Merge branch 'develop' of...

Merge branch 'develop' of https://github.com/ROCmSoftwarePlatform/composable_kernel into navi3x_mD_batchedGEMM_GroupConvFwd

client_example/01_gemm/gemm.cpp

@@ -83,7 +83,7 @@ int main(int argc, char* argv[])
         [](std::size_t nRow, std::size_t nCol, std::size_t stride, auto layout) {
             using Layout = decltype(layout);
 
-            if(std::is_same<Layout, ck::tensor_layout::gemm::RowMajor>::value)
+            if constexpr(std::is_same<Layout, ck::tensor_layout::gemm::RowMajor>::value)
             {
                 return (nRow - 1) * stride + nCol;
             }

client_example/02_gemm_add_add_fastgelu/gemm_add_add_fastgelu.cpp

@@ -92,7 +92,7 @@ int main(int argc, char* argv[])
         [](std::size_t nRow, std::size_t nCol, std::size_t stride, auto layout) {
             using Layout = decltype(layout);
 
-            if(std::is_same<Layout, ck::tensor_layout::gemm::RowMajor>::value)
+            if constexpr(std::is_same<Layout, ck::tensor_layout::gemm::RowMajor>::value)
             {
                 return (nRow - 1) * stride + nCol;
             }

client_example/02_gemm_add_add_fastgelu/gemm_add_fastgelu.cpp

@@ -88,7 +88,7 @@ int main(int argc, char* argv[])
         [](std::size_t nRow, std::size_t nCol, std::size_t stride, auto layout) {
             using Layout = decltype(layout);
 
-            if(std::is_same<Layout, ck::tensor_layout::gemm::RowMajor>::value)
+            if constexpr(std::is_same<Layout, ck::tensor_layout::gemm::RowMajor>::value)
             {
                 return (nRow - 1) * stride + nCol;
             }

client_example/02_gemm_add_add_fastgelu/gemm_fastgelu.cpp

@@ -84,7 +84,7 @@ int main(int argc, char* argv[])
         [](std::size_t nRow, std::size_t nCol, std::size_t stride, auto layout) {
             using Layout = decltype(layout);
 
-            if(std::is_same<Layout, ck::tensor_layout::gemm::RowMajor>::value)
+            if constexpr(std::is_same<Layout, ck::tensor_layout::gemm::RowMajor>::value)
             {
                 return (nRow - 1) * stride + nCol;
             }

client_example/03_gemm_layernorm/CMakeLists.txt

-add_executable(client_gemm_add_add_reduce_normalize gemm_add_add_layernorm.cpp)
-target_link_libraries(client_gemm_add_add_reduce_normalize PRIVATE composable_kernel::device_operations)
+add_executable(client_gemm_add_add_layernorm_naive gemm_add_add_layernorm_naive.cpp)
+target_link_libraries(client_gemm_add_add_layernorm_naive PRIVATE composable_kernel::device_operations)
+
+add_executable(client_gemm_add_relu_add_layernorm_welford gemm_add_relu_add_layernorm_welford.cpp)
+target_link_libraries(client_gemm_add_relu_add_layernorm_welford PRIVATE composable_kernel::device_operations)

client_example/03_gemm_layernorm/gemm_add_add_layernorm.cpp → client_example/03_gemm_layernorm/gemm_add_add_layernorm_naive.cpp

@@ -190,7 +190,7 @@ int main()
         [](std::size_t nRow, std::size_t nCol, std::size_t stride, auto layout) {
             using Layout = decltype(layout);
 
-            if(std::is_same<Layout, ck::tensor_layout::gemm::RowMajor>::value)
+            if constexpr(std::is_same<Layout, ck::tensor_layout::gemm::RowMajor>::value)
             {
                 return (nRow - 1) * stride + nCol;
             }

client_example/03_gemm_layernorm/gemm_add_relu_add_layernorm_welford.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <iostream>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d_layernorm.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+using F16 = ck::half_t;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using AddReluAdd  = ck::tensor_operation::element_wise::AddReluAdd;
+
+// DataType
+using ADataType     = F16;
+using BDataType     = F16;
+using D0DataType    = F16;
+using D1DataType    = F16;
+using GammaDataType = F16;
+using BetaDataType  = F16;
+using HDataType     = F16;
+
+// Layout
+using ALayout  = Row;
+using BLayout  = Col;
+using D0Layout = Row;
+using D1Layout = Row;
+using HLayout  = Row;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = AddReluAdd;
+using HElementOp   = PassThrough;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}, mMemSize_(mem_size)
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    void SetZero() const { (void)hipMemset(p_mem_, 0, mMemSize_); }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+    std::size_t mMemSize_;
+};
+
+int main(int argc, char* argv[])
+{
+    // GEMM shape
+    ck::index_t M = 1024;
+    ck::index_t N = 1024;
+    ck::index_t K = 1024;
+
+    ck::index_t StrideA  = K;
+    ck::index_t StrideB  = K;
+    ck::index_t StrideD0 = 0;
+    ck::index_t StrideD1 = N;
+    ck::index_t StrideH  = N;
+
+    float epsilon = 1e-5;
+
+    auto f_matrix_space_size =
+        [](std::size_t nRow, std::size_t nCol, std::size_t stride, auto layout) {
+            using Layout = decltype(layout);
+
+            if constexpr(std::is_same<Layout, ck::tensor_layout::gemm::RowMajor>::value)
+            {
+                return (nRow - 1) * stride + nCol;
+            }
+            else
+            {
+                return (nCol - 1) * stride + nRow;
+            }
+        };
+
+    SimpleDeviceMem a_device_buf(sizeof(ADataType) * f_matrix_space_size(M, K, StrideA, ALayout{}));
+    SimpleDeviceMem b_device_buf(sizeof(BDataType) * f_matrix_space_size(K, N, StrideB, BLayout{}));
+    SimpleDeviceMem d0_device_buf(sizeof(D0DataType) *
+                                  f_matrix_space_size(M, N, StrideD0, D0Layout{}));
+    SimpleDeviceMem d1_device_buf(sizeof(D1DataType) *
+                                  f_matrix_space_size(M, N, StrideD1, D1Layout{}));
+    SimpleDeviceMem gamma_device_buf(sizeof(GammaDataType) * N);
+    SimpleDeviceMem beta_device_buf(sizeof(BetaDataType) * N);
+    SimpleDeviceMem h_device_buf(sizeof(HDataType) * f_matrix_space_size(M, N, StrideH, HLayout{}));
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleDLayernorm<
+        ALayout,
+        BLayout,
+        ck::Tuple<D0Layout, D1Layout>,
+        HLayout,
+        ADataType,
+        BDataType,
+        ck::Tuple<D0DataType, D1DataType>,
+        GammaDataType,
+        BetaDataType,
+        HDataType,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::AddReluAdd,
+        ck::tensor_operation::element_wise::PassThrough>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    const auto a_element_op   = AElementOp{};
+    const auto b_element_op   = BElementOp{};
+    const auto cde_element_op = CDEElementOp{};
+    const auto h_element_op   = HElementOp{};
+
+    std::string best_op_name;
+    bool found            = false;
+    int best_op_id        = -1;
+    float best_ave_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr = op_ptrs[i];
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            a_device_buf.GetDeviceBuffer(),
+            b_device_buf.GetDeviceBuffer(),
+            {d0_device_buf.GetDeviceBuffer(), d1_device_buf.GetDeviceBuffer()},
+            gamma_device_buf.GetDeviceBuffer(),
+            beta_device_buf.GetDeviceBuffer(),
+            h_device_buf.GetDeviceBuffer(),
+            M,
+            N,
+            K,
+            StrideA,
+            StrideB,
+            {StrideD0, StrideD1},
+            StrideH,
+            epsilon,
+            a_element_op,
+            b_element_op,
+            cde_element_op,
+            h_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
+            SimpleDeviceMem workspace_dev(workspace_sz);
+            op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace_dev.GetDeviceBuffer());
+            h_device_buf.SetZero();
+
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t num_byte =
+                sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
+                (sizeof(D0DataType) + sizeof(D1DataType) + sizeof(HDataType)) * M * N +
+                (sizeof(GammaDataType) + sizeof(BetaDataType)) * N;
+
+            float gb_per_sec = num_byte / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << gb_per_sec << " GB/s, "
+                      << op_name << std::endl;
+
+            if(ave_time < best_ave_time)
+            {
+                found           = true;
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_gb_per_sec << " GB/s, "
+              << best_op_name << std::endl;
+
+    // run the best intance
+    {
+        auto& op_ptr = op_ptrs[best_op_id];
+
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            a_device_buf.GetDeviceBuffer(),
+            b_device_buf.GetDeviceBuffer(),
+            {d0_device_buf.GetDeviceBuffer(), d1_device_buf.GetDeviceBuffer()},
+            gamma_device_buf.GetDeviceBuffer(),
+            beta_device_buf.GetDeviceBuffer(),
+            h_device_buf.GetDeviceBuffer(),
+            M,
+            N,
+            K,
+            StrideA,
+            StrideB,
+            {StrideD0, StrideD1},
+            StrideH,
+            epsilon,
+            a_element_op,
+            b_element_op,
+            cde_element_op,
+            h_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
+            SimpleDeviceMem workspace_dev(workspace_sz);
+            op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace_dev.GetDeviceBuffer());
+            h_device_buf.SetZero();
+
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}
\ No newline at end of file

client_example/15_gemm_add_multiply/gemm_add_multiply.cpp

@@ -92,7 +92,7 @@ int main(int argc, char* argv[])
         [](std::size_t nRow, std::size_t nCol, std::size_t stride, auto layout) {
             using Layout = decltype(layout);
 
-            if(std::is_same<Layout, ck::tensor_layout::gemm::RowMajor>::value)
+            if constexpr(std::is_same<Layout, ck::tensor_layout::gemm::RowMajor>::value)
             {
                 return (nRow - 1) * stride + nCol;
             }

example/21_gemm_layernorm/gemm_bias_relu_add_layernorm_xdl_naive_fp16.cpp

@@ -4,7 +4,6 @@
 #include <iostream>
 #include <numeric>
 #include <initializer_list>
-#include <cstdlib>
 
 #include "ck/ck.hpp"
 #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
@@ -116,7 +115,7 @@ auto f_host_tensor_descriptor2d =
     [](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)
+        if constexpr(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
         {
             return HostTensorDescriptor({row, col}, {stride, 1_uz});
         }

example/21_gemm_layernorm/gemm_bias_relu_add_layernorm_xdl_welford_fp16.cpp

@@ -4,7 +4,6 @@
 #include <iostream>
 #include <numeric>
 #include <initializer_list>
-#include <cstdlib>
 
 #include "ck/ck.hpp"
 #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
@@ -15,6 +14,7 @@
 #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_gemm.hpp"
 #include "ck/library/reference_tensor_operation/cpu/reference_layernorm.hpp"
 #include "ck/library/utility/check_err.hpp"
@@ -69,21 +69,20 @@ using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleDLayern
 // clang-format on
 
 auto f_host_tensor_descriptor1d = [](std::size_t len, std::size_t stride) {
-    return HostTensorDescriptor(std::vector<std::size_t>({len}),
-                                std::vector<std::size_t>({stride}));
+    return HostTensorDescriptor({len}, {stride});
 };
 
 auto f_host_tensor_descriptor2d =
     [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
-        if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
+        using namespace ck::literals;
+
+        if constexpr(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
         {
-            return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
-                                        std::vector<std::size_t>({stride, 1}));
+            return HostTensorDescriptor({row, col}, {stride, 1_uz});
         }
         else
         {
-            return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
-                                        std::vector<std::size_t>({1, stride}));
+            return HostTensorDescriptor({row, col}, {1_uz, stride});
         }
     };
 
@@ -97,6 +96,7 @@ void host_gemm_layernorm(Tensor<HDataType>& h_m_n,
                          AElementOp a_element_op,
                          BElementOp b_element_op,
                          CDEElementOp cde_element_op,
+                         HElementOp h_element_op,
                          int M,
                          int N,
                          AccDataType epsilon = 1e-5)
@@ -145,7 +145,7 @@ void host_gemm_layernorm(Tensor<HDataType>& h_m_n,
     auto ref_layernorm_invoker = ref_layernorm.MakeInvoker();
 
     auto ref_layernorm_argument = ref_layernorm.MakeArgument(
-        e_m_n, gamma_n, beta_n, h_m_n, HElementOp{}, {M, N}, {1}, epsilon);
+        e_m_n, gamma_n, beta_n, h_m_n, h_element_op, {M, N}, {1}, epsilon);
     ref_layernorm_invoker.Run(ref_layernorm_argument);
 }
 
@@ -249,6 +249,7 @@ int main()
                             a_element_op,
                             b_element_op,
                             cde_element_op,
+                            h_element_op,
                             M,
                             N,
                             epsilon);

example/21_gemm_layernorm/gemm_layernorm_xdl_naive_fp16.cpp

@@ -4,7 +4,6 @@
 #include <iostream>
 #include <numeric>
 #include <initializer_list>
-#include <cstdlib>
 
 #include "ck/ck.hpp"
 #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
@@ -115,7 +114,7 @@ auto f_host_tensor_descriptor2d =
     [](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)
+        if constexpr(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
         {
             return HostTensorDescriptor({row, col}, {stride, 1_uz});
         }

example/21_gemm_layernorm/gemm_xdl_layernorm_naive_single_kernel_fp16.cpp

@@ -135,7 +135,7 @@ int main(int argc, char* argv[])
         [](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)
+            if constexpr(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
             {
                 return HostTensorDescriptor({row, col}, {stride, 1_uz});
             }

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

@@ -533,6 +533,11 @@ struct DeviceElementwiseNormalizationImpl
                 return (false);
         }
 
+        if(p_arg_->x_lds_size_ >= 65536)
+        {
+            return (false);
+        }
+
         return true;
     };
 

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

@@ -669,6 +669,9 @@ struct DeviceGemmMultipleDLayernorm_Xdl_CShuffle
             {
                 throw std::runtime_error("wrong! GridwiseGemmWelford has invalid setting");
             }
+            if(arg.p_workspace_e_grid_ == nullptr || arg.p_workspace_mean_ == nullptr ||
+               arg.p_workspace_var_ == nullptr || arg.p_workspace_count_ == nullptr)
+                throw std::runtime_error("wrong! WorkSpace pointer has not been set");
 
             index_t grid_size = arg.block_2_etile_map_.CalculateGridSize(arg.gemm_e_grid_desc_m_n_);
 
@@ -939,7 +942,11 @@ struct DeviceGemmMultipleDLayernorm_Xdl_CShuffle
             }
         }
 
-        return true;
+        return GridwiseGemmWelford::CheckValidity(arg.a_grid_desc_m_k_,
+                                                  arg.b_grid_desc_n_k_,
+                                                  arg.ds_grid_desc_m_n_,
+                                                  arg.gemm_e_grid_desc_m_n_,
+                                                  arg.block_2_etile_map_);
     }
 
     // polymorphic
@@ -1055,7 +1062,12 @@ struct DeviceGemmMultipleDLayernorm_Xdl_CShuffle
             << GemmKPerBlock << ", "
             << AK1 << ", "
             << BK1 << ", "
-            << getGemmSpecializationString(GemmSpec)
+            << getGemmSpecializationString(GemmSpec) << ", "
+            << PostShuffleThreadClusterSize_M_N::At(I0) << ", "
+            << PostShuffleThreadClusterSize_M_N::At(I1) << ", "
+            << LayernormThreadClusterSize_M_N::At(I0) << ", "
+            << LayernormThreadClusterSize_M_N::At(I1) << ", "
+            << LayernormThreadSliceSize_M
             << ">"
             << " LoopScheduler: "
             << LoopSchedToString[LoopSched] << ", "

include/ck/utility/math_v2.hpp

@@ -158,9 +158,9 @@ static inline __device__ bool isnan(half_t x)
     return (xx & 0x7FFF) > 0x7C00;
 };
 
-static inline __device__ float sqrt(float x) { return ::sqrtf(x); };
+static inline __device__ float sqrt(float x) { return __builtin_amdgcn_sqrtf(x); };
 
-static inline __device__ double sqrt(double x) { return ::sqrt(x); };
+static inline __device__ double sqrt(double x) { return __builtin_amdgcn_sqrt(x); };
 
 } // namespace math
 } // namespace ck

library/include/ck/library/tensor_operation_instance/device_operation_instance_factory.hpp

@@ -89,6 +89,7 @@ using Scale          = ck::tensor_operation::element_wise::Scale;
 using Bilinear       = ck::tensor_operation::element_wise::Bilinear;
 using AddAddFastGelu = ck::tensor_operation::element_wise::AddAddFastGelu;
 using AddFastGelu    = ck::tensor_operation::element_wise::AddFastGelu;
+using AddReluAdd     = ck::tensor_operation::element_wise::AddReluAdd;
 using FastGelu       = ck::tensor_operation::element_wise::FastGelu;
 using AddMultiply    = ck::tensor_operation::element_wise::AddMultiply;
 using ScaleAdd       = ck::tensor_operation::element_wise::ScaleAdd;

library/include/ck/library/tensor_operation_instance/gpu/gemm_add_add_fastgelu.hpp

@@ -10,7 +10,6 @@
 #include "ck/ck.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/element/element_wise_operation.hpp"
 
 #include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
 

library/include/ck/library/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+#include <memory>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d_layernorm.hpp"
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_kn_mn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleDLayernorm<Row,
+                                                             Row,
+                                                             Row_Row_Tuple,
+                                                             Row,
+                                                             F16,
+                                                             F16,
+                                                             F16_F16_Tuple,
+                                                             F16,
+                                                             F16,
+                                                             F16,
+                                                             PassThrough,
+                                                             PassThrough,
+                                                             AddReluAdd,
+                                                             PassThrough>>>&);
+
+void add_device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_nk_mn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleDLayernorm<Row,
+                                                             Col,
+                                                             Row_Row_Tuple,
+                                                             Row,
+                                                             F16,
+                                                             F16,
+                                                             F16_F16_Tuple,
+                                                             F16,
+                                                             F16,
+                                                             F16,
+                                                             PassThrough,
+                                                             PassThrough,
+                                                             AddReluAdd,
+                                                             PassThrough>>>&);
+
+void add_device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_kn_mn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleDLayernorm<Col,
+                                                             Row,
+                                                             Row_Row_Tuple,
+                                                             Row,
+                                                             F16,
+                                                             F16,
+                                                             F16_F16_Tuple,
+                                                             F16,
+                                                             F16,
+                                                             F16,
+                                                             PassThrough,
+                                                             PassThrough,
+                                                             AddReluAdd,
+                                                             PassThrough>>>&);
+
+void add_device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_nk_mn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleDLayernorm<Col,
+                                                             Col,
+                                                             Row_Row_Tuple,
+                                                             Row,
+                                                             F16,
+                                                             F16,
+                                                             F16_F16_Tuple,
+                                                             F16,
+                                                             F16,
+                                                             F16,
+                                                             PassThrough,
+                                                             PassThrough,
+                                                             AddReluAdd,
+                                                             PassThrough>>>&);
+
+// GEMM + Add + Relu + Add + Layernorm
+template <typename ALayout,
+          typename BLayout,
+          typename D0Layout,
+          typename D1Layout,
+          typename HLayout,
+          typename ADataType,
+          typename BDataType,
+          typename D0DataType,
+          typename D1DataType,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename HDataType>
+struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGemmMultipleDLayernorm<
+    ALayout,
+    BLayout,
+    ck::Tuple<D0Layout, D1Layout>,
+    HLayout,
+    ADataType,
+    BDataType,
+    ck::Tuple<D0DataType, D1DataType>,
+    GammaDataType,
+    BetaDataType,
+    HDataType,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::AddReluAdd,
+    ck::tensor_operation::element_wise::PassThrough>>
+{
+    using DeviceOp = DeviceGemmMultipleDLayernorm<ALayout,
+                                                  BLayout,
+                                                  ck::Tuple<D0Layout, D1Layout>,
+                                                  HLayout,
+                                                  ADataType,
+                                                  BDataType,
+                                                  ck::Tuple<D0DataType, D1DataType>,
+                                                  GammaDataType,
+                                                  BetaDataType,
+                                                  HDataType,
+                                                  ck::tensor_operation::element_wise::PassThrough,
+                                                  ck::tensor_operation::element_wise::PassThrough,
+                                                  ck::tensor_operation::element_wise::AddReluAdd,
+                                                  ck::tensor_operation::element_wise::PassThrough>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
+                     is_same_v<D0DataType, half_t> && is_same_v<D1DataType, half_t> &&
+                     is_same_v<GammaDataType, half_t> && is_same_v<BetaDataType, half_t> &&
+                     is_same_v<HDataType, half_t>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
+                         is_same_v<D0Layout, Row> && is_same_v<D1Layout, Row> &&
+                         is_same_v<HLayout, Row>)
+            {
+                add_device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_kn_mn_mn_mn_instances(
+                    op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
+                              is_same_v<D0Layout, Row> && is_same_v<D1Layout, Row> &&
+                              is_same_v<HLayout, Row>)
+            {
+                add_device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_nk_mn_mn_mn_instances(
+                    op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
+                              is_same_v<D0Layout, Row> && is_same_v<D1Layout, Row> &&
+                              is_same_v<HLayout, Row>)
+            {
+                add_device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_kn_mn_mn_mn_instances(
+                    op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> &&
+                              is_same_v<D0Layout, Row> && is_same_v<D1Layout, Row> &&
+                              is_same_v<HLayout, Row>)
+            {
+                add_device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_nk_mn_mn_mn_instances(
+                    op_ptrs);
+            }
+        }
+
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/elementwise_normalization/device_elementwise_normalization_f16_instance.cpp

@@ -23,6 +23,11 @@ template <typename XElementwise, typename YElementwise, index_t Rank, index_t Re
 using device_elementwise_normalization_f16_instances =
     std::tuple <
         // XDataType, GammaDataType, BetaDataType, AccDataType, YDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize>
+        DeviceElementwiseNormalizationImpl<ck::Tuple<F16, F16>, F16, F16, F32, F16, XElementwise ,YElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 1, 1, 1, 1, 1, 1>, // fallback kernel for large N
+        DeviceElementwiseNormalizationImpl<ck::Tuple<F16, F16>, F16, F16, F32, F16, XElementwise ,YElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 2, 1, 2, 1, 2, 2>, // fallback kernel for large N
+        DeviceElementwiseNormalizationImpl<ck::Tuple<F16, F16>, F16, F16, F32, F16, XElementwise ,YElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 8, 1, 8, 1, 8, 8>, // fallback kernel for large N
+        DeviceElementwiseNormalizationImpl<ck::Tuple<F16, F16>, F16, F16, F32, F16, XElementwise ,YElementwise, Rank, Reduce, 256, 2, 128, 1, 8, 1, 8, 1, 8, 1, 8, 8>, // fallback kernel for large N
+        DeviceElementwiseNormalizationImpl<ck::Tuple<F16, F16>, F16, F16, F32, F16, XElementwise ,YElementwise, Rank, Reduce, 256, 4, 64, 1, 8, 1, 1, 1, 1, 1, 1, 1>, // fallback kernel for large N
         DeviceElementwiseNormalizationImpl<ck::Tuple<F16, F16>, F16, F16, F32, F16, XElementwise ,YElementwise, Rank, Reduce, 256, 8, 32, 1, 8, 1, 1, 1, 1, 1, 1, 1>, // fallback kernel
         DeviceElementwiseNormalizationImpl<ck::Tuple<F16, F16>, F16, F16, F32, F16, XElementwise ,YElementwise, Rank, Reduce, 256, 8, 32, 1, 8, 1, 2, 1, 2, 1, 2, 2>, // fallback kernel
         DeviceElementwiseNormalizationImpl<ck::Tuple<F16, F16>, F16, F16, F32, F16, XElementwise ,YElementwise, Rank, Reduce, 256, 8, 32, 1, 8, 1, 4, 1, 4, 1, 4, 4>, // fallback kernel

library/src/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm/CMakeLists.txt

+add_instance_library(device_gemm_add_relu_add_layernorm_instance
+   device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_kn_mn_mn_mn_instance.cpp
+   device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_nk_mn_mn_mn_instance.cpp
+   device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_kn_mn_mn_mn_instance.cpp
+   device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_nk_mn_mn_mn_instance.cpp
+)