik) {
+                            a_thread_vec.template AsType<ComputeTypeA>()(ik) =
+                                a_thread_buf[Number<a_thread_desc_.CalculateOffset(
+                                    make_tuple(m0, I0, k0, ik))>{}];
+                            b_thread_vec.template AsType<ComputeTypeB>()(ik) =
+                                b_thread_bufs[I1][Number<b_thread_desc_.CalculateOffset(
+                                    make_tuple(n0, I0, k0, ik))>{}];
+                            b_thread_vec_up.template AsType<ComputeTypeB>()(ik) =
+                                b_thread_bufs_up[I1][Number<b_thread_desc_.CalculateOffset(
+                                    make_tuple(n0, I0, k0, ik))>{}];
+                        });
+
+                        constexpr index_t a_scale_offset =
+                            a_scale_thread_desc.CalculateOffset(make_tuple(m0, k0, I0));
+
+                        constexpr index_t b_scale_offset =
+                            b_scale_thread_desc.CalculateOffset(make_tuple(n0, k0, I0));
+
+                        vector_type<AScaleDataType, ScalesPerXdlopsRunPerThread> a_scale_thread_vec;
+                        vector_type<BScaleDataType, ScalesPerXdlopsRunPerThread> b_scale_thread_vec;
+                        vector_type<BScaleDataType, ScalesPerXdlopsRunPerThread>
+                            b_scale_thread_vec_up;
+
+                        // Pack b_scale_thread_buf into b_scale_thread_vec
+                        static_for<0, ScalesPerXdlopsRunPerThread, 1>{}([&](auto s) {
+                            a_scale_thread_vec.template AsType<AScaleDataType>()(s) =
+                                a_scale_thread_bufs[I1][Number<a_scale_offset + s>{}];
+                            b_scale_thread_vec.template AsType<BScaleDataType>()(s) =
+                                b_scale_thread_bufs[I1][Number<b_scale_offset + s>{}];
+                            b_scale_thread_vec_up.template AsType<BScaleDataType>()(s) =
+                                b_scale_thread_bufs_up[I1][Number<b_scale_offset + s>{}];
+                        });
+
+                        using mfma_input_type_a =
+                            typename vector_type<ComputeTypeA,
+                                                 xdlops_gemm.K1PerXdlops / APackedSize>::type;
+                        using mfma_input_type_b =
+                            typename vector_type<ComputeTypeB,
+                                                 xdlops_gemm.K1PerXdlops / BPackedSize>::type;
+
+                        constexpr index_t c_offset =
+                            c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                        // MFMA accumulation
+                        xdlops_gemm.template Run<>(
+                            a_thread_vec.template AsType<mfma_input_type_a>(),
+                            a_scale_thread_vec.template AsType<AScaleDataType>(),
+                            b_thread_vec.template AsType<mfma_input_type_b>(),
+                            b_scale_thread_vec.template AsType<BScaleDataType>(),
+                            c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                        xdlops_gemm.template Run<>(
+                            a_thread_vec.template AsType<mfma_input_type_a>(),
+                            a_scale_thread_vec.template AsType<AScaleDataType>(),
+                            b_thread_vec_up.template AsType<mfma_input_type_b>(),
+                            b_scale_thread_vec_up.template AsType<BScaleDataType>(),
+                            c_thread_buf_up.GetVectorTypeReference(Number<c_offset>{}));
+                    });
+                });
+            });
+        }
+        else if constexpr(TailNum == TailNumber::Odd)
+        {
+            static_for<0, MRepeat, 1>{}([&](auto m0) {
+                static_for<0, NRepeat, 1>{}([&](auto n0) {
+                    static_for<0, KRepeat, 1>{}([&](auto k0) {
+                        vector_type<ComputeTypeA, KPack> a_thread_vec;
+                        vector_type<ComputeTypeB, KPack> b_thread_vec;
+                        vector_type<ComputeTypeB, KPack> b_thread_vec_up;
+
+                        static_for<0, KPack / ComputePackedSize, 1>{}([&](auto ik) {
+                            a_thread_vec.template AsType<ComputeTypeA>()(ik) =
+                                a_thread_buf[Number<a_thread_desc_.CalculateOffset(
+                                    make_tuple(m0, I0, k0, ik))>{}];
+                            b_thread_vec.template AsType<ComputeTypeB>()(ik) =
+                                b_thread_bufs[I0][Number<b_thread_desc_.CalculateOffset(
+                                    make_tuple(n0, I0, k0, ik))>{}];
+                            b_thread_vec_up.template AsType<ComputeTypeB>()(ik) =
+                                b_thread_bufs_up[I0][Number<b_thread_desc_.CalculateOffset(
+                                    make_tuple(n0, I0, k0, ik))>{}];
+                        });
+
+                        constexpr index_t a_scale_offset =
+                            a_scale_thread_desc.CalculateOffset(make_tuple(m0, k0, I0));
+
+                        constexpr index_t b_scale_offset =
+                            b_scale_thread_desc.CalculateOffset(make_tuple(n0, k0, I0));
+
+                        vector_type<AScaleDataType, ScalesPerXdlopsRunPerThread> a_scale_thread_vec;
+                        vector_type<BScaleDataType, ScalesPerXdlopsRunPerThread> b_scale_thread_vec;
+                        vector_type<BScaleDataType, ScalesPerXdlopsRunPerThread>
+                            b_scale_thread_vec_up;
+
+                        // Pack b_scale_thread_buf into b_scale_thread_vec
+                        static_for<0, ScalesPerXdlopsRunPerThread, 1>{}([&](auto s) {
+                            a_scale_thread_vec.template AsType<AScaleDataType>()(s) =
+                                a_scale_thread_bufs[I0][Number<a_scale_offset + s>{}];
+                            b_scale_thread_vec.template AsType<BScaleDataType>()(s) =
+                                b_scale_thread_bufs[I0][Number<b_scale_offset + s>{}];
+                            b_scale_thread_vec_up.template AsType<BScaleDataType>()(s) =
+                                b_scale_thread_bufs_up[I0][Number<b_scale_offset + s>{}];
+                        });
+
+#if 0
+                        printf("bidx: %u, bidy: %u, tidx: %u, a_thread_vec: %02x, %02x, %02x, %02x, b_thread_vec: %02x, %02x, %02x, %02x,"
+                               "a_scale_thread_vec: %02x, b_scale_vec: %02x\n",
+                               blockIdx.x,
+                               blockIdx.y,
+                               threadIdx.x,
+
+                               *(reinterpret_cast<const uint8_t*>(&(a_thread_vec.template AsType<ComputeTypeA>()[Number<0>{}]))),
+                               *(reinterpret_cast<const uint8_t*>(&(a_thread_vec.template AsType<ComputeTypeA>()[Number<1>{}]))),
+                               *(reinterpret_cast<const uint8_t*>(&(a_thread_vec.template AsType<ComputeTypeA>()[Number<2>{}]))),
+                               *(reinterpret_cast<const uint8_t*>(&(a_thread_vec.template AsType<ComputeTypeA>()[Number<3>{}]))),
+                               *(reinterpret_cast<const uint8_t*>(&(b_thread_vec.template AsType<ComputeTypeB>()[Number<0>{}]))),
+                               *(reinterpret_cast<const uint8_t*>(&(b_thread_vec.template AsType<ComputeTypeB>()[Number<1>{}]))),
+                               *(reinterpret_cast<const uint8_t*>(&(b_thread_vec.template AsType<ComputeTypeB>()[Number<2>{}]))),
+                               *(reinterpret_cast<const uint8_t*>(&(b_thread_vec.template AsType<ComputeTypeB>()[Number<3>{}]))),
+
+                            //    type_convert<float>(a_thread_vec.template AsType<ComputeTypeA>()[Number<0>{}].unpack(Number<0>{})),
+                            //    type_convert<float>(a_thread_vec.template AsType<ComputeTypeA>()[Number<0>{}].unpack(Number<1>{})),
+                            //    type_convert<float>(a_thread_vec.template AsType<ComputeTypeA>()[Number<1>{}].unpack(Number<0>{})),
+                            //    type_convert<float>(a_thread_vec.template AsType<ComputeTypeA>()[Number<1>{}].unpack(Number<1>{})),
+                            //    type_convert<float>(b_thread_vec.template AsType<ComputeTypeB>()[Number<0>{}].unpack(Number<0>{})),
+                            //    type_convert<float>(b_thread_vec.template AsType<ComputeTypeB>()[Number<0>{}].unpack(Number<1>{})),
+                            //    type_convert<float>(b_thread_vec.template AsType<ComputeTypeB>()[Number<1>{}].unpack(Number<0>{})),
+                            //    type_convert<float>(b_thread_vec.template AsType<ComputeTypeB>()[Number<1>{}].unpack(Number<1>{})),
+                               *(reinterpret_cast<const uint8_t*>(&(a_scale_thread_vec.template AsType<AScaleDataType>()[Number<0>{}]))),
+                               *(reinterpret_cast<const uint8_t*>(&(b_scale_thread_vec.template AsType<BScaleDataType>()[Number<0>{}]))));
+#endif
+
+                        using mfma_input_type_a =
+                            typename vector_type<ComputeTypeA,
+                                                 xdlops_gemm.K1PerXdlops / APackedSize>::type;
+                        using mfma_input_type_b =
+                            typename vector_type<ComputeTypeB,
+                                                 xdlops_gemm.K1PerXdlops / BPackedSize>::type;
+
+                        constexpr index_t c_offset =
+                            c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                        // MFMA accumulation
+                        xdlops_gemm.template Run<>(
+                            a_thread_vec.template AsType<mfma_input_type_a>(),
+                            a_scale_thread_vec.template AsType<AScaleDataType>(),
+                            b_thread_vec.template AsType<mfma_input_type_b>(),
+                            b_scale_thread_vec.template AsType<BScaleDataType>(),
+                            c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                        xdlops_gemm.template Run<>(
+                            a_thread_vec.template AsType<mfma_input_type_a>(),
+                            a_scale_thread_vec.template AsType<AScaleDataType>(),
+                            b_thread_vec_up.template AsType<mfma_input_type_b>(),
+                            b_scale_thread_vec_up.template AsType<BScaleDataType>(),
+                            c_thread_buf_up.GetVectorTypeReference(Number<c_offset>{}));
+                    });
+#if 0
+                    printf("bidx: %u, bidx: %u, tidx: %u, c_thread_buf: %f, %f, %f, %f\n",
+                        blockIdx.x,
+                        blockIdx.y,
+                        threadIdx.x,
+                        (c_thread_buf[Number<0>{}]),
+                        (c_thread_buf[Number<1>{}]),
+                        (c_thread_buf[Number<2>{}]),
+                        (c_thread_buf[Number<3>{}]));
+#endif
+                });
+            });
+#if 0
+            printf("bidx: %u, bidx: %u, tidx: %u, c_thread_buf: %f, %f, %f, %f\n",
+                blockIdx.x,
+                blockIdx.y,
+                threadIdx.x,
+                type_convert<float>(c_thread_buf.GetVectorTypeReference(Number<0>{}).template AsType<float>()[Number<0>{}]),
+                type_convert<float>(c_thread_buf.GetVectorTypeReference(Number<0>{}).template AsType<float>()[Number<1>{}]),
+                type_convert<float>(c_thread_buf.GetVectorTypeReference(Number<0>{}).template AsType<float>()[Number<2>{}]),
+                type_convert<float>(c_thread_buf.GetVectorTypeReference(Number<0>{}).template AsType<float>()[Number<3>{}]));
+
+#endif
+        }
+    }
+
+    // TODO: make this field protected when a_scale_thread_copy_ is moved
+    // here
+    static constexpr auto a_scale_thread_desc = make_naive_tensor_descriptor_packed(
+        make_tuple(Number<MRepeat>{}, Number<KRepeat>{}, Number<ScalesPerXdlopsRunPerThread>{}));
+
+    // Is used to copy data from a_scale_grid to a_scale_thread
+    static constexpr auto a_scale_thread_desc_copy =
+        make_naive_tensor_descriptor_packed(make_tuple(Number<1>{}, Number<1>{}));
+
+    // TODO: make this field protected when b_scale_thread_copy_ is moved
+    // here
+    static constexpr auto b_scale_thread_desc = make_naive_tensor_descriptor_packed(
+        make_tuple(Number<NRepeat>{}, Number<KRepeat>{}, Number<ScalesPerXdlopsRunPerThread>{}));
+
+    // Is used to copy data from b_scale_grid to b_scale_thread_buf
+    static constexpr auto b_scale_thread_desc_copy =
+        make_naive_tensor_descriptor_packed(make_tuple(Number<1>{}, Number<1>{}));
+
+    protected:
+    static constexpr auto b_thread_desc_ = make_naive_tensor_descriptor_packed(
+        make_tuple(Number<NRepeat>{}, I1, Number<KRepeat>{}, Number<KPack>{}));
+    using Base::a_thread_copy_;
+    using Base::a_thread_desc_;
+    using Base::b_thread_copy_;
+    // using Base::b_thread_desc_;
+    using Base::c_thread_desc_;
+
+    static constexpr BTileDesc b_block_desc_n0_n1_k0_k1;
+};
+
+} // namespace ck

include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuffle_mx_moe_selector.hpp

@@ -0,0 +1,155 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuffle_mx_moe_v1.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuffle_mx_moe_gufusion_v1.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuffle_mx_moe_v3.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuffle_mx_moe_gufusion_v3.hpp"
+
+namespace ck {
+template <BlockGemmPipelineVersion BlkGemmPipelineVer,
+          BlockGemmPipelineScheduler BlkGemmPipeSche,
+          index_t ThreadBlockSize,
+          index_t ScaleBlockSize,
+          typename ADataType,
+          typename AScaleDataType,
+          typename BDataType,
+          typename BScaleDataType,
+          typename ComputeDataType, // TODO: remove this as in this pipeline ADataType and BDataType
+                                    // must be used for compute
+          typename AccDataType,
+          typename ATileDesc,
+          typename BTileDesc,
+          typename AMmaTileDesc,
+          typename BMmaTileDesc,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MRepeat,
+          index_t NRepeat,
+          index_t KPack,
+          bool GUFusion = false>
+constexpr auto BlockGemmMXBPreshufflePipeline_Selector()
+{
+
+    // Hardware MX GEMM pipeline
+    if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v1)
+    {
+        if constexpr(GUFusion)
+        {
+            return BlockwiseGemmXdlops_pipeline_bpreshuffle_mx_moe_gufusion_v1<
+                BlkGemmPipeSche,
+                ThreadBlockSize,
+                ScaleBlockSize,
+                ADataType,
+                AScaleDataType,
+                BDataType,
+                BScaleDataType,
+                ATileDesc,
+                BTileDesc,
+                AMmaTileDesc,
+                BMmaTileDesc,
+                ABlockTransferSrcScalarPerVector,
+                BBlockTransferSrcScalarPerVector,
+                MPerBlock,
+                NPerBlock,
+                KPerBlock,
+                MPerXDL,
+                NPerXDL,
+                MRepeat,
+                NRepeat,
+                KPack>{};
+            ;
+        }
+        else
+        {
+            return BlockwiseGemmXdlops_pipeline_bpreshuffle_mx_moe_v1<
+                BlkGemmPipeSche,
+                ThreadBlockSize,
+                ScaleBlockSize,
+                ADataType,
+                AScaleDataType,
+                BDataType,
+                BScaleDataType,
+                ATileDesc,
+                BTileDesc,
+                AMmaTileDesc,
+                BMmaTileDesc,
+                ABlockTransferSrcScalarPerVector,
+                BBlockTransferSrcScalarPerVector,
+                MPerBlock,
+                NPerBlock,
+                KPerBlock,
+                MPerXDL,
+                NPerXDL,
+                MRepeat,
+                NRepeat,
+                KPack>{};
+        }
+    }
+    else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v3)
+    {
+        if constexpr(GUFusion)
+        {
+            return BlockwiseGemmXdlops_pipeline_bpreshuffle_mx_moe_gufusion_v3<
+                BlkGemmPipeSche,
+                ThreadBlockSize,
+                ScaleBlockSize,
+                ADataType,
+                AScaleDataType,
+                BDataType,
+                BScaleDataType,
+                ATileDesc,
+                BTileDesc,
+                AMmaTileDesc,
+                BMmaTileDesc,
+                ABlockTransferSrcScalarPerVector,
+                BBlockTransferSrcScalarPerVector,
+                MPerBlock,
+                NPerBlock,
+                KPerBlock,
+                MPerXDL,
+                NPerXDL,
+                MRepeat,
+                NRepeat,
+                KPack>{};
+        }
+        else
+        {
+            return BlockwiseGemmXdlops_pipeline_bpreshuffle_mx_moe_v3<
+                BlkGemmPipeSche,
+                ThreadBlockSize,
+                ScaleBlockSize,
+                ADataType,
+                AScaleDataType,
+                BDataType,
+                BScaleDataType,
+                ATileDesc,
+                BTileDesc,
+                AMmaTileDesc,
+                BMmaTileDesc,
+                ABlockTransferSrcScalarPerVector,
+                BBlockTransferSrcScalarPerVector,
+                MPerBlock,
+                NPerBlock,
+                KPerBlock,
+                MPerXDL,
+                NPerXDL,
+                MRepeat,
+                NRepeat,
+                KPack>{};
+        }
+    }
+    else
+    {
+        std::cerr << "MX GEMM Pipeline configuration is not available" << std::endl;
+    }
+}
+
+} // namespace ck

include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuffle_mx_moe_v1.hpp

@@ -0,0 +1,862 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_mx_pipeline_xdlops_base.hpp"
+
+namespace ck {
+
+// Naive pipeline with lowest resource request per WGP
+// GlobalPrefetchStages: 2
+// LocalPreFillStages: 1
+// LocalPreFetchStages: 1
+// LocalSharedMemoryBuffer: 1
+
+template <BlockGemmPipelineScheduler BlkGemmPipelineVer,
+          index_t ThreadBlockSize,
+          index_t ScaleBlockSize,
+          typename ADataType,
+          typename AScaleDataType,
+          typename BDataType,
+          typename BScaleDataType,
+          typename ATileDesc,
+          typename BTileDesc,
+          typename AMmaTileDesc,
+          typename BMmaTileDesc,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MRepeat, // MXdlPerWave
+          index_t NRepeat, // NXdlPerWave
+          index_t KPack>
+struct BlockwiseGemmXdlops_pipeline_bpreshuffle_mx_moe_v1
+{
+};
+
+template <index_t ThreadBlockSize,
+          index_t ScaleBlockSize,
+          typename ADataType,
+          typename AScaleDataType,
+          typename BDataType,
+          typename BScaleDataType,
+          typename ATileDesc,
+          typename BTileDesc,
+          typename AMmaTileDesc,
+          typename BMmaTileDesc,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MRepeat, // MXdlPerWave
+          index_t NRepeat, // NXdlPerWave
+          index_t KPack>
+struct BlockwiseGemmXdlops_pipeline_bpreshuffle_mx_moe_v1<BlockGemmPipelineScheduler::Intrawave,
+                                                          ThreadBlockSize,
+                                                          ScaleBlockSize,
+                                                          ADataType,
+                                                          AScaleDataType,
+                                                          BDataType,
+                                                          BScaleDataType,
+                                                          ATileDesc,
+                                                          BTileDesc,
+                                                          AMmaTileDesc,
+                                                          BMmaTileDesc,
+                                                          ABlockTransferSrcScalarPerVector,
+                                                          BBlockTransferSrcScalarPerVector,
+                                                          MPerBlock,
+                                                          NPerBlock,
+                                                          KPerBlock,
+                                                          MPerXDL,
+                                                          NPerXDL,
+                                                          MRepeat,
+                                                          NRepeat,
+                                                          KPack>
+    : BlockwiseGemmXdlops_mx_pipeline_base<ThreadBlockSize,
+                                           ADataType,
+                                           BDataType,
+                                           ATileDesc,
+                                           BTileDesc,
+                                           AMmaTileDesc,
+                                           BMmaTileDesc,
+                                           ABlockTransferSrcScalarPerVector,
+                                           BBlockTransferSrcScalarPerVector,
+                                           MPerBlock,
+                                           NPerBlock,
+                                           KPerBlock,
+                                           MPerXDL,
+                                           NPerXDL,
+                                           MRepeat,
+                                           NRepeat,
+                                           KPack>
+
+{
+
+    using Base = BlockwiseGemmXdlops_mx_pipeline_base<ThreadBlockSize,
+                                                      ADataType,
+                                                      BDataType,
+                                                      ATileDesc,
+                                                      BTileDesc,
+                                                      AMmaTileDesc,
+                                                      BMmaTileDesc,
+                                                      ABlockTransferSrcScalarPerVector,
+                                                      BBlockTransferSrcScalarPerVector,
+                                                      MPerBlock,
+                                                      NPerBlock,
+                                                      KPerBlock,
+                                                      MPerXDL,
+                                                      NPerXDL,
+                                                      MRepeat,
+                                                      NRepeat,
+                                                      KPack>;
+    using Base::I0;
+    using Base::I1;
+    using Base::KRepeat;
+    using Base::MWaves;
+    using Base::NWaves;
+    using Base::WaveSize;
+    using Base::xdlops_gemm;
+
+    using Base::CalculateCThreadOriginDataIndex;
+    using Base::GetCBlockDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2;
+    using Base::GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2;
+    using Base::GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4;
+    using Base::GetCThreadBuffer;
+    using Base::GetCThreadDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2;
+    using Base::GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2;
+    using Base::GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4;
+    using Base::GetWaveIdx;
+    using Base::MakeCGridDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2;
+    using Base::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2;
+
+    using Base::a_block_desc_m0_m1_m2_k;
+    using Base::b_block_desc_n0_n1_n2_k;
+
+    using Base::AMmaKStride;
+    using Base::BMmaKStride;
+    using Base::KThreadChunk;
+
+    using Base::APackedSize;
+    using Base::BPackedSize;
+    using Base::ComputePackedSize;
+
+    using AccType      = typename Base::AccType;
+    using Tuple4       = typename Base::Tuple4;
+    using ComputeTypeA = typename Base::ComputeTypeA;
+    using ComputeTypeB = typename Base::ComputeTypeB;
+
+    static constexpr index_t PrefetchStages  = 2;
+    static constexpr index_t PrefillStages   = 1;
+    static constexpr index_t GlobalBufferNum = 2;
+
+    template <typename TileDesc_M0_M1_M2_K>
+    __host__ __device__ static constexpr auto MakeAGemmMmaTileDescriptor(const TileDesc_M0_M1_M2_K&)
+    {
+        constexpr index_t M0 = TileDesc_M0_M1_M2_K{}.GetLength(Number<0>{});
+        constexpr index_t M1 = TileDesc_M0_M1_M2_K{}.GetLength(Number<1>{});
+        constexpr index_t M2 = TileDesc_M0_M1_M2_K{}.GetLength(Number<2>{});
+        constexpr index_t K2 = KPack;
+        constexpr index_t K1 = 64 / NPerXDL;
+        constexpr index_t K0 = KRepeat;
+
+        return transform_tensor_descriptor(
+            TileDesc_M0_M1_M2_K{},
+            make_tuple(
+                make_pass_through_transform(Number<M0>{}),
+                make_pass_through_transform(Number<M1>{}),
+                make_pass_through_transform(Number<M2>{}),
+                make_unmerge_transform(make_tuple(Number<K0>{}, Number<K1>{}, Number<K2>{}))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3, 4, 5>{}));
+    }
+
+    static constexpr auto a_block_desc_m0_m1_m2_k0_k1_k2 =
+        MakeAGemmMmaTileDescriptor(a_block_desc_m0_m1_m2_k);
+
+    static constexpr auto ScalesPerKBlockSize =
+        KPerBlock / ScaleBlockSize; // How many mx-vectors per K block
+
+    //> How many mx-vectors in each row/col is processed in one call to xdlops_gemm.Run()
+    static constexpr auto ScalesPerXdlopsRun = (KPack * xdlops_gemm.K0PerXdlops) / ScaleBlockSize;
+
+    //> How many scales a thread must read to accommodate one call to xdlops_gemm.Run()
+    static constexpr auto ScalesPerXdlopsRunPerThread =
+        ScalesPerXdlopsRun / xdlops_gemm.mfma_instr.num_input_blks;
+
+    __host__ static constexpr bool BlockHasHotloop(index_t num_loop)
+    {
+        return num_loop > PrefetchStages;
+    }
+
+    __host__ static constexpr TailNumber BlockLoopTailNum(index_t num_loop)
+    {
+        return num_loop % 2 == 0 ? TailNumber::Even : TailNumber::Odd;
+    }
+
+    template <bool HasMainLoop,
+              TailNumber TailNum,
+              typename AGridDesc,
+              typename ABlockDesc,
+              typename ABlockTransfer,
+              typename AGridBuffer,
+              typename ABlockBuffer,
+              typename ABlockTransferStep,
+              typename BGridDesc,
+              typename BBlockDesc,
+              typename BBlockTransfer,
+              typename BGridBuffer,
+              typename BBlockBuffer,
+              typename BBlockTransferStep,
+              typename CThreadBuffer,
+              typename AScaleGridBuffer,
+              typename AScaleGridDesc,
+              typename AScaleThreadTransfer,
+              typename BScaleGridBuffer,
+              typename BScaleGridDesc,
+              typename BScaleThreadTransfer>
+    __device__ void Run(
+        // ABlockCopy
+        const AGridDesc& a_grid_desc,
+        const ABlockDesc& a_block_desc,
+        ABlockTransfer& a_blockwise_copy,
+        const AGridBuffer& a_grid_buf,
+        ABlockBuffer& a_block_buf,
+        const ABlockTransferStep& a_block_copy_step,
+        // BBlockCopy
+        const BGridDesc& b_grid_desc,
+        const BBlockDesc& b_block_desc,
+        BBlockTransfer& b_blockwise_copy,
+        const BGridBuffer& b_grid_buf,
+        BBlockBuffer& b_block_buf,
+        const BBlockTransferStep& b_block_copy_step,
+        // CThread
+        CThreadBuffer& c_thread_buf,
+        // A and B scales
+        const AScaleGridDesc& a_scale_grid_desc,
+        AScaleThreadTransfer& a_scale_thread_copy,
+        const AScaleGridBuffer& a_scale_grid_buf,
+        const BScaleGridDesc& b_scale_grid_desc,
+        BScaleThreadTransfer& b_scale_thread_copy,
+        const BScaleGridBuffer& b_scale_grid_buf,
+        index_t num_loop) const
+    {
+        ignore = b_block_desc;
+        ignore = b_block_buf;
+
+        auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeTypeA>(
+            a_thread_desc_.GetElementSpaceSize());
+        auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeTypeB>(
+            b_thread_desc_.GetElementSpaceSize());
+
+        StaticallyIndexedArray<decltype(b_thread_buf), Number<2>{}> b_thread_bufs;
+        constexpr auto b_block_origin_idx = make_tuple(I0, I0, I0, I0);
+
+        auto a_scale_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, AScaleDataType>(
+            a_scale_thread_desc.GetElementSpaceSize());
+        auto b_scale_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, BScaleDataType>(
+            b_scale_thread_desc.GetElementSpaceSize());
+
+        StaticallyIndexedArray<decltype(a_scale_thread_buf), Number<2>{}> a_scale_thread_bufs;
+        StaticallyIndexedArray<decltype(b_scale_thread_buf), Number<2>{}> b_scale_thread_bufs;
+
+        // Global prefetch A1 B1
+        a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, I0);
+        b_blockwise_copy.Run(b_grid_desc,
+                             b_grid_buf,
+                             b_block_desc_n0_n1_k0_k1,
+                             b_block_origin_idx,
+                             b_thread_bufs(I0));
+
+        a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+        b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+        // Prefetch a_scales
+        static_for<0, MRepeat, 1>{}([&](auto m0) {
+            static_for<0, KRepeat, 1>{}([&](auto k0) {
+                static_for<0, ScalesPerXdlopsRunPerThread, 1>{}([&](auto s) {
+                    constexpr auto a_scale_offset =
+                        a_scale_thread_desc.CalculateOffset(make_tuple(m0, k0, s));
+                    auto a_scale_thread_buf_copy =
+                        make_static_buffer<AddressSpaceEnum::Vgpr, AScaleDataType>(
+                            a_scale_thread_desc_copy.GetElementSpaceSize());
+                    a_scale_thread_copy.Run(a_scale_grid_desc,
+                                            a_scale_grid_buf,
+                                            a_scale_thread_desc_copy,
+                                            make_tuple(I0, I0),
+                                            a_scale_thread_buf_copy);
+
+                    a_scale_thread_buf(I0)(Number<a_scale_offset>{}) =
+                        a_scale_thread_buf_copy[Number<0>{}];
+                    a_scale_thread_copy.MoveSrcSliceWindow(
+                        a_scale_grid_desc,
+                        make_multi_index(0, xdlops_gemm.KPerXdlops / ScaleBlockSize));
+                });
+            });
+            a_scale_thread_copy.MoveSrcSliceWindow(
+                a_scale_grid_desc, make_multi_index(MWaves * MPerXDL, -ScalesPerKBlockSize));
+        });
+
+        // restore row id and advance to the next set of scales
+        a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc,
+                                               make_multi_index(-MPerBlock, ScalesPerKBlockSize));
+
+        // Prefetch b_scales to buf 0
+        static_for<0, NRepeat, 1>{}([&](auto n0) {
+            static_for<0, KRepeat, 1>{}([&](auto k0) {
+                static_for<0, ScalesPerXdlopsRunPerThread, 1>{}([&](auto s) {
+                    constexpr auto b_scale_offset =
+                        b_scale_thread_desc.CalculateOffset(make_tuple(n0, k0, s));
+                    auto b_scale_thread_buf_copy =
+                        make_static_buffer<AddressSpaceEnum::Vgpr, BScaleDataType>(
+                            b_scale_thread_desc_copy.GetElementSpaceSize());
+                    b_scale_thread_copy.Run(b_scale_grid_desc,
+                                            b_scale_grid_buf,
+                                            b_scale_thread_desc_copy,
+                                            make_tuple(I0, I0),
+                                            b_scale_thread_buf_copy);
+
+                    b_scale_thread_bufs(I0)(Number<b_scale_offset>{}) =
+                        b_scale_thread_buf_copy[Number<0>{}];
+                    b_scale_thread_copy.MoveSrcSliceWindow(
+                        b_scale_grid_desc,
+                        make_multi_index(0, xdlops_gemm.KPerXdlops / ScaleBlockSize));
+                });
+            });
+            b_scale_thread_copy.MoveSrcSliceWindow(
+                b_scale_grid_desc, make_multi_index(NWaves * NPerXDL, -ScalesPerKBlockSize));
+        });
+
+        // restore col id and advance to the next set of scales
+        // NWaves * NPerXDL * NRepeat == NPerBlock
+        b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc,
+                                               make_multi_index(-NPerBlock, ScalesPerKBlockSize));
+
+        __builtin_amdgcn_sched_barrier(0);
+
+        // Local prefill A1
+        a_blockwise_copy.RunWrite(a_block_desc, a_block_buf, I0);
+
+        // Global prefetch A2
+        a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, I0);
+        a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+
+        // Prefetch a_scales to buf 1
+        static_for<0, MRepeat, 1>{}([&](auto m0) {
+            static_for<0, KRepeat, 1>{}([&](auto k0) {
+                static_for<0, ScalesPerXdlopsRunPerThread, 1>{}([&](auto s) {
+                    constexpr auto a_scale_offset =
+                        a_scale_thread_desc.CalculateOffset(make_tuple(m0, k0, s));
+                    auto a_scale_thread_buf_copy =
+                        make_static_buffer<AddressSpaceEnum::Vgpr, AScaleDataType>(
+                            a_scale_thread_desc_copy.GetElementSpaceSize());
+                    a_scale_thread_copy.Run(a_scale_grid_desc,
+                                            a_scale_grid_buf,
+                                            a_scale_thread_desc_copy,
+                                            make_tuple(I0, I0),
+                                            a_scale_thread_buf_copy);
+
+                    a_scale_thread_buf(I1)(Number<a_scale_offset>{}) =
+                        a_scale_thread_buf_copy[Number<0>{}];
+                    a_scale_thread_copy.MoveSrcSliceWindow(
+                        a_scale_grid_desc,
+                        make_multi_index(0, xdlops_gemm.KPerXdlops / ScaleBlockSize));
+                });
+            });
+            a_scale_thread_copy.MoveSrcSliceWindow(
+                a_scale_grid_desc, make_multi_index(MWaves * MPerXDL, -ScalesPerKBlockSize));
+        });
+
+        // restore row id and advance to the next set of scales
+        a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc,
+                                               make_multi_index(-MPerBlock, ScalesPerKBlockSize));
+
+        // Prefetch b_scales to buf 1
+        static_for<0, NRepeat, 1>{}([&](auto n0) {
+            static_for<0, KRepeat, 1>{}([&](auto k0) {
+                static_for<0, ScalesPerXdlopsRunPerThread, 1>{}([&](auto s) {
+                    constexpr auto b_scale_offset =
+                        b_scale_thread_desc.CalculateOffset(make_tuple(n0, k0, s));
+                    auto b_scale_thread_buf_copy =
+                        make_static_buffer<AddressSpaceEnum::Vgpr, BScaleDataType>(
+                            b_scale_thread_desc_copy.GetElementSpaceSize());
+                    b_scale_thread_copy.Run(b_scale_grid_desc,
+                                            b_scale_grid_buf,
+                                            b_scale_thread_desc_copy,
+                                            make_tuple(I0, I0),
+                                            b_scale_thread_buf_copy);
+
+                    b_scale_thread_bufs(I1)(Number<b_scale_offset>{}) =
+                        b_scale_thread_buf_copy[Number<0>{}];
+                    b_scale_thread_copy.MoveSrcSliceWindow(
+                        b_scale_grid_desc,
+                        make_multi_index(0, xdlops_gemm.KPerXdlops / ScaleBlockSize));
+                });
+            });
+            b_scale_thread_copy.MoveSrcSliceWindow(
+                b_scale_grid_desc, make_multi_index(NWaves * NPerXDL, -ScalesPerKBlockSize));
+        });
+
+        b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc,
+                                               make_multi_index(-NPerBlock, ScalesPerKBlockSize));
+
+        // Local prefetch A1
+        block_sync_lds();
+        static_for<0, KRepeat, 1>{}([&](auto k) {
+            constexpr auto k_step = k * xdlops_gemm.KPerXdlops * (KPack / xdlops_gemm.K1PerXdlops);
+
+            static_for<0, MRepeat, 1>{}([&](auto m0) {
+                static_for<0, xdlops_gemm.K1PerXdlops / KThreadChunk, 1>{}([&](auto chunk) {
+                    constexpr auto a_k_step_chunk =
+                        k_step + chunk * KThreadChunk * xdlops_gemm.mfma_instr.num_input_blks;
+                    a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
+                                       make_tuple(m0, I0, I0, Number<a_k_step_chunk>{}),
+                                       a_block_buf,
+                                       a_thread_desc_,
+                                       make_tuple(m0, I0, k, Number<chunk * KThreadChunk>{}),
+                                       a_thread_buf);
+                });
+            });
+        });
+
+        // Initialize C
+        c_thread_buf.Clear();
+
+        // main body
+        if constexpr(HasMainLoop)
+        {
+            // loop over k with the step KPerBlock
+            index_t i = 0;
+            do
+            {
+                auto LoopFunc = [&](auto mfma_reg_buf, auto local_read_buf) {
+                    b_blockwise_copy.Run(b_grid_desc,
+                                         b_grid_buf,
+                                         b_block_desc_n0_n1_k0_k1,
+                                         b_block_origin_idx,
+                                         b_thread_bufs(local_read_buf));
+                    b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+                    block_sync_lds();
+                    a_blockwise_copy.RunWrite(a_block_desc, a_block_buf, mfma_reg_buf);
+
+                    a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, local_read_buf);
+                    a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+
+                    static_for<0, MRepeat, 1>{}([&](auto m0) {
+                        static_for<0, NRepeat, 1>{}([&](auto n0) {
+                            static_for<0, KRepeat, 1>{}([&](auto k0) {
+                                vector_type<ComputeTypeA, KPack> a_thread_vec;
+                                vector_type<ComputeTypeB, KPack> b_thread_vec;
+
+                                static_for<0, KPack / ComputePackedSize, 1>{}([&](auto ik) {
+                                    a_thread_vec.template AsType<ComputeTypeA>()(ik) =
+                                        a_thread_buf[Number<a_thread_desc_.CalculateOffset(
+                                            make_tuple(m0, I0, k0, ik))>{}];
+                                    b_thread_vec.template AsType<ComputeTypeB>()(ik) =
+                                        b_thread_bufs[mfma_reg_buf]
+                                                     [Number<b_thread_desc_.CalculateOffset(
+                                                         make_tuple(n0, I0, k0, ik))>{}];
+                                });
+
+                                constexpr index_t a_scale_offset =
+                                    a_scale_thread_desc.CalculateOffset(make_tuple(m0, k0, I0));
+                                constexpr index_t b_scale_offset =
+                                    b_scale_thread_desc.CalculateOffset(make_tuple(n0, k0, I0));
+
+                                static_assert(
+                                    0 < ScalesPerXdlopsRunPerThread,
+                                    "Must have at least one scale per Xdlops per Thread.");
+
+                                vector_type<AScaleDataType, ScalesPerXdlopsRunPerThread>
+                                    a_scale_thread_vec;
+                                vector_type<BScaleDataType, ScalesPerXdlopsRunPerThread>
+                                    b_scale_thread_vec;
+
+                                // Pack scale_thread_buf into scale_thread_vec
+                                static_for<0, ScalesPerXdlopsRunPerThread, 1>{}([&](auto s) {
+                                    a_scale_thread_vec.template AsType<AScaleDataType>()(s) =
+                                        a_scale_thread_bufs[mfma_reg_buf]
+                                                           [Number<a_scale_offset + s>{}];
+                                    b_scale_thread_vec.template AsType<BScaleDataType>()(s) =
+                                        b_scale_thread_bufs[mfma_reg_buf]
+                                                           [Number<b_scale_offset + s>{}];
+                                });
+
+                                using mfma_input_type_a =
+                                    typename vector_type<ComputeTypeA,
+                                                         xdlops_gemm.K1PerXdlops /
+                                                             APackedSize>::type;
+                                using mfma_input_type_b =
+                                    typename vector_type<ComputeTypeB,
+                                                         xdlops_gemm.K1PerXdlops /
+                                                             BPackedSize>::type;
+
+                                constexpr index_t c_offset =
+                                    c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                                // MFMA accumulation
+                                xdlops_gemm.template Run<>(
+                                    a_thread_vec.template AsType<mfma_input_type_a>(),
+                                    a_scale_thread_vec.template AsType<AScaleDataType>(),
+                                    b_thread_vec.template AsType<mfma_input_type_b>(),
+                                    b_scale_thread_vec.template AsType<BScaleDataType>(),
+                                    c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                            });
+                        });
+                    });
+
+                    block_sync_lds();
+
+                    // a thread copy
+                    static_for<0, KRepeat, 1>{}([&](auto k) {
+                        constexpr auto k_step =
+                            k * xdlops_gemm.KPerXdlops * (KPack / xdlops_gemm.K1PerXdlops);
+
+                        static_for<0, MRepeat, 1>{}([&](auto m0) {
+                            static_for<0, xdlops_gemm.K1PerXdlops / KThreadChunk, 1>{}(
+                                [&](auto chunk) {
+                                    constexpr auto a_k_step_chunk =
+                                        k_step + chunk * KThreadChunk *
+                                                     xdlops_gemm.mfma_instr.num_input_blks;
+                                    a_thread_copy_.Run(
+                                        a_block_desc_m0_m1_m2_k,
+                                        make_tuple(m0, I0, I0, Number<a_k_step_chunk>{}),
+                                        a_block_buf,
+                                        a_thread_desc_,
+                                        make_tuple(m0, I0, k, Number<chunk * KThreadChunk>{}),
+                                        a_thread_buf);
+                                });
+                        });
+                    });
+
+                    // Prefetch a_scales
+                    a_scale_thread_copy.Run(a_scale_grid_desc,
+                                            a_scale_grid_buf,
+                                            a_scale_thread_desc,
+                                            make_tuple(I0, I0, I0),
+                                            a_scale_thread_bufs(mfma_reg_buf));
+
+                    // restore row id and advance to the next set of scales
+                    a_scale_thread_copy.MoveSrcSliceWindow(
+                        a_scale_grid_desc, make_multi_index(0, ScalesPerKBlockSize, 0));
+
+                    // Prefetch b_scales
+                    static_for<0, NRepeat, 1>{}([&](auto n0) {
+                        static_for<0, KRepeat, 1>{}([&](auto k0) {
+                            static_for<0, ScalesPerXdlopsRunPerThread, 1>{}([&](auto s) {
+                                constexpr auto b_scale_offset =
+                                    b_scale_thread_desc.CalculateOffset(make_tuple(n0, k0, s));
+                                auto b_scale_thread_buf_copy =
+                                    make_static_buffer<AddressSpaceEnum::Vgpr, BScaleDataType>(
+                                        b_scale_thread_desc_copy.GetElementSpaceSize());
+                                b_scale_thread_copy.Run(b_scale_grid_desc,
+                                                        b_scale_grid_buf,
+                                                        b_scale_thread_desc_copy,
+                                                        make_tuple(I0, I0),
+                                                        b_scale_thread_buf_copy);
+
+                                b_scale_thread_bufs(mfma_reg_buf)(Number<b_scale_offset>{}) =
+                                    b_scale_thread_buf_copy[Number<0>{}];
+                                b_scale_thread_copy.MoveSrcSliceWindow(
+                                    b_scale_grid_desc,
+                                    make_multi_index(0, xdlops_gemm.KPerXdlops / ScaleBlockSize));
+                            });
+                        });
+                        b_scale_thread_copy.MoveSrcSliceWindow(
+                            b_scale_grid_desc,
+                            make_multi_index(NWaves * NPerXDL, -ScalesPerKBlockSize));
+                    });
+
+                    b_scale_thread_copy.MoveSrcSliceWindow(
+                        b_scale_grid_desc, make_multi_index(-NPerBlock, ScalesPerKBlockSize));
+                };
+
+                LoopFunc(I0, I1);
+                LoopFunc(I1, I0);
+
+                i += 2;
+            } while(i < (num_loop - 2));
+        }
+
+        // tail
+        if constexpr(TailNum == TailNumber::Even)
+        {
+            b_blockwise_copy.Run(b_grid_desc,
+                                 b_grid_buf,
+                                 b_block_desc_n0_n1_k0_k1,
+                                 b_block_origin_idx,
+                                 b_thread_bufs(I1));
+            block_sync_lds();
+            a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
+
+            static_for<0, MRepeat, 1>{}([&](auto m0) {
+                static_for<0, NRepeat, 1>{}([&](auto n0) {
+                    static_for<0, KRepeat, 1>{}([&](auto k0) {
+                        vector_type<ComputeTypeA, KPack> a_thread_vec;
+                        vector_type<ComputeTypeB, KPack> b_thread_vec;
+
+                        static_for<0, KPack / ComputePackedSize, 1>{}([&](auto ik) {
+                            a_thread_vec.template AsType<ComputeTypeA>()(ik) =
+                                a_thread_buf[Number<a_thread_desc_.CalculateOffset(
+                                    make_tuple(m0, I0, k0, ik))>{}];
+                            b_thread_vec.template AsType<ComputeTypeB>()(ik) =
+                                b_thread_bufs[I0][Number<b_thread_desc_.CalculateOffset(
+                                    make_tuple(n0, I0, k0, ik))>{}];
+                        });
+
+                        constexpr index_t a_scale_offset =
+                            a_scale_thread_desc.CalculateOffset(make_tuple(m0, k0, I0));
+
+                        constexpr index_t b_scale_offset =
+                            b_scale_thread_desc.CalculateOffset(make_tuple(n0, k0, I0));
+
+                        vector_type<AScaleDataType, ScalesPerXdlopsRunPerThread> a_scale_thread_vec;
+                        vector_type<BScaleDataType, ScalesPerXdlopsRunPerThread> b_scale_thread_vec;
+
+                        // Pack b_scale_thread_buf into b_scale_thread_vec
+                        static_for<0, ScalesPerXdlopsRunPerThread, 1>{}([&](auto s) {
+                            a_scale_thread_vec.template AsType<AScaleDataType>()(s) =
+                                a_scale_thread_bufs[I0][Number<a_scale_offset + s>{}];
+                            b_scale_thread_vec.template AsType<BScaleDataType>()(s) =
+                                b_scale_thread_bufs[I0][Number<b_scale_offset + s>{}];
+                        });
+
+                        using mfma_input_type_a =
+                            typename vector_type<ComputeTypeA,
+                                                 xdlops_gemm.K1PerXdlops / APackedSize>::type;
+                        using mfma_input_type_b =
+                            typename vector_type<ComputeTypeB,
+                                                 xdlops_gemm.K1PerXdlops / BPackedSize>::type;
+
+                        constexpr index_t c_offset =
+                            c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                        // MFMA accumulation
+                        xdlops_gemm.template Run<>(
+                            a_thread_vec.template AsType<mfma_input_type_a>(),
+                            a_scale_thread_vec.template AsType<AScaleDataType>(),
+                            b_thread_vec.template AsType<mfma_input_type_b>(),
+                            b_scale_thread_vec.template AsType<BScaleDataType>(),
+                            c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                    });
+                });
+            });
+
+            block_sync_lds();
+
+            // a thread copy
+            static_for<0, KRepeat, 1>{}([&](auto k) {
+                constexpr auto k_step =
+                    k * xdlops_gemm.KPerXdlops * (KPack / xdlops_gemm.K1PerXdlops);
+
+                static_for<0, MRepeat, 1>{}([&](auto m0) {
+                    static_for<0, xdlops_gemm.K1PerXdlops / KThreadChunk, 1>{}([&](auto chunk) {
+                        constexpr auto a_k_step_chunk =
+                            k_step + chunk * KThreadChunk * xdlops_gemm.mfma_instr.num_input_blks;
+                        a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
+                                           make_tuple(m0, I0, I0, Number<a_k_step_chunk>{}),
+                                           a_block_buf,
+                                           a_thread_desc_,
+                                           make_tuple(m0, I0, k, Number<chunk * KThreadChunk>{}),
+                                           a_thread_buf);
+                    });
+                });
+            });
+
+            static_for<0, MRepeat, 1>{}([&](auto m0) {
+                static_for<0, NRepeat, 1>{}([&](auto n0) {
+                    static_for<0, KRepeat, 1>{}([&](auto k0) {
+                        vector_type<ComputeTypeA, KPack> a_thread_vec;
+                        vector_type<ComputeTypeB, KPack> b_thread_vec;
+
+                        static_for<0, KPack / ComputePackedSize, 1>{}([&](auto ik) {
+                            a_thread_vec.template AsType<ComputeTypeA>()(ik) =
+                                a_thread_buf[Number<a_thread_desc_.CalculateOffset(
+                                    make_tuple(m0, I0, k0, ik))>{}];
+                            b_thread_vec.template AsType<ComputeTypeB>()(ik) =
+                                b_thread_bufs[I1][Number<b_thread_desc_.CalculateOffset(
+                                    make_tuple(n0, I0, k0, ik))>{}];
+                        });
+
+                        constexpr index_t a_scale_offset =
+                            a_scale_thread_desc.CalculateOffset(make_tuple(m0, k0, I0));
+
+                        constexpr index_t b_scale_offset =
+                            b_scale_thread_desc.CalculateOffset(make_tuple(n0, k0, I0));
+
+                        vector_type<AScaleDataType, ScalesPerXdlopsRunPerThread> a_scale_thread_vec;
+                        vector_type<BScaleDataType, ScalesPerXdlopsRunPerThread> b_scale_thread_vec;
+
+                        // Pack b_scale_thread_buf into b_scale_thread_vec
+                        static_for<0, ScalesPerXdlopsRunPerThread, 1>{}([&](auto s) {
+                            a_scale_thread_vec.template AsType<AScaleDataType>()(s) =
+                                a_scale_thread_bufs[I1][Number<a_scale_offset + s>{}];
+                            b_scale_thread_vec.template AsType<BScaleDataType>()(s) =
+                                b_scale_thread_bufs[I1][Number<b_scale_offset + s>{}];
+                        });
+
+                        using mfma_input_type_a =
+                            typename vector_type<ComputeTypeA,
+                                                 xdlops_gemm.K1PerXdlops / APackedSize>::type;
+                        using mfma_input_type_b =
+                            typename vector_type<ComputeTypeB,
+                                                 xdlops_gemm.K1PerXdlops / BPackedSize>::type;
+
+                        constexpr index_t c_offset =
+                            c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                        // MFMA accumulation
+                        xdlops_gemm.template Run<>(
+                            a_thread_vec.template AsType<mfma_input_type_a>(),
+                            a_scale_thread_vec.template AsType<AScaleDataType>(),
+                            b_thread_vec.template AsType<mfma_input_type_b>(),
+                            b_scale_thread_vec.template AsType<BScaleDataType>(),
+                            c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                    });
+                });
+            });
+        }
+        else if constexpr(TailNum == TailNumber::Odd)
+        {
+            static_for<0, MRepeat, 1>{}([&](auto m0) {
+                static_for<0, NRepeat, 1>{}([&](auto n0) {
+                    static_for<0, KRepeat, 1>{}([&](auto k0) {
+                        vector_type<ComputeTypeA, KPack> a_thread_vec;
+                        vector_type<ComputeTypeB, KPack> b_thread_vec;
+
+                        static_for<0, KPack / ComputePackedSize, 1>{}([&](auto ik) {
+                            a_thread_vec.template AsType<ComputeTypeA>()(ik) =
+                                a_thread_buf[Number<a_thread_desc_.CalculateOffset(
+                                    make_tuple(m0, I0, k0, ik))>{}];
+                            b_thread_vec.template AsType<ComputeTypeB>()(ik) =
+                                b_thread_bufs[I0][Number<b_thread_desc_.CalculateOffset(
+                                    make_tuple(n0, I0, k0, ik))>{}];
+                        });
+
+                        constexpr index_t a_scale_offset =
+                            a_scale_thread_desc.CalculateOffset(make_tuple(m0, k0, I0));
+
+                        constexpr index_t b_scale_offset =
+                            b_scale_thread_desc.CalculateOffset(make_tuple(n0, k0, I0));
+
+                        vector_type<AScaleDataType, ScalesPerXdlopsRunPerThread> a_scale_thread_vec;
+                        vector_type<BScaleDataType, ScalesPerXdlopsRunPerThread> b_scale_thread_vec;
+
+                        // Pack b_scale_thread_buf into b_scale_thread_vec
+                        static_for<0, ScalesPerXdlopsRunPerThread, 1>{}([&](auto s) {
+                            a_scale_thread_vec.template AsType<AScaleDataType>()(s) =
+                                a_scale_thread_bufs[I0][Number<a_scale_offset + s>{}];
+                            b_scale_thread_vec.template AsType<BScaleDataType>()(s) =
+                                b_scale_thread_bufs[I0][Number<b_scale_offset + s>{}];
+                        });
+
+#if 0
+                        printf("bidx: %u, bidy: %u, tidx: %u, a_thread_vec: %02x, %02x, %02x, %02x, b_thread_vec: %02x, %02x, %02x, %02x,"
+                               "a_scale_thread_vec: %02x, b_scale_vec: %02x\n",
+                               blockIdx.x,
+                               blockIdx.y,
+                               threadIdx.x,
+
+                               *(reinterpret_cast<const uint8_t*>(&(a_thread_vec.template AsType<ComputeTypeA>()[Number<0>{}]))),
+                               *(reinterpret_cast<const uint8_t*>(&(a_thread_vec.template AsType<ComputeTypeA>()[Number<1>{}]))),
+                               *(reinterpret_cast<const uint8_t*>(&(a_thread_vec.template AsType<ComputeTypeA>()[Number<2>{}]))),
+                               *(reinterpret_cast<const uint8_t*>(&(a_thread_vec.template AsType<ComputeTypeA>()[Number<3>{}]))),
+                               *(reinterpret_cast<const uint8_t*>(&(b_thread_vec.template AsType<ComputeTypeB>()[Number<0>{}]))),
+                               *(reinterpret_cast<const uint8_t*>(&(b_thread_vec.template AsType<ComputeTypeB>()[Number<1>{}]))),
+                               *(reinterpret_cast<const uint8_t*>(&(b_thread_vec.template AsType<ComputeTypeB>()[Number<2>{}]))),
+                               *(reinterpret_cast<const uint8_t*>(&(b_thread_vec.template AsType<ComputeTypeB>()[Number<3>{}]))),
+
+                            //    type_convert<float>(a_thread_vec.template AsType<ComputeTypeA>()[Number<0>{}].unpack(Number<0>{})),
+                            //    type_convert<float>(a_thread_vec.template AsType<ComputeTypeA>()[Number<0>{}].unpack(Number<1>{})),
+                            //    type_convert<float>(a_thread_vec.template AsType<ComputeTypeA>()[Number<1>{}].unpack(Number<0>{})),
+                            //    type_convert<float>(a_thread_vec.template AsType<ComputeTypeA>()[Number<1>{}].unpack(Number<1>{})),
+                            //    type_convert<float>(b_thread_vec.template AsType<ComputeTypeB>()[Number<0>{}].unpack(Number<0>{})),
+                            //    type_convert<float>(b_thread_vec.template AsType<ComputeTypeB>()[Number<0>{}].unpack(Number<1>{})),
+                            //    type_convert<float>(b_thread_vec.template AsType<ComputeTypeB>()[Number<1>{}].unpack(Number<0>{})),
+                            //    type_convert<float>(b_thread_vec.template AsType<ComputeTypeB>()[Number<1>{}].unpack(Number<1>{})),
+                               *(reinterpret_cast<const uint8_t*>(&(a_scale_thread_vec.template AsType<AScaleDataType>()[Number<0>{}]))),
+                               *(reinterpret_cast<const uint8_t*>(&(b_scale_thread_vec.template AsType<BScaleDataType>()[Number<0>{}]))));
+#endif
+
+                        using mfma_input_type_a =
+                            typename vector_type<ComputeTypeA,
+                                                 xdlops_gemm.K1PerXdlops / APackedSize>::type;
+                        using mfma_input_type_b =
+                            typename vector_type<ComputeTypeB,
+                                                 xdlops_gemm.K1PerXdlops / BPackedSize>::type;
+
+                        constexpr index_t c_offset =
+                            c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                        // MFMA accumulation
+                        xdlops_gemm.template Run<>(
+                            a_thread_vec.template AsType<mfma_input_type_a>(),
+                            a_scale_thread_vec.template AsType<AScaleDataType>(),
+                            b_thread_vec.template AsType<mfma_input_type_b>(),
+                            b_scale_thread_vec.template AsType<BScaleDataType>(),
+                            c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                    });
+#if 0
+                    printf("bidx: %u, bidx: %u, tidx: %u, c_thread_buf: %f, %f, %f, %f\n",
+                        blockIdx.x,
+                        blockIdx.y,
+                        threadIdx.x,
+                        (c_thread_buf[Number<0>{}]),
+                        (c_thread_buf[Number<1>{}]),
+                        (c_thread_buf[Number<2>{}]),
+                        (c_thread_buf[Number<3>{}]));
+#endif
+                });
+            });
+#if 0
+            printf("bidx: %u, bidx: %u, tidx: %u, c_thread_buf: %f, %f, %f, %f\n",
+                blockIdx.x,
+                blockIdx.y,
+                threadIdx.x,
+                type_convert<float>(c_thread_buf.GetVectorTypeReference(Number<0>{}).template AsType<float>()[Number<0>{}]),
+                type_convert<float>(c_thread_buf.GetVectorTypeReference(Number<0>{}).template AsType<float>()[Number<1>{}]),
+                type_convert<float>(c_thread_buf.GetVectorTypeReference(Number<0>{}).template AsType<float>()[Number<2>{}]),
+                type_convert<float>(c_thread_buf.GetVectorTypeReference(Number<0>{}).template AsType<float>()[Number<3>{}]));
+
+#endif
+        }
+    }
+
+    // TODO: make this field protected when a_scale_thread_copy_ is moved
+    // here
+    static constexpr auto a_scale_thread_desc = make_naive_tensor_descriptor_packed(
+        make_tuple(Number<MRepeat>{}, Number<KRepeat>{}, Number<ScalesPerXdlopsRunPerThread>{}));
+
+    // Is used to copy data from a_scale_grid to a_scale_thread
+    static constexpr auto a_scale_thread_desc_copy =
+        make_naive_tensor_descriptor_packed(make_tuple(Number<1>{}, Number<1>{}));
+
+    // TODO: make this field protected when b_scale_thread_copy_ is moved
+    // here
+    static constexpr auto b_scale_thread_desc = make_naive_tensor_descriptor_packed(
+        make_tuple(Number<NRepeat>{}, Number<KRepeat>{}, Number<ScalesPerXdlopsRunPerThread>{}));
+
+    // Is used to copy data from b_scale_grid to b_scale_thread_buf
+    static constexpr auto b_scale_thread_desc_copy =
+        make_naive_tensor_descriptor_packed(make_tuple(Number<1>{}, Number<1>{}));
+
+    protected:
+    static constexpr auto b_thread_desc_ = make_naive_tensor_descriptor_packed(
+        make_tuple(Number<NRepeat>{}, I1, Number<KRepeat>{}, Number<KPack>{}));
+    using Base::a_thread_copy_;
+    using Base::a_thread_desc_;
+    using Base::b_thread_copy_;
+    // using Base::b_thread_desc_;
+    using Base::c_thread_desc_;
+
+    static constexpr BTileDesc b_block_desc_n0_n1_k0_k1;
+};
+
+} // namespace ck

include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuffle_mx_selector.hpp

@@ -0,0 +1,94 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuflle_v1_mx.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuflle_v3_mx.hpp"
+
+namespace ck {
+template <BlockGemmPipelineVersion BlkGemmPipelineVer,
+          BlockGemmPipelineScheduler BlkGemmPipeSche,
+          index_t ThreadBlockSize,
+          index_t ScaleBlockSize,
+          typename ADataType,
+          typename AScaleDataType,
+          typename BDataType,
+          typename BScaleDataType,
+          typename ComputeDataType, // TODO: remove this as in this pipeline ADataType and BDataType
+                                    // must be used for compute
+          typename AccDataType,
+          typename ATileDesc,
+          typename BTileDesc,
+          typename AMmaTileDesc,
+          typename BMmaTileDesc,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MRepeat,
+          index_t NRepeat,
+          index_t KPack>
+constexpr auto BlockGemmMXBPreshufflePipeline_Selector()
+{
+
+    // Hardware MX GEMM pipeline
+    if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v1)
+    {
+        return BlockwiseGemmXdlops_pipeline_bpreshuffle_v1_mx<BlkGemmPipeSche,
+                                                              ThreadBlockSize,
+                                                              ScaleBlockSize,
+                                                              ADataType,
+                                                              AScaleDataType,
+                                                              BDataType,
+                                                              BScaleDataType,
+                                                              ATileDesc,
+                                                              BTileDesc,
+                                                              AMmaTileDesc,
+                                                              BMmaTileDesc,
+                                                              ABlockTransferSrcScalarPerVector,
+                                                              BBlockTransferSrcScalarPerVector,
+                                                              MPerBlock,
+                                                              NPerBlock,
+                                                              KPerBlock,
+                                                              MPerXDL,
+                                                              NPerXDL,
+                                                              MRepeat,
+                                                              NRepeat,
+                                                              KPack>{};
+    }
+    else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v3)
+    {
+        static_assert(MRepeat >= 4, "MRepeat should at least be 4 in BlockGemmPipelineVersion::v3");
+        return BlockwiseGemmXdlops_pipeline_bpreshuffle_v3_mx<BlkGemmPipeSche,
+                                                              ThreadBlockSize,
+                                                              ScaleBlockSize,
+                                                              ADataType,
+                                                              AScaleDataType,
+                                                              BDataType,
+                                                              BScaleDataType,
+                                                              ATileDesc,
+                                                              BTileDesc,
+                                                              AMmaTileDesc,
+                                                              BMmaTileDesc,
+                                                              ABlockTransferSrcScalarPerVector,
+                                                              BBlockTransferSrcScalarPerVector,
+                                                              MPerBlock,
+                                                              NPerBlock,
+                                                              KPerBlock,
+                                                              MPerXDL,
+                                                              NPerXDL,
+                                                              MRepeat,
+                                                              NRepeat,
+                                                              KPack>{};
+    }
+    else
+    {
+        std::cerr << "MX GEMM Pipeline configuration is not available" << std::endl;
+    }
+}
+
+} // namespace ck

include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuffle_selector.hpp

@@ -8,6 +8,7 @@
 #include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuffle_dequant_v1.hpp"
 #include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuffle_gufusion_dequant_v1.hpp"
 #include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuffle_v2.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuffle_gufusion_v3.hpp"
 #include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuffle_v3.hpp"
 #include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuffle_dequant_v3.hpp"
 #include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v4.hpp"
@@ -171,26 +172,54 @@ constexpr auto BlockGemmBPreshufflePipeline_Selector()
         static_assert(MRepeat >= 4, "MRepeat should at least be 4 in BlockGemmPipelineVersion::v3");
         if constexpr(std::is_same<ADataType, BDataType>::value)
         {
-            return BlockwiseGemmXdlops_pipeline_bpreshuffle_v3<BlkGemmPipeSche,
-                                                               BlockSize,
-                                                               ADataType,
-                                                               BDataType,
-                                                               ComputeDataType,
-                                                               AccDataType,
-                                                               ATileDesc,
-                                                               BTileDesc,
-                                                               AMmaTileDesc,
-                                                               BMmaTileDesc,
-                                                               ABlockTransferSrcScalarPerVector,
-                                                               BBlockTransferSrcScalarPerVector,
-                                                               MPerBlock,
-                                                               NPerBlock,
-                                                               KPerBlock,
-                                                               MPerXDL,
-                                                               NPerXDL,
-                                                               MRepeat,
-                                                               NRepeat,
-                                                               KPack>{};
+            if constexpr(GUFusion)
+            {
+                return BlockwiseGemmXdlops_pipeline_bpreshuffle_gufusion_v3<
+                    BlkGemmPipeSche,
+                    BlockSize,
+                    ADataType,
+                    BDataType,
+                    ComputeDataType,
+                    AccDataType,
+                    ATileDesc,
+                    BTileDesc,
+                    AMmaTileDesc,
+                    BMmaTileDesc,
+                    ABlockTransferSrcScalarPerVector,
+                    BBlockTransferSrcScalarPerVector,
+                    MPerBlock,
+                    NPerBlock,
+                    KPerBlock,
+                    MPerXDL,
+                    NPerXDL,
+                    MRepeat,
+                    NRepeat,
+                    KPack>{};
+            }
+            else 
+            {
+                    
+                return BlockwiseGemmXdlops_pipeline_bpreshuffle_v3<BlkGemmPipeSche,
+                                                                BlockSize,
+                                                                ADataType,
+                                                                BDataType,
+                                                                ComputeDataType,
+                                                                AccDataType,
+                                                                ATileDesc,
+                                                                BTileDesc,
+                                                                AMmaTileDesc,
+                                                                BMmaTileDesc,
+                                                                ABlockTransferSrcScalarPerVector,
+                                                                BBlockTransferSrcScalarPerVector,
+                                                                MPerBlock,
+                                                                NPerBlock,
+                                                                KPerBlock,
+                                                                MPerXDL,
+                                                                NPerXDL,
+                                                                MRepeat,
+                                                                NRepeat,
+                                                                KPack>{};
+            }
         }
         else
         {

include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuffle_v1.hpp

@@ -1,5 +1,5 @@
 // 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
 

include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuffle_v2.hpp

@@ -1,5 +1,5 @@
 // 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
 

include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuflle_v1_mx.hpp

@@ -0,0 +1,825 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_mx_pipeline_xdlops_base.hpp"
+
+namespace ck {
+
+// Naive pipeline with lowest resource request per WGP
+// GlobalPrefetchStages: 2
+// LocalPreFillStages: 1
+// LocalPreFetchStages: 1
+// LocalSharedMemoryBuffer: 1
+
+template <BlockGemmPipelineScheduler BlkGemmPipelineVer,
+          index_t ThreadBlockSize,
+          index_t ScaleBlockSize,
+          typename ADataType,
+          typename AScaleDataType,
+          typename BDataType,
+          typename BScaleDataType,
+          typename ATileDesc,
+          typename BTileDesc,
+          typename AMmaTileDesc,
+          typename BMmaTileDesc,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MRepeat, // MXdlPerWave
+          index_t NRepeat, // NXdlPerWave
+          index_t KPack>
+struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v1_mx
+{
+};
+
+template <index_t ThreadBlockSize,
+          index_t ScaleBlockSize,
+          typename ADataType,
+          typename AScaleDataType,
+          typename BDataType,
+          typename BScaleDataType,
+          typename ATileDesc,
+          typename BTileDesc,
+          typename AMmaTileDesc,
+          typename BMmaTileDesc,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MRepeat, // MXdlPerWave
+          index_t NRepeat, // NXdlPerWave
+          index_t KPack>
+struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v1_mx<BlockGemmPipelineScheduler::Intrawave,
+                                                      ThreadBlockSize,
+                                                      ScaleBlockSize,
+                                                      ADataType,
+                                                      AScaleDataType,
+                                                      BDataType,
+                                                      BScaleDataType,
+                                                      ATileDesc,
+                                                      BTileDesc,
+                                                      AMmaTileDesc,
+                                                      BMmaTileDesc,
+                                                      ABlockTransferSrcScalarPerVector,
+                                                      BBlockTransferSrcScalarPerVector,
+                                                      MPerBlock,
+                                                      NPerBlock,
+                                                      KPerBlock,
+                                                      MPerXDL,
+                                                      NPerXDL,
+                                                      MRepeat,
+                                                      NRepeat,
+                                                      KPack>
+    : BlockwiseGemmXdlops_mx_pipeline_base<ThreadBlockSize,
+                                           ADataType,
+                                           BDataType,
+                                           ATileDesc,
+                                           BTileDesc,
+                                           AMmaTileDesc,
+                                           BMmaTileDesc,
+                                           ABlockTransferSrcScalarPerVector,
+                                           BBlockTransferSrcScalarPerVector,
+                                           MPerBlock,
+                                           NPerBlock,
+                                           KPerBlock,
+                                           MPerXDL,
+                                           NPerXDL,
+                                           MRepeat,
+                                           NRepeat,
+                                           KPack>
+
+{
+
+    using Base = BlockwiseGemmXdlops_mx_pipeline_base<ThreadBlockSize,
+                                                      ADataType,
+                                                      BDataType,
+                                                      ATileDesc,
+                                                      BTileDesc,
+                                                      AMmaTileDesc,
+                                                      BMmaTileDesc,
+                                                      ABlockTransferSrcScalarPerVector,
+                                                      BBlockTransferSrcScalarPerVector,
+                                                      MPerBlock,
+                                                      NPerBlock,
+                                                      KPerBlock,
+                                                      MPerXDL,
+                                                      NPerXDL,
+                                                      MRepeat,
+                                                      NRepeat,
+                                                      KPack>;
+    using Base::I0;
+    using Base::I1;
+    using Base::KRepeat;
+    using Base::MWaves;
+    using Base::NWaves;
+    using Base::WaveSize;
+    using Base::xdlops_gemm;
+
+    using Base::CalculateCThreadOriginDataIndex;
+    using Base::GetCBlockDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2;
+    using Base::GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2;
+    using Base::GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4;
+    using Base::GetCThreadBuffer;
+    using Base::GetCThreadDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2;
+    using Base::GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2;
+    using Base::GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4;
+    using Base::GetWaveIdx;
+    using Base::MakeCGridDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2;
+    using Base::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2;
+
+    using Base::a_block_desc_m0_m1_m2_k;
+    using Base::b_block_desc_n0_n1_n2_k;
+
+    using Base::AMmaKStride;
+    using Base::BMmaKStride;
+    using Base::KThreadChunk;
+
+    using Base::ComputePackedSize;
+    using Base::APackedSize;
+    using Base::BPackedSize;
+
+    using AccType      = typename Base::AccType;
+    using Tuple4       = typename Base::Tuple4;
+    using ComputeTypeA = typename Base::ComputeTypeA;
+    using ComputeTypeB = typename Base::ComputeTypeB;
+
+    static constexpr index_t PrefetchStages  = 2;
+    static constexpr index_t PrefillStages   = 1;
+    static constexpr index_t GlobalBufferNum = 2;
+
+    template <typename TileDesc_M0_M1_M2_K>
+    __host__ __device__ static constexpr auto MakeAGemmMmaTileDescriptor(const TileDesc_M0_M1_M2_K&)
+    {
+        constexpr index_t M0 = TileDesc_M0_M1_M2_K{}.GetLength(Number<0>{});
+        constexpr index_t M1 = TileDesc_M0_M1_M2_K{}.GetLength(Number<1>{});
+        constexpr index_t M2 = TileDesc_M0_M1_M2_K{}.GetLength(Number<2>{});
+        constexpr index_t K2 = KPack;
+        constexpr index_t K1 = 64 / NPerXDL;
+        constexpr index_t K0 = KRepeat;
+
+        return transform_tensor_descriptor(
+            TileDesc_M0_M1_M2_K{},
+            make_tuple(
+                make_pass_through_transform(Number<M0>{}),
+                make_pass_through_transform(Number<M1>{}),
+                make_pass_through_transform(Number<M2>{}),
+                make_unmerge_transform(make_tuple(Number<K0>{}, Number<K1>{}, Number<K2>{}))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3, 4, 5>{}));
+    }
+
+    static constexpr auto a_block_desc_m0_m1_m2_k0_k1_k2 =
+        MakeAGemmMmaTileDescriptor(a_block_desc_m0_m1_m2_k);
+
+    static constexpr auto ScalesPerKBlockSize =
+        KPerBlock / ScaleBlockSize; // How many mx-vectors per K block
+
+    //> How many mx-vectors in each row/col is processed in one call to xdlops_gemm.Run()
+    static constexpr auto ScalesPerXdlopsRun = (KPack * xdlops_gemm.K0PerXdlops) / ScaleBlockSize;
+
+    //> How many scales a thread must read to accommodate one call to xdlops_gemm.Run()
+    static constexpr auto ScalesPerXdlopsRunPerThread =
+        ScalesPerXdlopsRun / xdlops_gemm.mfma_instr.num_input_blks;
+
+    __host__ static constexpr bool BlockHasHotloop(index_t num_loop)
+    {
+        return num_loop > PrefetchStages;
+    }
+
+    __host__ static constexpr TailNumber BlockLoopTailNum(index_t num_loop)
+    {
+        return num_loop % 2 == 0 ? TailNumber::Even : TailNumber::Odd;
+    }
+
+    template <bool HasMainLoop,
+              TailNumber TailNum,
+              typename AGridDesc,
+              typename ABlockDesc,
+              typename ABlockTransfer,
+              typename AGridBuffer,
+              typename ABlockBuffer,
+              typename ABlockTransferStep,
+              typename BGridDesc,
+              typename BBlockDesc,
+              typename BBlockTransfer,
+              typename BGridBuffer,
+              typename BBlockBuffer,
+              typename BBlockTransferStep,
+              typename CThreadBuffer,
+              typename AScaleGridBuffer,
+              typename AScaleGridDesc,
+              typename AScaleThreadTransfer,
+              typename BScaleGridBuffer,
+              typename BScaleGridDesc,
+              typename BScaleThreadTransfer>
+    __device__ void Run(
+        // ABlockCopy
+        const AGridDesc& a_grid_desc,
+        const ABlockDesc& a_block_desc,
+        ABlockTransfer& a_blockwise_copy,
+        const AGridBuffer& a_grid_buf,
+        ABlockBuffer& a_block_buf,
+        const ABlockTransferStep& a_block_copy_step,
+        // BBlockCopy
+        const BGridDesc& b_grid_desc,
+        const BBlockDesc& b_block_desc,
+        BBlockTransfer& b_blockwise_copy,
+        const BGridBuffer& b_grid_buf,
+        BBlockBuffer& b_block_buf,
+        const BBlockTransferStep& b_block_copy_step,
+        // CThread
+        CThreadBuffer& c_thread_buf,
+        // A and B scales
+        const AScaleGridDesc& a_scale_grid_desc,
+        AScaleThreadTransfer& a_scale_thread_copy,
+        const AScaleGridBuffer& a_scale_grid_buf,
+        const BScaleGridDesc& b_scale_grid_desc,
+        BScaleThreadTransfer& b_scale_thread_copy,
+        const BScaleGridBuffer& b_scale_grid_buf,
+        index_t num_loop) const
+    {
+        ignore = b_block_desc;
+        ignore = b_block_buf;
+        auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeTypeA>(
+            a_thread_desc_.GetElementSpaceSize());
+        auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeTypeB>(
+            b_thread_desc_.GetElementSpaceSize());
+
+        StaticallyIndexedArray<decltype(b_thread_buf), Number<2>{}> b_thread_bufs;
+        constexpr auto b_block_origin_idx = make_tuple(I0, I0, I0, I0);
+
+        auto a_scale_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, AScaleDataType>(
+            a_scale_thread_desc.GetElementSpaceSize());
+        auto b_scale_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, BScaleDataType>(
+            b_scale_thread_desc.GetElementSpaceSize());
+
+        StaticallyIndexedArray<decltype(a_scale_thread_buf), Number<2>{}> a_scale_thread_bufs;
+        StaticallyIndexedArray<decltype(b_scale_thread_buf), Number<2>{}> b_scale_thread_bufs;
+
+        // Global prefetch A1 B1
+        a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, I0);
+        b_blockwise_copy.Run(b_grid_desc,
+                             b_grid_buf,
+                             b_block_desc_n0_n1_k0_k1,
+                             b_block_origin_idx,
+                             b_thread_bufs(I0));
+
+        a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+        b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+        // Prefetch a_scales to buf 0
+        static_for<0, MRepeat, 1>{}([&](auto m0) {
+            static_for<0, KRepeat, 1>{}([&](auto k0) {
+                static_for<0, ScalesPerXdlopsRunPerThread, 1>{}([&](auto s) {
+                    constexpr auto a_scale_offset =
+                        a_scale_thread_desc.CalculateOffset(make_tuple(m0, k0, s));
+                    auto a_scale_thread_buf_copy =
+                        make_static_buffer<AddressSpaceEnum::Vgpr, AScaleDataType>(
+                            a_scale_thread_desc_copy.GetElementSpaceSize());
+                    a_scale_thread_copy.Run(a_scale_grid_desc,
+                                            a_scale_grid_buf,
+                                            a_scale_thread_desc_copy,
+                                            make_tuple(I0, I0),
+                                            a_scale_thread_buf_copy);
+
+                    a_scale_thread_bufs(I0)(Number<a_scale_offset>{}) =
+                        a_scale_thread_buf_copy[Number<0>{}];
+                    a_scale_thread_copy.MoveSrcSliceWindow(
+                        a_scale_grid_desc,
+                        make_multi_index(0, xdlops_gemm.KPerXdlops / ScaleBlockSize));
+                });
+            });
+            a_scale_thread_copy.MoveSrcSliceWindow(
+                a_scale_grid_desc, make_multi_index(MWaves * MPerXDL, -ScalesPerKBlockSize));
+        });
+
+        // restore row id and advance to the next set of scales
+        a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc,
+                                               make_multi_index(-MPerBlock, ScalesPerKBlockSize));
+
+        // Prefetch b_scales to buf 0
+        static_for<0, NRepeat, 1>{}([&](auto n0) {
+            static_for<0, KRepeat, 1>{}([&](auto k0) {
+                static_for<0, ScalesPerXdlopsRunPerThread, 1>{}([&](auto s) {
+                    constexpr auto b_scale_offset =
+                        b_scale_thread_desc.CalculateOffset(make_tuple(n0, k0, s));
+                    auto b_scale_thread_buf_copy =
+                        make_static_buffer<AddressSpaceEnum::Vgpr, BScaleDataType>(
+                            b_scale_thread_desc_copy.GetElementSpaceSize());
+                    b_scale_thread_copy.Run(b_scale_grid_desc,
+                                            b_scale_grid_buf,
+                                            b_scale_thread_desc_copy,
+                                            make_tuple(I0, I0),
+                                            b_scale_thread_buf_copy);
+
+                    b_scale_thread_bufs(I0)(Number<b_scale_offset>{}) =
+                        b_scale_thread_buf_copy[Number<0>{}];
+                    b_scale_thread_copy.MoveSrcSliceWindow(
+                        b_scale_grid_desc,
+                        make_multi_index(0, xdlops_gemm.KPerXdlops / ScaleBlockSize));
+                });
+            });
+            b_scale_thread_copy.MoveSrcSliceWindow(
+                b_scale_grid_desc, make_multi_index(NWaves * NPerXDL, -ScalesPerKBlockSize));
+        });
+
+        // restore col id and advance to the next set of scales
+        // NWaves * NPerXDL * NRepeat == NPerBlock
+        b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc,
+                                               make_multi_index(-NPerBlock, ScalesPerKBlockSize));
+
+        __builtin_amdgcn_sched_barrier(0);
+
+        // Local prefill A1
+        a_blockwise_copy.RunWrite(a_block_desc, a_block_buf, I0);
+
+        // Global prefetch A2
+        a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, I0);
+        a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+
+        // Prefetch a_scales to buf 1
+        static_for<0, MRepeat, 1>{}([&](auto m0) {
+            static_for<0, KRepeat, 1>{}([&](auto k0) {
+                static_for<0, ScalesPerXdlopsRunPerThread, 1>{}([&](auto s) {
+                    constexpr auto a_scale_offset =
+                        a_scale_thread_desc.CalculateOffset(make_tuple(m0, k0, s));
+                    auto a_scale_thread_buf_copy =
+                        make_static_buffer<AddressSpaceEnum::Vgpr, AScaleDataType>(
+                            a_scale_thread_desc_copy.GetElementSpaceSize());
+                    a_scale_thread_copy.Run(a_scale_grid_desc,
+                                            a_scale_grid_buf,
+                                            a_scale_thread_desc_copy,
+                                            make_tuple(I0, I0),
+                                            a_scale_thread_buf_copy);
+
+                    a_scale_thread_bufs(I1)(Number<a_scale_offset>{}) =
+                        a_scale_thread_buf_copy[Number<0>{}];
+                    a_scale_thread_copy.MoveSrcSliceWindow(
+                        a_scale_grid_desc,
+                        make_multi_index(0, xdlops_gemm.KPerXdlops / ScaleBlockSize));
+                });
+            });
+            a_scale_thread_copy.MoveSrcSliceWindow(
+                a_scale_grid_desc, make_multi_index(MWaves * MPerXDL, -ScalesPerKBlockSize));
+        });
+
+        // restore row id and advance to the next set of scales
+        a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc,
+                                               make_multi_index(-MPerBlock, ScalesPerKBlockSize));
+
+        // Prefetch b_scales to buf 1
+        static_for<0, NRepeat, 1>{}([&](auto n0) {
+            static_for<0, KRepeat, 1>{}([&](auto k0) {
+                static_for<0, ScalesPerXdlopsRunPerThread, 1>{}([&](auto s) {
+                    constexpr auto b_scale_offset =
+                        b_scale_thread_desc.CalculateOffset(make_tuple(n0, k0, s));
+                    auto b_scale_thread_buf_copy =
+                        make_static_buffer<AddressSpaceEnum::Vgpr, BScaleDataType>(
+                            b_scale_thread_desc_copy.GetElementSpaceSize());
+                    b_scale_thread_copy.Run(b_scale_grid_desc,
+                                            b_scale_grid_buf,
+                                            b_scale_thread_desc_copy,
+                                            make_tuple(I0, I0),
+                                            b_scale_thread_buf_copy);
+
+                    b_scale_thread_bufs(I1)(Number<b_scale_offset>{}) =
+                        b_scale_thread_buf_copy[Number<0>{}];
+                    b_scale_thread_copy.MoveSrcSliceWindow(
+                        b_scale_grid_desc,
+                        make_multi_index(0, xdlops_gemm.KPerXdlops / ScaleBlockSize));
+                });
+            });
+            b_scale_thread_copy.MoveSrcSliceWindow(
+                b_scale_grid_desc, make_multi_index(NWaves * NPerXDL, -ScalesPerKBlockSize));
+        });
+
+        b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc,
+                                               make_multi_index(-NPerBlock, ScalesPerKBlockSize));
+
+        // Local prefetch A1
+        block_sync_lds();
+        static_for<0, KRepeat, 1>{}([&](auto k) {
+            constexpr auto k_step = k * xdlops_gemm.KPerXdlops * (KPack / xdlops_gemm.K1PerXdlops);
+
+            static_for<0, MRepeat, 1>{}([&](auto m0) {
+                static_for<0, xdlops_gemm.K1PerXdlops / KThreadChunk, 1>{}([&](auto chunk) {
+                    constexpr auto a_k_step_chunk =
+                        k_step + chunk * KThreadChunk * xdlops_gemm.mfma_instr.num_input_blks;
+                    a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
+                                       make_tuple(m0, I0, I0, Number<a_k_step_chunk>{}),
+                                       a_block_buf,
+                                       a_thread_desc_,
+                                       make_tuple(m0, I0, k, Number<chunk * KThreadChunk>{}),
+                                       a_thread_buf);
+                });
+            });
+        });
+        
+        // Initialize C
+        c_thread_buf.Clear();
+
+        // main body
+        if constexpr(HasMainLoop)
+        {
+            // loop over k with the step KPerBlock
+            index_t i = 0;
+            do
+            {
+                auto LoopFunc = [&](auto mfma_reg_buf, auto local_read_buf) {
+                    b_blockwise_copy.Run(b_grid_desc,
+                                         b_grid_buf,
+                                         b_block_desc_n0_n1_k0_k1,
+                                         b_block_origin_idx,
+                                         b_thread_bufs(local_read_buf));
+                    b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+                    block_sync_lds();
+                    a_blockwise_copy.RunWrite(a_block_desc, a_block_buf, mfma_reg_buf);
+
+                    a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, local_read_buf);
+                    a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+
+                    static_for<0, MRepeat, 1>{}([&](auto m0) {
+                        static_for<0, NRepeat, 1>{}([&](auto n0) {
+                            static_for<0, KRepeat, 1>{}([&](auto k0) {
+                                vector_type<ComputeTypeA, KPack> a_thread_vec;
+                                vector_type<ComputeTypeB, KPack> b_thread_vec;
+
+                                static_for<0, KPack / ComputePackedSize, 1>{}([&](auto ik) {
+                                    a_thread_vec.template AsType<ComputeTypeA>()(ik) =
+                                        a_thread_buf[Number<a_thread_desc_.CalculateOffset(
+                                            make_tuple(m0, I0, k0, ik))>{}];
+                                    b_thread_vec.template AsType<ComputeTypeB>()(ik) =
+                                        b_thread_bufs[mfma_reg_buf]
+                                                     [Number<b_thread_desc_.CalculateOffset(
+                                                         make_tuple(n0, I0, k0, ik))>{}];
+                                });
+
+                                constexpr index_t a_scale_offset =
+                                    a_scale_thread_desc.CalculateOffset(make_tuple(m0, k0, I0));
+                                constexpr index_t b_scale_offset =
+                                    b_scale_thread_desc.CalculateOffset(make_tuple(n0, k0, I0));
+
+                                static_assert(
+                                    0 < ScalesPerXdlopsRunPerThread,
+                                    "Must have at least one scale per Xdlops per Thread.");
+
+                                vector_type<AScaleDataType, ScalesPerXdlopsRunPerThread>
+                                    a_scale_thread_vec;
+                                vector_type<BScaleDataType, ScalesPerXdlopsRunPerThread>
+                                    b_scale_thread_vec;
+
+                                // Pack scale_thread_buf into scale_thread_vec
+                                static_for<0, ScalesPerXdlopsRunPerThread, 1>{}([&](auto s) {
+                                    a_scale_thread_vec.template AsType<AScaleDataType>()(s) =
+                                        a_scale_thread_bufs[mfma_reg_buf]
+                                                           [Number<a_scale_offset + s>{}];
+                                    b_scale_thread_vec.template AsType<BScaleDataType>()(s) =
+                                        b_scale_thread_bufs[mfma_reg_buf]
+                                                           [Number<b_scale_offset + s>{}];
+                                });
+
+                                using mfma_input_type_a =
+                                    typename vector_type<ComputeTypeA,
+                                                         xdlops_gemm.K1PerXdlops / APackedSize>::type;
+                                using mfma_input_type_b =
+                                    typename vector_type<ComputeTypeB,
+                                                         xdlops_gemm.K1PerXdlops / BPackedSize>::type;
+
+                                constexpr index_t c_offset =
+                                    c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                                // MFMA accumulation
+                                xdlops_gemm.template Run<>(
+                                    a_thread_vec.template AsType<mfma_input_type_a>(),
+                                    a_scale_thread_vec.template AsType<AScaleDataType>(),
+                                    b_thread_vec.template AsType<mfma_input_type_b>(),
+                                    b_scale_thread_vec.template AsType<BScaleDataType>(),
+                                    c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                            });
+                        });
+                    });
+
+                    block_sync_lds();
+
+                    // a thread copy
+                    static_for<0, KRepeat, 1>{}([&](auto k) {
+                        constexpr auto k_step =
+                            k * xdlops_gemm.KPerXdlops * (KPack / xdlops_gemm.K1PerXdlops);
+
+                        static_for<0, MRepeat, 1>{}([&](auto m0) {
+                            static_for<0, xdlops_gemm.K1PerXdlops / KThreadChunk, 1>{}(
+                                [&](auto chunk) {
+                                    constexpr auto a_k_step_chunk =
+                                        k_step + chunk * KThreadChunk *
+                                                     xdlops_gemm.mfma_instr.num_input_blks;
+                                    a_thread_copy_.Run(
+                                        a_block_desc_m0_m1_m2_k,
+                                        make_tuple(m0, I0, I0, Number<a_k_step_chunk>{}),
+                                        a_block_buf,
+                                        a_thread_desc_,
+                                        make_tuple(m0, I0, k, Number<chunk * KThreadChunk>{}),
+                                        a_thread_buf);
+                                });
+                        });
+                    });
+
+                    // Prefetch a_scales
+                    static_for<0, MRepeat, 1>{}([&](auto m0) {
+                        static_for<0, KRepeat, 1>{}([&](auto k0) {
+                            static_for<0, ScalesPerXdlopsRunPerThread, 1>{}([&](auto s) {
+                                constexpr auto a_scale_offset =
+                                    a_scale_thread_desc.CalculateOffset(make_tuple(m0, k0, s));
+                                auto a_scale_thread_buf_copy =
+                                    make_static_buffer<AddressSpaceEnum::Vgpr, AScaleDataType>(
+                                        a_scale_thread_desc_copy.GetElementSpaceSize());
+                                a_scale_thread_copy.Run(a_scale_grid_desc,
+                                                        a_scale_grid_buf,
+                                                        a_scale_thread_desc_copy,
+                                                        make_tuple(I0, I0),
+                                                        a_scale_thread_buf_copy);
+
+                                a_scale_thread_bufs(mfma_reg_buf)(Number<a_scale_offset>{}) =
+                                    a_scale_thread_buf_copy[Number<0>{}];
+                                a_scale_thread_copy.MoveSrcSliceWindow(
+                                    a_scale_grid_desc,
+                                    make_multi_index(0, xdlops_gemm.KPerXdlops / ScaleBlockSize));
+                            });
+                        });
+                        a_scale_thread_copy.MoveSrcSliceWindow(
+                            a_scale_grid_desc,
+                            make_multi_index(MWaves * MPerXDL, -ScalesPerKBlockSize));
+                    });
+
+                    // restore row id and advance to the next set of scales
+                    a_scale_thread_copy.MoveSrcSliceWindow(
+                        a_scale_grid_desc, make_multi_index(-MPerBlock, ScalesPerKBlockSize));
+
+                    // Prefetch b_scales
+                    static_for<0, NRepeat, 1>{}([&](auto n0) {
+                        static_for<0, KRepeat, 1>{}([&](auto k0) {
+                            static_for<0, ScalesPerXdlopsRunPerThread, 1>{}([&](auto s) {
+                                constexpr auto b_scale_offset =
+                                    b_scale_thread_desc.CalculateOffset(make_tuple(n0, k0, s));
+                                auto b_scale_thread_buf_copy =
+                                    make_static_buffer<AddressSpaceEnum::Vgpr, BScaleDataType>(
+                                        b_scale_thread_desc_copy.GetElementSpaceSize());
+                                b_scale_thread_copy.Run(b_scale_grid_desc,
+                                                        b_scale_grid_buf,
+                                                        b_scale_thread_desc_copy,
+                                                        make_tuple(I0, I0),
+                                                        b_scale_thread_buf_copy);
+
+                                b_scale_thread_bufs(mfma_reg_buf)(Number<b_scale_offset>{}) =
+                                    b_scale_thread_buf_copy[Number<0>{}];
+                                b_scale_thread_copy.MoveSrcSliceWindow(
+                                    b_scale_grid_desc,
+                                    make_multi_index(0, xdlops_gemm.KPerXdlops / ScaleBlockSize));
+                            });
+                        });
+                        b_scale_thread_copy.MoveSrcSliceWindow(
+                            b_scale_grid_desc,
+                            make_multi_index(NWaves * NPerXDL, -ScalesPerKBlockSize));
+                    });
+
+                    b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc,
+                                               make_multi_index(-NPerBlock, ScalesPerKBlockSize));
+                };
+
+                LoopFunc(I0, I1);
+                LoopFunc(I1, I0);
+
+                i += 2;
+            } while(i < (num_loop - 2));
+        }
+
+        // tail
+        if constexpr(TailNum == TailNumber::Even)
+        {
+            b_blockwise_copy.Run(b_grid_desc,
+                                 b_grid_buf,
+                                 b_block_desc_n0_n1_k0_k1,
+                                 b_block_origin_idx,
+                                 b_thread_bufs(I1));
+            block_sync_lds();
+            a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
+
+            static_for<0, MRepeat, 1>{}([&](auto m0) {
+                static_for<0, NRepeat, 1>{}([&](auto n0) {
+                    static_for<0, KRepeat, 1>{}([&](auto k0) {
+                        vector_type<ComputeTypeA, KPack> a_thread_vec;
+                        vector_type<ComputeTypeB, KPack> b_thread_vec;
+
+                        static_for<0, KPack / ComputePackedSize, 1>{}([&](auto ik) {
+                            a_thread_vec.template AsType<ComputeTypeA>()(ik) =
+                                a_thread_buf[Number<a_thread_desc_.CalculateOffset(
+                                    make_tuple(m0, I0, k0, ik))>{}];
+                            b_thread_vec.template AsType<ComputeTypeB>()(ik) =
+                                b_thread_bufs[I0][Number<b_thread_desc_.CalculateOffset(
+                                    make_tuple(n0, I0, k0, ik))>{}];
+                        });
+
+                        constexpr index_t a_scale_offset =
+                            a_scale_thread_desc.CalculateOffset(make_tuple(m0, k0, I0));
+
+                        constexpr index_t b_scale_offset =
+                            b_scale_thread_desc.CalculateOffset(make_tuple(n0, k0, I0));
+
+                        vector_type<AScaleDataType, ScalesPerXdlopsRunPerThread> a_scale_thread_vec;
+                        vector_type<BScaleDataType, ScalesPerXdlopsRunPerThread> b_scale_thread_vec;
+
+                        // Pack b_scale_thread_buf into b_scale_thread_vec
+                        static_for<0, ScalesPerXdlopsRunPerThread, 1>{}([&](auto s) {
+                            a_scale_thread_vec.template AsType<AScaleDataType>()(s) =
+                                a_scale_thread_bufs[I0][Number<a_scale_offset + s>{}];
+                            b_scale_thread_vec.template AsType<BScaleDataType>()(s) =
+                                b_scale_thread_bufs[I0][Number<b_scale_offset + s>{}];
+                        });
+
+                        using mfma_input_type_a =
+                            typename vector_type<ComputeTypeA, xdlops_gemm.K1PerXdlops / APackedSize>::type;
+                        using mfma_input_type_b =
+                            typename vector_type<ComputeTypeB, xdlops_gemm.K1PerXdlops / BPackedSize>::type;
+
+                        constexpr index_t c_offset =
+                            c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                        // MFMA accumulation
+                        xdlops_gemm.template Run<>(
+                            a_thread_vec.template AsType<mfma_input_type_a>(),
+                            a_scale_thread_vec.template AsType<AScaleDataType>(),
+                            b_thread_vec.template AsType<mfma_input_type_b>(),
+                            b_scale_thread_vec.template AsType<BScaleDataType>(),
+                            c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                    });
+                });
+            });
+
+            block_sync_lds();
+
+            // a thread copy
+            static_for<0, KRepeat, 1>{}([&](auto k) {
+                constexpr auto k_step =
+                    k * xdlops_gemm.KPerXdlops * (KPack / xdlops_gemm.K1PerXdlops);
+
+                static_for<0, MRepeat, 1>{}([&](auto m0) {
+                    static_for<0, xdlops_gemm.K1PerXdlops / KThreadChunk, 1>{}([&](auto chunk) {
+                        constexpr auto a_k_step_chunk =
+                            k_step + chunk * KThreadChunk * xdlops_gemm.mfma_instr.num_input_blks;
+                        a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
+                                           make_tuple(m0, I0, I0, Number<a_k_step_chunk>{}),
+                                           a_block_buf,
+                                           a_thread_desc_,
+                                           make_tuple(m0, I0, k, Number<chunk * KThreadChunk>{}),
+                                           a_thread_buf);
+                    });
+                });
+            });
+
+            static_for<0, MRepeat, 1>{}([&](auto m0) {
+                static_for<0, NRepeat, 1>{}([&](auto n0) {
+                    static_for<0, KRepeat, 1>{}([&](auto k0) {
+                        vector_type<ComputeTypeA, KPack> a_thread_vec;
+                        vector_type<ComputeTypeB, KPack> b_thread_vec;
+
+                        static_for<0, KPack / ComputePackedSize, 1>{}([&](auto ik) {
+                            a_thread_vec.template AsType<ComputeTypeA>()(ik) =
+                                a_thread_buf[Number<a_thread_desc_.CalculateOffset(
+                                    make_tuple(m0, I0, k0, ik))>{}];
+                            b_thread_vec.template AsType<ComputeTypeB>()(ik) =
+                                b_thread_bufs[I1][Number<b_thread_desc_.CalculateOffset(
+                                    make_tuple(n0, I0, k0, ik))>{}];
+                        });
+
+                        constexpr index_t a_scale_offset =
+                            a_scale_thread_desc.CalculateOffset(make_tuple(m0, k0, I0));
+
+                        constexpr index_t b_scale_offset =
+                            b_scale_thread_desc.CalculateOffset(make_tuple(n0, k0, I0));
+
+                        vector_type<AScaleDataType, ScalesPerXdlopsRunPerThread> a_scale_thread_vec;
+                        vector_type<BScaleDataType, ScalesPerXdlopsRunPerThread> b_scale_thread_vec;
+
+                        // Pack b_scale_thread_buf into b_scale_thread_vec
+                        static_for<0, ScalesPerXdlopsRunPerThread, 1>{}([&](auto s) {
+                            a_scale_thread_vec.template AsType<AScaleDataType>()(s) =
+                                a_scale_thread_bufs[I1][Number<a_scale_offset + s>{}];
+                            b_scale_thread_vec.template AsType<BScaleDataType>()(s) =
+                                b_scale_thread_bufs[I1][Number<b_scale_offset + s>{}];
+                        });
+
+                        using mfma_input_type_a =
+                            typename vector_type<ComputeTypeA, xdlops_gemm.K1PerXdlops / APackedSize>::type;
+                        using mfma_input_type_b =
+                            typename vector_type<ComputeTypeB, xdlops_gemm.K1PerXdlops / BPackedSize>::type;
+
+                        constexpr index_t c_offset =
+                            c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                        // MFMA accumulation
+                        xdlops_gemm.template Run<>(
+                            a_thread_vec.template AsType<mfma_input_type_a>(),
+                            a_scale_thread_vec.template AsType<AScaleDataType>(),
+                            b_thread_vec.template AsType<mfma_input_type_b>(),
+                            b_scale_thread_vec.template AsType<BScaleDataType>(),
+                            c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                    });
+                });
+            });
+        }
+        else if constexpr(TailNum == TailNumber::Odd)
+        {
+            static_for<0, MRepeat, 1>{}([&](auto m0) {
+                static_for<0, NRepeat, 1>{}([&](auto n0) {
+                    static_for<0, KRepeat, 1>{}([&](auto k0) {
+                        vector_type<ComputeTypeA, KPack> a_thread_vec;
+                        vector_type<ComputeTypeB, KPack> b_thread_vec;
+
+                        static_for<0, KPack / ComputePackedSize, 1>{}([&](auto ik) {
+                            a_thread_vec.template AsType<ComputeTypeA>()(ik) =
+                                a_thread_buf[Number<a_thread_desc_.CalculateOffset(
+                                    make_tuple(m0, I0, k0, ik))>{}];
+                            b_thread_vec.template AsType<ComputeTypeB>()(ik) =
+                                b_thread_bufs[I0][Number<b_thread_desc_.CalculateOffset(
+                                    make_tuple(n0, I0, k0, ik))>{}];
+                        });
+
+                        constexpr index_t a_scale_offset =
+                            a_scale_thread_desc.CalculateOffset(make_tuple(m0, k0, I0));
+
+                        constexpr index_t b_scale_offset =
+                            b_scale_thread_desc.CalculateOffset(make_tuple(n0, k0, I0));
+
+                        vector_type<AScaleDataType, ScalesPerXdlopsRunPerThread> a_scale_thread_vec;
+                        vector_type<BScaleDataType, ScalesPerXdlopsRunPerThread> b_scale_thread_vec;
+
+                        // Pack b_scale_thread_buf into b_scale_thread_vec
+                        static_for<0, ScalesPerXdlopsRunPerThread, 1>{}([&](auto s) {
+                            a_scale_thread_vec.template AsType<AScaleDataType>()(s) =
+                                a_scale_thread_bufs[I0][Number<a_scale_offset + s>{}];
+                            b_scale_thread_vec.template AsType<BScaleDataType>()(s) =
+                                b_scale_thread_bufs[I0][Number<b_scale_offset + s>{}];
+                        });
+
+                        using mfma_input_type_a =
+                            typename vector_type<ComputeTypeA, xdlops_gemm.K1PerXdlops / APackedSize>::type;
+                        using mfma_input_type_b =
+                            typename vector_type<ComputeTypeB, xdlops_gemm.K1PerXdlops / BPackedSize>::type;
+
+                        constexpr index_t c_offset =
+                            c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                        // MFMA accumulation
+                        xdlops_gemm.template Run<>(
+                            a_thread_vec.template AsType<mfma_input_type_a>(),
+                            a_scale_thread_vec.template AsType<AScaleDataType>(),
+                            b_thread_vec.template AsType<mfma_input_type_b>(),
+                            b_scale_thread_vec.template AsType<BScaleDataType>(),
+                            c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                    });
+                });
+            });
+        }
+    }
+
+    // TODO: make this field protected when a_scale_thread_copy_ is moved
+    // here
+    static constexpr auto a_scale_thread_desc = make_naive_tensor_descriptor_packed(
+        make_tuple(Number<MRepeat>{}, Number<KRepeat>{}, Number<ScalesPerXdlopsRunPerThread>{}));
+
+    // Is used to copy data from a_scale_grid to a_scale_thread
+    static constexpr auto a_scale_thread_desc_copy =
+        make_naive_tensor_descriptor_packed(make_tuple(Number<1>{}, Number<1>{}));
+
+    // TODO: make this field protected when b_scale_thread_copy_ is moved
+    // here
+    static constexpr auto b_scale_thread_desc = make_naive_tensor_descriptor_packed(
+        make_tuple(Number<NRepeat>{}, Number<KRepeat>{}, Number<ScalesPerXdlopsRunPerThread>{}));
+
+    // Is used to copy data from b_scale_grid to b_scale_thread_buf
+    static constexpr auto b_scale_thread_desc_copy =
+        make_naive_tensor_descriptor_packed(make_tuple(Number<1>{}, Number<1>{}));
+
+    protected:
+
+    static constexpr auto b_thread_desc_ = make_naive_tensor_descriptor_packed(
+        make_tuple(Number<NRepeat>{}, I1, Number<KRepeat>{}, Number<KPack>{}));
+    using Base::a_thread_copy_;
+    using Base::a_thread_desc_;
+    using Base::b_thread_copy_;
+    // using Base::b_thread_desc_;
+    using Base::c_thread_desc_;
+
+    static constexpr BTileDesc b_block_desc_n0_n1_k0_k1;
+};
+
+} // namespace ck

include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuflle_v3_mx.hpp

@@ -0,0 +1,932 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_mx_pipeline_xdlops_base.hpp"
+
+namespace ck {
+
+// Naive pipeline with lowest resource request per WGP
+// GlobalPrefetchStages: 2
+// LocalPreFillStages: 1
+// LocalPreFetchStages: 1
+// LocalSharedMemoryBuffer: 1
+
+template <BlockGemmPipelineScheduler BlkGemmPipelineVer,
+          index_t ThreadBlockSize,
+          index_t ScaleBlockSize,
+          typename ADataType,
+          typename AScaleDataType,
+          typename BDataType,
+          typename BScaleDataType,
+          typename ATileDesc,
+          typename BTileDesc,
+          typename AMmaTileDesc,
+          typename BMmaTileDesc,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MRepeat, // MXdlPerWave
+          index_t NRepeat, // NXdlPerWave
+          index_t KPack>
+struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v3_mx
+{
+};
+
+template <index_t ThreadBlockSize,
+          index_t ScaleBlockSize,
+          typename ADataType,
+          typename AScaleDataType,
+          typename BDataType,
+          typename BScaleDataType,
+          typename ATileDesc,
+          typename BTileDesc,
+          typename AMmaTileDesc,
+          typename BMmaTileDesc,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MRepeat, // MXdlPerWave
+          index_t NRepeat, // NXdlPerWave
+          index_t KPack>
+struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v3_mx<BlockGemmPipelineScheduler::Intrawave,
+                                                      ThreadBlockSize,
+                                                      ScaleBlockSize,
+                                                      ADataType,
+                                                      AScaleDataType,
+                                                      BDataType,
+                                                      BScaleDataType,
+                                                      ATileDesc,
+                                                      BTileDesc,
+                                                      AMmaTileDesc,
+                                                      BMmaTileDesc,
+                                                      ABlockTransferSrcScalarPerVector,
+                                                      BBlockTransferSrcScalarPerVector,
+                                                      MPerBlock,
+                                                      NPerBlock,
+                                                      KPerBlock,
+                                                      MPerXDL,
+                                                      NPerXDL,
+                                                      MRepeat,
+                                                      NRepeat,
+                                                      KPack>
+    : BlockwiseGemmXdlops_mx_pipeline_base<ThreadBlockSize,
+                                           ADataType,
+                                           BDataType,
+                                           ATileDesc,
+                                           BTileDesc,
+                                           AMmaTileDesc,
+                                           BMmaTileDesc,
+                                           ABlockTransferSrcScalarPerVector,
+                                           BBlockTransferSrcScalarPerVector,
+                                           MPerBlock,
+                                           NPerBlock,
+                                           KPerBlock,
+                                           MPerXDL,
+                                           NPerXDL,
+                                           MRepeat,
+                                           NRepeat,
+                                           KPack>
+
+{
+
+    using Base = BlockwiseGemmXdlops_mx_pipeline_base<ThreadBlockSize,
+                                                      ADataType,
+                                                      BDataType,
+                                                      ATileDesc,
+                                                      BTileDesc,
+                                                      AMmaTileDesc,
+                                                      BMmaTileDesc,
+                                                      ABlockTransferSrcScalarPerVector,
+                                                      BBlockTransferSrcScalarPerVector,
+                                                      MPerBlock,
+                                                      NPerBlock,
+                                                      KPerBlock,
+                                                      MPerXDL,
+                                                      NPerXDL,
+                                                      MRepeat,
+                                                      NRepeat,
+                                                      KPack>;
+    using Base::I0;
+    using Base::I1;
+    using Base::I2;
+    using Base::KRepeat;
+    using Base::MWaves;
+    using Base::NWaves;
+    using Base::WaveSize;
+    using Base::xdlops_gemm;
+    using typename Base::HotLoopInstList;
+
+    using Base::CalculateCThreadOriginDataIndex;
+    using Base::GetCBlockDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2;
+    using Base::GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2;
+    using Base::GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4;
+    using Base::GetCThreadBuffer;
+    using Base::GetCThreadDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2;
+    using Base::GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2;
+    using Base::GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4;
+    using Base::GetWaveIdx;
+    using Base::MakeCGridDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2;
+    using Base::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2;
+
+    using Base::a_block_desc_m0_m1_m2_k;
+    using Base::b_block_desc_n0_n1_n2_k;
+
+    using Base::AMmaKStride;
+    using Base::BMmaKStride;
+    using Base::KThreadChunk;
+
+    using Base::APackedSize;
+    using Base::BPackedSize;
+    using Base::ComputePackedSize;
+
+    using AccType      = typename Base::AccType;
+    using Tuple4       = typename Base::Tuple4;
+    using ComputeTypeA = typename Base::ComputeTypeA;
+    using ComputeTypeB = typename Base::ComputeTypeB;
+
+    static constexpr index_t PrefetchStages        = 2;
+    static constexpr index_t PrefillStages         = 1;
+    static constexpr index_t GlobalBufferNum       = 2;
+    static constexpr index_t HotloopLocalBufSwitch = MRepeat % 2 == 0 ? 0 : 1;
+
+    template <typename TileDesc_M0_M1_M2_K>
+    __host__ __device__ static constexpr auto MakeAGemmMmaTileDescriptor(const TileDesc_M0_M1_M2_K&)
+    {
+        constexpr index_t M0 = TileDesc_M0_M1_M2_K{}.GetLength(Number<0>{});
+        constexpr index_t M1 = TileDesc_M0_M1_M2_K{}.GetLength(Number<1>{});
+        constexpr index_t M2 = TileDesc_M0_M1_M2_K{}.GetLength(Number<2>{});
+        constexpr index_t K2 = KPack;
+        constexpr index_t K1 = 64 / NPerXDL;
+        constexpr index_t K0 = KRepeat;
+
+        return transform_tensor_descriptor(
+            TileDesc_M0_M1_M2_K{},
+            make_tuple(
+                make_pass_through_transform(Number<M0>{}),
+                make_pass_through_transform(Number<M1>{}),
+                make_pass_through_transform(Number<M2>{}),
+                make_unmerge_transform(make_tuple(Number<K0>{}, Number<K1>{}, Number<K2>{}))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3, 4, 5>{}));
+    }
+
+    static constexpr auto a_block_desc_m0_m1_m2_k0_k1_k2 =
+        MakeAGemmMmaTileDescriptor(a_block_desc_m0_m1_m2_k);
+
+    static constexpr auto ScalesPerKBlockSize =
+        KPerBlock / ScaleBlockSize; // How many mx-vectors per K block
+
+    //> How many mx-vectors in each row/col is processed in one call to xdlops_gemm.Run()
+    static constexpr auto ScalesPerXdlopsRun = (KPack * xdlops_gemm.K0PerXdlops) / ScaleBlockSize;
+
+    //> How many scales a thread must read to accommodate one call to xdlops_gemm.Run()
+    static constexpr auto ScalesPerXdlopsRunPerThread =
+        ScalesPerXdlopsRun / xdlops_gemm.mfma_instr.num_input_blks;
+
+    __host__ static constexpr bool BlockHasHotloop(index_t num_loop)
+    {
+        return num_loop > PrefetchStages;
+    }
+
+    __device__ static constexpr auto HotLoopScheduler()
+    {
+        // A/B split schedule
+        // compiler is likely to use ds_read2 when instruction width smaller than 16bytes
+        constexpr auto num_ds_read_inst_a =
+            HotLoopInstList::A_LDS_Read_Width * sizeof(ADataType) == 16
+                ? HotLoopInstList::A_LDS_Read_Inst_Num
+                : HotLoopInstList::A_LDS_Read_Inst_Num / 2;
+        constexpr auto num_ds_read_inst_b =
+            HotLoopInstList::B_LDS_Read_Width * sizeof(BDataType) == 16
+                ? HotLoopInstList::B_LDS_Read_Inst_Num
+                : HotLoopInstList::B_LDS_Read_Inst_Num / 2;
+
+        constexpr auto num_ds_write_inst_a = HotLoopInstList::A_LDS_Write_Inst_Num;
+        constexpr auto num_ds_write_inst_b = HotLoopInstList::B_LDS_Write_Inst_Num;
+
+        constexpr auto num_buffer_load_inst_a = HotLoopInstList::A_Buffer_Load_Inst_Num;
+        constexpr auto num_buffer_load_inst_b = HotLoopInstList::B_Buffer_Load_Inst_Num;
+
+        constexpr auto num_mfma_inst = HotLoopInstList::C_MFMA_Inst_Num;
+
+        constexpr auto mfma_cycle = HotLoopInstList::C_MFMA_Inst_Cycle;
+        constexpr auto ds_read_a_issue_cycle =
+            HotLoopInstList::A_LDS_Read_Width * sizeof(ADataType) == 16 ? 8 : 4;
+        constexpr auto ds_read_b_issue_cycle =
+            HotLoopInstList::B_LDS_Read_Width * sizeof(BDataType) == 16 ? 8 : 4;
+        constexpr auto ds_read_a_mfma_rate =
+            (mfma_cycle - 4 + 2 * ds_read_a_issue_cycle - 1) / (2 * ds_read_a_issue_cycle);
+        constexpr auto ds_read_b_mfma_rate =
+            (mfma_cycle - 4 + 2 * ds_read_b_issue_cycle - 1) / (2 * ds_read_b_issue_cycle);
+
+        constexpr auto num_dsread_a_mfma =
+            (num_ds_read_inst_a + ds_read_a_mfma_rate - 1) / ds_read_a_mfma_rate;
+        constexpr auto num_dsread_b_mfma =
+            (num_ds_read_inst_b + ds_read_b_mfma_rate - 1) / ds_read_b_mfma_rate;
+
+        // stage 1
+        // Separate this part?
+        // constexpr auto num_mfma_per_ds_read = sizeof(ComputeDataType) / sizeof(ADataType) >
+        //                                               sizeof(ComputeDataType) / sizeof(BDataType)
+        //                                           ? sizeof(ComputeDataType) / sizeof(ADataType)
+        //                                           : sizeof(ComputeDataType) / sizeof(BDataType);
+        constexpr auto num_mfma_stage1 = num_mfma_inst - (num_dsread_a_mfma + num_dsread_b_mfma);
+        constexpr auto num_mfma_per_issue =
+            num_mfma_stage1 / (num_buffer_load_inst_a + num_buffer_load_inst_b);
+        constexpr auto num_dswrite_per_issue_a = num_ds_write_inst_a / num_buffer_load_inst_a;
+        constexpr auto num_dswrite_per_issue_b = num_ds_write_inst_b / num_buffer_load_inst_b;
+
+        static_for<0, num_buffer_load_inst_a, 1>{}([&](auto i) {
+            ignore = i;
+            static_for<0, num_dswrite_per_issue_a, 1>{}([&](auto idswrite) {
+                ignore = idswrite;
+                __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS write
+                __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+            });
+            __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read
+            __builtin_amdgcn_sched_group_barrier(
+                0x008, num_mfma_per_issue - num_dswrite_per_issue_a, 0); // MFMA
+        });
+        static_for<0, num_buffer_load_inst_b, 1>{}([&](auto i) {
+            ignore = i;
+            static_for<0, num_dswrite_per_issue_b, 1>{}([&](auto idswrite) {
+                ignore = idswrite;
+                __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS write
+                __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+            });
+            __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read
+            __builtin_amdgcn_sched_group_barrier(
+                0x008, num_mfma_per_issue - num_dswrite_per_issue_b, 0); // MFMA
+        });
+
+        // stage 2
+        static_for<0, num_dsread_a_mfma, 1>{}([&](auto i) {
+            if constexpr((num_ds_read_inst_a - (i + 1) * ds_read_a_mfma_rate) >=
+                         ds_read_a_mfma_rate)
+            {
+                __builtin_amdgcn_sched_group_barrier(0x100, ds_read_a_mfma_rate, 0); // DS read
+            }
+            else
+            {
+                __builtin_amdgcn_sched_group_barrier(0x100,
+                                                     num_ds_read_inst_a - (num_dsread_a_mfma - 1) *
+                                                                              ds_read_a_mfma_rate,
+                                                     0); // DS read
+            }
+            __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+        });
+
+        static_for<0, num_dsread_b_mfma, 1>{}([&](auto i) {
+            if constexpr((num_ds_read_inst_b - (i + 1) * ds_read_b_mfma_rate) >=
+                         ds_read_b_mfma_rate)
+            {
+                __builtin_amdgcn_sched_group_barrier(0x100, ds_read_b_mfma_rate, 0); // DS read
+            }
+            else
+            {
+                __builtin_amdgcn_sched_group_barrier(0x100,
+                                                     num_ds_read_inst_b - (num_dsread_b_mfma - 1) *
+                                                                              ds_read_b_mfma_rate,
+                                                     0); // DS read
+            }
+            __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+        });
+    }
+
+    __host__ static constexpr TailNumber BlockLoopTailNum(index_t num_loop)
+    {
+        return num_loop % 2 == 0 ? TailNumber::Even : TailNumber::Odd;
+    }
+
+    template <bool HasMainLoop,
+              TailNumber TailNum,
+              typename AGridDesc,
+              typename ABlockDesc,
+              typename ABlockTransfer,
+              typename AGridBuffer,
+              typename ABlockBuffer,
+              typename ABlockTransferStep,
+              typename BGridDesc,
+              typename BBlockDesc,
+              typename BBlockTransfer,
+              typename BGridBuffer,
+              typename BBlockBuffer,
+              typename BBlockTransferStep,
+              typename CThreadBuffer,
+              typename AScaleGridBuffer,
+              typename AScaleGridDesc,
+              typename AScaleThreadTransfer,
+              typename BScaleGridBuffer,
+              typename BScaleGridDesc,
+              typename BScaleThreadTransfer>
+    __device__ void Run(
+        // ABlockCopy
+        const AGridDesc& a_grid_desc,
+        const ABlockDesc& a_block_desc,
+        ABlockTransfer& a_blockwise_copy,
+        const AGridBuffer& a_grid_buf,
+        ABlockBuffer& a_block_buf,
+        const ABlockTransferStep& a_block_copy_step,
+        // BBlockCopy
+        const BGridDesc& b_grid_desc,
+        const BBlockDesc& b_block_desc,
+        BBlockTransfer& b_blockwise_copy,
+        const BGridBuffer& b_grid_buf,
+        BBlockBuffer& b_block_buf,
+        const BBlockTransferStep& b_block_copy_step,
+        // CThread
+        CThreadBuffer& c_thread_buf,
+        // A and B scales
+        const AScaleGridDesc& a_scale_grid_desc,
+        AScaleThreadTransfer& a_scale_thread_copy,
+        const AScaleGridBuffer& a_scale_grid_buf,
+        const BScaleGridDesc& b_scale_grid_desc,
+        BScaleThreadTransfer& b_scale_thread_copy,
+        const BScaleGridBuffer& b_scale_grid_buf,
+        index_t num_loop) const
+    {
+        auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeTypeA>(
+            a_thread_desc_.GetElementSpaceSize());
+        auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeTypeB>(
+            b_thread_desc_.GetElementSpaceSize());
+
+        StaticallyIndexedArray<decltype(b_thread_buf), Number<2>{}> b_thread_bufs;
+        constexpr auto b_block_origin_idx = make_tuple(I0, I0, I0, I0);
+
+        auto a_scale_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, AScaleDataType>(
+            a_scale_thread_desc.GetElementSpaceSize());
+        auto b_scale_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, BScaleDataType>(
+            b_scale_thread_desc.GetElementSpaceSize());
+
+        StaticallyIndexedArray<decltype(a_scale_thread_buf), Number<2>{}> a_scale_thread_bufs;
+        StaticallyIndexedArray<decltype(b_scale_thread_buf), Number<2>{}> b_scale_thread_bufs;
+
+        // Global prefetch B1
+        b_blockwise_copy.Run(b_grid_desc,
+                             b_grid_buf,
+                             b_block_desc_n0_n1_k0_k1,
+                             b_block_origin_idx,
+                             b_thread_bufs(I0));
+        b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+        // Global prefetch A1
+        a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
+        a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+
+        // Prefetch a_scales 1
+        static_for<0, MRepeat, 1>{}([&](auto m0) {
+            static_for<0, KRepeat, 1>{}([&](auto k0) {
+                static_for<0, ScalesPerXdlopsRunPerThread, 1>{}([&](auto s) {
+                    constexpr auto a_scale_offset =
+                        a_scale_thread_desc.CalculateOffset(make_tuple(m0, k0, s));
+                    auto a_scale_thread_buf_copy =
+                        make_static_buffer<AddressSpaceEnum::Vgpr, AScaleDataType>(
+                            a_scale_thread_desc_copy.GetElementSpaceSize());
+                    a_scale_thread_copy.Run(a_scale_grid_desc,
+                                            a_scale_grid_buf,
+                                            a_scale_thread_desc_copy,
+                                            make_tuple(I0, I0),
+                                            a_scale_thread_buf_copy);
+
+                    a_scale_thread_bufs(I0)(Number<a_scale_offset>{}) =
+                        a_scale_thread_buf_copy[Number<0>{}];
+                    a_scale_thread_copy.MoveSrcSliceWindow(
+                        a_scale_grid_desc,
+                        make_multi_index(0, xdlops_gemm.KPerXdlops / ScaleBlockSize));
+                });
+            });
+            a_scale_thread_copy.MoveSrcSliceWindow(
+                a_scale_grid_desc, make_multi_index(MWaves * MPerXDL, -ScalesPerKBlockSize));
+        });
+        // restore row id and advance to the next set of scales
+        a_scale_thread_copy.MoveSrcSliceWindow(a_scale_grid_desc,
+                                               make_multi_index(-MPerBlock, ScalesPerKBlockSize));
+
+        // Prefetch b_scales 1
+        static_for<0, NRepeat, 1>{}([&](auto n0) {
+            static_for<0, KRepeat, 1>{}([&](auto k0) {
+                static_for<0, ScalesPerXdlopsRunPerThread, 1>{}([&](auto s) {
+                    constexpr auto b_scale_offset =
+                        b_scale_thread_desc.CalculateOffset(make_tuple(n0, k0, s));
+                    auto b_scale_thread_buf_copy =
+                        make_static_buffer<AddressSpaceEnum::Vgpr, BScaleDataType>(
+                            b_scale_thread_desc_copy.GetElementSpaceSize());
+                    b_scale_thread_copy.Run(b_scale_grid_desc,
+                                            b_scale_grid_buf,
+                                            b_scale_thread_desc_copy,
+                                            make_tuple(I0, I0),
+                                            b_scale_thread_buf_copy);
+
+                    b_scale_thread_bufs(I0)(Number<b_scale_offset>{}) =
+                        b_scale_thread_buf_copy[Number<0>{}];
+                    b_scale_thread_copy.MoveSrcSliceWindow(
+                        b_scale_grid_desc,
+                        make_multi_index(0, xdlops_gemm.KPerXdlops / ScaleBlockSize));
+                });
+            });
+            b_scale_thread_copy.MoveSrcSliceWindow(
+                b_scale_grid_desc, make_multi_index(NWaves * NPerXDL, -ScalesPerKBlockSize));
+        });
+        // restore col id and advance to the next set of scales
+        b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc,
+                                               make_multi_index(-NPerBlock, ScalesPerKBlockSize));
+
+        // Local prefill A1
+        a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(I0)); // vmem->vgpr-> lds0
+
+        // Global prefetch A2
+        a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
+        a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+
+        // Initialize C
+        c_thread_buf.Clear();
+
+        // Local prefetch A1
+        block_sync_lds();
+        static_for<0, KRepeat, 1>{}([&](auto k) {
+            constexpr auto k_step = k * xdlops_gemm.KPerXdlops * (KPack / xdlops_gemm.K1PerXdlops);
+
+            static_for<0, MRepeat, 1>{}([&](auto m0) {
+                static_for<0, xdlops_gemm.K1PerXdlops / KThreadChunk, 1>{}([&](auto chunk) {
+                    constexpr auto a_k_step_chunk =
+                        k_step + chunk * KThreadChunk * xdlops_gemm.mfma_instr.num_input_blks;
+                    a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
+                                       make_tuple(m0, I0, I0, Number<a_k_step_chunk>{}),
+                                       a_block_buf.At(I0),
+                                       a_thread_desc_,
+                                       make_tuple(m0, I0, k, Number<chunk * KThreadChunk>{}),
+                                       a_thread_buf);
+                });
+            });
+        });
+
+        // main body
+        if constexpr(HasMainLoop)
+        {
+            // loop over k with the step KPerBlock
+            index_t i = 0;
+            do
+            {
+                auto LoopFunc = [&](auto mfma_reg_buf, auto local_read_buf, auto a_buf) {
+                    // Prefetch a_scales 2
+                    static_for<0, MRepeat, 1>{}([&](auto m0) {
+                        static_for<0, KRepeat, 1>{}([&](auto k0) {
+                            static_for<0, ScalesPerXdlopsRunPerThread, 1>{}([&](auto s) {
+                                constexpr auto a_scale_offset =
+                                    a_scale_thread_desc.CalculateOffset(make_tuple(m0, k0, s));
+                                auto a_scale_thread_buf_copy =
+                                    make_static_buffer<AddressSpaceEnum::Vgpr, AScaleDataType>(
+                                        a_scale_thread_desc_copy.GetElementSpaceSize());
+                                a_scale_thread_copy.Run(a_scale_grid_desc,
+                                                        a_scale_grid_buf,
+                                                        a_scale_thread_desc_copy,
+                                                        make_tuple(I0, I0),
+                                                        a_scale_thread_buf_copy);
+
+                                a_scale_thread_bufs(local_read_buf)(Number<a_scale_offset>{}) =
+                                    a_scale_thread_buf_copy[Number<0>{}];
+                                a_scale_thread_copy.MoveSrcSliceWindow(
+                                    a_scale_grid_desc,
+                                    make_multi_index(0, xdlops_gemm.KPerXdlops / ScaleBlockSize));
+                            });
+                        });
+                        a_scale_thread_copy.MoveSrcSliceWindow(
+                            a_scale_grid_desc,
+                            make_multi_index(MWaves * MPerXDL, -ScalesPerKBlockSize));
+                    });
+                    // restore row id and advance to the next set of scales
+                    a_scale_thread_copy.MoveSrcSliceWindow(
+                        a_scale_grid_desc, make_multi_index(-MPerBlock, ScalesPerKBlockSize));
+
+                    // Prefetch b_scales 2
+                    static_for<0, NRepeat, 1>{}([&](auto n0) {
+                        static_for<0, KRepeat, 1>{}([&](auto k0) {
+                            static_for<0, ScalesPerXdlopsRunPerThread, 1>{}([&](auto s) {
+                                constexpr auto b_scale_offset =
+                                    b_scale_thread_desc.CalculateOffset(make_tuple(n0, k0, s));
+                                auto b_scale_thread_buf_copy =
+                                    make_static_buffer<AddressSpaceEnum::Vgpr, BScaleDataType>(
+                                        b_scale_thread_desc_copy.GetElementSpaceSize());
+                                b_scale_thread_copy.Run(b_scale_grid_desc,
+                                                        b_scale_grid_buf,
+                                                        b_scale_thread_desc_copy,
+                                                        make_tuple(I0, I0),
+                                                        b_scale_thread_buf_copy);
+
+                                b_scale_thread_bufs(local_read_buf)(Number<b_scale_offset>{}) =
+                                    b_scale_thread_buf_copy[Number<0>{}];
+                                b_scale_thread_copy.MoveSrcSliceWindow(
+                                    b_scale_grid_desc,
+                                    make_multi_index(0, xdlops_gemm.KPerXdlops / ScaleBlockSize));
+                            });
+                        });
+                        b_scale_thread_copy.MoveSrcSliceWindow(
+                            b_scale_grid_desc,
+                            make_multi_index(NWaves * NPerXDL, -ScalesPerKBlockSize));
+                    });
+                    // restore col id and advance to the next set of scales
+                    b_scale_thread_copy.MoveSrcSliceWindow(
+                        b_scale_grid_desc, make_multi_index(-NPerBlock, ScalesPerKBlockSize));
+
+                    // Local prefill A2
+                    block_sync_lds();
+                    a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(local_read_buf));
+
+                    // Global prefetch A1
+                    a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
+                    a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+
+                    // Global prefetch B2
+                    b_blockwise_copy.Run(b_grid_desc,
+                                         b_grid_buf,
+                                         b_block_desc_n0_n1_k0_k1,
+                                         b_block_origin_idx,
+                                         b_thread_bufs(local_read_buf));
+                    b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+                    // A1 * B1
+                    static_for<0, MRepeat, 1>{}([&](auto m0) {
+                        static_for<0, NRepeat, 1>{}([&](auto n0) {
+                            static_for<0, KRepeat, 1>{}([&](auto k0) {
+                                vector_type<ComputeTypeA, KPack> a_thread_vec;
+                                vector_type<ComputeTypeB, KPack> b_thread_vec;
+
+                                static_for<0, KPack / ComputePackedSize, 1>{}([&](auto ik) {
+                                    a_thread_vec.template AsType<ComputeTypeA>()(ik) =
+                                        a_thread_buf[Number<a_thread_desc_.CalculateOffset(
+                                            make_tuple(m0, I0, k0, ik))>{}];
+                                    b_thread_vec.template AsType<ComputeTypeB>()(ik) =
+                                        b_thread_bufs[mfma_reg_buf]
+                                                     [Number<b_thread_desc_.CalculateOffset(
+                                                         make_tuple(n0, I0, k0, ik))>{}];
+                                });
+
+                                constexpr index_t a_scale_offset =
+                                    a_scale_thread_desc.CalculateOffset(make_tuple(m0, k0, I0));
+                                constexpr index_t b_scale_offset =
+                                    b_scale_thread_desc.CalculateOffset(make_tuple(n0, k0, I0));
+
+                                vector_type<AScaleDataType, ScalesPerXdlopsRunPerThread>
+                                    a_scale_thread_vec;
+                                vector_type<BScaleDataType, ScalesPerXdlopsRunPerThread>
+                                    b_scale_thread_vec;
+
+                                // Pack scale_thread_buf into scale_thread_vec
+                                static_for<0, ScalesPerXdlopsRunPerThread, 1>{}([&](auto s) {
+                                    a_scale_thread_vec.template AsType<AScaleDataType>()(s) =
+                                        a_scale_thread_bufs[mfma_reg_buf]
+                                                           [Number<a_scale_offset + s>{}];
+                                    b_scale_thread_vec.template AsType<BScaleDataType>()(s) =
+                                        b_scale_thread_bufs[mfma_reg_buf]
+                                                           [Number<b_scale_offset + s>{}];
+                                });
+
+                                using mfma_input_type_a =
+                                    typename vector_type<ComputeTypeA,
+                                                         xdlops_gemm.K1PerXdlops /
+                                                             APackedSize>::type;
+                                using mfma_input_type_b =
+                                    typename vector_type<ComputeTypeB,
+                                                         xdlops_gemm.K1PerXdlops /
+                                                             BPackedSize>::type;
+
+                                constexpr index_t c_offset =
+                                    c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                                // MFMA accumulation
+                                xdlops_gemm.template Run<>(
+                                    a_thread_vec.template AsType<mfma_input_type_a>(),
+                                    a_scale_thread_vec.template AsType<AScaleDataType>(),
+                                    b_thread_vec.template AsType<mfma_input_type_b>(),
+                                    b_scale_thread_vec.template AsType<BScaleDataType>(),
+                                    c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                            }); // KRepeat
+                        });     // NRepeat
+                    });         // MRepeat
+
+                    // Local prefetch A2
+                    block_sync_lds();
+                    static_for<0, KRepeat, 1>{}([&](auto k) {
+                        constexpr auto k_step =
+                            k * xdlops_gemm.KPerXdlops * (KPack / xdlops_gemm.K1PerXdlops);
+
+                        static_for<0, MRepeat, 1>{}([&](auto m0) {
+                            static_for<0, xdlops_gemm.K1PerXdlops / KThreadChunk, 1>{}(
+                                [&](auto chunk) {
+                                    constexpr auto a_k_step_chunk =
+                                        k_step + chunk * KThreadChunk *
+                                                     xdlops_gemm.mfma_instr.num_input_blks;
+                                    a_thread_copy_.Run(
+                                        a_block_desc_m0_m1_m2_k,
+                                        make_tuple(m0, I0, I0, Number<a_k_step_chunk>{}),
+                                        a_block_buf.At(local_read_buf),
+                                        a_thread_desc_,
+                                        make_tuple(m0, I0, k, Number<chunk * KThreadChunk>{}),
+                                        a_thread_buf);
+                                });
+                        });
+                    });
+
+                     HotLoopScheduler();
+                    __builtin_amdgcn_sched_barrier(0);
+                }; // LoopFunc
+
+                LoopFunc(I0, I1, I0);
+                LoopFunc(I1, I0, I1);
+
+                i += 2;
+            } while(i < (num_loop - 2));
+        }
+
+        // tail
+        if constexpr(TailNum == TailNumber::Even)
+        {
+            // Prefetch a_scales 2
+            static_for<0, MRepeat, 1>{}([&](auto m0) {
+                static_for<0, KRepeat, 1>{}([&](auto k0) {
+                    static_for<0, ScalesPerXdlopsRunPerThread, 1>{}([&](auto s) {
+                        constexpr auto a_scale_offset =
+                            a_scale_thread_desc.CalculateOffset(make_tuple(m0, k0, s));
+                        auto a_scale_thread_buf_copy =
+                            make_static_buffer<AddressSpaceEnum::Vgpr, AScaleDataType>(
+                                a_scale_thread_desc_copy.GetElementSpaceSize());
+                        a_scale_thread_copy.Run(a_scale_grid_desc,
+                                                a_scale_grid_buf,
+                                                a_scale_thread_desc_copy,
+                                                make_tuple(I0, I0),
+                                                a_scale_thread_buf_copy);
+
+                        a_scale_thread_bufs(I1)(Number<a_scale_offset>{}) =
+                            a_scale_thread_buf_copy[Number<0>{}];
+                        a_scale_thread_copy.MoveSrcSliceWindow(
+                            a_scale_grid_desc,
+                            make_multi_index(0, xdlops_gemm.KPerXdlops / ScaleBlockSize));
+                    });
+                });
+                a_scale_thread_copy.MoveSrcSliceWindow(
+                    a_scale_grid_desc, make_multi_index(MWaves * MPerXDL, -ScalesPerKBlockSize));
+            });
+
+            // Prefetch b_scales 2
+            static_for<0, NRepeat, 1>{}([&](auto n0) {
+                static_for<0, KRepeat, 1>{}([&](auto k0) {
+                    static_for<0, ScalesPerXdlopsRunPerThread, 1>{}([&](auto s) {
+                        constexpr auto b_scale_offset =
+                            b_scale_thread_desc.CalculateOffset(make_tuple(n0, k0, s));
+                        auto b_scale_thread_buf_copy =
+                            make_static_buffer<AddressSpaceEnum::Vgpr, BScaleDataType>(
+                                b_scale_thread_desc_copy.GetElementSpaceSize());
+                        b_scale_thread_copy.Run(b_scale_grid_desc,
+                                                b_scale_grid_buf,
+                                                b_scale_thread_desc_copy,
+                                                make_tuple(I0, I0),
+                                                b_scale_thread_buf_copy);
+
+                        b_scale_thread_bufs(I1)(Number<b_scale_offset>{}) =
+                            b_scale_thread_buf_copy[Number<0>{}];
+                        b_scale_thread_copy.MoveSrcSliceWindow(
+                            b_scale_grid_desc,
+                            make_multi_index(0, xdlops_gemm.KPerXdlops / ScaleBlockSize));
+                    });
+                });
+                b_scale_thread_copy.MoveSrcSliceWindow(
+                    b_scale_grid_desc, make_multi_index(NWaves * NPerXDL, -ScalesPerKBlockSize));
+            });
+
+            // Local prefill A2
+            block_sync_lds();
+            a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(I1));
+
+            // Global prefetch B2
+            b_blockwise_copy.Run(b_grid_desc,
+                                 b_grid_buf,
+                                 b_block_desc_n0_n1_k0_k1,
+                                 b_block_origin_idx,
+                                 b_thread_bufs(I1));
+
+            // A1 * B1
+            static_for<0, MRepeat, 1>{}([&](auto m0) {
+                static_for<0, NRepeat, 1>{}([&](auto n0) {
+                    static_for<0, KRepeat, 1>{}([&](auto k0) {
+                        vector_type<ComputeTypeA, KPack> a_thread_vec;
+                        vector_type<ComputeTypeB, KPack> b_thread_vec;
+
+                        static_for<0, KPack / ComputePackedSize, 1>{}([&](auto ik) {
+                            a_thread_vec.template AsType<ComputeTypeA>()(ik) =
+                                a_thread_buf[Number<a_thread_desc_.CalculateOffset(
+                                    make_tuple(m0, I0, k0, ik))>{}];
+                            b_thread_vec.template AsType<ComputeTypeB>()(ik) =
+                                b_thread_bufs[I0][Number<b_thread_desc_.CalculateOffset(
+                                    make_tuple(n0, I0, k0, ik))>{}];
+                        });
+
+                        constexpr index_t a_scale_offset =
+                            a_scale_thread_desc.CalculateOffset(make_tuple(m0, k0, I0));
+
+                        constexpr index_t b_scale_offset =
+                            b_scale_thread_desc.CalculateOffset(make_tuple(n0, k0, I0));
+
+                        vector_type<AScaleDataType, ScalesPerXdlopsRunPerThread> a_scale_thread_vec;
+                        vector_type<BScaleDataType, ScalesPerXdlopsRunPerThread> b_scale_thread_vec;
+
+                        // Pack b_scale_thread_buf into b_scale_thread_vec
+                        static_for<0, ScalesPerXdlopsRunPerThread, 1>{}([&](auto s) {
+                            a_scale_thread_vec.template AsType<AScaleDataType>()(s) =
+                                a_scale_thread_bufs[I0][Number<a_scale_offset + s>{}];
+                            b_scale_thread_vec.template AsType<BScaleDataType>()(s) =
+                                b_scale_thread_bufs[I0][Number<b_scale_offset + s>{}];
+                        });
+
+                        using mfma_input_type_a =
+                            typename vector_type<ComputeTypeA,
+                                                 xdlops_gemm.K1PerXdlops / APackedSize>::type;
+                        using mfma_input_type_b =
+                            typename vector_type<ComputeTypeB,
+                                                 xdlops_gemm.K1PerXdlops / BPackedSize>::type;
+
+                        constexpr index_t c_offset =
+                            c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                        // MFMA accumulation
+                        xdlops_gemm.template Run<>(
+                            a_thread_vec.template AsType<mfma_input_type_a>(),
+                            a_scale_thread_vec.template AsType<AScaleDataType>(),
+                            b_thread_vec.template AsType<mfma_input_type_b>(),
+                            b_scale_thread_vec.template AsType<BScaleDataType>(),
+                            c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                    }); // KRepeat
+                });     // NRepeat
+            });         // MRepeat
+
+            // Local prefetch A2
+            block_sync_lds();
+            static_for<0, KRepeat, 1>{}([&](auto k) {
+                constexpr auto k_step =
+                    k * xdlops_gemm.KPerXdlops * (KPack / xdlops_gemm.K1PerXdlops);
+
+                static_for<0, MRepeat, 1>{}([&](auto m0) {
+                    static_for<0, xdlops_gemm.K1PerXdlops / KThreadChunk, 1>{}([&](auto chunk) {
+                        constexpr auto a_k_step_chunk =
+                            k_step + chunk * KThreadChunk * xdlops_gemm.mfma_instr.num_input_blks;
+                        a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
+                                           make_tuple(m0, I0, I0, Number<a_k_step_chunk>{}),
+                                           a_block_buf.At(I1),
+                                           a_thread_desc_,
+                                           make_tuple(m0, I0, k, Number<chunk * KThreadChunk>{}),
+                                           a_thread_buf);
+                    });
+                });
+            });
+
+            // A2 * B2
+            static_for<0, MRepeat, 1>{}([&](auto m0) {
+                static_for<0, NRepeat, 1>{}([&](auto n0) {
+                    static_for<0, KRepeat, 1>{}([&](auto k0) {
+                        vector_type<ComputeTypeA, KPack> a_thread_vec;
+                        vector_type<ComputeTypeB, KPack> b_thread_vec;
+
+                        static_for<0, KPack / ComputePackedSize, 1>{}([&](auto ik) {
+                            a_thread_vec.template AsType<ComputeTypeA>()(ik) =
+                                a_thread_buf[Number<a_thread_desc_.CalculateOffset(
+                                    make_tuple(m0, I0, k0, ik))>{}];
+                            b_thread_vec.template AsType<ComputeTypeB>()(ik) =
+                                b_thread_bufs[I1][Number<b_thread_desc_.CalculateOffset(
+                                    make_tuple(n0, I0, k0, ik))>{}];
+                        });
+
+                        constexpr index_t a_scale_offset =
+                            a_scale_thread_desc.CalculateOffset(make_tuple(m0, k0, I0));
+
+                        constexpr index_t b_scale_offset =
+                            b_scale_thread_desc.CalculateOffset(make_tuple(n0, k0, I0));
+
+                        vector_type<AScaleDataType, ScalesPerXdlopsRunPerThread> a_scale_thread_vec;
+                        vector_type<BScaleDataType, ScalesPerXdlopsRunPerThread> b_scale_thread_vec;
+
+                        // Pack b_scale_thread_buf into b_scale_thread_vec
+                        static_for<0, ScalesPerXdlopsRunPerThread, 1>{}([&](auto s) {
+                            a_scale_thread_vec.template AsType<AScaleDataType>()(s) =
+                                a_scale_thread_bufs[I1][Number<a_scale_offset + s>{}];
+                            b_scale_thread_vec.template AsType<BScaleDataType>()(s) =
+                                b_scale_thread_bufs[I1][Number<b_scale_offset + s>{}];
+                        });
+
+                        using mfma_input_type_a =
+                            typename vector_type<ComputeTypeA,
+                                                 xdlops_gemm.K1PerXdlops / APackedSize>::type;
+                        using mfma_input_type_b =
+                            typename vector_type<ComputeTypeB,
+                                                 xdlops_gemm.K1PerXdlops / BPackedSize>::type;
+
+                        constexpr index_t c_offset =
+                            c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                        // MFMA accumulation
+                        xdlops_gemm.template Run<>(
+                            a_thread_vec.template AsType<mfma_input_type_a>(),
+                            a_scale_thread_vec.template AsType<AScaleDataType>(),
+                            b_thread_vec.template AsType<mfma_input_type_b>(),
+                            b_scale_thread_vec.template AsType<BScaleDataType>(),
+                            c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                    }); // KRepeat
+                });     // NRepeat
+            });         // MRepeat
+        }
+        else if constexpr(TailNum == TailNumber::Odd)
+        {
+            static_for<0, MRepeat, 1>{}([&](auto m0) {
+                static_for<0, NRepeat, 1>{}([&](auto n0) {
+                    static_for<0, KRepeat, 1>{}([&](auto k0) {
+                        vector_type<ComputeTypeA, KPack> a_thread_vec;
+                        vector_type<ComputeTypeB, KPack> b_thread_vec;
+
+                        static_for<0, KPack / ComputePackedSize, 1>{}([&](auto ik) {
+                            a_thread_vec.template AsType<ComputeTypeA>()(ik) =
+                                a_thread_buf[Number<a_thread_desc_.CalculateOffset(
+                                    make_tuple(m0, I0, k0, ik))>{}];
+                            b_thread_vec.template AsType<ComputeTypeB>()(ik) =
+                                b_thread_bufs[I0][Number<b_thread_desc_.CalculateOffset(
+                                    make_tuple(n0, I0, k0, ik))>{}];
+                        });
+
+                        constexpr index_t a_scale_offset =
+                            a_scale_thread_desc.CalculateOffset(make_tuple(m0, k0, I0));
+
+                        constexpr index_t b_scale_offset =
+                            b_scale_thread_desc.CalculateOffset(make_tuple(n0, k0, I0));
+
+                        vector_type<AScaleDataType, ScalesPerXdlopsRunPerThread> a_scale_thread_vec;
+                        vector_type<BScaleDataType, ScalesPerXdlopsRunPerThread> b_scale_thread_vec;
+
+                        // Pack b_scale_thread_buf into b_scale_thread_vec
+                        static_for<0, ScalesPerXdlopsRunPerThread, 1>{}([&](auto s) {
+                            a_scale_thread_vec.template AsType<AScaleDataType>()(s) =
+                                a_scale_thread_bufs[I0][Number<a_scale_offset + s>{}];
+                            b_scale_thread_vec.template AsType<BScaleDataType>()(s) =
+                                b_scale_thread_bufs[I0][Number<b_scale_offset + s>{}];
+                        });
+
+                        using mfma_input_type_a =
+                            typename vector_type<ComputeTypeA,
+                                                 xdlops_gemm.K1PerXdlops / APackedSize>::type;
+                        using mfma_input_type_b =
+                            typename vector_type<ComputeTypeB,
+                                                 xdlops_gemm.K1PerXdlops / BPackedSize>::type;
+
+                        constexpr index_t c_offset =
+                            c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+
+                        // MFMA accumulation
+                        xdlops_gemm.template Run<>(
+                            a_thread_vec.template AsType<mfma_input_type_a>(),
+                            a_scale_thread_vec.template AsType<AScaleDataType>(),
+                            b_thread_vec.template AsType<mfma_input_type_b>(),
+                            b_scale_thread_vec.template AsType<BScaleDataType>(),
+                            c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
+                    }); // KRepeat
+                });     // NRepeat
+            });         // MRepeat
+        }
+    }
+
+    // TODO: make this field protected when a_scale_thread_copy_ is moved
+    // here
+    static constexpr auto a_scale_thread_desc = make_naive_tensor_descriptor_packed(
+        make_tuple(Number<MRepeat>{}, Number<KRepeat>{}, Number<ScalesPerXdlopsRunPerThread>{}));
+
+    // Is used to copy data from a_scale_grid to a_scale_thread
+    static constexpr auto a_scale_thread_desc_copy =
+        make_naive_tensor_descriptor_packed(make_tuple(Number<1>{}, Number<1>{}));
+
+    // TODO: make this field protected when b_scale_thread_copy_ is moved
+    // here
+    static constexpr auto b_scale_thread_desc = make_naive_tensor_descriptor_packed(
+        make_tuple(Number<NRepeat>{}, Number<KRepeat>{}, Number<ScalesPerXdlopsRunPerThread>{}));
+
+    // Is used to copy data from b_scale_grid to b_scale_thread_buf
+    static constexpr auto b_scale_thread_desc_copy =
+        make_naive_tensor_descriptor_packed(make_tuple(Number<1>{}, Number<1>{}));
+
+    protected:
+    static constexpr auto b_thread_desc_ = make_naive_tensor_descriptor_packed(
+        make_tuple(Number<NRepeat>{}, I1, Number<KRepeat>{}, Number<KPack>{}));
+    using Base::a_thread_copy_;
+    using Base::a_thread_desc_;
+    using Base::b_thread_copy_;
+    // using Base::b_thread_desc_;
+    using Base::c_thread_desc_;
+
+    static constexpr BTileDesc b_block_desc_n0_n1_k0_k1;
+};
+
+} // namespace ck

include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_mx_moe_nbs_selector.hpp

@@ -0,0 +1,132 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v1_mx.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_mx_moe_nbs_v3.hpp"
+// #include
+// "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_mx_moe_nbs_gufusion_v1.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_mx_moe_nbs_gufusion_v3.hpp"
+
+namespace ck {
+template <BlockGemmPipelineVersion BlkGemmPipelineVer,
+          BlockGemmPipelineScheduler BlkGemmPipeSche,
+          index_t ThreadBlockSize,
+          index_t ScaleBlockSize,
+          typename ADataType,
+          typename AScaleDataType,
+          typename BDataType,
+          typename BScaleDataType,
+          typename ComputeDataType, // TODO: remove this as in this pipeline ADataType and BDataType
+                                    // must be used for compute
+          typename AccDataType,
+          typename ATileDesc,
+          typename BTileDesc,
+          typename AMmaTileDesc,
+          typename BMmaTileDesc,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MRepeat,
+          index_t NRepeat,
+          index_t KPack,
+          bool GUFusion = false>
+constexpr auto BlockGemmMXNBSPipeline_Selector()
+{
+
+    // Hardware MX GEMM pipeline
+    if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v1)
+    {
+        if constexpr(GUFusion)
+        {
+            return nullptr;
+        }
+        else
+        {
+            return BlockwiseGemmXdlops_pipeline_v1_mx<BlkGemmPipeSche,
+                                                      ThreadBlockSize,
+                                                      ScaleBlockSize,
+                                                      ADataType,
+                                                      AScaleDataType,
+                                                      BDataType,
+                                                      BScaleDataType,
+                                                      ATileDesc,
+                                                      BTileDesc,
+                                                      AMmaTileDesc,
+                                                      BMmaTileDesc,
+                                                      ABlockTransferSrcScalarPerVector,
+                                                      BBlockTransferSrcScalarPerVector,
+                                                      MPerBlock,
+                                                      NPerBlock,
+                                                      KPerBlock,
+                                                      MPerXDL,
+                                                      NPerXDL,
+                                                      MRepeat,
+                                                      NRepeat,
+                                                      KPack>{};
+        }
+    }
+    else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v3)
+    {
+        if constexpr(GUFusion)
+        {
+            return BlockwiseGemmXdlops_pipeline_mx_moe_bns_gufusion_v3<
+                BlkGemmPipeSche,
+                ThreadBlockSize,
+                ScaleBlockSize,
+                ADataType,
+                AScaleDataType,
+                BDataType,
+                BScaleDataType,
+                ATileDesc,
+                BTileDesc,
+                AMmaTileDesc,
+                BMmaTileDesc,
+                ABlockTransferSrcScalarPerVector,
+                BBlockTransferSrcScalarPerVector,
+                MPerBlock,
+                NPerBlock,
+                KPerBlock,
+                MPerXDL,
+                NPerXDL,
+                MRepeat,
+                NRepeat,
+                KPack>{};
+        }
+        else
+        {
+            return BlockwiseGemmXdlops_pipeline_mx_moe_nbs_v3<BlkGemmPipeSche,
+                                                              ThreadBlockSize,
+                                                              ScaleBlockSize,
+                                                              ADataType,
+                                                              AScaleDataType,
+                                                              BDataType,
+                                                              BScaleDataType,
+                                                              ATileDesc,
+                                                              BTileDesc,
+                                                              AMmaTileDesc,
+                                                              BMmaTileDesc,
+                                                              ABlockTransferSrcScalarPerVector,
+                                                              BBlockTransferSrcScalarPerVector,
+                                                              MPerBlock,
+                                                              NPerBlock,
+                                                              KPerBlock,
+                                                              MPerXDL,
+                                                              NPerXDL,
+                                                              MRepeat,
+                                                              NRepeat,
+                                                              KPack>{};
+        }
+    }
+    else
+    {
+        std::cerr << "MX GEMM Pipeline configuration is not available" << std::endl;
+    }
+}
+
+} // namespace ck

include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v1_ab_scale.hpp

@@ -205,7 +205,7 @@ struct BlockwiseGemmXdlops_pipeline_v1_ab_scale<BlockGemmPipelineScheduler::Intr
 
         constexpr auto num_mfma_inst = HotLoopInstList::C_MFMA_Inst_Num;
 
-        constexpr auto mfma_cycle = NPerXDL == 16 ? 16 : 32;
+        constexpr auto mfma_cycle = HotLoopInstList::C_MFMA_Inst_Cycle;
         constexpr auto ds_read_a_issue_cycle =
             HotLoopInstList::A_LDS_Read_Width * sizeof(ADataType) == 16 ? 8 : 4;
         constexpr auto ds_read_b_issue_cycle =

include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v3.hpp

@@ -177,8 +177,8 @@ struct BlockwiseGemmXdlops_pipeline_v3<BlockGemmPipelineScheduler::Intrawave,
         constexpr auto num_buffer_load_inst_b = HotLoopInstList::B_Buffer_Load_Inst_Num;
 
         constexpr auto num_mfma_inst = HotLoopInstList::C_MFMA_Inst_Num;
+        constexpr auto mfma_cycle    = HotLoopInstList::C_MFMA_Inst_Cycle;
 
-        constexpr auto mfma_cycle = NPerXDL == 16 ? 16 : 32;
         constexpr auto ds_read_a_issue_cycle =
             HotLoopInstList::A_LDS_Read_Width * sizeof(ADataType) == 16 ? 8 : 4;
         constexpr auto ds_read_b_issue_cycle =

include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v3_ab_scale.hpp

@@ -179,7 +179,7 @@ struct BlockwiseGemmXdlops_pipeline_v3_ab_scale<BlockGemmPipelineScheduler::Intr
 
         constexpr auto num_mfma_inst = HotLoopInstList::C_MFMA_Inst_Num;
 
-        constexpr auto mfma_cycle = NPerXDL == 16 ? 16 : 32;
+        constexpr auto mfma_cycle = HotLoopInstList::C_MFMA_Inst_Cycle;
         constexpr auto ds_read_a_issue_cycle =
             HotLoopInstList::A_LDS_Read_Width * sizeof(ADataType) == 16 ? 8 : 4;
         constexpr auto ds_read_b_issue_cycle =

include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v3_b_scale.hpp

@@ -178,7 +178,7 @@ struct BlockwiseGemmXdlops_pipeline_v3_b_scale<BlockGemmPipelineScheduler::Intra
 
         constexpr auto num_mfma_inst = HotLoopInstList::C_MFMA_Inst_Num;
 
-        constexpr auto mfma_cycle = NPerXDL == 16 ? 16 : 32;
+        constexpr auto mfma_cycle = HotLoopInstList::C_MFMA_Inst_Cycle;
         constexpr auto ds_read_a_issue_cycle =
             HotLoopInstList::A_LDS_Read_Width * sizeof(ADataType) == 16 ? 8 : 4;
         constexpr auto ds_read_b_issue_cycle =

include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v5.hpp

@@ -188,7 +188,7 @@ struct BlockwiseGemmXdlops_pipeline_v5<BlockGemmPipelineScheduler::Intrawave,
 
         constexpr auto num_mfma_inst = HotLoopInstList::C_MFMA_Inst_Num;
 
-        constexpr auto mfma_cycle = NPerXDL == 16 ? 16 : 32;
+        constexpr auto mfma_cycle = HotLoopInstList::C_MFMA_Inst_Cycle;
         constexpr auto ds_read_a_issue_cycle =
             HotLoopInstList::A_LDS_Read_Width * sizeof(ADataType) == 16 ? 8 : 4;
         constexpr auto ds_read_b_issue_cycle =

include/ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp

@@ -149,6 +149,52 @@ struct DeviceGemmMultipleDSplitKBPreShuffle : public BaseOperator
 #endif
 };
 
+template <typename ALayout,
+          typename BLayout,
+          typename DsLayout,
+          typename ELayout,
+          typename ADataType,
+          typename AScaleDataType,
+          typename BDataType,
+          typename BScaleDataType,
+          typename DsDataType,
+          typename EDataType,
+          index_t ScaleBlockSize,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation>
+struct DeviceMoEGemmMXBPreShuffle : public BaseOperator
+{
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+#ifndef CK_CODE_GEN_RTC
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_a,
+                        const void* p_a_scale,
+                        const void* p_b,
+                        const void* p_b_scale,
+                        std::array<const void*, NumDTensor> p_ds,
+                        void* p_e,
+                        ck::index_t M,
+                        ck::index_t N,
+                        ck::index_t K,
+                        ck::index_t StrideA,
+                        ck::index_t StrideAScale,
+                        ck::index_t StrideB,
+                        ck::index_t StrideBScale,
+                        std::array<ck::index_t, NumDTensor> StrideDs,
+                        ck::index_t StrideE,
+                        ck::index_t KBatch,
+                        AElementwiseOperation a_element_op,
+                        BElementwiseOperation b_element_op,
+                        CDEElementwiseOperation cde_element_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+
+    virtual int GetPreShuffleParameters() = 0;
+#endif
+};
+
 } // namespace device
 } // namespace tensor_operation
 } // namespace ck

include/ck/tensor_operation/gpu/device/device_gemm_mx.hpp

@@ -45,6 +45,44 @@ struct DeviceGemmMX : public BaseOperator
     virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
 };
 
+template <typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          typename ADataType,
+          typename AScaleDataType,
+          typename BDataType,
+          typename BScaleDataType,
+          typename CDataType,
+          index_t ScaleBlockSize,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation>
+struct DeviceGemmMX_BPreshuffle : public BaseOperator
+{
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_a,
+                        const void* p_a_scale,
+                        const void* p_b,
+                        const void* p_b_scale,
+                        void* p_c,
+                        ck::index_t M,
+                        ck::index_t N,
+                        ck::index_t K,
+                        ck::index_t StrideA,
+                        ck::index_t StrideAScale,
+                        ck::index_t StrideB,
+                        ck::index_t StrideBScale,
+                        ck::index_t StrideC,
+                        ck::index_t KBatch,
+                        AElementwiseOperation a_element_op,
+                        BElementwiseOperation b_element_op,
+                        CElementwiseOperation c_element_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+
+    virtual int GetPreShuffleParameters() = 0;
+};
+
 } // namespace device
 } // namespace tensor_operation
 } // namespace ck

include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_v3_mx_b_preshuffle.hpp

@@ -0,0 +1,605 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+
+#include "ck/host_utility/flush_cache.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_mx.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_mx_b_preshuffle.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// clang-format off
+/**
+ * \brief WIP: Implements XDL CShuffle V3 GEMM for microscale-compliant data types
+ *
+ * This class is a work-in-progress implementation of the XDL CShuffle V3 GEMM for
+ * microscale-compliant data types.
+ *
+ * Assumptions:
+ * - A and B data types are compliant with the OCP Microscaling Formats (MX) Specification
+ * - Each scale applies to ScaleBlockSize elements in K direction
+ * - A scale matrix is a row-major
+ * - B scale matrix is a column-major
+ * - Scale data types must have get_exponent_value() specialization, whereas lowest 8 bits of the
+ * exponent will be interpreted as conventional biased Float32 exponent (E8M0)
+ *
+ * Tunable parameters.
+ * The CK instance includes a series of tunable template parameters to control the parallel
+ * granularity of the workload to achieve load balancing on different hardware platforms. These
+ * parameters include Block Size, M/N/K Per Block, M/N per XDL, AK1, BK1, etc.
+ *  - Block Size determines the number of threads in the thread block.
+ *  - M/N/K Per Block determines the size of tile that each thread block is responsible for
+ * calculating.
+ *  - M/N Per XDL refers to M/N size for Instinct accelerator Matrix Fused Multiply Add (MFMA)
+ * instructions operating on a per-wavefront basis.
+ *  - A/B K1 is related to the data type. It can be any value ranging from 1 to K Per Block. To
+ * achieve the optimal load/store performance, 128bit per load is suggested. In addition, the A/B
+ * loading parameters must be changed accordingly to match the A/B K1 value; otherwise, it will
+ * result in compilation errors.
+ *
+ * Conditions for achieving computational load balancing on different hardware platforms can vary.
+ *
+ * Serialized version of the algorithm:
+ * \code
+ * // E = A * B + C
+ * // Loop over E[MPerBlock,NPerBlock] tiles
+ * for(int mb = 0; mb < M; mb += MPerBlock){
+ *    for(int nb = 0; nb < N; nb += NPerBlock){
+ *       // initialize E[MPerBlock,NPerBlock] tile
+ *       for(int mt = mb; mt < mb + MPerBlock; mt++){
+ *          for(int nt = nb; nt < nb + NPerBlock; nt++){
+ *             E[mt,nt] = C[mt,nt];
+ *          }
+ *       }
+ *
+ *       // multiply-accumulate per tile
+ *       for(int kb = 0; kb < K; kb += KPerBlock){
+ *         for(int m0 = mb; m0 < mb + MPerBlock; m0 += MWaves * MPerXDL){
+ *           for(int n0 = nb; n0 < nb + NPerBlock; n0 += NWaves * NPerXDL){
+ *             for(int mw = m0; mw < m0 + MWaves * MPerXDL; mw += MPerXDL){
+ *               for(int nw = n0; nw < n0 + NWaves * NPerXDL; nw += NPerXDL){
+ *                 for(int k0 = kb; k0 < kb + KPerBlock; k0 += mfma.num_input_blks*KPack){
+ *                   // MFMA accumulation
+ *                   for(int k_pack = k0; k_pack < k0 + mfma.num_input_blks*KPack; k_pack += KPerXdlops){
+ *                     // MFMA instruction
+ *                     for(int k_mfma = k_pack; k_mfma < k_pack + KPerXdlops; k_mfma += mfma.k_per_blk){
+ *                       for(int m = mw; m < mw + MPerXDL; m++){
+ *                         for(int n = nw; n < nw + NPerXDL; n++){
+ *                           for(int k = k_mfma; k < k_mfma + mfma.k_per_blk; k++){
+ *                            E[m,n] += A[m,k] * B[k,n];
+ *                           }
+ *                         }
+ *                       }
+ *                     }
+ *                   }
+ *                 }
+ *               }
+ *             }
+ *           }
+ *         }
+ *       }
+ *    }
+ * }
+ * \endcode
+ *
+ */
+// clang-format on
+template <typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          typename ADataType,
+          typename AScaleDataType,
+          typename BDataType,
+          typename BScaleDataType,
+          typename CDataType,
+          typename GemmAccDataType, // TODO: always float
+          typename CShuffleDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          GemmSpecialization GemmSpec,
+          index_t ScaleBlockSize, // Scaling block size
+          index_t BlockSize,      // Thread block size
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t AK1,
+          index_t BK1,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
+          BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave,
+          BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1,
+          typename ComputeTypeA =
+              ADataType, // XXX: These should always be the same as ADataType and BDataType
+          typename ComputeTypeB =
+              BDataType // TODO: Hardcode them and remove from the list of template parameters
+          >
+struct DeviceGemmMX_Xdl_CShuffleV3_BPreShuffle
+    : public DeviceGemmMX_BPreshuffle<ALayout,
+                                      BLayout,
+                                      CLayout,
+                                      ADataType,
+                                      AScaleDataType,
+                                      BDataType,
+                                      BScaleDataType,
+                                      CDataType,
+                                      ScaleBlockSize,
+                                      AElementwiseOperation,
+                                      BElementwiseOperation,
+                                      CElementwiseOperation>
+{
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseGemmMX_xdl_cshuffle_v3_b_preshuffle<
+        ALayout,
+        BLayout,
+        CLayout,
+        ADataType,
+        AScaleDataType,
+        BDataType,
+        BScaleDataType,
+        GemmAccDataType,
+        CShuffleDataType,
+        CDataType,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CElementwiseOperation,
+        GemmSpec,
+        ScaleBlockSize,
+        BlockSize,
+        MPerBlock,
+        NPerBlock,
+        KPerBlock,
+        AK1,
+        BK1,
+        MPerXDL,
+        NPerXDL,
+        MXdlPerWave,
+        NXdlPerWave,
+        ABlockTransferThreadClusterLengths_AK0_M_AK1,
+        ABlockTransferThreadClusterArrangeOrder,
+        ABlockTransferSrcAccessOrder,
+        ABlockTransferSrcVectorDim,
+        ABlockTransferSrcScalarPerVector,
+        ABlockTransferDstScalarPerVector_AK1,
+        false,
+        ABlockLdsExtraM,
+        BBlockTransferThreadClusterLengths_BK0_N_BK1,
+        BBlockTransferThreadClusterArrangeOrder,
+        BBlockTransferSrcAccessOrder,
+        BBlockTransferSrcVectorDim,
+        BBlockTransferSrcScalarPerVector,
+        BBlockTransferDstScalarPerVector_BK1,
+        false,
+        BBlockLdsExtraN,
+        CShuffleMXdlPerWavePerShuffle,
+        CShuffleNXdlPerWavePerShuffle,
+        CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+        CShuffleBlockTransferScalarPerVector_NPerBlock,
+        BlkGemmPipeSched,
+        BlkGemmPipelineVer,
+        ComputeTypeA,
+        ComputeTypeB>;
+
+    using Argument = typename GridwiseGemm::Argument;
+
+    int GetPreShuffleParameters() override { return NPerXDL; }
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            if(stream_config.log_level_ > 0)
+            {
+                arg.Print();
+                GridwiseGemm::BlockwiseGemmPipe::HotLoopInstList::Print();
+            }
+
+            if(!GridwiseGemm::CheckValidity(arg))
+            {
+                throw std::runtime_error("wrong! GridwiseGemm has invalid setting");
+            }
+
+            index_t gdx, gdy, gdz;
+            std::tie(gdx, gdy, gdz) = GridwiseGemm::CalculateGridSize(arg.M, arg.N, arg.KBatch);
+
+            float ave_time = 0;
+
+            index_t k_grain = arg.KBatch * KPerBlock;
+            index_t K_split = (arg.K + k_grain - 1) / k_grain * KPerBlock;
+
+            const bool has_main_k_block_loop = GridwiseGemm::CalculateHasMainKBlockLoop(K_split);
+
+            const auto Run = [&](const auto& kernel) {
+                if(stream_config.flush_cache)
+                {
+                    Argument arg_ = arg;
+
+                    const auto a_grid_desc_ak0_m_ak1 = GridwiseGemm::MakeAGridDescriptor_AK0_M_AK1(
+                        arg_.M, arg_.MPadded, arg_.K, arg_.KPadded, arg_.StrideA, arg_.AK0);
+                    const auto b_grid_desc_bk0_n_bk1 = GridwiseGemm::MakeBGridDescriptor_BK0_N_BK1(
+                        arg_.K, arg_.KPadded, arg_.N, arg_.NPadded, arg_.StrideB, arg_.BK0);
+
+                    auto size_a_buffer =
+                        a_grid_desc_ak0_m_ak1.GetElementSpaceSize() * sizeof(ADataType);
+                    auto size_b_buffer =
+                        b_grid_desc_bk0_n_bk1.GetElementSpaceSize() * sizeof(BDataType);
+
+                    ck::utility::RotatingMemWrapper<Argument> rotating_mem(
+                        arg_, stream_config.rotating_count, size_a_buffer, size_b_buffer);
+                    rotating_mem.Print();
+
+                    auto run_flush_cache = [&]() {
+                        // flush icache
+                        ck::utility::flush_icache();
+                        // rotating mem
+                        rotating_mem.Next();
+                        // clear c mem
+                        if(arg_.KBatch > 1)
+                            hipGetErrorString(hipMemsetAsync(arg_.p_c_grid,
+                                                             0,
+                                                             arg_.M * arg_.N * sizeof(CDataType),
+                                                             stream_config.stream_id_));
+                    };
+
+                    ave_time = ck::utility::launch_and_time_kernel_with_preprocess<false>(
+                        stream_config,
+                        run_flush_cache,
+                        kernel,
+                        dim3(gdx, gdy, gdz),
+                        dim3(BlockSize),
+                        0,
+                        arg_);
+                }
+                else
+                {
+                    if(arg.KBatch > 1)
+                        hipGetErrorString(hipMemsetAsync(arg.p_c_grid,
+                                                         0,
+                                                         arg.M * arg.N * sizeof(CDataType),
+                                                         stream_config.stream_id_));
+
+                    ave_time = launch_and_time_kernel(
+                        stream_config, kernel, dim3(gdx, gdy, gdz), dim3(BlockSize), 0, arg);
+                }
+            };
+
+            // TODO: Check if this is the right algorithm for minimum_occupancy
+            constexpr index_t minimum_occupancy =
+                BlkGemmPipeSched == BlockGemmPipelineScheduler::Intrawave
+                    ? (BlkGemmPipelineVer == BlockGemmPipelineVersion::v3 &&
+                       MPerBlock * NPerBlock * KPerBlock * sizeof(ADataType) <= 128 * 128 * 64 * 2)
+                          ? 2
+                          : 1
+                    : 2;
+
+            if(has_main_k_block_loop)
+            {
+                // Tail number always full
+                if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v1)
+                {
+                    if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd)
+                    {
+                        const auto kernel =
+                            kernel_gemm_xdl_cshuffle_v3_b_preshuffle<GridwiseGemm,
+                                                                     true,
+                                                                     InMemoryDataOperationEnum::Set,
+                                                                     minimum_occupancy,
+                                                                     TailNumber::Odd>;
+                        Run(kernel);
+                    }
+                    else
+                    {
+                        const auto kernel =
+                            kernel_gemm_xdl_cshuffle_v3_b_preshuffle<GridwiseGemm,
+                                                                     true,
+                                                                     InMemoryDataOperationEnum::Set,
+                                                                     minimum_occupancy,
+                                                                     TailNumber::Even>;
+                        Run(kernel);
+                    }
+                }
+                else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v3)
+                {
+                    if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd)
+                    {
+                        const auto kernel = kernel_gemm_xdl_cshuffle_v3_b_preshuffle_2lds<
+                            GridwiseGemm,
+                            true,
+                            InMemoryDataOperationEnum::Set,
+                            minimum_occupancy,
+                            TailNumber::Odd>;
+                        Run(kernel);
+                    }
+                    else
+                    {
+                        const auto kernel = kernel_gemm_xdl_cshuffle_v3_b_preshuffle_2lds<
+                            GridwiseGemm,
+                            true,
+                            InMemoryDataOperationEnum::Set,
+                            minimum_occupancy,
+                            TailNumber::Even>;
+                        Run(kernel);
+                    }
+                }
+            }
+            else
+            {
+                // Tail number always 1
+                if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v1)
+                {
+                    if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd)
+                    {
+                        const auto kernel =
+                            kernel_gemm_xdl_cshuffle_v3_b_preshuffle<GridwiseGemm,
+                                                                     false,
+                                                                     InMemoryDataOperationEnum::Set,
+                                                                     minimum_occupancy,
+                                                                     TailNumber::Odd>;
+                        Run(kernel);
+                    }
+                    else
+                    {
+                        const auto kernel =
+                            kernel_gemm_xdl_cshuffle_v3_b_preshuffle<GridwiseGemm,
+                                                                     false,
+                                                                     InMemoryDataOperationEnum::Set,
+                                                                     minimum_occupancy,
+                                                                     TailNumber::Even>;
+                        Run(kernel);
+                    }
+                }
+                else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v3)
+                {
+                    if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd)
+                    {
+                        const auto kernel = kernel_gemm_xdl_cshuffle_v3_b_preshuffle_2lds<
+                            GridwiseGemm,
+                            false,
+                            InMemoryDataOperationEnum::Set,
+                            minimum_occupancy,
+                            TailNumber::Odd>;
+                        Run(kernel);
+                    }
+                    else
+                    {
+                        const auto kernel = kernel_gemm_xdl_cshuffle_v3_b_preshuffle_2lds<
+                            GridwiseGemm,
+                            false,
+                            InMemoryDataOperationEnum::Set,
+                            minimum_occupancy,
+                            TailNumber::Even>;
+                        Run(kernel);
+                    }
+                }
+            }
+
+            return ave_time;
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        static_assert(is_scale_mfma_data_type<ADataType>() && is_scale_mfma_data_type<BDataType>(),
+                      "Only microscaling formats are supported for ADataType and BDataType");
+
+        static_assert(ScaleBlockSize == 32, "Only ScaleBlockSize 32 is supported");
+
+        static_assert(is_same_v<ComputeTypeA, ADataType> && is_same_v<ComputeTypeB, BDataType>,
+                      "ComputeTypeA and ComputeTypeB must be the same as ADataType and BDataType");
+
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if constexpr(!IsValidCompilationParameter())
+        {
+            return false;
+        }
+
+        if(!ck::is_xdl_supported())
+        {
+            return false;
+        }
+
+        if(!is_bf16_atomic_supported() && std::is_same_v<CDataType, ck::bhalf_t> && arg.KBatch > 1)
+        {
+            return false;
+        }
+
+        if((arg.K % AK1 != 0 || arg.K % BK1 != 0) && !(GemmSpec == GemmSpecialization::MKPadding ||
+                                                       GemmSpec == GemmSpecialization::NKPadding ||
+                                                       GemmSpec == GemmSpecialization::MNKPadding ||
+                                                       GemmSpec == GemmSpecialization::KPadding))
+        {
+            return false;
+        }
+
+        return GridwiseGemm::CheckValidity(arg);
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(const ADataType* p_a,
+                             const AScaleDataType* p_a_scale,
+                             const BDataType* p_b,
+                             const BScaleDataType* p_b_scale,
+                             CDataType* p_c,
+                             index_t M,
+                             index_t N,
+                             index_t K,
+                             index_t StrideA,
+                             index_t StrideScaleA,
+                             index_t StrideB,
+                             index_t StrideScaleB,
+                             index_t StrideC,
+                             index_t KBatch,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             CElementwiseOperation c_element_op)
+    {
+        return Argument{p_a,
+                        p_a_scale,
+                        p_b,
+                        p_b_scale,
+                        p_c,
+                        M,
+                        N,
+                        K,
+                        StrideA,
+                        StrideScaleA,
+                        StrideB,
+                        StrideScaleB,
+                        StrideC,
+                        KBatch,
+                        a_element_op,
+                        b_element_op,
+                        c_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument> MakeArgumentPointer(const void* p_a,
+                                                      const void* p_a_scale,
+                                                      const void* p_b,
+                                                      const void* p_b_scale,
+                                                      void* p_c,
+                                                      ck::index_t M,
+                                                      ck::index_t N,
+                                                      ck::index_t K,
+                                                      ck::index_t StrideA,
+                                                      ck::index_t StrideScaleA,
+                                                      ck::index_t StrideB,
+                                                      ck::index_t StrideScaleB,
+                                                      ck::index_t StrideC,
+                                                      ck::index_t KBatch,
+                                                      AElementwiseOperation a_element_op,
+                                                      BElementwiseOperation b_element_op,
+                                                      CElementwiseOperation c_element_op) override
+    {
+        return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
+                                          static_cast<const AScaleDataType*>(p_a_scale),
+                                          static_cast<const BDataType*>(p_b),
+                                          static_cast<const BScaleDataType*>(p_b_scale),
+                                          static_cast<CDataType*>(p_c),
+                                          M,
+                                          N,
+                                          K,
+                                          StrideA,
+                                          StrideScaleA,
+                                          StrideB,
+                                          StrideScaleB,
+                                          StrideC,
+                                          KBatch,
+                                          a_element_op,
+                                          b_element_op,
+                                          c_element_op);
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    // polymorphic
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        std::map<BlockGemmPipelineScheduler, std::string> BlkGemmPipelineSchedulerToString{
+            {BlockGemmPipelineScheduler::Intrawave, "Intrawave"},
+            {BlockGemmPipelineScheduler::Interwave, "Interwave"}};
+
+        std::map<BlockGemmPipelineVersion, std::string> BlkGemmPipelineVersionToString{
+            {BlockGemmPipelineVersion::v1, "v1"},
+            {BlockGemmPipelineVersion::v2, "v2"},
+            {BlockGemmPipelineVersion::v3, "v3"},
+            {BlockGemmPipelineVersion::v4, "v4"},
+            {BlockGemmPipelineVersion::v5, "v5"}};
+
+        // clang-format off
+        str << "DeviceGemmMX_Xdl_CShuffleV3"
+            << "<"
+            << getGemmSpecializationString(GemmSpec) << ", "
+            << std::string(ALayout::name)[0]
+            << std::string(BLayout::name)[0]
+            << std::string(CLayout::name)[0]
+            << ">"
+            << " BlkSize: "
+            << BlockSize << ", "
+            << "BlkTile: "
+            << MPerBlock<<"x"<<NPerBlock<<"x"<<KPerBlock << ", "
+            << "WaveTile: "
+            << MPerXDL<<"x"<<NPerXDL << ", "
+            << "WaveMap: "
+            << MXdlPerWave<<"x" << NXdlPerWave<<", "
+            << "VmemReadVec: "
+            << ABlockTransferSrcScalarPerVector<<"x"<<BBlockTransferSrcScalarPerVector<<", "
+            << "BlkGemmPipelineScheduler: "
+            << BlkGemmPipelineSchedulerToString[BlkGemmPipeSched] << ", "
+            << "BlkGemmPipelineVersion: "
+            << BlkGemmPipelineVersionToString[BlkGemmPipelineVer] << ", "
+            << "BlkGemmPipelinePrefetchStages: "
+            << GridwiseGemm::BlockwiseGemmPipe::PrefetchStages << ", "
+            << "Kpack: "
+            << GridwiseGemm::BlockwiseGemmPipe::AMmaKStride << ", "
+            << "ScaleBlockSize: "
+            << ScaleBlockSize;
+        // clang-format on
+
+        return str.str();
+    }
+    REGISTER_EXTRA_PRINTING_METHODS
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

include/ck/tensor_operation/gpu/device/impl/device_moe_mx_gemm.hpp

@@ -0,0 +1,584 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_moe_mx_gemm.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+#include "ck/host_utility/flush_cache.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename ALayout,
+          typename BLayout,
+          typename DsLayout,
+          typename CLayout,
+          typename ADataType,
+          typename AScaleDataType,
+          typename BDataType,
+          typename BScaleDataType,
+          typename DsDataType,
+          typename CDataType,
+          typename GemmAccDataType,
+          typename CShuffleDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          GemmSpecialization GemmSpec,
+          index_t ScaleBlockSize,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t AK1,
+          index_t BK1,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename CDEShuffleBlockTransferScalarPerVectors,
+          BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave,
+          BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1,
+          index_t ActivationOP                        = 0,
+          bool NSwizzle                               = false,
+          bool IsInputGemm                            = true,
+          bool MulRoutedWeight                        = true,
+          typename IndexType                          = index_t,
+          typename ComputeTypeA                       = ADataType,
+          typename ComputeTypeB                       = BDataType>
+struct DeviceMoeGemmMX : public DeviceMoEGemmMXBPreShuffle<ALayout,
+                                                           BLayout,
+                                                           DsLayout,
+                                                           CLayout,
+                                                           ADataType,
+                                                           AScaleDataType,
+                                                           BDataType,
+                                                           BScaleDataType,
+                                                           DsDataType,
+                                                           CDataType,
+                                                           ScaleBlockSize,
+                                                           AElementwiseOperation,
+                                                           BElementwiseOperation,
+                                                           CElementwiseOperation>
+{
+    static constexpr index_t NumDTensor = DsDataType::Size();
+    using GridwiseGemm =
+        GridwiseMoeGemmMX<ALayout,
+                          BLayout,
+                          DsLayout,
+                          CLayout,
+                          ADataType,
+                          AScaleDataType,
+                          BDataType,
+                          BScaleDataType,
+                          GemmAccDataType,
+                          CShuffleDataType,
+                          DsDataType,
+                          CDataType,
+                          AElementwiseOperation,
+                          BElementwiseOperation,
+                          CElementwiseOperation,
+                          GemmSpec,
+                          ScaleBlockSize,
+                          BlockSize,
+                          MPerBlock,
+                          NPerBlock,
+                          KPerBlock,
+                          AK1,
+                          BK1,
+                          MPerXDL,
+                          NPerXDL,
+                          MXdlPerWave,
+                          NXdlPerWave,
+                          ABlockTransferThreadClusterLengths_AK0_M_AK1,
+                          ABlockTransferThreadClusterArrangeOrder,
+                          ABlockTransferSrcAccessOrder,
+                          ABlockTransferSrcVectorDim,
+                          ABlockTransferSrcScalarPerVector,
+                          ABlockTransferDstScalarPerVector_AK1,
+                          false,
+                          ABlockLdsExtraM,
+                          BBlockTransferThreadClusterLengths_BK0_N_BK1,
+                          BBlockTransferThreadClusterArrangeOrder,
+                          BBlockTransferSrcAccessOrder,
+                          BBlockTransferSrcVectorDim,
+                          BBlockTransferSrcScalarPerVector,
+                          BBlockTransferDstScalarPerVector_BK1,
+                          false,
+                          BBlockLdsExtraN,
+                          CShuffleMXdlPerWavePerShuffle,
+                          CShuffleNXdlPerWavePerShuffle,
+                          CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+                          CDEShuffleBlockTransferScalarPerVectors,
+                          BlkGemmPipeSched,
+                          BlkGemmPipelineVer,
+                          ActivationOP,
+                          NSwizzle,
+                          IsInputGemm,
+                          MulRoutedWeight,
+                          IndexType,
+                          ComputeTypeA,
+                          ComputeTypeB>;
+
+    using Argument = typename GridwiseGemm::Argument;
+
+    static constexpr index_t APackedSize = []() {
+        if constexpr(is_same_v<remove_cvref_t<ADataType>, pk_i4_t> ||
+                     is_same_v<remove_cvref_t<ADataType>, f4x2_pk_t>)
+            return 2;
+        else
+            return 1;
+    }();
+
+    static constexpr index_t BPackedSize = []() {
+        if constexpr(is_same_v<remove_cvref_t<BDataType>, pk_i4_t> ||
+                     is_same_v<remove_cvref_t<BDataType>, f4x2_pk_t>)
+            return 2;
+        else
+            return 1;
+    }();
+
+    int GetPreShuffleParameters() override { return NPerXDL; }
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            if(stream_config.log_level_ > 0)
+            {
+                arg.Print();
+            }
+
+            if(!GridwiseGemm::CheckValidity(arg))
+            {
+                throw std::runtime_error("wrong! GridwiseGemm has invalid setting");
+            }
+
+            index_t gdx, gdy, gdz;
+            std::tie(gdx, gdy, gdz) = GridwiseGemm::CalculateGridSize(arg.M, arg.N);
+
+            float ave_time = 0;
+
+            index_t k_grain = arg.KBatch * KPerBlock;
+            index_t K_split = (arg.K + k_grain - 1) / k_grain * KPerBlock;
+
+            const bool has_main_k_block_loop = GridwiseGemm::CalculateHasMainKBlockLoop(K_split);
+
+            const auto RunKernel = [&](const auto& kernel) {
+                if(stream_config.flush_cache)
+                {
+
+                    std::array<std::size_t, NumDTensor> DsSize;
+
+                    Argument arg_ = arg;
+
+                    const auto a_grid_desc_ak0_m_ak1 = GridwiseGemm::MakeAGridDescriptor_AK0_M_AK1(
+                        arg_.M, arg_.MPadded, arg_.K, arg_.KPadded, arg_.StrideA, arg_.AK0);
+                    const auto b_grid_desc_bk0_n_bk1 = GridwiseGemm::MakeBGridDescriptor_BK0_N_BK1(
+                        arg_.K, arg_.KPadded, arg_.N, arg_.NPadded, arg_.StrideB, arg_.BK0);
+
+                    auto size_a_buffer = a_grid_desc_ak0_m_ak1.GetElementSpaceSize() *
+                                         sizeof(ADataType) / APackedSize;
+                    auto size_b_buffer = b_grid_desc_bk0_n_bk1.GetElementSpaceSize() *
+                                         sizeof(BDataType) / BPackedSize;
+
+                    const auto ds_grid_desc_m_n = GridwiseGemm::MakeDsGridDescriptor_M_N(
+                        arg_.M, arg_.MPadded, arg_.N, arg_.NPadded, arg_.StrideDs);
+
+                    static_for<0, NumDTensor, 1>{}([&](auto i) {
+                        using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
+                        DsSize[i] = ds_grid_desc_m_n[i].GetElementSpaceSize() * sizeof(DDataType);
+                    });
+                    ck::utility::RotatingMemWrapperMultiD<Argument, DsDataType> rotating_mem(
+                        arg_, stream_config.rotating_count, size_a_buffer, size_b_buffer, DsSize);
+                    rotating_mem.Print();
+
+                    auto run_flush_cache = [&]() {
+                        // flush icache
+                        ck::utility::flush_icache();
+                        // rotating mem
+                        rotating_mem.Next();
+                        // clear c mem
+                        if(arg_.KBatch > 1)
+                            hipGetErrorString(hipMemsetAsync(arg_.p_c_grid,
+                                                             0,
+                                                             arg_.M * arg_.N * sizeof(CDataType),
+                                                             stream_config.stream_id_));
+                    };
+
+                    ave_time = ck::utility::launch_and_time_kernel_with_preprocess<false>(
+                        stream_config,
+                        run_flush_cache,
+                        kernel,
+                        dim3(gdx, gdy, gdz),
+                        dim3(BlockSize),
+                        0,
+                        arg_);
+                }
+                else
+                {
+                    if(arg.KBatch > 1)
+                        hipGetErrorString(hipMemsetAsync(arg.p_c_grid,
+                                                         0,
+                                                         arg.M * arg.N * sizeof(CDataType),
+                                                         stream_config.stream_id_));
+
+                    ave_time = launch_and_time_kernel(
+                        stream_config, kernel, dim3(gdx, gdy, gdz), dim3(BlockSize), 0, arg);
+                }
+            };
+
+            constexpr auto estimated_reg_a = MPerBlock * KPerBlock * sizeof(ADataType) /
+                                             APackedSize / BlockSize / 4 *
+                                             (1 + GridwiseGemm::NWave);
+            constexpr auto estimated_reg_b = NPerBlock * KPerBlock * sizeof(BDataType) /
+                                             BPackedSize / BlockSize / 4 * (2) *
+                                             (IsInputGemm ? 2 : 1);
+            constexpr auto estimated_reg_c = MPerBlock * NPerBlock * sizeof(GemmAccDataType) /
+                                             BlockSize / 4 * (IsInputGemm ? 2 : 1);
+            constexpr auto estimated_reg_total =
+                estimated_reg_a + estimated_reg_b + estimated_reg_c;
+
+            constexpr index_t minimum_occupancy = (estimated_reg_total >= 256) ? 1 : 2;
+
+            constexpr auto MemoryDataOp =
+                IsInputGemm ? InMemoryDataOperationEnum::Set : InMemoryDataOperationEnum::AtomicAdd;
+            if(has_main_k_block_loop)
+            {
+                // Tail number always full
+                if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v1)
+                {
+                    {
+                        if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd)
+                        {
+                            const auto kernel = kernel_moe_mxgemm<GridwiseGemm,
+                                                                  true,
+                                                                  MemoryDataOp,
+                                                                  minimum_occupancy,
+                                                                  TailNumber::Odd>;
+                            RunKernel(kernel);
+                        }
+                        else
+                        {
+                            const auto kernel = kernel_moe_mxgemm<GridwiseGemm,
+                                                                  true,
+                                                                  MemoryDataOp,
+                                                                  minimum_occupancy,
+                                                                  TailNumber::Even>;
+                            RunKernel(kernel);
+                        }
+                    }
+                }
+                else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v2 ||
+                                  BlkGemmPipelineVer == BlockGemmPipelineVersion::v3)
+                {
+                    if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd)
+                    {
+                        const auto kernel = kernel_moe_mxgemm_2lds<GridwiseGemm,
+                                                                   true,
+                                                                   MemoryDataOp,
+                                                                   minimum_occupancy,
+                                                                   TailNumber::Odd>;
+                        RunKernel(kernel);
+                    }
+                    else
+                    {
+                        const auto kernel = kernel_moe_mxgemm_2lds<GridwiseGemm,
+                                                                   true,
+                                                                   MemoryDataOp,
+                                                                   minimum_occupancy,
+                                                                   TailNumber::Even>;
+                        RunKernel(kernel);
+                    }
+                }
+                else
+                {
+                    throw std::runtime_error("todo: only v1 & v3 support now");
+                }
+            }
+            else
+            {
+                if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v1)
+                {
+                    if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd)
+                    {
+                        const auto kernel = kernel_moe_mxgemm<GridwiseGemm,
+                                                              false,
+                                                              MemoryDataOp,
+                                                              minimum_occupancy,
+                                                              TailNumber::Odd>;
+                        RunKernel(kernel);
+                    }
+                    else
+                    {
+                        const auto kernel = kernel_moe_mxgemm<GridwiseGemm,
+                                                              false,
+                                                              MemoryDataOp,
+                                                              minimum_occupancy,
+                                                              TailNumber::Even>;
+                        RunKernel(kernel);
+                    }
+                }
+                else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v3)
+                {
+                    if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd)
+                    {
+                        const auto kernel = kernel_moe_mxgemm_2lds<GridwiseGemm,
+                                                                   false,
+                                                                   MemoryDataOp,
+                                                                   minimum_occupancy,
+                                                                   TailNumber::Odd>;
+                        RunKernel(kernel);
+                    }
+                    else
+                    {
+                        const auto kernel = kernel_moe_mxgemm_2lds<GridwiseGemm,
+                                                                   false,
+                                                                   MemoryDataOp,
+                                                                   minimum_occupancy,
+                                                                   TailNumber::Even>;
+                        RunKernel(kernel);
+                    }
+                }
+            }
+
+            return ave_time;
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        // only impl kbatch 1 now
+        if(arg.KBatch > 1)
+        {
+            return false;
+        }
+        if(!ck::is_xdl_supported())
+        {
+            return false;
+        }
+
+        if(!is_bf16_atomic_supported() && std::is_same_v<CDataType, ck::bhalf_t> && arg.KBatch > 1)
+        {
+            return false;
+        }
+
+        if((arg.K % AK1 != 0 || arg.K % BK1 != 0) && !(GemmSpec == GemmSpecialization::MKPadding ||
+                                                       GemmSpec == GemmSpecialization::NKPadding ||
+                                                       GemmSpec == GemmSpecialization::MNKPadding ||
+                                                       GemmSpec == GemmSpecialization::KPadding))
+        {
+            return false;
+        }
+        if(arg.N % NPerBlock != 0 || arg.K % KPerBlock != 0)
+        {
+            return false;
+        }
+
+        return GridwiseGemm::CheckValidity(arg);
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(const void* p_sorted_token_ids,
+                             const void* p_sorted_expert_ids,
+                             const void* p_max_token_id,
+                             const void* p_a,
+                             const void* p_a_scale,
+                             const void* p_b,
+                             const void* p_b_scale,
+                             std::array<const void*, NumDTensor> p_ds,
+                             void* p_c,
+                             index_t NumTokens,
+                             index_t TopK,
+                             index_t M,
+                             index_t N,
+                             index_t K,
+                             index_t StrideA,
+                             index_t StrideScaleA,
+                             index_t StrideB,
+                             index_t StrideScaleB,
+                             std::array<index_t, NumDTensor> StrideDs,
+                             index_t StrideC,
+                             index_t KBatch,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             CElementwiseOperation c_element_op)
+    {
+        return Argument{static_cast<const index_t*>(p_sorted_token_ids),
+                        static_cast<const index_t*>(p_sorted_expert_ids),
+                        static_cast<const index_t*>(p_max_token_id),
+                        static_cast<const ADataType*>(p_a),
+                        static_cast<const AScaleDataType*>(p_a_scale),
+                        static_cast<const BDataType*>(p_b),
+                        static_cast<const BScaleDataType*>(p_b_scale),
+                        p_ds,
+                        static_cast<CDataType*>(p_c),
+                        NumTokens,
+                        TopK,
+                        M,
+                        N,
+                        K,
+                        StrideA,
+                        StrideScaleA,
+                        StrideB,
+                        StrideScaleB,
+                        StrideDs,
+                        StrideC,
+                        KBatch,
+                        a_element_op,
+                        b_element_op,
+                        c_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument> MakeArgumentPointer(const void* p_a,
+                                                      const void* p_a_scale,
+                                                      const void* p_b,
+                                                      const void* p_b_scale,
+                                                      std::array<const void*, NumDTensor> p_ds,
+                                                      void* p_c,
+                                                      index_t M,
+                                                      index_t N,
+                                                      index_t K,
+                                                      index_t StrideA,
+                                                      index_t StrideScaleA,
+                                                      index_t StrideB,
+                                                      index_t StrideScaleB,
+                                                      std::array<ck::index_t, NumDTensor> StrideDs,
+                                                      index_t StrideC,
+                                                      index_t KBatch,
+                                                      AElementwiseOperation a_element_op,
+                                                      BElementwiseOperation b_element_op,
+                                                      CElementwiseOperation c_element_op) override
+    {
+        return std::make_unique<Argument>(nullptr,
+                                          nullptr,
+                                          nullptr,
+                                          static_cast<const ADataType*>(p_a),
+                                          static_cast<const AScaleDataType*>(p_a_scale),
+                                          static_cast<const BDataType*>(p_b),
+                                          static_cast<const BScaleDataType*>(p_b_scale),
+                                          p_ds,
+                                          static_cast<CDataType*>(p_c),
+                                          M, // randoms set, no use
+                                          0,
+                                          M,
+                                          N,
+                                          K,
+                                          StrideA,
+                                          StrideScaleA,
+                                          StrideB,
+                                          StrideScaleB,
+                                          StrideDs,
+                                          StrideC,
+                                          KBatch,
+                                          a_element_op,
+                                          b_element_op,
+                                          c_element_op);
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    // polymorphic
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        std::map<BlockGemmPipelineScheduler, std::string> BlkGemmPipelineSchedulerToString{
+            {BlockGemmPipelineScheduler::Intrawave, "Intrawave"},
+            {BlockGemmPipelineScheduler::Interwave, "Interwave"}};
+
+        std::map<BlockGemmPipelineVersion, std::string> BlkGemmPipelineVersionToString{
+            {BlockGemmPipelineVersion::v1, "v1"},
+            {BlockGemmPipelineVersion::v2, "v2"},
+            {BlockGemmPipelineVersion::v3, "v3"},
+            {BlockGemmPipelineVersion::v4, "v4"},
+            {BlockGemmPipelineVersion::v5, "v5"}};
+
+        // clang-format off
+        str << "DeviceMoeGEmmMx"
+            << "<"
+            << getGemmSpecializationString(GemmSpec) << ", "
+            << std::string(ALayout::name)[0]
+            << std::string(BLayout::name)[0]
+            << std::string(CLayout::name)[0]
+            << ">"
+            << " BlkSize: "
+            << BlockSize << ", "
+            << "BlkTile: "
+            << MPerBlock<<"x"<<NPerBlock<<"x"<<KPerBlock << ", "
+            << "WaveTile: "
+            << MPerXDL<<"x"<<NPerXDL << ", "
+            << "WaveMap: "
+            << MXdlPerWave<<"x" << NXdlPerWave<<", "
+            << "VmemReadVec: "
+            << ABlockTransferSrcScalarPerVector<<"x"<<BBlockTransferSrcScalarPerVector<<", "
+            << "BlkGemmPipelineScheduler: "
+            << BlkGemmPipelineSchedulerToString[BlkGemmPipeSched] << ", "
+            << "BlkGemmPipelineVersion: "
+            << BlkGemmPipelineVersionToString[BlkGemmPipelineVer] << ", "
+            << "BlkGemmPipelinePrefetchStages: "
+            << GridwiseGemm::BlockwiseGemmPipe::PrefetchStages;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

include/ck/tensor_operation/gpu/device/impl/device_moe_mx_gemm_bns.hpp

@@ -0,0 +1,579 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_moe_mx_gemm_bns.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+#include "ck/host_utility/flush_cache.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename ALayout,
+          typename BLayout,
+          typename DsLayout,
+          typename CLayout,
+          typename ADataType,
+          typename AScaleDataType,
+          typename BDataType,
+          typename BScaleDataType,
+          typename DsDataType,
+          typename CDataType,
+          typename GemmAccDataType,
+          typename CShuffleDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          GemmSpecialization GemmSpec,
+          index_t ScaleBlockSize,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t AK1,
+          index_t BK1,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename CDEShuffleBlockTransferScalarPerVectors,
+          BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave,
+          BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1,
+          index_t ActivationOP                        = 0,
+          bool NSwizzle                               = false,
+          bool IsInputGemm                            = true,
+          bool MulRoutedWeight                        = true,
+          typename IndexType                          = index_t,
+          typename ComputeTypeA                       = ADataType,
+          typename ComputeTypeB                       = BDataType>
+struct DeviceMoeGemmMXBNS : public DeviceMoEGemmMXBPreShuffle<ALayout,
+                                                              BLayout,
+                                                              DsLayout,
+                                                              CLayout,
+                                                              ADataType,
+                                                              AScaleDataType,
+                                                              BDataType,
+                                                              BScaleDataType,
+                                                              DsDataType,
+                                                              CDataType,
+                                                              ScaleBlockSize,
+                                                              AElementwiseOperation,
+                                                              BElementwiseOperation,
+                                                              CElementwiseOperation>
+{
+    static constexpr index_t NumDTensor = DsDataType::Size();
+    using GridwiseGemm =
+        GridwiseMoeGemmMXBNS<ALayout,
+                             BLayout,
+                             DsLayout,
+                             CLayout,
+                             ADataType,
+                             AScaleDataType,
+                             BDataType,
+                             BScaleDataType,
+                             GemmAccDataType,
+                             CShuffleDataType,
+                             DsDataType,
+                             CDataType,
+                             AElementwiseOperation,
+                             BElementwiseOperation,
+                             CElementwiseOperation,
+                             GemmSpec,
+                             ScaleBlockSize,
+                             BlockSize,
+                             MPerBlock,
+                             NPerBlock,
+                             KPerBlock,
+                             AK1,
+                             BK1,
+                             MPerXDL,
+                             NPerXDL,
+                             MXdlPerWave,
+                             NXdlPerWave,
+                             ABlockTransferThreadClusterLengths_AK0_M_AK1,
+                             ABlockTransferThreadClusterArrangeOrder,
+                             ABlockTransferSrcAccessOrder,
+                             ABlockTransferSrcVectorDim,
+                             ABlockTransferSrcScalarPerVector,
+                             ABlockTransferDstScalarPerVector_AK1,
+                             false,
+                             ABlockLdsExtraM,
+                             BBlockTransferThreadClusterLengths_BK0_N_BK1,
+                             BBlockTransferThreadClusterArrangeOrder,
+                             BBlockTransferSrcAccessOrder,
+                             BBlockTransferSrcVectorDim,
+                             BBlockTransferSrcScalarPerVector,
+                             BBlockTransferDstScalarPerVector_BK1,
+                             false,
+                             BBlockLdsExtraN,
+                             CShuffleMXdlPerWavePerShuffle,
+                             CShuffleNXdlPerWavePerShuffle,
+                             CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+                             CDEShuffleBlockTransferScalarPerVectors,
+                             BlkGemmPipeSched,
+                             BlkGemmPipelineVer,
+                             ActivationOP,
+                             NSwizzle,
+                             IsInputGemm,
+                             MulRoutedWeight,
+                             IndexType,
+                             ComputeTypeA,
+                             ComputeTypeB>;
+
+    using Argument = typename GridwiseGemm::Argument;
+
+    static constexpr index_t APackedSize = []() {
+        if constexpr(is_same_v<remove_cvref_t<ADataType>, pk_i4_t> ||
+                     is_same_v<remove_cvref_t<ADataType>, f4x2_pk_t>)
+            return 2;
+        else
+            return 1;
+    }();
+
+    static constexpr index_t BPackedSize = []() {
+        if constexpr(is_same_v<remove_cvref_t<BDataType>, pk_i4_t> ||
+                     is_same_v<remove_cvref_t<BDataType>, f4x2_pk_t>)
+            return 2;
+        else
+            return 1;
+    }();
+
+    int GetPreShuffleParameters() override { return NPerXDL; }
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            if(stream_config.log_level_ > 0)
+            {
+                arg.Print();
+            }
+
+            if(!GridwiseGemm::CheckValidity(arg))
+            {
+                throw std::runtime_error("wrong! GridwiseGemm has invalid setting");
+            }
+
+            index_t gdx, gdy, gdz;
+            std::tie(gdx, gdy, gdz) = GridwiseGemm::CalculateGridSize(arg.M, arg.N);
+
+            float ave_time = 0;
+
+            index_t k_grain = arg.KBatch * KPerBlock;
+            index_t K_split = (arg.K + k_grain - 1) / k_grain * KPerBlock;
+
+            const bool has_main_k_block_loop = GridwiseGemm::CalculateHasMainKBlockLoop(K_split);
+
+            const auto RunKernel = [&](const auto& kernel) {
+                if(stream_config.flush_cache)
+                {
+
+                    std::array<std::size_t, NumDTensor> DsSize;
+
+                    Argument arg_ = arg;
+
+                    const auto a_grid_desc_ak0_m_ak1 = GridwiseGemm::MakeAGridDescriptor_AK0_M_AK1(
+                        arg_.M, arg_.MPadded, arg_.K, arg_.KPadded, arg_.StrideA, arg_.AK0);
+                    const auto b_grid_desc_bk0_n_bk1 = GridwiseGemm::MakeBGridDescriptor_BK0_N_BK1(
+                        arg_.K, arg_.KPadded, arg_.N, arg_.NPadded, arg_.StrideB, arg_.BK0);
+
+                    auto size_a_buffer =
+                        a_grid_desc_ak0_m_ak1.GetElementSpaceSize() * sizeof(ADataType);
+                    auto size_b_buffer =
+                        b_grid_desc_bk0_n_bk1.GetElementSpaceSize() * sizeof(BDataType);
+
+                    const auto ds_grid_desc_m_n = GridwiseGemm::MakeDsGridDescriptor_M_N(
+                        arg_.M, arg_.MPadded, arg_.N, arg_.NPadded, arg_.StrideDs);
+
+                    static_for<0, NumDTensor, 1>{}([&](auto i) {
+                        using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
+                        DsSize[i] = ds_grid_desc_m_n[i].GetElementSpaceSize() * sizeof(DDataType);
+                    });
+                    ck::utility::RotatingMemWrapperMultiD<Argument, DsDataType> rotating_mem(
+                        arg_, stream_config.rotating_count, size_a_buffer, size_b_buffer, DsSize);
+                    rotating_mem.Print();
+
+                    auto run_flush_cache = [&]() {
+                        // flush icache
+                        ck::utility::flush_icache();
+                        // rotating mem
+                        rotating_mem.Next();
+                        // clear c mem
+                        if(arg_.KBatch > 1)
+                            hipGetErrorString(hipMemsetAsync(arg_.p_c_grid,
+                                                             0,
+                                                             arg_.M * arg_.N * sizeof(CDataType),
+                                                             stream_config.stream_id_));
+                    };
+
+                    ave_time = ck::utility::launch_and_time_kernel_with_preprocess<false>(
+                        stream_config,
+                        run_flush_cache,
+                        kernel,
+                        dim3(gdx, gdy, gdz),
+                        dim3(BlockSize),
+                        0,
+                        arg_);
+                }
+                else
+                {
+                    if(arg.KBatch > 1)
+                        hipGetErrorString(hipMemsetAsync(arg.p_c_grid,
+                                                         0,
+                                                         arg.M * arg.N * sizeof(CDataType),
+                                                         stream_config.stream_id_));
+
+                    ave_time = launch_and_time_kernel(
+                        stream_config, kernel, dim3(gdx, gdy, gdz), dim3(BlockSize), 0, arg);
+                }
+            };
+
+            // TODO: Check if this is the right algorithm for minimum_occupancy
+            constexpr index_t minimum_occupancy =
+                BlkGemmPipeSched == BlockGemmPipelineScheduler::Intrawave
+                    ? (BlkGemmPipelineVer == BlockGemmPipelineVersion::v3 &&
+                       MPerBlock * NPerBlock * KPerBlock * sizeof(ADataType) <= 128 * 128 * 64 * 2)
+                          ? 2
+                          : 1
+                    : 2;
+
+            constexpr auto MemoryDataOp =
+                IsInputGemm ? InMemoryDataOperationEnum::Set : InMemoryDataOperationEnum::AtomicAdd;
+            if(has_main_k_block_loop)
+            {
+                // Tail number always full
+                if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v1)
+                {
+                    {
+                        if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd)
+                        {
+                            const auto kernel = kernel_moe_mxgemm<GridwiseGemm,
+                                                                  true,
+                                                                  MemoryDataOp,
+                                                                  minimum_occupancy,
+                                                                  TailNumber::Odd>;
+                            RunKernel(kernel);
+                        }
+                        else
+                        {
+                            const auto kernel = kernel_moe_mxgemm<GridwiseGemm,
+                                                                  true,
+                                                                  MemoryDataOp,
+                                                                  minimum_occupancy,
+                                                                  TailNumber::Even>;
+                            RunKernel(kernel);
+                        }
+                    }
+                }
+                else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v3)
+                {
+                    if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd)
+                    {
+                        const auto kernel = kernel_moe_mxgemm<GridwiseGemm,
+                                                              true,
+                                                              MemoryDataOp,
+                                                              minimum_occupancy,
+                                                              TailNumber::Odd>;
+                        RunKernel(kernel);
+                    }
+                    else
+                    {
+                        const auto kernel = kernel_moe_mxgemm<GridwiseGemm,
+                                                              true,
+                                                              MemoryDataOp,
+                                                              minimum_occupancy,
+                                                              TailNumber::Even>;
+                        RunKernel(kernel);
+                    }
+                }
+                else
+                {
+                    throw std::runtime_error("todo: only v1 & v3 support now");
+                }
+            }
+            else
+            {
+                if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v1)
+                {
+                    if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd)
+                    {
+                        const auto kernel = kernel_moe_mxgemm<GridwiseGemm,
+                                                              false,
+                                                              MemoryDataOp,
+                                                              minimum_occupancy,
+                                                              TailNumber::Odd>;
+                        RunKernel(kernel);
+                    }
+                    else
+                    {
+                        const auto kernel = kernel_moe_mxgemm<GridwiseGemm,
+                                                              false,
+                                                              MemoryDataOp,
+                                                              minimum_occupancy,
+                                                              TailNumber::Even>;
+                        RunKernel(kernel);
+                    }
+                }
+                else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v3)
+                {
+                    if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd)
+                    {
+                        const auto kernel = kernel_moe_mxgemm<GridwiseGemm,
+                                                              false,
+                                                              MemoryDataOp,
+                                                              minimum_occupancy,
+                                                              TailNumber::Odd>;
+                        RunKernel(kernel);
+                    }
+                    else
+                    {
+                        const auto kernel = kernel_moe_mxgemm<GridwiseGemm,
+                                                              false,
+                                                              MemoryDataOp,
+                                                              minimum_occupancy,
+                                                              TailNumber::Even>;
+                        RunKernel(kernel);
+                    }
+                }
+            }
+
+            return ave_time;
+        }
+
+        // polymorphic
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        // only impl kbatch 1 now
+        if(arg.KBatch > 1)
+        {
+            return false;
+        }
+        if(!ck::is_xdl_supported())
+        {
+            return false;
+        }
+
+        if(!is_bf16_atomic_supported() && std::is_same_v<CDataType, ck::bhalf_t> && arg.KBatch > 1)
+        {
+            return false;
+        }
+
+        if((arg.K % AK1 != 0 || arg.K % BK1 != 0) && !(GemmSpec == GemmSpecialization::MKPadding ||
+                                                       GemmSpec == GemmSpecialization::NKPadding ||
+                                                       GemmSpec == GemmSpecialization::MNKPadding ||
+                                                       GemmSpec == GemmSpecialization::KPadding))
+        {
+            return false;
+        }
+        if(arg.N % NPerBlock != 0 || arg.K % KPerBlock != 0)
+        {
+            return false;
+        }
+
+        return GridwiseGemm::CheckValidity(arg);
+    }
+
+    // polymorphic
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(const void* p_sorted_token_ids,
+                             const void* p_sorted_expert_ids,
+                             const void* p_max_token_id,
+                             const void* p_a,
+                             const void* p_a_scale,
+                             const void* p_b,
+                             const void* p_b_scale,
+                             std::array<const void*, NumDTensor> p_ds,
+                             void* p_c,
+                             index_t NumTokens,
+                             index_t TopK,
+                             index_t M,
+                             index_t N,
+                             index_t K,
+                             index_t StrideA,
+                             index_t StrideScaleA,
+                             index_t StrideB,
+                             index_t StrideScaleB,
+                             std::array<index_t, NumDTensor> StrideDs,
+                             index_t StrideC,
+                             index_t KBatch,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             CElementwiseOperation c_element_op)
+    {
+        return Argument{static_cast<const index_t*>(p_sorted_token_ids),
+                        static_cast<const index_t*>(p_sorted_expert_ids),
+                        static_cast<const index_t*>(p_max_token_id),
+                        static_cast<const ADataType*>(p_a),
+                        static_cast<const AScaleDataType*>(p_a_scale),
+                        static_cast<const BDataType*>(p_b),
+                        static_cast<const BScaleDataType*>(p_b_scale),
+                        p_ds,
+                        static_cast<CDataType*>(p_c),
+                        NumTokens,
+                        TopK,
+                        M,
+                        N,
+                        K,
+                        StrideA,
+                        StrideScaleA,
+                        StrideB,
+                        StrideScaleB,
+                        StrideDs,
+                        StrideC,
+                        KBatch,
+                        a_element_op,
+                        b_element_op,
+                        c_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    // polymorphic
+    std::unique_ptr<BaseArgument> MakeArgumentPointer(const void* p_a,
+                                                      const void* p_a_scale,
+                                                      const void* p_b,
+                                                      const void* p_b_scale,
+                                                      std::array<const void*, NumDTensor> p_ds,
+                                                      void* p_c,
+                                                      index_t M,
+                                                      index_t N,
+                                                      index_t K,
+                                                      index_t StrideA,
+                                                      index_t StrideScaleA,
+                                                      index_t StrideB,
+                                                      index_t StrideScaleB,
+                                                      std::array<ck::index_t, NumDTensor> StrideDs,
+                                                      index_t StrideC,
+                                                      index_t KBatch,
+                                                      AElementwiseOperation a_element_op,
+                                                      BElementwiseOperation b_element_op,
+                                                      CElementwiseOperation c_element_op) override
+    {
+        return std::make_unique<Argument>(nullptr,
+                                          nullptr,
+                                          nullptr,
+                                          static_cast<const ADataType*>(p_a),
+                                          static_cast<const AScaleDataType*>(p_a_scale),
+                                          static_cast<const BDataType*>(p_b),
+                                          static_cast<const BScaleDataType*>(p_b_scale),
+                                          p_ds,
+                                          static_cast<CDataType*>(p_c),
+                                          M, // randoms set, no use
+                                          0,
+                                          M,
+                                          N,
+                                          K,
+                                          StrideA,
+                                          StrideScaleA,
+                                          StrideB,
+                                          StrideScaleB,
+                                          StrideDs,
+                                          StrideC,
+                                          KBatch,
+                                          a_element_op,
+                                          b_element_op,
+                                          c_element_op);
+    }
+
+    // polymorphic
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    // polymorphic
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        std::map<BlockGemmPipelineScheduler, std::string> BlkGemmPipelineSchedulerToString{
+            {BlockGemmPipelineScheduler::Intrawave, "Intrawave"},
+            {BlockGemmPipelineScheduler::Interwave, "Interwave"}};
+
+        std::map<BlockGemmPipelineVersion, std::string> BlkGemmPipelineVersionToString{
+            {BlockGemmPipelineVersion::v1, "v1"},
+            {BlockGemmPipelineVersion::v2, "v2"},
+            {BlockGemmPipelineVersion::v3, "v3"},
+            {BlockGemmPipelineVersion::v4, "v4"},
+            {BlockGemmPipelineVersion::v5, "v5"}};
+
+        // clang-format off
+        str << "DeviceMoeGEmmMx"
+            << "<"
+            << getGemmSpecializationString(GemmSpec) << ", "
+            << std::string(ALayout::name)[0]
+            << std::string(BLayout::name)[0]
+            << std::string(CLayout::name)[0]
+            << ">"
+            << " BlkSize: "
+            << BlockSize << ", "
+            << "BlkTile: "
+            << MPerBlock<<"x"<<NPerBlock<<"x"<<KPerBlock << ", "
+            << "WaveTile: "
+            << MPerXDL<<"x"<<NPerXDL << ", "
+            << "WaveMap: "
+            << MXdlPerWave<<"x" << NXdlPerWave<<", "
+            << "VmemReadVec: "
+            << ABlockTransferSrcScalarPerVector<<"x"<<BBlockTransferSrcScalarPerVector<<", "
+            << "BlkGemmPipelineScheduler: "
+            << BlkGemmPipelineSchedulerToString[BlkGemmPipeSched] << ", "
+            << "BlkGemmPipelineVersion: "
+            << BlkGemmPipelineVersionToString[BlkGemmPipelineVer] << ", "
+            << "BlkGemmPipelinePrefetchStages: "
+            << GridwiseGemm::BlockwiseGemmPipe::PrefetchStages;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_multi_d_b_preshuffle.hpp

@@ -1,5 +1,5 @@
 // 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
 

include/ck/tensor_operation/gpu/grid/gridwise_moe_gemm.hpp

@@ -262,7 +262,7 @@ struct GridwiseMoeGemm
     }
     __host__ __device__ static auto CalculateBK0Shuffled(index_t K)
     {
-        return math::integer_divide_ceil(K, KLane * KPack);
+        return math::integer_divide_ceil(K, KLane * KPack / KGroup);
     }
 
     __host__ __device__ static auto CalculateKPadded(index_t K)
@@ -404,7 +404,7 @@ struct GridwiseMoeGemm
 
     __host__ __device__ static auto MakeBGridDescriptor_Preshuffled(index_t N0, index_t K0)
     {
-        constexpr index_t NkSwizzleNumber = Number<warpSize * KPack>{};
+        constexpr index_t NkSwizzleNumber = Number<warpSize * KPack / KGroup>{};
         return make_naive_tensor_descriptor(
             make_tuple(N0 / NWave, NWave, K0, NkSwizzleNumber),
             make_tuple(NWave * K0 * NkSwizzleNumber, K0 * NkSwizzleNumber, NkSwizzleNumber, I1));
@@ -1314,7 +1314,7 @@ struct GridwiseMoeGemm
                   make_multi_index(n_block_data_idx_on_grid,
                                    get_warp_local_1d_id() % NWave,
                                    0,
-                                   KPack * (get_thread_local_1d_id() % warpSize)));
+                                   KPack / KGroup * (get_thread_local_1d_id() % warpSize)));
 
         // LDS allocation for A and B: be careful of alignment
         // Cast after lds
@@ -1360,7 +1360,7 @@ struct GridwiseMoeGemm
                       make_multi_index(n_block_data_idx_on_grid,
                                        get_warp_local_1d_id() % NWave,
                                        0,
-                                       KPack * (get_thread_local_1d_id() % warpSize)));
+                                       KPack / KGroup * (get_thread_local_1d_id() % warpSize)));
             blockwise_gemm_pipeline.template Run<HasMainKBlockLoop, TailNum>(
                 a_grid_desc_ak0_m_ak1,
                 a_block_desc_ak0_m_ak1,
@@ -1899,7 +1899,8 @@ struct GridwiseMoeGemm
         const auto c_grid_desc_mblock_mperblock_nblock_nperblock =
             MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
                 c_grid_desc_m_n, problem.MBlock, problem.NBlock);
-        const index_t max_token_id    = __builtin_amdgcn_readfirstlane(p_max_token_id[0]);
+        const index_t max_token_id = __builtin_amdgcn_readfirstlane(p_max_token_id[0]);
+        // static_assert(NSwizzle == false, "to do fix: need another pr in sorting merged");
         const index_t expert_block_id = NSwizzle ? blockIdx.x / problem.NBlock : blockIdx.y;
         if(expert_block_id * MPerBlock >= max_token_id)
             return;
@@ -1908,12 +1909,13 @@ struct GridwiseMoeGemm
         const auto block_mn = [&]() -> std::pair<int, int> {
             if constexpr(NSwizzle)
             {
-                const index_t ecnt_prefix    = p_max_token_id[1 + expert_id];
-                const index_t prefix_block   = ecnt_prefix * problem.NBlock;
-                const index_t ecnt           = p_max_token_id[2 + expert_id] - ecnt_prefix;
-                const index_t expert_swizzle = ecnt > 0 ? ecnt : 1;
-                const index_t bid_new        = blockIdx.x - prefix_block;
-                const index_t nid            = __builtin_amdgcn_readfirstlane(
+                const index_t ecnt_prefix  = p_max_token_id[1 + expert_id];
+                const index_t prefix_block = ecnt_prefix * problem.NBlock;
+                const index_t ecnt         = p_max_token_id[2 + expert_id] - ecnt_prefix;
+                const index_t expert_swizzle =
+                    ecnt > 0 ? ecnt : 1; // p_max_token_id[expert_id + 1]; // 2
+                const index_t bid_new = blockIdx.x - prefix_block;
+                const index_t nid     = __builtin_amdgcn_readfirstlane(
                     bid_new % 8 + bid_new / (8 * expert_swizzle) * 8);
                 const index_t mid =
                     __builtin_amdgcn_readfirstlane(ecnt_prefix + bid_new / 8 % expert_swizzle);
@@ -1924,9 +1926,9 @@ struct GridwiseMoeGemm
                 return {blockIdx.x, blockIdx.y};
             }
         }();
+
         const index_t block_n_id = block_mn.first;
         const index_t block_m_id = block_mn.second;
-
         const index_t token0 =
             __builtin_amdgcn_readfirstlane(p_sorted_token_ids[block_m_id * MPerBlock] & 0xffffff);
 
@@ -1938,11 +1940,9 @@ struct GridwiseMoeGemm
         constexpr auto AMRepeats  = MPerBlock / AMThreads;
         const index_t token_pos   = block_m_id * MPerBlock + threadIdx.x / AKThreads * AMRepeats;
 
-        if(token_pos >= max_token_id || expert_block_id * MPerBlock >= max_token_id ||
-           token0 >= problem.NumTokens)
+        if(token_pos >= max_token_id || token0 >= problem.NumTokens)
             return;
-        StaticallyIndexedArray<IndexType, AMRepeats>
-            gather_offsets; //= p_sorted_token_ids[token_pos];
+        StaticallyIndexedArray<IndexType, AMRepeats> gather_offsets;
         static_for<0, AMRepeats, 1>{}([&](auto m0) {
             const index_t fused_token = p_sorted_token_ids[token_pos + m0];
             index_t token_offset      = fused_token & 0xffffff;
@@ -1952,7 +1952,8 @@ struct GridwiseMoeGemm
             }
             gather_offsets(m0) = static_cast<IndexType>(token_offset) * problem.K;
         });
-        const index_t expert_stride = __builtin_amdgcn_readfirstlane(problem.N * problem.K);
+        const index_t expert_stride =
+            __builtin_amdgcn_readfirstlane(problem.N * problem.K * (IsInputGemm ? 2 : 1));
 
         // N0, K0, Blocksize*KPack
         const index_t n_block_data_idx_on_grid =
@@ -1963,7 +1964,6 @@ struct GridwiseMoeGemm
         const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
             p_b_grid + expert_id * expert_stride / BPackedSize,
             b_grid_desc_bpreshuffled.GetElementSpaceSize());
-
         // A matrix in LDS memory, dst of blockwise copy
         constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
 
@@ -2025,7 +2025,7 @@ struct GridwiseMoeGemm
                   make_multi_index(n_block_data_idx_on_grid,
                                    get_warp_local_1d_id() % NWave,
                                    0,
-                                   KPack * (get_thread_local_1d_id() % warpSize)));
+                                   KPack / KGroup * (get_thread_local_1d_id() % warpSize)));
 
         // LDS allocation for A and B: be careful of alignment
         // Cast after lds
@@ -2042,24 +2042,76 @@ struct GridwiseMoeGemm
         static_assert(std::is_default_constructible_v<BlockwiseGemmPipe>);
         auto blockwise_gemm_pipeline = BlockwiseGemmPipe{};
         auto c_thread_buf            = blockwise_gemm_pipeline.GetCThreadBuffer();
+        decltype(c_thread_buf) c_thread_buf_up;
+
+        StaticBufferTupleOfVector<AddressSpaceEnum::Vgpr,
+                                  float,
+                                  c_thread_buf.num_of_v_,
+                                  c_thread_buf.s_per_v,
+                                  true>
+            c_thread_buf_fp32;
 
         const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane(
             (a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2)) /
             KPerBlock);
 
-        blockwise_gemm_pipeline.template Run<HasMainKBlockLoop, TailNum>(a_grid_desc_ak0_m_ak1,
-                                                                         a_block_desc_ak0_m_ak1,
-                                                                         a_blockwise_copy,
-                                                                         a_grid_buf,
-                                                                         a_block_bufs,
-                                                                         a_block_slice_copy_step,
-                                                                         b_grid_desc_bpreshuffled,
-                                                                         b_blockwise_copy,
-                                                                         b_grid_buf,
-                                                                         b_block_bufs,
-                                                                         b_block_slice_copy_step,
-                                                                         c_thread_buf,
-                                                                         num_k_block_main_loop);
+        if constexpr(IsInputGemm)
+        {
+            const BDataType* p_b_grid_up = p_b_grid + expert_stride / 2 / BPackedSize;
+            const auto b_grid_buf_up     = make_dynamic_buffer<AddressSpaceEnum::Global>(
+                p_b_grid_up + expert_id * expert_stride / BPackedSize,
+                b_grid_desc_bpreshuffled.GetElementSpaceSize());
+            auto b_blockwise_copy_up = ThreadwiseTensorSliceTransfer_v2<
+                BDataType,
+                BDataType,
+                decltype(b_grid_desc_bpreshuffled),
+                decltype(b_block_desc_bk0_n_bk1),
+                Sequence<Number<NXdlPerWave>{}, I1, Number<KRepeat>{}, Number<BK1Value>{}>,
+                Sequence<1, 2, 0, 3>,
+                3,
+                BBlockTransferSrcScalarPerVector,
+                BThreadTransferSrcResetCoordinateAfterRun,
+                true>(b_grid_desc_bpreshuffled,
+                      make_multi_index(n_block_data_idx_on_grid,
+                                       get_warp_local_1d_id() % NWave,
+                                       0,
+                                       KPack / KGroup * (get_thread_local_1d_id() % warpSize)));
+            blockwise_gemm_pipeline.template Run<HasMainKBlockLoop, TailNum>(
+                a_grid_desc_ak0_m_ak1,
+                a_block_desc_ak0_m_ak1,
+                a_blockwise_copy,
+                a_grid_buf,
+                a_block_bufs,
+                a_block_slice_copy_step,
+                b_grid_desc_bpreshuffled,
+                b_blockwise_copy,
+                b_blockwise_copy_up,
+                b_grid_buf,
+                b_grid_buf_up,
+                b_block_bufs,
+                b_block_slice_copy_step,
+                c_thread_buf,
+                c_thread_buf_up,
+                num_k_block_main_loop);
+        }
+        else
+        {
+
+            blockwise_gemm_pipeline.template Run<HasMainKBlockLoop, TailNum>(
+                a_grid_desc_ak0_m_ak1,
+                a_block_desc_ak0_m_ak1,
+                a_blockwise_copy,
+                a_grid_buf,
+                a_block_bufs,
+                a_block_slice_copy_step,
+                b_grid_desc_bpreshuffled,
+                b_blockwise_copy,
+                b_grid_buf,
+                b_block_bufs,
+                b_block_slice_copy_step,
+                c_thread_buf,
+                num_k_block_main_loop);
+        }
 
         // shuffle C and write out
         {
@@ -2087,6 +2139,185 @@ struct GridwiseMoeGemm
             constexpr auto M4 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I6);
             constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I7);
 
+            // mul scales
+            const float* p_sorted_weights_0 = p_ds_grid[I0];
+            const float* p_scale_b          = p_ds_grid[I1];
+
+            static_assert(M0 * M1 * M2 * M3 * M4 == MPerBlock);
+            static_assert(M4 == 4);
+            const index_t m1 = get_warp_local_1d_id() / NWave;
+            const index_t m3 = threadIdx.x % get_warp_size() / MPerXdl;
+
+            if(p_sorted_weights_0 != nullptr && p_scale_b != nullptr)
+            {
+                if constexpr(PerTokenQuant)
+                {
+                    constexpr index_t scale_stride = (IsInputGemm ? 2 : 1);
+                    p_scale_b += expert_id * problem.N * scale_stride + block_n_id * NPerBlock +
+                                 get_warp_local_1d_id() % NWave * NPerXdl + threadIdx.x % NPerXdl;
+                }
+                else
+                {
+                    p_scale_b += expert_id;
+                }
+
+                vector_type<int32_t, 4> scale_token_ids;
+                vector_type<float, 4> topk_weights;
+                static_for<0, NXdlPerWave, 1>{}([&](auto n0) {
+                    const float scale_b = p_scale_b[n0 * NWave * NPerXdl * PerTokenQuant];
+                    static_for<0, MXdlPerWave, 1>{}([&](auto m0) { // MXDLPerWave
+                        static_for<0, M2, 1>{}([&](auto m2) {      // m_inst_num_groups_per_blk
+                            const index_t m_pos = block_m_id * MPerBlock + m0 * M1 * M2 * M3 * M4 +
+                                                  m1 * M2 * M3 * M4 + m2 * M3 * M4 + m3 * M4;
+                            if constexpr(PerTokenQuant)
+                            {
+                                scale_token_ids =
+                                    *c_style_pointer_cast<const vector_type<int32_t, M4>*>(
+                                        p_sorted_token_ids + m_pos);
+                            }
+                            if constexpr(MulRoutedWeight)
+                            {
+                                topk_weights = *c_style_pointer_cast<const vector_type<float, M4>*>(
+                                    p_ds_grid[I2] + m_pos);
+                            }
+                            static_for<0, M4, 1>{}([&](auto m4) { // m_inst_group_size
+                                float scale_a = [&]() {
+                                    if constexpr(PerTokenQuant)
+                                    {
+                                        index_t fused_token = scale_token_ids.AsType<index_t>()[m4];
+                                        const index_t token_offset = fused_token & 0xffffff;
+                                        return token_offset < problem.NumTokens
+                                                   ? p_sorted_weights_0[token_offset]
+                                                   : 0.0;
+                                    }
+                                    else
+                                    {
+                                        return p_sorted_weights_0[0];
+                                    }
+                                }();
+                                constexpr index_t c_offset =
+                                    blockwise_gemm_pipeline.GetCThreadDesc().CalculateOffset(
+                                        make_tuple(m0, n0, m2 * M4 + m4));
+                                constexpr auto cidx = Number<c_offset>{};
+                                if constexpr(IsInputGemm) // gu fusion
+                                {
+                                    if constexpr(ActivationOperation == Activation::silu_and_mul)
+                                    {
+                                        const float scale_up =
+                                            p_scale_b[(n0 * NWave * NPerXdl + problem.N) *
+                                                      PerTokenQuant];
+                                        float gate = scale_a * scale_b * c_thread_buf[cidx];
+                                        float up   = scale_a * scale_up * c_thread_buf_up[cidx];
+                                        if constexpr(MulRoutedWeight)
+                                        {
+                                            gate = gate * topk_weights.AsType<float>()[m4];
+                                            up   = up * topk_weights.AsType<float>()[m4];
+                                        }
+                                        if constexpr(is_same_v<remove_cvref_t<BDataType>, pk_i4_t>)
+                                        {
+                                            gate *= 16;
+                                            up *= 16;
+                                        }
+                                        tensor_operation::element_wise::Silu{}(gate, gate);
+                                        c_thread_buf_fp32(cidx) = gate * up;
+                                    }
+                                    else if(ActivationOperation == Activation::gelu_and_mul)
+                                    {
+                                        const float scale_up =
+                                            p_scale_b[(n0 * NWave * NPerXdl + problem.N) *
+                                                      PerTokenQuant];
+                                        float gate = scale_a * scale_b * c_thread_buf[cidx];
+                                        float up   = scale_a * scale_up * c_thread_buf_up[cidx];
+                                        if constexpr(MulRoutedWeight)
+                                        {
+                                            gate = gate * topk_weights.AsType<float>()[m4];
+                                            up   = up * topk_weights.AsType<float>()[m4];
+                                        }
+                                        if constexpr(is_same_v<remove_cvref_t<BDataType>, pk_i4_t>)
+                                        {
+                                            gate *= 16;
+                                            up *= 16;
+                                        }
+                                        tensor_operation::element_wise::Gelu{}(gate, gate);
+                                        c_thread_buf_fp32(cidx) = gate * up;
+                                    }
+                                }
+                                else
+                                {
+                                    c_thread_buf_fp32(cidx) =
+                                        scale_a * scale_b * c_thread_buf[cidx];
+                                    if constexpr(MulRoutedWeight)
+                                    {
+                                        c_thread_buf_fp32(cidx) = c_thread_buf_fp32(cidx) *
+                                                                  topk_weights.AsType<float>()[m4];
+                                    }
+                                }
+                            });
+                        });
+                    });
+                });
+            }
+            else
+            {
+                vector_type<float, 4> topk_weights; // for gemm2 only
+                static_for<0, NXdlPerWave, 1>{}([&](auto n0) {
+                    static_for<0, MXdlPerWave, 1>{}([&](auto m0) { // MXDLPerWave
+                        static_for<0, M2, 1>{}([&](auto m2) {      // m_inst_num_groups_per_blk
+                            const index_t m_pos = block_m_id * MPerBlock + m0 * M1 * M2 * M3 * M4 +
+                                                  m1 * M2 * M3 * M4 + m2 * M3 * M4 + m3 * M4;
+                            if constexpr(MulRoutedWeight)
+                            {
+                                topk_weights = *c_style_pointer_cast<const vector_type<float, M4>*>(
+                                    p_ds_grid[I2] + m_pos);
+                            }
+                            static_for<0, M4, 1>{}([&](auto m4) { // m_inst_group_size
+                                constexpr index_t c_offset =
+                                    blockwise_gemm_pipeline.GetCThreadDesc().CalculateOffset(
+                                        make_tuple(m0, n0, m2 * M4 + m4));
+                                constexpr auto cidx = Number<c_offset>{};
+
+                                if constexpr(IsInputGemm) // gu fusion
+                                {
+                                    if constexpr(ActivationOperation == Activation::silu_and_mul)
+                                    {
+                                        float gate = c_thread_buf[cidx];
+                                        float up   = c_thread_buf_up[cidx];
+                                        if constexpr(MulRoutedWeight)
+                                        {
+                                            gate = gate * topk_weights.AsType<float>()[m4];
+                                            up   = up * topk_weights.AsType<float>()[m4];
+                                        }
+                                        tensor_operation::element_wise::Silu{}(gate, gate);
+                                        c_thread_buf_fp32(cidx) = gate * up;
+                                    }
+                                    else if(ActivationOperation == Activation::gelu_and_mul)
+                                    {
+                                        float gate = c_thread_buf[cidx];
+                                        float up   = c_thread_buf_up[cidx];
+                                        if constexpr(MulRoutedWeight)
+                                        {
+                                            gate = gate * topk_weights.AsType<float>()[m4];
+                                            up   = up * topk_weights.AsType<float>()[m4];
+                                        }
+                                        tensor_operation::element_wise::Gelu{}(gate, gate);
+                                        c_thread_buf_fp32(cidx) = gate * up;
+                                    }
+                                }
+                                else
+                                {
+                                    c_thread_buf_fp32(cidx) = c_thread_buf[cidx];
+                                    if constexpr(MulRoutedWeight)
+                                    {
+                                        c_thread_buf_fp32(cidx) = topk_weights.AsType<float>()[m4] *
+                                                                  c_thread_buf_fp32[cidx];
+                                    }
+                                }
+                            });
+                        });
+                    });
+                });
+            }
+
             constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
                 GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
 
@@ -2184,18 +2415,8 @@ struct GridwiseMoeGemm
 
             const auto ds_grid_buf = generate_tuple(
                 [&](auto i) {
-                    using DDataType       = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
-                    const DDataType* ptr_ = p_ds_grid[i];
-                    // hack logic here to support different kind of strides. todo fix it.
-                    // ascale t, 1; bscale E, N, 1, move ptr to E
-                    // if(i.value == 1)
-                    // {
-                    //     ptr_ +=
-                    //         expert_id * (problem.StrideDs[1] ? problem.StrideDs[1] * problem.N :
-                    //         1);
-                    // }
                     return make_dynamic_buffer<AddressSpaceEnum::Global>(
-                        ptr_, ds_grid_desc_m_n[i].GetElementSpaceSize());
+                        p_ds_grid[i], ds_grid_desc_m_n[i].GetElementSpaceSize());
                 },
                 Number<NumDTensor>{});
 
@@ -2271,7 +2492,6 @@ struct GridwiseMoeGemm
 
             auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
                 p_c_grid, c_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
-            // space filling curve for threadwise C in VGPR
             constexpr auto sfc_c_vgpr =
                 SpaceFillingCurve<Sequence<MXdlPerWave, NXdlPerWave, 1, 1, M2, 1, M4, 1>,
                                   Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
@@ -2310,7 +2530,7 @@ struct GridwiseMoeGemm
                     block_m_id * MPerBlock + threadIdx.x / ENThreads * EMRepeats + dstidx(I1);
                 static_for<0, EMRepeats, 1>{}([&](auto m0) {
                     const index_t fused_token = p_sorted_token_ids[c_token_pos + m0];
-                    index_t token_offset      = fused_token & 0xffffff;
+                    IndexType token_offset    = fused_token & 0xffffff;
                     if constexpr(IsInputGemm)
                     {
                         token_offset = token_offset * problem.TopK + (fused_token >> 24);
@@ -2323,7 +2543,7 @@ struct GridwiseMoeGemm
                 // each thread write its data from VGPR to LDS
                 c_thread_copy_vgpr_to_lds.Run(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2,
                                               sfc_c_vgpr.GetIndexTupleOfNumber(access_id),
-                                              c_thread_buf,
+                                              c_thread_buf_fp32,
                                               c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
                                               c_shuffle_block_buf);
 

include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp

@@ -423,6 +423,253 @@ struct ThreadwiseTensorSliceTransfer_v2
     SrcCoord src_coord_;
 }; // namespace ck
 
+template <typename SrcData,
+          typename DstData,
+          typename SrcDesc,
+          typename DstDesc,
+          typename SliceLengths,
+          typename DimAccessOrder,
+          index_t SrcVectorDim,
+          index_t SrcScalarPerVector,
+          index_t SrcScalarStrideInVector,
+          bool SrcResetCoordinateAfterRun,
+          index_t scale_gather_num,
+          index_t KRepeat,
+          bool InvalidElementAsNaN                                        = false,
+          typename enable_if<DstDesc::IsKnownAtCompileTime(), bool>::type = false>
+struct ThreadwiseTensorSliceTransfer_v2_gather
+{
+    static_assert((InvalidElementAsNaN && !ck::is_integral<DstData>::value) ||
+                      (!InvalidElementAsNaN),
+                  "Filling invalid element as NaN is only for floating point types");
+
+    static constexpr index_t nDim = SliceLengths::Size();
+
+    using Index = MultiIndex<nDim>;
+
+    using SrcCoord = decltype(make_tensor_coordinate(SrcDesc{}, Index{}));
+
+    using SrcCoordStep = decltype(make_tensor_coordinate_step(SrcDesc{}, Index{}));
+
+    static constexpr index_t PackedSize = []() {
+        if constexpr(is_same_v<remove_cvref_t<SrcData>, pk_i4_t> ||
+                     is_same_v<remove_cvref_t<SrcData>, f4x2_pk_t>)
+            return 2;
+        else
+            return 1;
+    }();
+
+    __device__ constexpr ThreadwiseTensorSliceTransfer_v2_gather(
+        const SrcDesc& src_desc,
+        const Index& src_slice_origin_idx,
+        const StaticallyIndexedArray<index_t, scale_gather_num>& scale_gather_offsets)
+        : src_coord_(make_tensor_coordinate(src_desc, src_slice_origin_idx)),
+          scale_gather_offsets_(scale_gather_offsets)
+    {
+        static_assert(DstDesc::IsKnownAtCompileTime(),
+                      "wrong! SrcDesc need to known at compile-time");
+        static_assert(SliceLengths::At(Number<SrcVectorDim>{}) % SrcScalarPerVector == 0,
+                      "wrong! Not divisible");
+
+        if constexpr(is_same_v<remove_cvref_t<SrcData>, pk_i4_t> ||
+                     is_same_v<remove_cvref_t<SrcData>, f4x2_pk_t>)
+        {
+            static_assert(SrcScalarPerVector % PackedSize == 0, "pk data N cannot be 1");
+        }
+    }
+
+    __device__ void SetSrcSliceOrigin(const SrcDesc& src_desc, const Index& src_slice_origin_idx)
+    {
+        auto adjusted_origin_idx = [&]() {
+            Index idx;
+
+            static_for<0, nDim, 1>{}(
+                [&](auto i) { idx(i) = i.value == 0 ? 0 : src_slice_origin_idx[Number<i>{}]; });
+
+            return idx;
+        }();
+
+        src_coord_ = make_tensor_coordinate(src_desc, adjusted_origin_idx);
+    }
+
+    template <typename SrcBuffer, typename DstBuffer, typename DstSliceOriginIdx>
+    __device__ void Run(const SrcDesc& src_desc,
+                        const SrcBuffer& src_buf,
+                        const DstDesc&,
+                        const DstSliceOriginIdx&,
+                        DstBuffer& dst_buf)
+    {
+        static_assert(DstDesc::IsKnownAtCompileTime(),
+                      "wrong! DstDesc need to known at compile-time");
+
+        static_assert(is_known_at_compile_time<remove_cvref_t<DstSliceOriginIdx>>::value,
+                      "wrong! DstSliceOrigin need to known at compile-time");
+
+        static_assert(
+            is_same<remove_cvref_t<typename DstBuffer::type>, remove_cvref_t<DstData>>::value &&
+            "wrong! inconsistent type");
+
+        // DstDesc and dst_slice_origin_idx are known at compile-time
+        constexpr auto dst_desc             = remove_cvref_t<DstDesc>{};
+        constexpr auto dst_slice_origin_idx = DstSliceOriginIdx{};
+
+        // scalar per access on each dim
+        // TODO: don't use lambda_scalar_per_access
+        constexpr auto src_scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<SrcVectorDim, SrcScalarPerVector>{}, Number<nDim>{});
+
+        constexpr auto src_scalar_step_in_vector =
+            generate_sequence(detail::lambda_scalar_step_in_vector<SrcVectorDim>{}, Number<nDim>{});
+
+        using SpaceFillingCurve = SpaceFillingCurve<SliceLengths,
+                                                    DimAccessOrder,
+                                                    remove_cv_t<decltype(src_scalar_per_access)>>;
+
+        // loop over tensor and copy
+        constexpr auto num_access = SpaceFillingCurve::GetNumOfAccess();
+
+        static_for<0, scale_gather_num, 1>{}([&](auto gather_idx) { // MRepeat
+            static_for<0, KRepeat, 1>{}([&](auto k0) {              // KRepeat
+                constexpr auto current_dst_origin =
+                    to_multi_index(dst_slice_origin_idx) + make_multi_index(gather_idx, k0, 0);
+                MoveSrcSliceWindow(src_desc, make_multi_index(0, 0, 0));
+
+                static_for<0, num_access, 1>{}([&](auto idx_1d) {
+                    typename vector_type_maker<SrcData, SrcScalarPerVector / PackedSize>::type
+                        src_vector;
+
+                    using src_vector_t =
+                        typename vector_type_maker<SrcData,
+                                                   SrcScalarPerVector / PackedSize>::type::type;
+                    constexpr auto src_data_idx = SpaceFillingCurve::GetIndex(idx_1d);
+
+                    const bool is_src_valid =
+                        coordinate_has_valid_offset_assuming_visible_index_is_valid(src_desc,
+                                                                                    src_coord_);
+
+                    // copy data from src_buf into src_vector
+                    src_vector.template AsType<src_vector_t>()(Number<0>{}) =
+                        src_buf.template Get<src_vector_t>(src_coord_.GetOffset() / PackedSize +
+                                                               scale_gather_offsets_(gather_idx),
+                                                           is_src_valid);
+
+                    // copy data from src_vector into dst_buf
+                    static_for<0, SrcScalarPerVector / PackedSize, 1>{}([&](auto i) {
+                        constexpr index_t dst_offset =
+                            dst_desc.CalculateOffset(to_multi_index(dst_slice_origin_idx) +
+                                                     src_data_idx + i * src_scalar_step_in_vector);
+                        constexpr auto full_dst_offset =
+                            dst_desc.CalculateOffset(current_dst_origin) + dst_offset;
+
+                        if constexpr(InvalidElementAsNaN)
+                        {
+                            dst_buf(full_dst_offset) =
+                                is_src_valid ? type_convert<DstData>(
+                                                   src_vector.template AsType<SrcData>()[i])
+                                             : NumericLimits<DstData>::QuietNaN();
+                        }
+                        else
+                        {
+                            dst_buf(Number<full_dst_offset>{}) =
+                                type_convert<DstData>(src_vector.template AsType<SrcData>()[i]);
+                        }
+                    });
+
+                    if constexpr(idx_1d.value != num_access - 1)
+                    {
+                        constexpr auto forward_step = SpaceFillingCurve::GetForwardStep(idx_1d);
+
+                        move_tensor_coordinate(src_desc,
+                                               src_coord_,
+                                               make_tensor_coordinate_step(src_desc, forward_step));
+                    }
+
+                    MoveSrcSliceWindow(
+                        src_desc,
+                        make_multi_index(
+                            0, 4, 0)); // hacky fix: 4 means xdlops_gemm.KPerXdlops / ScaleBlockSize
+                });
+            });
+            MoveSrcSliceWindow(src_desc, make_multi_index(0, -(KRepeat * 4), 0));
+        });
+
+        // printf("blockIdx.y: %d, tid: %d, dst_buf<%f>\n",
+        //        blockIdx.y,
+        //        threadIdx.x,
+        //        dst_buf(Number<0>{}));
+
+        // move src coordinate back to slice origin (or not)
+        if constexpr(SrcResetCoordinateAfterRun)
+        {
+            const auto src_reset_step =
+                make_tensor_coordinate_step(src_desc, GetSrcCoordinateResetStep());
+
+            move_tensor_coordinate(src_desc, src_coord_, src_reset_step);
+        }
+    }
+
+    __device__ static constexpr auto GetSrcCoordinateResetStep()
+    {
+        constexpr auto src_scalar_per_access = generate_sequence(
+            detail::lambda_scalar_per_access<SrcVectorDim, SrcScalarPerVector>{}, Number<nDim>{});
+
+        using SpaceFillingCurve = SpaceFillingCurve<SliceLengths,
+                                                    DimAccessOrder,
+                                                    remove_cv_t<decltype(src_scalar_per_access)>>;
+
+        constexpr auto num_access = SpaceFillingCurve::GetNumOfAccess();
+        if constexpr(num_access == 0)
+        {
+            return typename SpaceFillingCurve::Index{};
+        }
+        else
+        {
+            constexpr auto reset_step =
+                SpaceFillingCurve::GetStepBetween(Number<num_access - 1>{}, Number<0>{});
+
+            return reset_step;
+        }
+    }
+
+    // dst_slice_origin_step_idx need to be known at compile-time, for performance reason
+    __device__ void MoveSrcSliceWindow(const SrcDesc& src_desc,
+                                       const Index& src_slice_origin_step_idx)
+    {
+        // if src coord was not reset by Run(), then need to adjust the step here
+        const auto adjusted_step_idx =
+            SrcResetCoordinateAfterRun ? src_slice_origin_step_idx
+                                       : src_slice_origin_step_idx + GetSrcCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(src_desc, adjusted_step_idx);
+
+        move_tensor_coordinate(src_desc, src_coord_, adjusted_step);
+    }
+
+    // src_slice_origin_step_idx need to be known at compile-time, for performance reason
+    template <typename SrcMoveSliceWindowStepHack>
+    __device__ void
+    MoveSrcSliceWindow(const SrcDesc& src_desc,
+                       const Index& src_slice_origin_step_idx,
+                       const SrcMoveSliceWindowStepHack& src_move_slice_window_step_hack)
+    {
+        // if src coord was not reset by RunRead(), then need to adjust the step here
+        const auto adjusted_step_idx =
+            SrcResetCoordinateAfterRun ? src_slice_origin_step_idx
+                                       : src_slice_origin_step_idx + GetSrcCoordinateResetStep();
+
+        // is it OK to construct a new step every time?
+        const auto adjusted_step = make_tensor_coordinate_step(
+            src_desc, adjusted_step_idx, src_move_slice_window_step_hack);
+
+        move_tensor_coordinate(src_desc, src_coord_, adjusted_step);
+    }
+
+    private:
+    SrcCoord src_coord_;
+    StaticallyIndexedArray<index_t, scale_gather_num> scale_gather_offsets_;
+}; // namespace ck
+
 // Assume:
 //   1. src_desc and dst_desc are not known at compile-time
 //   2. SrcBuffer and DstBuffer are DynamicBuffer

include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v3r1_gather.hpp

@@ -96,7 +96,8 @@ struct ThreadwiseTensorSliceTransfer_v3r1_gather
           dst_element_op_(dst_element_op),
           gather_offsets_(gather_offsets)
     {
-        if constexpr(is_same_v<remove_cvref_t<SrcData>, pk_i4_t>)
+        if constexpr(is_same_v<remove_cvref_t<SrcData>, pk_i4_t> ||
+                     is_same_v<remove_cvref_t<SrcData>, f4x2_pk_t>)
         {
             static_assert(is_same_v<remove_cvref_t<SrcData>, remove_cvref_t<DstData>>,
                           "SrcData != DstData");
@@ -105,7 +106,8 @@ struct ThreadwiseTensorSliceTransfer_v3r1_gather
                 SrcScalarPerVector_ % PackedSize == 0 && DstScalarPerVector_ % PackedSize == 0,
                 "SrcScalarPerVector_ and DstScalarPerVector_ cannot be 1 for packed data type");
 
-            static_assert(SrcVectorDim == DstVectorDim, "pk_i4_t does not support transpose");
+            static_assert(SrcVectorDim == DstVectorDim,
+                          "pk_i4_t or f4x2_pk_t does not support transpose");
         }
     }
 
@@ -222,10 +224,19 @@ struct ThreadwiseTensorSliceTransfer_v3r1_gather
             auto gather_offset =
                 gather_offsets_(ordered_src_access_idx[Number<ordered_gather_dim>{}]);
 
-            const IndexType ld_offset = src_coord_.GetOffset() + gather_offset;
+            const IndexType ld_offset = src_coord_.GetOffset() / PackedSize + gather_offset;
             src_oob_thread_scratch_tuple_(thread_scratch_id)
                 .template SetAsType<bool>(src_data_idx_seq, true);
 
+#if 0
+            printf("blkx: %u, blky: %u, tidx: %u, gather_offset: %d, cal_offset: %d\n",
+                   blockIdx.x,
+                   blockIdx.y,
+                   threadIdx.x,
+                   gather_offset,
+                   src_coord_.GetOffset() / PackedSize);
+#endif
+
             using src_vector_type = vector_type_maker_t<SrcData, SrcScalarPerVector>;
             using src_vector_t    = typename src_vector_type::type;
 
@@ -276,6 +287,19 @@ struct ThreadwiseTensorSliceTransfer_v3r1_gather
                 .template SetAsType<dst_vector_t>(src_data_idx_seq,
                                                   op_r_v.template AsType<dst_vector_t>()[I0]);
 
+#if 0
+            auto data_print = src_thread_scratch_tuple_(thread_scratch_id).data_;
+            printf("blkx: %u, blky: %u, tidx: %u, src_thread_scratch_tuple_=<%02x, %02x, %02x, "
+                   "%02x>\n",
+                   blockIdx.x,
+                   blockIdx.y,
+                   threadIdx.x,
+                   *reinterpret_cast<const uint8_t*>(&data_print[Number<0>{}]),
+                   *reinterpret_cast<const uint8_t*>(&data_print[Number<1>{}]),
+                   *reinterpret_cast<const uint8_t*>(&data_print[Number<2>{}]),
+                   *reinterpret_cast<const uint8_t*>(&data_print[Number<3>{}]));
+#endif
+
             auto move_on_dim = [&]() constexpr
             {
                 StaticallyIndexedArray<bool, nDim> move_on_dim_;

include/ck/utility/dtype_vector.hpp

@@ -1036,11 +1036,11 @@ struct vector_type<T, 128, typename ck::enable_if_t<is_native_type<T>()>>
         StaticallyIndexedArray<d32_t, 4> d32x4_;
         StaticallyIndexedArray<d64_t, 2> d64x2_;
         StaticallyIndexedArray<d128_t, 1> d128x1_;
-    } data_;
+    } data_ = {d128_t{0}};
 
-    __host__ __device__ constexpr vector_type() : data_{type{0}} {}
+    __attribute__((host)) __attribute__((device)) constexpr vector_type() {}
 
-    __host__ __device__ constexpr vector_type(type v) : data_{v} {}
+    __attribute__((host)) __attribute__((device)) constexpr vector_type(type v) { (void)v; }
 
     template <typename X>
     __host__ __device__ constexpr const auto& AsType() const
@@ -1164,11 +1164,11 @@ struct vector_type<T, 256, typename ck::enable_if_t<is_native_type<T>()>>
         StaticallyIndexedArray<d64_t, 4> d64x4_;
         StaticallyIndexedArray<d128_t, 2> d128x2_;
         StaticallyIndexedArray<d256_t, 1> d256x1_;
-    } data_;
+    } data_ = {d256_t{0}};
 
-    __host__ __device__ constexpr vector_type() : data_{type{0}} {}
+    __attribute__((host)) __attribute__((device)) constexpr vector_type() {}
 
-    __host__ __device__ constexpr vector_type(type v) : data_{v} {}
+    __attribute__((host)) __attribute__((device)) constexpr vector_type(type v) { (void)v; }
 
     template <typename X>
     __host__ __device__ constexpr const auto& AsType() const

library/include/ck/library/reference_tensor_operation/cpu/reference_moe_mx_gemm1.hpp

@@ -0,0 +1,264 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+#include <unordered_map>
+
+#include "ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace host {
+
+template <typename ADataType,
+          typename AScaleDataType,
+          typename BDataType,
+          typename BScaleDataType,
+          typename CDataType,
+          typename D0DataType, // expert weight
+          typename AccDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          index_t ActivationType_ = 0,
+          bool MulRoutedWeight    = true,
+          typename ComputeTypeA   = CDataType,
+          typename ComputeTypeB   = ComputeTypeA>
+struct ReferenceMoeMXGemm1 : public device::BaseOperator
+{
+    // Argument
+    static constexpr auto ActivationType = ActivationType_;
+    struct Argument : public device::BaseArgument
+    {
+        Argument(const Tensor<ck::index_t>& sorted_token_ids,
+                 const Tensor<ck::index_t>& expert_ids,
+                 const Tensor<ck::index_t>& max_token_id,
+                 const index_t sorted_tile_size,
+                 const Tensor<ADataType>& a_t_k,
+                 const Tensor<AScaleDataType>& a_t_k_scale,
+                 const Tensor<BDataType>& b_e_n_k,
+                 const Tensor<BScaleDataType>& b_e_n_k_scale,
+                 const Tensor<D0DataType>& d2,
+                 Tensor<CDataType>& c_t_k_n,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CElementwiseOperation c_element_op)
+            : sorted_token_ids_{sorted_token_ids},
+              expert_ids_{expert_ids},
+              max_token_id_{max_token_id},
+              sorted_tile_size_{sorted_tile_size},
+              a_t_k_{a_t_k},
+              a_t_k_scale_{a_t_k_scale},
+              b_e_n_k_{b_e_n_k},
+              b_e_n_k_scale_{b_e_n_k_scale},
+              d2_{d2},
+              c_t_k_n_{c_t_k_n},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              c_element_op_{c_element_op}
+        {
+        }
+
+        const Tensor<ck::index_t>& sorted_token_ids_;
+        const Tensor<ck::index_t>& expert_ids_;
+        const Tensor<ck::index_t>& max_token_id_;
+        index_t sorted_tile_size_;
+        const Tensor<ADataType>& a_t_k_;
+        const Tensor<AScaleDataType>& a_t_k_scale_;
+        const Tensor<BDataType>& b_e_n_k_;
+        const Tensor<BScaleDataType>& b_e_n_k_scale_;
+        const Tensor<D0DataType>& d2_;
+        Tensor<CDataType>& c_t_k_n_;
+
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CElementwiseOperation c_element_op_;
+    };
+
+    // Invoker
+    struct Invoker : public device::BaseInvoker
+    {
+        using Argument = ReferenceMoeMXGemm1::Argument;
+
+        float Run(const Argument& arg)
+        {
+            static_assert(ActivationType < 2, "Not supported activation type");
+            const int full_n = arg.c_t_k_n_.mDesc.GetLengths()[2];
+            arg.c_t_k_n_.SetZero();
+            auto f_mk_kn_mn = [&](auto m, auto n) {
+                const int K                   = arg.a_t_k_.mDesc.GetLengths()[1];
+                const ck::index_t SCALE_BLOCK = K / arg.b_e_n_k_scale_.mDesc.GetLengths()[1];
+                AccDataType v_acc{0};
+                AccDataType v_acc_up{0};
+                ComputeTypeA v_a{0};
+                ComputeTypeB v_b{0};
+                ComputeTypeB v_b_up{0};
+                const int t         = arg.sorted_token_ids_(m) & 0xffffff;
+                const int topk_id   = arg.sorted_token_ids_(m) >> 24;
+                const int e         = arg.expert_ids_(m / arg.sorted_tile_size_);
+                const int token_cnt = arg.c_t_k_n_.mDesc.GetLengths()[0];
+                D0DataType v_topk_w = arg.d2_(m, 0); // expert
+
+                if(t < token_cnt)
+                {
+                    for(int k = 0; k < K; ++k)
+                    {
+                        auto a_f4x2  = arg.a_t_k_(t, k).data;
+                        auto a_scale = arg.a_t_k_scale_(t, k / SCALE_BLOCK);
+                        if constexpr(is_same_v<ADataType, f4x2_pk_t>)
+                        {
+
+                            f4_t f4 = 0;
+                            if(k % 2 == 1)
+                                f4 = (a_f4x2 >> 0) & 0xf;
+                            else
+                                f4 = (a_f4x2 >> 4) & 0xf;
+                            v_a = type_convert<ComputeTypeA>(f4) *
+                                  type_convert<ComputeTypeA>(a_scale);
+                        }
+                        else
+                        {
+                            v_a = type_convert<ComputeTypeA>(a_f4x2) *
+                                  type_convert<ComputeTypeA>(a_scale);
+                            arg.a_element_op_(v_a, v_a);
+                        }
+                        auto b_f4x2     = arg.b_e_n_k_(e, k, n).data;
+                        auto b_f4x2_up  = arg.b_e_n_k_(e, k, n + full_n).data;
+                        auto b_scale    = arg.b_e_n_k_scale_(e, k / SCALE_BLOCK, n);
+                        auto b_scale_up = arg.b_e_n_k_scale_(e, k / SCALE_BLOCK, n + full_n);
+                        if constexpr(is_same_v<BDataType, f4x2_pk_t>)
+                        {
+
+                            f4_t f4    = 0;
+                            f4_t f4_up = 0;
+                            if(k % 2 == 1)
+                            {
+                                f4    = (b_f4x2 >> 0) & 0xf;
+                                f4_up = (b_f4x2_up >> 0) & 0xf;
+                            }
+                            else
+                            {
+                                f4    = (b_f4x2 >> 4) & 0xf;
+                                f4_up = (b_f4x2_up >> 4) & 0xf;
+                            }
+                            v_b = type_convert<ComputeTypeB>(f4) *
+                                  type_convert<ComputeTypeB>(b_scale);
+                            v_b_up = type_convert<ComputeTypeB>(f4_up) *
+                                     type_convert<ComputeTypeB>(b_scale_up);
+                        }
+                        else
+                        {
+                            v_b = type_convert<ComputeTypeB>(b_f4x2) *
+                                  type_convert<ComputeTypeB>(b_scale);
+                            v_b_up = type_convert<ComputeTypeB>(b_f4x2_up) *
+                                     type_convert<ComputeTypeB>(b_scale_up);
+                            arg.b_element_op_(v_b, v_b);
+                            arg.b_element_op_(v_b_up, v_b_up);
+                        }
+
+                        v_acc +=
+                            ck::type_convert<AccDataType>(v_a) * ck::type_convert<AccDataType>(v_b);
+                        v_acc_up += ck::type_convert<AccDataType>(v_a) *
+                                    ck::type_convert<AccDataType>(v_b_up);
+                    }
+                    CDataType v_c{0};
+                    CDataType v_c_up{0};
+                    if constexpr(MulRoutedWeight)
+                    {
+                        v_acc *= v_topk_w;
+                        v_acc_up *= v_topk_w;
+                    }
+                    arg.c_element_op_(v_c, v_acc);
+                    arg.c_element_op_(v_c_up, v_acc_up);
+                    if constexpr(ActivationType == 1)
+                    {
+                        tensor_operation::element_wise::Silu{}(v_c, v_c);
+                        arg.c_t_k_n_(t, topk_id, n) = v_c * v_c_up;
+                    }
+                    else if constexpr(ActivationType == 0)
+                    {
+                        tensor_operation::element_wise::Gelu{}(v_c, v_c);
+                        arg.c_t_k_n_(t, topk_id, n) = v_c * v_c_up;
+                    }
+                }
+            };
+
+            const ck::index_t max_token_id = arg.max_token_id_(0);
+            make_ParallelTensorFunctor(f_mk_kn_mn, max_token_id, full_n)(
+                std::thread::hardware_concurrency());
+
+            return 0;
+        }
+
+        float Run(const device::BaseArgument* p_arg,
+                  const StreamConfig& /* stream_config */ = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg));
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    bool IsSupportedArgument(const device::BaseArgument*) override { return true; }
+
+    static auto MakeArgument(const Tensor<ck::index_t>& sorted_token_ids,
+                             const Tensor<ck::index_t>& expert_ids,
+                             const Tensor<ck::index_t>& max_token_id,
+                             const index_t sorted_tile_size,
+                             const Tensor<ADataType>& a_t_k,
+                             const Tensor<AScaleDataType>& a_t_k_scale,
+                             const Tensor<BDataType>& b_e_n_k,
+                             const Tensor<BScaleDataType>& b_e_n_k_scale,
+                             const Tensor<D0DataType>& d2,
+                             Tensor<CDataType>& c_t_k_n,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             CElementwiseOperation c_element_op)
+    {
+        return Argument{sorted_token_ids,
+                        expert_ids,
+                        max_token_id,
+                        sorted_tile_size,
+                        a_t_k,
+                        a_t_k_scale,
+                        b_e_n_k,
+                        b_e_n_k_scale,
+                        d2,
+                        c_t_k_n,
+                        a_element_op,
+                        b_element_op,
+                        c_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    virtual std::unique_ptr<device::BaseInvoker> MakeInvokerPointer()
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "ReferenceMoeMxGemm1"
+            << std::endl;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace host
+} // namespace tensor_operation
+} // namespace ck

library/include/ck/library/reference_tensor_operation/cpu/reference_moe_mx_gemm2.hpp

@@ -0,0 +1,233 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+#include <unordered_map>
+
+#include "ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace host {
+
+template <typename ADataType,
+          typename AScaleDataType,
+          typename BDataType,
+          typename BScaleDataType,
+          typename D0DataType, // expert weight
+          typename CDataType,
+          typename AccDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          bool MulRoutedWeight  = true,
+          typename ComputeTypeA = CDataType,
+          typename ComputeTypeB = ComputeTypeA>
+struct ReferenceMoeMXGemm2 : public device::BaseOperator
+{
+    // Argument
+    struct Argument : public device::BaseArgument
+    {
+        Argument(const Tensor<ck::index_t>& sorted_token_ids,
+                 const Tensor<ck::index_t>& expert_ids,
+                 const Tensor<ck::index_t>& max_token_id,
+                 const index_t sorted_tile_size,
+                 const Tensor<ADataType>& a_t_k_k,
+                 const Tensor<AScaleDataType>& a_t_k_k_scale,
+                 const Tensor<BDataType>& b_e_n_k,
+                 const Tensor<BScaleDataType>& b_e_n_k_scale,
+                 const Tensor<D0DataType>& d2,
+                 Tensor<CDataType>& c_t_n,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CElementwiseOperation c_element_op)
+            : sorted_token_ids_{sorted_token_ids},
+              expert_ids_{expert_ids},
+              max_token_id_{max_token_id},
+              sorted_tile_size_{sorted_tile_size},
+              a_t_k_k_{a_t_k_k},
+              a_t_k_k_scale_{a_t_k_k_scale},
+              b_e_n_k_{b_e_n_k},
+              b_e_n_k_scale_{b_e_n_k_scale},
+              d2_{d2},
+              c_t_n_{c_t_n},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              c_element_op_{c_element_op}
+        {
+        }
+
+        const Tensor<ck::index_t>& sorted_token_ids_;
+        const Tensor<ck::index_t>& expert_ids_;
+        const Tensor<ck::index_t>& max_token_id_;
+        index_t sorted_tile_size_;
+        const Tensor<ADataType>& a_t_k_k_;
+        const Tensor<AScaleDataType>& a_t_k_k_scale_;
+        const Tensor<BDataType>& b_e_n_k_;
+        const Tensor<BScaleDataType>& b_e_n_k_scale_;
+        const Tensor<D0DataType>& d2_;
+        Tensor<CDataType>& c_t_n_;
+
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CElementwiseOperation c_element_op_;
+    };
+
+    // Invoker
+    struct Invoker : public device::BaseInvoker
+    {
+        using Argument = ReferenceMoeMXGemm2::Argument;
+
+        float Run(const Argument& arg)
+        {
+            arg.c_t_n_.SetZero();
+            auto f_mk_kn_mn = [&](auto m, auto n) {
+                const int K                   = arg.a_t_k_k_.mDesc.GetLengths()[2];
+                const ck::index_t SCALE_BLOCK = K / arg.b_e_n_k_scale_.mDesc.GetLengths()[1];
+                AccDataType v_acc{0};
+                ComputeTypeA v_a{0};
+                ComputeTypeB v_b{0};
+                const int t         = arg.sorted_token_ids_(m) & 0xffffff;
+                const int topk_id   = arg.sorted_token_ids_(m) >> 24;
+                const int e         = arg.expert_ids_(m / arg.sorted_tile_size_);
+                const int token_cnt = arg.c_t_n_.mDesc.GetLengths()[0];
+                D0DataType v_topk_w = arg.d2_(m, 0); // expert
+
+                if(t < token_cnt)
+                {
+                    for(int k = 0; k < K; ++k)
+                    {
+                        if constexpr(is_same_v<ADataType, f4x2_pk_t>)
+                        {
+                            auto f4x2    = arg.a_t_k_k_(t, topk_id, k).data;
+                            auto a_scale = arg.a_t_k_k_scale_(t, topk_id, k / SCALE_BLOCK);
+
+                            f4_t f4 = 0;
+                            if(k % 2 == 1)
+                                f4 = (f4x2 >> 0) & 0xf;
+                            else
+                                f4 = (f4x2 >> 4) & 0xf;
+
+                            v_a = type_convert<ComputeTypeA>(f4) *
+                                  type_convert<ComputeTypeA>(a_scale);
+                        }
+                        else
+                        {
+                            arg.a_element_op_(
+                                v_a, type_convert<ComputeTypeA>(arg.a_t_k_k_(t, topk_id, k)));
+                        }
+                        if constexpr(is_same_v<BDataType, f4x2_pk_t>)
+                        {
+                            auto f4x2    = arg.b_e_n_k_(e, k, n).data;
+                            auto b_scale = arg.b_e_n_k_scale_(e, k / SCALE_BLOCK, n);
+
+                            f4_t f4 = 0;
+                            if(k % 2 == 1)
+                                f4 = (f4x2 >> 0) & 0xf;
+                            else
+                                f4 = (f4x2 >> 4) & 0xf;
+
+                            v_b = type_convert<ComputeTypeB>(f4) *
+                                  type_convert<ComputeTypeB>(b_scale);
+                        }
+                        else
+                        {
+                            arg.b_element_op_(v_b,
+                                              type_convert<ComputeTypeB>(arg.b_e_n_k_(e, k, n)));
+                        }
+
+                        v_acc +=
+                            ck::type_convert<AccDataType>(v_a) * ck::type_convert<AccDataType>(v_b);
+                    }
+                    CDataType v_c{0};
+                    if constexpr(MulRoutedWeight)
+                    {
+                        arg.c_element_op_(v_c, v_acc, 1.f, 1.f, v_topk_w); // hacky, need to fix
+                    }
+                    else
+                    {
+                        arg.c_element_op_(v_c, v_acc, 1.f, 1.f, 1.f);
+                    }
+                    arg.c_t_n_(t, n) += v_c;
+                }
+            };
+
+            const ck::index_t max_token_id = arg.max_token_id_(0);
+            make_ParallelTensorFunctor(f_mk_kn_mn, max_token_id, arg.c_t_n_.mDesc.GetLengths()[1])(
+                std::thread::hardware_concurrency());
+
+            return 0;
+        }
+
+        float Run(const device::BaseArgument* p_arg,
+                  const StreamConfig& /* stream_config */ = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg));
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    bool IsSupportedArgument(const device::BaseArgument*) override { return true; }
+
+    static auto MakeArgument(const Tensor<ck::index_t>& sorted_token_ids,
+                             const Tensor<ck::index_t>& expert_ids,
+                             const Tensor<ck::index_t>& max_token_id,
+                             const index_t sorted_tile_size,
+                             const Tensor<ADataType>& a_t_k_k,
+                             const Tensor<AScaleDataType>& a_t_k_k_scale,
+                             const Tensor<BDataType>& b_e_n_k,
+                             const Tensor<BScaleDataType>& b_e_n_k_scale,
+                             const Tensor<D0DataType>& d2,
+                             Tensor<CDataType>& c_t_n,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             CElementwiseOperation c_element_op)
+    {
+        return Argument{sorted_token_ids,
+                        expert_ids,
+                        max_token_id,
+                        sorted_tile_size,
+                        a_t_k_k,
+                        a_t_k_k_scale,
+                        b_e_n_k,
+                        b_e_n_k_scale,
+                        d2,
+                        c_t_n,
+                        a_element_op,
+                        b_element_op,
+                        c_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    virtual std::unique_ptr<device::BaseInvoker> MakeInvokerPointer()
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "ReferenceMoeGemm2"
+            << std::endl;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace host
+} // namespace tensor_operation
+} // namespace ck