merge develop

example/11_conv2d_bwd_weight/conv2d_bwd_weight_xdl.cpp

@@ -82,9 +82,9 @@ using ReferenceConvBwdWeightInstance =
 
 int main(int argc, char* argv[])
 {
-    bool do_verification = 0;
-    int init_method      = 0;
-    int nrepeat          = 5;
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
     int do_log           = 0;
     int split_k          = 4;
 
@@ -109,7 +109,7 @@ int main(int argc, char* argv[])
     {
         do_verification = std::stoi(argv[1]);
         init_method     = std::stoi(argv[2]);
-        nrepeat         = std::stoi(argv[3]);
+        time_kernel     = std::stoi(argv[3]);
         do_log          = std::stoi(argv[4]);
         split_k         = std::stoi(argv[5]);
     }
@@ -117,7 +117,7 @@ int main(int argc, char* argv[])
     {
         do_verification = std::stoi(argv[1]);
         init_method     = std::stoi(argv[2]);
-        nrepeat         = std::stoi(argv[3]);
+        time_kernel     = std::stoi(argv[3]);
         do_log          = std::stoi(argv[4]);
         split_k         = std::stoi(argv[5]);
 
@@ -141,7 +141,7 @@ int main(int argc, char* argv[])
     {
         printf("arg1: verification (0=no, 1=yes)\n");
         printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
-        printf("arg3: run kernel # of times (>1)\n");
+        printf("arg3: time kernel (0=n0, 1=yes)\n");
         printf("arg4: is show log (0=no, 1=yes)\n");
         printf("arg5: split-k \n");
         printf("arg6 to 19: N, K, C, Y, X, Hi, Wi, Sy, Sx, Dy, Dx, LeftPy, LeftPx, RightPy, "
@@ -246,7 +246,7 @@ int main(int argc, char* argv[])
         return 1;
     }
 
-    float ave_time = invoker.Run(argument, nrepeat);
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
 
     std::size_t flop = std::size_t(2) * N * K * Ho * Wo * C * Y * X;
 
@@ -291,6 +291,9 @@ int main(int argc, char* argv[])
             LogRangeAsType<float>(std::cout << "wei_host  : ", wei_k_c_y_x_host_result.mData, ",")
                 << std::endl;
         }
-        ck::utils::check_err(wei_k_c_y_x_device_result.mData, wei_k_c_y_x_host_result.mData);
+        return ck::utils::check_err(wei_k_c_y_x_device_result.mData, wei_k_c_y_x_host_result.mData)
+                   ? 0
+                   : 1;
     }
+    return 0;
 }

example/12_reduce/CMakeLists.txt

-add_example_executable(example_reduce_blockwise reduce_blockwise.cpp)
+add_example_executable(example_reduce_blockwise reduce_blockwise.cpp -D 16,64,32,960 -v 1 1 10)

example/12_reduce/reduce_blockwise.cpp

@@ -116,10 +116,9 @@ class SimpleAppArgs
     std::vector<size_t> inLengths;
     std::vector<float> scales;
 
-    bool do_verification = false;
-
+    bool do_verification = true;
     int init_method      = 1;
-    int nrepeat     = 5;
+    bool time_kernel     = false;
 
     public:
     void show_usage(const char* cmd)
@@ -135,7 +134,7 @@ class SimpleAppArgs
         std::cout << "Arg1 -- init method (0=no init, 1=single integer value, 2=scope integer "
                      "value, 3=decimal value)"
                   << std::endl;
-        std::cout << "Arg2 -- number of repeats to run the kernel" << std::endl;
+        std::cout << "Arg2 -- time kernel (0=n0, 1=yes)" << std::endl;
     };
 
     int processArgs(int argc, char* argv[])
@@ -182,7 +181,7 @@ class SimpleAppArgs
             throw std::runtime_error("Invalid cmd-line arguments, more argumetns are needed!");
 
         init_method = std::atoi(argv[optind++]);
-        nrepeat     = std::atoi(argv[optind]);
+        time_kernel = std::atoi(argv[optind]);
 
         if(scales.empty())
         {
@@ -352,7 +351,7 @@ int main(int argc, char* argv[])
 
     auto invoker_ptr = reduce.MakeInvokerPointer();
 
-    float avg_time = invoker_ptr->Run(argument_ptr.get(), args.nrepeat);
+    float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, args.time_kernel});
 
     std::size_t num_bytes = invariant_total_length * reduce_total_length * sizeof(InDataType) +
                             invariant_total_length * sizeof(OutDataType);
@@ -362,16 +361,17 @@ int main(int argc, char* argv[])
     std::cout << "Perf: " << avg_time << " ms, " << gb_per_sec << " GB/s, " << reduce_name
               << std::endl;
 
+    bool pass = true;
     if(args.do_verification)
     {
         out_dev.FromDevice(out.mData.data());
-        ck::utils::check_err(out.mData, out_ref.mData);
+        pass &= ck::utils::check_err(out.mData, out_ref.mData);
 
         if(NeedIndices)
         {
             out_indices_dev.FromDevice(out_indices.mData.data());
-            ck::utils::check_err(out_indices.mData, out_indices_ref.mData);
-            ;
+            pass &= ck::utils::check_err(out_indices.mData, out_indices_ref.mData);
         };
     };
+    return pass ? 0 : 1;
 }

example/13_pool2d_fwd/pool2d_fwd.cpp

@@ -149,9 +149,9 @@ int main(int argc, char* argv[])
 {
     using namespace ck::host_reduce;
 
-    bool do_verification = 0;
-    int init_method      = 0;
-    int nrepeat          = 5;
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
 
     // Pool shape
     ck::index_t N               = 128;
@@ -171,13 +171,13 @@ int main(int argc, char* argv[])
     {
         do_verification = std::stoi(argv[1]);
         init_method     = std::stoi(argv[2]);
-        nrepeat         = std::stoi(argv[3]);
+        time_kernel     = std::stoi(argv[3]);
     }
     else if(argc == 16)
     {
         do_verification = std::stoi(argv[1]);
         init_method     = std::stoi(argv[2]);
-        nrepeat         = std::stoi(argv[3]);
+        time_kernel     = std::stoi(argv[3]);
 
         N               = std::stoi(argv[4]);
         C               = std::stoi(argv[5]);
@@ -196,7 +196,7 @@ int main(int argc, char* argv[])
     {
         printf("arg1: verification (0=no, 1=yes)\n");
         printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
-        printf("arg3: run kernel # of times (>1)\n");
+        printf("arg3: time kernel (0=n0, 1=yes)\n");
         printf("arg4 to 15: N, C, Y, X, Hi, Wi, Sy, Sx, LeftPy, LeftPx, RightPy, "
                "RightPx\n");
         exit(0);
@@ -271,7 +271,7 @@ int main(int argc, char* argv[])
                                  "not support this problem");
     }
 
-    float ave_time = invoker_ptr->Run(argument_ptr.get(), nrepeat);
+    float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
 
     std::size_t flop = std::size_t(2) * N * C * Ho * Wo * Y * X;
 
@@ -285,6 +285,7 @@ int main(int argc, char* argv[])
     std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
               << std::endl;
 
+    bool pass = true;
     if(do_verification)
     {
         pool_host_verify<InDataType,
@@ -302,14 +303,15 @@ int main(int argc, char* argv[])
 
         out_device_buf.FromDevice(out_n_c_ho_wo_device.mData.data());
 
-        ck::utils::check_err(out_n_c_ho_wo_device.mData, out_n_c_ho_wo_host.mData);
+        pass &= ck::utils::check_err(out_n_c_ho_wo_device.mData, out_n_c_ho_wo_host.mData);
 
         if constexpr(NeedIndices)
         {
             out_indices_device_buf.FromDevice(out_indices_n_c_ho_wo_device.mData.data());
 
-            //          ck::utils::check_err(out_indices_n_c_ho_wo_device.mData,
-            //          out_indices_n_c_ho_wo_host.mData);;
+            pass &= ck::utils::check_err(out_indices_n_c_ho_wo_device.mData,
+                                         out_indices_n_c_ho_wo_host.mData);
         };
     }
+    return pass ? 0 : 1;
 }

example/14_gemm_xdl_requant_relu_requant/gemm_xdl_requant_relu_requant_int8.cpp

@@ -105,9 +105,9 @@ using ReferenceGemmInstance = ck::tensor_operation::host::
 
 int main(int argc, char* argv[])
 {
-    bool do_verification = 0;
-    int init_method      = 0;
-    int nrepeat          = 5;
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
 
     // GEMM shape
     ck::index_t M = 3840;
@@ -125,13 +125,13 @@ int main(int argc, char* argv[])
     {
         do_verification = std::stoi(argv[1]);
         init_method     = std::stoi(argv[2]);
-        nrepeat         = std::stoi(argv[3]);
+        time_kernel     = std::stoi(argv[3]);
     }
     else if(argc == 10)
     {
         do_verification = std::stoi(argv[1]);
         init_method     = std::stoi(argv[2]);
-        nrepeat         = std::stoi(argv[3]);
+        time_kernel     = std::stoi(argv[3]);
 
         M = std::stoi(argv[4]);
         N = std::stoi(argv[5]);
@@ -145,7 +145,7 @@ int main(int argc, char* argv[])
     {
         printf("arg1: verification (0=no, 1=yes)\n");
         printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
-        printf("arg3: run kernel # of times (>1)\n");
+        printf("arg3: time kernel (0=n0, 1=yes)\n");
         printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC\n");
         exit(0);
     }
@@ -219,7 +219,7 @@ int main(int argc, char* argv[])
             "not support this GEMM problem");
     }
 
-    float ave_time = invoker.Run(argument, nrepeat);
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
 
     std::size_t flop = std::size_t(2) * M * N * K;
     std::size_t num_btype =
@@ -244,7 +244,7 @@ int main(int argc, char* argv[])
 
         ref_invoker.Run(ref_argument);
 
-        ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData);
+        return ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData) ? 0 : 1;
     }
 
     return 0;

example/15_grouped_gemm/grouped_gemm_xdl_fp16.cpp

@@ -60,21 +60,21 @@ using ReferenceGemmInstance = ck::tensor_operation::host::
 
 int main(int argc, char* argv[])
 {
-    bool do_verification = 0;
-    int init_method      = 0;
-    int nrepeat          = 5;
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
 
     if(argc == 4)
     {
         do_verification = std::stoi(argv[1]);
         init_method     = std::stoi(argv[2]);
-        nrepeat         = std::stoi(argv[3]);
+        time_kernel     = std::stoi(argv[3]);
     }
     else
     {
         printf("arg1: verification (0=no, 1=yes)\n");
         printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
-        printf("arg3: run kernel # of times (>1)\n");
+        printf("arg3: time kernel (0=n0, 1=yes)\n");
         exit(0);
     }
 
@@ -202,7 +202,7 @@ int main(int argc, char* argv[])
             "not support this GEMM problem");
     }
 
-    float ave_time = invoker.Run(argument, nrepeat);
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
 
     float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
 
@@ -211,6 +211,7 @@ int main(int argc, char* argv[])
     std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
               << gemm.GetTypeString() << std::endl;
 
+    bool pass = true;
     if(do_verification)
     {
         for(std::size_t i = 0; i < gemm_shapes.size(); i++)
@@ -227,9 +228,9 @@ int main(int argc, char* argv[])
                                                       c_element_op);
 
             ref_invoker.Run(ref_argument);
-            ck::utils::check_err(c_device_tensors[i].mData, c_host_tensors[i].mData);
+            pass &= ck::utils::check_err(c_device_tensors[i].mData, c_host_tensors[i].mData);
         }
     }
 
-    return 0;
+    return pass ? 0 : 1;
 }

example/16_gemm_reduce/CMakeLists.txt

-add_example_executable(example_gemm_reduce_xdl_fp16 gemm_reduce_xdl_fp16.cpp)
+add_example_executable(example_gemm_reduce_xdl_max_fp16 gemm_reduce_xdl_max_fp16.cpp)
+add_example_executable(example_gemm_reduce_xdl_sum_squaresum_fp16 gemm_reduce_xdl_sum_squaresum_fp16.cpp)

example/16_gemm_reduce/gemm_reduce_xdl_max_fp16.cpp

+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+#include <stdlib.h>
+
+#include "check_err.hpp"
+#include "config.hpp"
+#include "device.hpp"
+#include "host_tensor.hpp"
+#include "host_tensor_generator.hpp"
+#include "device_tensor.hpp"
+#include "device_gemm_reduce_xdl_cshuffle.hpp"
+#include "element_wise_operation.hpp"
+#include "reference_gemm.hpp"
+#include "gemm_specialization.hpp"
+#include "element_wise_reduce_operation.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16 = ck::half_t;
+using F32 = float;
+using F64 = double;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using ADataType         = F16;
+using BDataType         = F16;
+using CDataType         = F16;
+using ReduceAccDataType = F32;
+using DDataType         = F64;
+using DPtrsGlobal       = ck::Tuple<DDataType*>;
+
+using ALayout = ck::tensor_layout::gemm::RowMajor;
+using BLayout = ck::tensor_layout::gemm::ColumnMajor;
+using CLayout = ck::tensor_layout::gemm::RowMajor;
+
+using AElementOp  = ck::tensor_operation::element_wise::PassThrough;
+using BElementOp  = ck::tensor_operation::element_wise::PassThrough;
+using CElementOp  = ck::tensor_operation::element_wise::PassThrough;
+using DsReduceOp  = ck::Tuple<ck::reduce::Max<ReduceAccDataType>>;
+using DsElementOp = ck::Tuple<
+    ck::tensor_operation::element_wise::UnaryIdentic<ReduceAccDataType, ReduceAccDataType, false>>;
+using DGlobalMemOp =
+    ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicMax>;
+
+static constexpr auto GemmSpecialization =
+    ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceGemmReduceInstance = ck::tensor_operation::device::DeviceGemmReduce_Xdl_CShuffle
+//######| ALayout| BLayout| CLayout|AData| BData| CData|  GemmAcc| CShuffle|         ReduceAcc|         DData|           A|           B|           C|         Dxs|   DxsInEleOp|  DxsOutEleOp|             D|               GEMM| NumGemmK| 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|              CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
+//######|        |        |        | Type|  Type|  Type| DataType| DataType|          DataType|    Type Tuple| Elementwise| Elementwise| Elementwise|      Reduce|             |             |    MemoryData|     Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar|    ExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar|    ExtraN| MXdlPerWave| NXdlPerWave|            _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths|     SrcDstScalarPerVector|        SrcDstScalarPerVector|
+//######|        |        |        |     |      |      |         |         |                  |              |   Operation|   Operation|   Operation|   Operation|             |             |     Operation|                   |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|            _NBlock_NPerBlock|      _NPerBlock| _MPerBlock_NPerBlock|                _NPerBlock|                   _MPerBlock|
+//######|        |        |        |     |      |      |         |         |                  |              |            |            |            |            |             |             |              |                   |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                     |                          |                             |
+        <     Row,     Col,     Row,  F16,   F16,   F16,      F32,      F32, ReduceAccDataType,   DPtrsGlobal,  AElementOp,  BElementOp,  CElementOp, DsReduceOp,   DsElementOp,  DsElementOp,  DGlobalMemOp, GemmSpecialization,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>;
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::
+    ReferenceGemm<ADataType, BDataType, CDataType, AElementOp, BElementOp, CElementOp>;
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    // GEMM shape
+    ck::index_t M = 3840;
+    ck::index_t N = 4096;
+    ck::index_t K = 4096;
+
+    ck::index_t StrideA = 4096;
+    ck::index_t StrideB = 4096;
+    ck::index_t StrideC = 4096;
+
+    if(argc == 1)
+    {
+        // do nothing
+    }
+    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 == 10)
+    {
+        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]);
+
+        StrideA = std::stoi(argv[7]);
+        StrideB = std::stoi(argv[8]);
+        StrideC = std::stoi(argv[9]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: run kernel # of times (>1)\n");
+        printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC\n");
+        exit(0);
+    }
+
+    auto f_host_tensor_descriptor =
+        [](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)
+            {
+                return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
+                                            std::vector<std::size_t>({stride, 1}));
+            }
+            else
+            {
+                return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
+                                            std::vector<std::size_t>({1, stride}));
+            }
+        };
+
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
+
+    Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+    Tensor<DDataType> d_m_host_result(
+        HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(M)})));
+
+    Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+    Tensor<DDataType> d_m_device_result(
+        HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(M)})));
+
+    std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
+    std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
+    std::cout << "c_m_n: " << c_m_n_host_result.mDesc << std::endl;
+    std::cout << "d_m: " << d_m_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        break;
+    default:
+        a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        break;
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpace());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpace());
+    DeviceMem c_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpace());
+    DeviceMem d_device_buf(sizeof(DDataType) * d_m_device_result.mDesc.GetElementSpace());
+
+    a_device_buf.ToDevice(a_m_k.mData.data());
+    b_device_buf.ToDevice(b_k_n.mData.data());
+
+    auto a_element_op  = AElementOp{};
+    auto b_element_op  = BElementOp{};
+    auto c_element_op  = CElementOp{};
+    auto ds_element_op = DsElementOp{};
+    auto p_ds_global   = ck::make_tuple(static_cast<DDataType*>(d_device_buf.GetDeviceBuffer()));
+
+    // do GEMM
+    auto gemm     = DeviceGemmReduceInstance{};
+    auto invoker  = gemm.MakeInvoker();
+    auto argument = gemm.MakeArgument(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
+                                      static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
+                                      static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
+                                      p_ds_global,
+                                      M,
+                                      N,
+                                      K,
+                                      StrideA,
+                                      StrideB,
+                                      StrideC,
+                                      a_element_op,
+                                      b_element_op,
+                                      c_element_op,
+                                      ds_element_op,
+                                      ds_element_op);
+
+    if(!gemm.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_gemm with the specified compilation parameters does "
+            "not support this GEMM problem");
+    }
+
+    // init D
+    d_device_buf.SetValue(ck::NumericLimits<DDataType>::Lowest());
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    std::size_t flop = std::size_t(2) * M * N * K;
+    std::size_t num_btype =
+        sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(CDataType) * M * 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, "
+              << gemm.GetTypeString() << std::endl;
+
+    bool pass = true;
+
+    if(do_verification)
+    {
+        c_device_buf.FromDevice(c_m_n_device_result.mData.data());
+        d_device_buf.FromDevice(d_m_device_result.mData.data());
+
+        auto ref_gemm    = ReferenceGemmInstance{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        auto ref_argument = ref_gemm.MakeArgument(
+            a_m_k, b_k_n, c_m_n_host_result, a_element_op, b_element_op, c_element_op);
+
+        ref_invoker.Run(ref_argument);
+
+        auto d_reduce_op = DsReduceOp{}[ck::Number<0>{}];
+
+        for(int m = 0; m < M; ++m)
+        {
+            ReduceAccDataType d_acc = d_reduce_op.GetReductionZeroVal();
+
+            for(int n = 0; n < N; ++n)
+                d_reduce_op(d_acc, c_m_n_host_result(m, n));
+
+            d_m_host_result(m) = d_acc;
+        }
+
+        pass = ck::utils::check_err(c_m_n_device_result.mData,
+                                    c_m_n_host_result.mData,
+                                    "Error: Incorrect results c") &&
+               ck::utils::check_err(d_m_device_result.mData,
+                                    d_m_host_result.mData,
+                                    "Error: Incorrect results d",
+                                    1e-3,
+                                    1e-3);
+    }
+
+    return pass ? 0 : 1;
+}

example/16_gemm_reduce/gemm_reduce_xdl_fp16.cpp → example/16_gemm_reduce/gemm_reduce_xdl_sum_squaresum_fp16.cpp

@@ -3,7 +3,8 @@
 #include <initializer_list>
 #include <cstdlib>
 #include <stdlib.h>
-#include <half.hpp>
+
+#include "check_err.hpp"
 #include "config.hpp"
 #include "device.hpp"
 #include "host_tensor.hpp"
@@ -28,7 +29,9 @@ using Col = ck::tensor_layout::gemm::ColumnMajor;
 using ADataType         = F16;
 using BDataType         = F16;
 using CDataType         = F16;
+using ReduceAccDataType = F32;
 using DDataType         = F32;
+using DPtrsGlobal       = ck::Tuple<DDataType*, DDataType*>;
 
 using ALayout = ck::tensor_layout::gemm::RowMajor;
 using BLayout = ck::tensor_layout::gemm::ColumnMajor;
@@ -37,20 +40,31 @@ using CLayout = ck::tensor_layout::gemm::RowMajor;
 using AElementOp  = ck::tensor_operation::element_wise::PassThrough;
 using BElementOp  = ck::tensor_operation::element_wise::PassThrough;
 using CElementOp  = ck::tensor_operation::element_wise::PassThrough;
-using D0ReduceOp  = ck::reduce::Add<float>;
-using D1ReduceOp  = ck::reduce::Add<float>;
-using D1ElementOp = ck::tensor_operation::element_wise::UnarySquare<float, float, false>;
+using D0ReduceOp  = ck::reduce::Add<ReduceAccDataType>;
+using D1ReduceOp  = ck::reduce::Add<ReduceAccDataType>;
+using DxsReduceOp = ck::Tuple<D0ReduceOp, D1ReduceOp>;
+
+using UnaryIdenticElementOp =
+    ck::tensor_operation::element_wise::UnaryIdentic<ReduceAccDataType, ReduceAccDataType, false>;
+using UnarySquareElementOp =
+    ck::tensor_operation::element_wise::UnarySquare<ReduceAccDataType, ReduceAccDataType, false>;
+using DxsInElementOp  = ck::Tuple<UnaryIdenticElementOp, UnarySquareElementOp>;
+using DxsOutElementOp = ck::Tuple<UnaryIdenticElementOp, UnaryIdenticElementOp>;
+
+using DGlobalMemOp =
+    ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd,
+                                          ck::InMemoryDataOperationEnum::AtomicAdd>;
 
 static constexpr auto GemmSpecialization =
     ck::tensor_operation::device::GemmSpecialization::Default;
 
 // clang-format off
 using DeviceGemmReduceInstance = ck::tensor_operation::device::DeviceGemmReduce_Xdl_CShuffle
-//######| ALayout| BLayout| CLayout|AData| BData| CData|  GemmAcc| CShuffle| ReduceAcc| DData|           A|           B|           C|         D0|         D1|    D1EleOp|    GEMM| NumGemmK| 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|              CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
-//######|        |        |        | Type|  Type|  Type| DataType| DataType|  DataType|  Type| Elementwise| Elementwise| Elementwise|     Reduce|     Reduce|     Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar|    ExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar|    ExtraN| MXdlPerWave| NXdlPerWave|            _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths|     SrcDstScalarPerVector|        SrcDstScalarPerVector|
-//######|        |        |        |     |      |      |         |         |          |      |   Operation|   Operation|   Operation|  Operation|  Operation|                   |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|            _NBlock_NPerBlock|      _NPerBlock| _MPerBlock_NPerBlock|                _NPerBlock|                   _MPerBlock|
-//######|        |        |        |     |      |      |         |         |          |      |            |            |            |           |           |                   |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                     |                          |                             |
-        <     Row,     Col,     Row,  F16,   F16,   F16,      F32,      F32,       F32,   F32,  AElementOp,  BElementOp,  CElementOp, D0ReduceOp, D1ReduceOp, D1ElementOp, GemmSpecialization,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>;
+//######| ALayout| BLayout| CLayout|AData| BData| CData|  GemmAcc| CShuffle| ReduceAcc|         DData|           A|           B|           C|         Dxs|     DxsInEleOp|     DxsOutEleOp|             D|               GEMM| NumGemmK| 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|              CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
+//######|        |        |        | Type|  Type|  Type| DataType| DataType|  DataType|    Type Tuple| Elementwise| Elementwise| Elementwise|      Reduce|               |                |    MemoryData|     Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar|    ExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar|    ExtraN| MXdlPerWave| NXdlPerWave|            _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths|     SrcDstScalarPerVector|        SrcDstScalarPerVector|
+//######|        |        |        |     |      |      |         |         |          |              |   Operation|   Operation|   Operation|   Operation|               |                |     Operation|                   |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|            _NBlock_NPerBlock|      _NPerBlock| _MPerBlock_NPerBlock|                _NPerBlock|                   _MPerBlock|
+//######|        |        |        |     |      |      |         |         |          |              |            |            |            |            |               |                |              |                   |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                     |                          |                             |
+        <     Row,     Col,     Row,  F16,   F16,   F16,      F32,      F32,       F32,   DPtrsGlobal,  AElementOp,  BElementOp,  CElementOp, DxsReduceOp, DxsInElementOp, DxsOutElementOp,  DGlobalMemOp, GemmSpecialization,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>;
 // clang-format on
 
 using ReferenceGemmInstance = ck::tensor_operation::host::
@@ -58,9 +72,9 @@ using ReferenceGemmInstance = ck::tensor_operation::host::
 
 int main(int argc, char* argv[])
 {
-    bool do_verification = 1;
+    bool do_verification = true;
     int init_method      = 1;
-    int nrepeat          = 5;
+    bool time_kernel     = false;
 
     // GEMM shape
     ck::index_t M = 3840;
@@ -79,13 +93,13 @@ int main(int argc, char* argv[])
     {
         do_verification = std::stoi(argv[1]);
         init_method     = std::stoi(argv[2]);
-        nrepeat         = std::stoi(argv[3]);
+        time_kernel     = std::stoi(argv[3]);
     }
     else if(argc == 10)
     {
         do_verification = std::stoi(argv[1]);
         init_method     = std::stoi(argv[2]);
-        nrepeat         = std::stoi(argv[3]);
+        time_kernel     = std::stoi(argv[3]);
 
         M = std::stoi(argv[4]);
         N = std::stoi(argv[5]);
@@ -99,7 +113,7 @@ int main(int argc, char* argv[])
     {
         printf("arg1: verification (0=no, 1=yes)\n");
         printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
-        printf("arg3: run kernel # of times (>1)\n");
+        printf("arg3: time kernel (0=n0, 1=yes)\n");
         printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC\n");
         exit(0);
     }
@@ -164,7 +178,8 @@ int main(int argc, char* argv[])
     auto a_element_op = AElementOp{};
     auto b_element_op = BElementOp{};
     auto c_element_op = CElementOp{};
-    auto d1_element_op = D1ElementOp{};
+    auto dxs_global   = ck::make_tuple(static_cast<DDataType*>(d0_device_buf.GetDeviceBuffer()),
+                                     static_cast<DDataType*>(d1_device_buf.GetDeviceBuffer()));
 
     // do GEMM
     auto gemm     = DeviceGemmReduceInstance{};
@@ -172,8 +187,7 @@ int main(int argc, char* argv[])
     auto argument = gemm.MakeArgument(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
                                       static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
                                       static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
-                                      static_cast<DDataType*>(d0_device_buf.GetDeviceBuffer()),
-                                      static_cast<DDataType*>(d1_device_buf.GetDeviceBuffer()),
+                                      dxs_global,
                                       M,
                                       N,
                                       K,
@@ -183,7 +197,8 @@ int main(int argc, char* argv[])
                                       a_element_op,
                                       b_element_op,
                                       c_element_op,
-                                      d1_element_op);
+                                      DxsInElementOp{},
+                                      DxsOutElementOp{});
 
     if(!gemm.IsSupportedArgument(argument))
     {
@@ -192,30 +207,13 @@ int main(int argc, char* argv[])
             "not support this GEMM problem");
     }
 
-    // warm up
-    invoker.Run(argument);
-
-    // timing
-    float total_time = 0;
-
-    for(int i = 0; i < nrepeat; ++i)
-    {
     // init DO, D1 to 0
     d0_device_buf.SetZero();
     d1_device_buf.SetZero();
 
-        KernelTimer timer;
-
-        timer.Start();
-
-        invoker.Run(argument);
-
-        timer.End();
-
-        total_time += timer.GetElapsedTime();
-    }
-
-    float ave_time = total_time / nrepeat;
+    // if time_kernel == true, kernel will run multiple times. This kernel use atomic-add so result
+    // will not be correct. need to set time_kernel = false for correctness test
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
 
     std::size_t flop = std::size_t(2) * M * N * K;
     std::size_t num_btype =
@@ -228,6 +226,8 @@ int main(int argc, char* argv[])
     std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
               << gemm.GetTypeString() << std::endl;
 
+    bool pass = true;
+
     if(do_verification)
     {
         c_device_buf.FromDevice(c_m_n_device_result.mData.data());
@@ -252,10 +252,12 @@ int main(int argc, char* argv[])
 
             for(int n = 0; n < N; ++n)
             {
-                float d0_val = ck::type_convert<float>(c_m_n_host_result(m, n));
-                float d1_val;
+                float c_val  = ck::type_convert<float>(c_m_n_host_result(m, n));
+                float d0_val = 0;
+                float d1_val = 0;
 
-                d1_element_op(d1_val, d0_val);
+                UnaryIdenticElementOp{}(d0_val, c_val);
+                UnarySquareElementOp{}(d1_val, c_val);
                 d0_reduce_op(d0_acc, d0_val);
                 d1_reduce_op(d1_acc, d1_val);
             }
@@ -264,10 +266,20 @@ int main(int argc, char* argv[])
             d1_m_host_result(m) = ck::type_convert<DDataType>(d1_acc);
         }
 
-        check_error(c_m_n_host_result, c_m_n_device_result);
-        check_error(d0_m_host_result, d0_m_device_result);
-        check_error(d1_m_host_result, d1_m_device_result);
+        pass = ck::utils::check_err(c_m_n_device_result.mData,
+                                    c_m_n_host_result.mData,
+                                    "Error: Incorrect results c") &&
+               ck::utils::check_err(d0_m_device_result.mData,
+                                    d0_m_host_result.mData,
+                                    "Error: Incorrect results d0",
+                                    1e-4,
+                                    1e-5) &&
+               ck::utils::check_err(d1_m_device_result.mData,
+                                    d1_m_host_result.mData,
+                                    "Error: Incorrect results d1",
+                                    1e-3,
+                                    1e-5);
     }
 
-    return 0;
+    return pass ? 0 : 1;
 }

example/17_convnd_bwd_data_xdl/convnd_bwd_data_xdl.cpp

@@ -87,7 +87,7 @@ void print_use_msg()
 {
     std::cout << "arg1: verification (0=no, 1=yes)\n"
               << "arg2: initialization (0=no init, 1=random value, 2= init to 1 )\n"
-              << "arg3: run kernel # of times (>1)\n"
+              << "arg3: time kernel (0=n0, 1=yes)\n"
               << "arg4: N spatial dimensions (default 2)\n"
               << "Following arguments (depending on number of spatial dims):\n"
               << " N, K, C, \n"
@@ -165,9 +165,9 @@ DeviceConvBwdDataBasePtr get_conv_instance(int num_dim_spatial)
 
 int main(int argc, char* argv[])
 {
-    bool do_verification = 0;
-    int init_method      = 0;
-    int nrepeat          = 5;
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
     int num_dim_spatial  = 2;
 
     ck::utils::conv::ConvParams params;
@@ -177,13 +177,13 @@ int main(int argc, char* argv[])
     {
         do_verification = std::stoi(argv[1]);
         init_method     = std::stoi(argv[2]);
-        nrepeat         = std::stoi(argv[3]);
+        time_kernel     = std::stoi(argv[3]);
     }
     else if(argc > 4)
     {
         do_verification = std::stoi(argv[1]);
         init_method     = std::stoi(argv[2]);
-        nrepeat         = std::stoi(argv[3]);
+        time_kernel     = std::stoi(argv[3]);
         num_dim_spatial = std::stoi(argv[4]);
         // check args number
         int conv_args     = 3 + num_dim_spatial * 6;
@@ -284,7 +284,7 @@ int main(int argc, char* argv[])
             "not support this Conv problem");
     }
 
-    float ave_time = invoker->Run(argument.get(), nrepeat);
+    float ave_time = invoker->Run(argument.get(), StreamConfig{nullptr, time_kernel});
 
     std::size_t flop = ck::utils::conv::get_flops(
         params.N_, params.C_, params.K_, params.filter_spatial_lengths_, output_spatial_lengths);
@@ -322,7 +322,10 @@ int main(int argc, char* argv[])
 
             in_device_buf.FromDevice(in_n_c_hi_wi_device_result.mData.data());
 
-            check_error(in_n_c_hi_wi_host_result, in_n_c_hi_wi_device_result);
+            return ck::utils::check_err(in_n_c_hi_wi_device_result.mData,
+                                        in_n_c_hi_wi_host_result.mData)
+                       ? 0
+                       : 1;
         };
 
         switch(num_dim_spatial)
@@ -347,4 +350,5 @@ int main(int argc, char* argv[])
         }
         }
     }
+    return 0;
 }

example/18_batched_gemm_reduce/batched_gemm_reduce_xdl_fp16.cpp

@@ -4,6 +4,7 @@
 #include <cstdlib>
 #include <stdlib.h>
 #include <half.hpp>
+#include "check_err.hpp"
 #include "config.hpp"
 #include "device.hpp"
 #include "host_tensor.hpp"
@@ -27,7 +28,9 @@ using Col = ck::tensor_layout::gemm::ColumnMajor;
 using ADataType         = F16;
 using BDataType         = F16;
 using CDataType         = F16;
+using ReduceAccDataType = F32;
 using DDataType         = F32;
+using DPtrsGlobal       = ck::Tuple<DDataType*, DDataType*>;
 
 using ALayout = ck::tensor_layout::gemm::RowMajor;
 using BLayout = ck::tensor_layout::gemm::ColumnMajor;
@@ -36,20 +39,31 @@ using CLayout = ck::tensor_layout::gemm::RowMajor;
 using AElementOp  = ck::tensor_operation::element_wise::PassThrough;
 using BElementOp  = ck::tensor_operation::element_wise::PassThrough;
 using CElementOp  = ck::tensor_operation::element_wise::PassThrough;
-using D0ReduceOp  = ck::reduce::Add<float>;
-using D1ReduceOp  = ck::reduce::Add<float>;
-using D1ElementOp = ck::tensor_operation::element_wise::UnarySquare<float, float, false>;
+using D0ReduceOp  = ck::reduce::Add<ReduceAccDataType>;
+using D1ReduceOp  = ck::reduce::Add<ReduceAccDataType>;
+using DxsReduceOp = ck::Tuple<D0ReduceOp, D1ReduceOp>;
+
+using UnaryIdenticElementOp =
+    ck::tensor_operation::element_wise::UnaryIdentic<ReduceAccDataType, ReduceAccDataType, false>;
+using UnarySquareElementOp =
+    ck::tensor_operation::element_wise::UnarySquare<ReduceAccDataType, ReduceAccDataType, false>;
+using DxsInElementOp  = ck::Tuple<UnaryIdenticElementOp, UnarySquareElementOp>;
+using DxsOutElementOp = ck::Tuple<UnaryIdenticElementOp, UnaryIdenticElementOp>;
+
+using DGlobalMemOp =
+    ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd,
+                                          ck::InMemoryDataOperationEnum::AtomicAdd>;
 
 static constexpr auto GemmSpecialization =
     ck::tensor_operation::device::GemmSpecialization::Default;
 
 // clang-format off
 using DeviceBatchedGemmReduceInstance = ck::tensor_operation::device::DeviceBatchedGemmReduce_Xdl_CShuffle
-//######| ALayout| BLayout| CLayout|AData| BData| CData|  GemmAcc| CShuffle| ReduceAcc| DData|           A|           B|           C|         D0|         D1|     D1EleOp|      GEMM| NumGemmK| 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|              CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
-//######|        |        |        | Type|  Type|  Type| DataType| DataType|  DataType|  Type| Elementwise| Elementwise| Elementwise|     Reduce|     Reduce|     Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar|    ExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar|    ExtraN| MXdlPerWave| NXdlPerWave|            _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths|     SrcDstScalarPerVector|        SrcDstScalarPerVector|
-//######|        |        |        |     |      |      |         |         |          |      |   Operation|   Operation|   Operation|  Operation|  Operation|                   |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|            _NBlock_NPerBlock|      _NPerBlock| _MPerBlock_NPerBlock|                _NPerBlock|                   _MPerBlock|
-//######|        |        |        |     |      |      |         |         |          |      |            |            |            |           |           |                   |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                     |                          |                             |
-        <     Row,     Col,     Row,  F16,   F16,   F16,      F32,      F32,       F32,   F32,  AElementOp,  BElementOp,  CElementOp, D0ReduceOp, D1ReduceOp, D1ElementOp, GemmSpecialization,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>;
+//######| ALayout| BLayout| CLayout|AData| BData| CData|  GemmAcc| CShuffle| ReduceAcc|         DData|           A|           B|           C|         Dxs|     DxsInEleOp|     DxsOutEleOp|            D|               GEMM| NumGemmK| 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|              CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
+//######|        |        |        | Type|  Type|  Type| DataType| DataType|  DataType|    Type Tuple| Elementwise| Elementwise| Elementwise|      Reduce|               |                |   MemoryData|     Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar|    ExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar|    ExtraN| MXdlPerWave| NXdlPerWave|            _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths|     SrcDstScalarPerVector|        SrcDstScalarPerVector|
+//######|        |        |        |     |      |      |         |         |          |              |   Operation|   Operation|   Operation|   Operation|               |                |    Operation|                   |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|            _NBlock_NPerBlock|      _NPerBlock| _MPerBlock_NPerBlock|                _NPerBlock|                   _MPerBlock|
+//######|        |        |        |     |      |      |         |         |          |              |            |            |            |            |               |                |             |                   |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                     |                          |                             |
+        <     Row,     Col,     Row,  F16,   F16,   F16,      F32,      F32,       F32,   DPtrsGlobal,  AElementOp,  BElementOp,  CElementOp, DxsReduceOp, DxsInElementOp, DxsOutElementOp, DGlobalMemOp, GemmSpecialization,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8,             S<64, 4>,                         4,                            1>;
 // clang-format on
 
 using ReferenceBatchedGemmInstance = ck::tensor_operation::host::
@@ -57,18 +71,18 @@ using ReferenceBatchedGemmInstance = ck::tensor_operation::host::
 
 int main(int argc, char* argv[])
 {
-    bool do_verification = 1;
+    bool do_verification = true;
     int init_method      = 1;
-    int nrepeat          = 5;
+    bool time_kernel     = false;
 
     // GEMM shape
-    ck::index_t M = 3840;
-    ck::index_t N = 4096;
-    ck::index_t K = 4096;
+    ck::index_t M = 2048;
+    ck::index_t N = 1920;
+    ck::index_t K = 2048;
 
-    ck::index_t StrideA = 4096;
-    ck::index_t StrideB = 4096;
-    ck::index_t StrideC = 4096;
+    ck::index_t StrideA = 2048;
+    ck::index_t StrideB = 2048;
+    ck::index_t StrideC = 1920;
 
     ck::index_t BatchCount = 4;
 
@@ -80,13 +94,13 @@ int main(int argc, char* argv[])
     {
         do_verification = std::stoi(argv[1]);
         init_method     = std::stoi(argv[2]);
-        nrepeat         = std::stoi(argv[3]);
+        time_kernel     = std::stoi(argv[3]);
     }
     else if(argc == 11)
     {
         do_verification = std::stoi(argv[1]);
         init_method     = std::stoi(argv[2]);
-        nrepeat         = std::stoi(argv[3]);
+        time_kernel     = std::stoi(argv[3]);
 
         M = std::stoi(argv[4]);
         N = std::stoi(argv[5]);
@@ -96,13 +110,13 @@ int main(int argc, char* argv[])
         StrideB = std::stoi(argv[8]);
         StrideC = std::stoi(argv[9]);
 
-        BatchCount = std::stoi(argv[9]);
+        BatchCount = std::stoi(argv[10]);
     }
     else
     {
         printf("arg1: verification (0=no, 1=yes)\n");
         printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
-        printf("arg3: run kernel # of times (>1)\n");
+        printf("arg3: time kernel (0=n0, 1=yes)\n");
         printf("arg4 to 10: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC, BatchCount\n");
         exit(0);
     }
@@ -172,9 +186,8 @@ int main(int argc, char* argv[])
     auto a_element_op = AElementOp{};
     auto b_element_op = BElementOp{};
     auto c_element_op = CElementOp{};
-    auto d0_reduce_op  = D0ReduceOp{};
-    auto d1_reduce_op  = D1ReduceOp{};
-    auto d1_element_op = D1ElementOp{};
+    auto dxs_global   = ck::make_tuple(static_cast<DDataType*>(d0_device_buf.GetDeviceBuffer()),
+                                     static_cast<DDataType*>(d1_device_buf.GetDeviceBuffer()));
 
     // do GEMM
     auto batched_gemm = DeviceBatchedGemmReduceInstance{};
@@ -183,8 +196,7 @@ int main(int argc, char* argv[])
         batched_gemm.MakeArgument(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
                                   static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
                                   static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
-                                  static_cast<DDataType*>(d0_device_buf.GetDeviceBuffer()),
-                                  static_cast<DDataType*>(d1_device_buf.GetDeviceBuffer()),
+                                  dxs_global,
                                   M,
                                   N,
                                   K,
@@ -194,7 +206,8 @@ int main(int argc, char* argv[])
                                   a_element_op,
                                   b_element_op,
                                   c_element_op,
-                                  d1_element_op,
+                                  DxsInElementOp{},
+                                  DxsOutElementOp{},
                                   BatchCount);
 
     if(!batched_gemm.IsSupportedArgument(argument))
@@ -204,30 +217,13 @@ int main(int argc, char* argv[])
             "not support this GEMM problem");
     }
 
-    // warm up
-    invoker.Run(argument);
-
-    // timing
-    float total_time = 0;
-
-    for(int i = 0; i < nrepeat; ++i)
-    {
     // init DO, D1 to 0
     d0_device_buf.SetZero();
     d1_device_buf.SetZero();
 
-        KernelTimer timer;
-
-        timer.Start();
-
-        invoker.Run(argument);
-
-        timer.End();
-
-        total_time += timer.GetElapsedTime();
-    }
-
-    float ave_time = total_time / nrepeat;
+    // if time_kernel == true, kernel will run multiple times. This kernel use atomic-add so result
+    // will not be correct. need to set time_kernel = false for correctness test
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
 
     std::size_t flop      = std::size_t(2) * BatchCount * M * N * K;
     std::size_t num_btype = sizeof(ADataType) * BatchCount * M * K +
@@ -241,6 +237,7 @@ int main(int argc, char* argv[])
     std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
               << batched_gemm.GetTypeString() << std::endl;
 
+    bool pass = true;
     if(do_verification)
     {
         c_device_buf.FromDevice(c_g_m_n_device_result.mData.data());
@@ -255,6 +252,9 @@ int main(int argc, char* argv[])
 
         ref_invoker.Run(ref_argument);
 
+        auto d0_reduce_op = D0ReduceOp{};
+        auto d1_reduce_op = D1ReduceOp{};
+
         for(int batch = 0; batch < BatchCount; ++batch)
         {
             for(int m = 0; m < M; ++m)
@@ -264,10 +264,12 @@ int main(int argc, char* argv[])
 
                 for(int n = 0; n < N; ++n)
                 {
-                    float d0_val = ck::type_convert<float>(c_g_m_n_host_result(m, n));
-                    float d1_val;
+                    float c_val  = ck::type_convert<float>(c_g_m_n_host_result(batch, m, n));
+                    float d0_val = 0;
+                    float d1_val = 0;
 
-                    d1_element_op(d1_val, d0_val);
+                    UnaryIdenticElementOp{}(d0_val, c_val);
+                    UnarySquareElementOp{}(d1_val, c_val);
                     d0_reduce_op(d0_acc, d0_val);
                     d1_reduce_op(d1_acc, d1_val);
                 }
@@ -277,10 +279,20 @@ int main(int argc, char* argv[])
             }
         }
 
-        check_error(c_g_m_n_host_result, c_g_m_n_device_result);
-        check_error(d0_g_m_host_result, d0_g_m_device_result);
-        check_error(d1_g_m_host_result, d1_g_m_device_result);
+        pass = ck::utils::check_err(c_g_m_n_host_result.mData,
+                                    c_g_m_n_device_result.mData,
+                                    "Error: Incorrect results c") &&
+               ck::utils::check_err(d0_g_m_device_result.mData,
+                                    d0_g_m_host_result.mData,
+                                    "Error: Incorrect results! D0",
+                                    1e-4,
+                                    1e-5) &&
+               ck::utils::check_err(d1_g_m_device_result.mData,
+                                    d1_g_m_host_result.mData,
+                                    "Error: Incorrect results! D1",
+                                    1e-3,
+                                    1e-5);
     }
 
-    return 0;
+    return pass ? 0 : 1;
 }

example/19_binary_elementwise/CMakeLists.txt

+add_example_executable(example_broadcast_add_2d broadcast_add_2d.cpp)
+add_example_executable(example_elementwise_add_1d elementwise_add_1d.cpp)
+add_example_executable(example_elementwise_add_4d elementwise_add_4d.cpp)
\ No newline at end of file

example/19_binary_elementwise/broadcast_add_2d.cpp

+#include <iostream>
+#include <cstdlib>
+#include "check_err.hpp"
+#include "config.hpp"
+#include "device.hpp"
+#include "host_tensor.hpp"
+#include "host_tensor_generator.hpp"
+
+#include "device_tensor.hpp"
+#include "binary_element_wise_operation.hpp"
+#include "device_binary_elementwise.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using ABDataType             = F16;
+using CDataType              = F16;
+using EltwiseComputeDataType = F32;
+
+using Add = ck::tensor_operation::binary_element_wise::Add;
+
+using DeviceElementwiseAddInstance = ck::tensor_operation::device::
+    DeviceBinaryElementwise<ABDataType, ABDataType, CDataType, EltwiseComputeDataType, Add, 2, 8>;
+
+template <typename HostTensorA,
+          typename HostTensorB,
+          typename HostTensorC,
+          typename ComputeDataType,
+          typename Functor,
+          int broadcastDim>
+void host_broadcast2D(
+    HostTensorC& C, const HostTensorA& A, const HostTensorB& B, int M, int N, Functor functor)
+{
+    using ctype = ck::remove_reference_t<decltype(C(0, 0))>;
+
+    for(int m = 0; m < M; ++m)
+    {
+        for(int n = 0; n < N; ++n)
+        {
+            ComputeDataType Amn = static_cast<ComputeDataType>(A(m, n));
+            ComputeDataType Cmn = 0;
+            if constexpr(broadcastDim == 0)
+            {
+                ComputeDataType Bn = static_cast<ComputeDataType>(B(n));
+                functor(Cmn, Amn, Bn);
+            }
+            else
+            {
+                ComputeDataType Bm = static_cast<ComputeDataType>(B(m));
+                functor(Cmn, Amn, Bm);
+            }
+            C(m, n) = static_cast<ctype>(Cmn);
+        }
+    }
+}
+
+int main()
+{
+    bool do_verification = true;
+    bool time_kernel     = false;
+
+    ck::index_t M      = 1024;
+    ck::index_t N      = 1024;
+    ck::index_t Stride = 1024;
+
+    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}));
+    };
+
+    auto f_host_tensor_descriptor2d = [](std::size_t row, std::size_t col, std::size_t stride) {
+        return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
+                                    std::vector<std::size_t>({stride, 1}));
+    };
+
+    Tensor<ABDataType> a_m_n(f_host_tensor_descriptor2d(M, N, Stride));
+    Tensor<ABDataType> b_n(f_host_tensor_descriptor1d(N, 1));
+    Tensor<CDataType> c_m_n(f_host_tensor_descriptor2d(M, N, Stride));
+
+    a_m_n.GenerateTensorValue(GeneratorTensor_3<ABDataType>{0.0, 1.0});
+    b_n.GenerateTensorValue(GeneratorTensor_3<ABDataType>{0.0, 1.0});
+
+    DeviceMem a_m_n_device_buf(sizeof(ABDataType) * a_m_n.mDesc.GetElementSpace());
+    DeviceMem b_n_device_buf(sizeof(ABDataType) * b_n.mDesc.GetElementSpace());
+    DeviceMem c_m_n_device_buf(sizeof(CDataType) * c_m_n.mDesc.GetElementSpace());
+
+    a_m_n_device_buf.ToDevice(a_m_n.mData.data());
+    b_n_device_buf.ToDevice(b_n.mData.data());
+
+    auto broadcastAdd = DeviceElementwiseAddInstance{};
+    auto argument     = broadcastAdd.MakeArgumentPointer(a_m_n_device_buf.GetDeviceBuffer(),
+                                                     b_n_device_buf.GetDeviceBuffer(),
+                                                     c_m_n_device_buf.GetDeviceBuffer(),
+                                                     {M, N},
+                                                     {Stride, 1},
+                                                     {0, 1}, // broadcast in first dimension
+                                                     {Stride, 1},
+                                                     Add{});
+
+    if(!broadcastAdd.IsSupportedArgument(argument.get()))
+    {
+        throw std::runtime_error("The runtime parameters seems not supported by the "
+                                 "DeviceBinaryElementwise_2D instance, exiting!");
+    };
+
+    auto broadcastAdd_invoker_ptr = broadcastAdd.MakeInvokerPointer();
+    float ave_time =
+        broadcastAdd_invoker_ptr->Run(argument.get(), StreamConfig{nullptr, time_kernel});
+
+    std::cout << "Perf: " << ave_time << " ms" << std::endl;
+
+    bool pass = true;
+    if(do_verification)
+    {
+        c_m_n_device_buf.FromDevice(c_m_n.mData.data());
+        Tensor<CDataType> host_c_m_n(f_host_tensor_descriptor2d(M, N, Stride));
+
+        host_broadcast2D<Tensor<ABDataType>,
+                         Tensor<ABDataType>,
+                         Tensor<CDataType>,
+                         EltwiseComputeDataType,
+                         Add,
+                         0>(host_c_m_n, a_m_n, b_n, M, N, Add{});
+
+        pass &= ck::utils::check_err(
+            c_m_n.mData, host_c_m_n.mData, "Error: Incorrect results d1", 1e-3, 1e-3);
+    }
+
+    return pass ? 0 : 1;
+}

example/19_binary_elementwise/elementwise_add_1d.cpp

+#include <iostream>
+#include <cstdlib>
+#include "check_err.hpp"
+#include "config.hpp"
+#include "device.hpp"
+#include "host_tensor.hpp"
+#include "host_tensor_generator.hpp"
+
+#include "device_tensor.hpp"
+#include "binary_element_wise_operation.hpp"
+#include "device_binary_elementwise.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using ABDataType             = F16;
+using CDataType              = F16;
+using EltwiseComputeDataType = F32;
+
+using Add = ck::tensor_operation::binary_element_wise::Add;
+
+using DeviceElementwiseAddInstance = ck::tensor_operation::device::
+    DeviceBinaryElementwise<ABDataType, ABDataType, CDataType, EltwiseComputeDataType, Add, 1, 8>;
+
+template <typename HostTensorA,
+          typename HostTensorB,
+          typename HostTensorC,
+          typename ComputeDataType,
+          typename Functor>
+void host_elementwise1D(
+    HostTensorC& C, const HostTensorA& A, const HostTensorB& B, int M, Functor functor)
+{
+    using ctype = ck::remove_reference_t<decltype(C(0))>;
+
+    for(int m = 0; m < M; ++m)
+    {
+        ComputeDataType Am = static_cast<ComputeDataType>(A(m));
+        ComputeDataType Bm = static_cast<ComputeDataType>(B(m));
+        ComputeDataType Cm = 0;
+        functor(Cm, Am, Bm);
+        C(m) = static_cast<ctype>(Cm);
+    }
+}
+
+int main()
+{
+    bool do_verification = true;
+    bool time_kernel     = false;
+
+    ck::index_t M = 1024;
+
+    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}));
+    };
+
+    Tensor<ABDataType> a_m(f_host_tensor_descriptor1d(M, 1));
+    Tensor<ABDataType> b_m(f_host_tensor_descriptor1d(M, 1));
+    Tensor<CDataType> c_m(f_host_tensor_descriptor1d(M, 1));
+
+    a_m.GenerateTensorValue(GeneratorTensor_3<ABDataType>{0.0, 1.0});
+    b_m.GenerateTensorValue(GeneratorTensor_3<ABDataType>{0.0, 1.0});
+
+    DeviceMem a_m_device_buf(sizeof(ABDataType) * a_m.mDesc.GetElementSpace());
+    DeviceMem b_m_device_buf(sizeof(ABDataType) * b_m.mDesc.GetElementSpace());
+    DeviceMem c_m_device_buf(sizeof(CDataType) * c_m.mDesc.GetElementSpace());
+
+    a_m_device_buf.ToDevice(a_m.mData.data());
+    b_m_device_buf.ToDevice(b_m.mData.data());
+
+    auto broadcastAdd = DeviceElementwiseAddInstance{};
+    auto argument     = broadcastAdd.MakeArgumentPointer(a_m_device_buf.GetDeviceBuffer(),
+                                                     b_m_device_buf.GetDeviceBuffer(),
+                                                     c_m_device_buf.GetDeviceBuffer(),
+                                                     {M},
+                                                     {1},
+                                                     {1},
+                                                     {1},
+                                                     Add{});
+
+    if(!broadcastAdd.IsSupportedArgument(argument.get()))
+    {
+        throw std::runtime_error("The runtime parameters seems not supported by the "
+                                 "DeviceBinaryElementwise_2D instance, exiting!");
+    };
+
+    auto broadcastAdd_invoker_ptr = broadcastAdd.MakeInvokerPointer();
+    float ave_time =
+        broadcastAdd_invoker_ptr->Run(argument.get(), StreamConfig{nullptr, time_kernel});
+
+    std::cout << "Perf: " << ave_time << " ms" << std::endl;
+
+    bool pass = true;
+    if(do_verification)
+    {
+        c_m_device_buf.FromDevice(c_m.mData.data());
+        Tensor<CDataType> host_c_m(f_host_tensor_descriptor1d(M, 1));
+
+        host_elementwise1D<Tensor<ABDataType>,
+                           Tensor<ABDataType>,
+                           Tensor<CDataType>,
+                           EltwiseComputeDataType,
+                           Add>(host_c_m, a_m, b_m, M, Add{});
+
+        pass &= ck::utils::check_err(
+            c_m.mData, host_c_m.mData, "Error: Incorrect results d1", 1e-3, 1e-3);
+    }
+
+    return pass ? 0 : 1;
+}

example/19_binary_elementwise/elementwise_add_4d.cpp

+#include <iostream>
+#include <cstdlib>
+#include "check_err.hpp"
+#include "config.hpp"
+#include "device.hpp"
+#include "host_tensor.hpp"
+#include "host_tensor_generator.hpp"
+
+#include "device_tensor.hpp"
+#include "binary_element_wise_operation.hpp"
+#include "device_binary_elementwise.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using ABDataType             = F16;
+using CDataType              = F16;
+using EltwiseComputeDataType = F32;
+
+using Add = ck::tensor_operation::binary_element_wise::Add;
+
+using DeviceElementwiseAddInstance = ck::tensor_operation::device::
+    DeviceBinaryElementwise<ABDataType, ABDataType, CDataType, EltwiseComputeDataType, Add, 4, 8>;
+
+template <typename HostTensorA,
+          typename HostTensorB,
+          typename HostTensorC,
+          typename ComputeDataType,
+          typename Functor>
+void host_elementwise4D(HostTensorC& C,
+                        const HostTensorA& A,
+                        const HostTensorB& B,
+                        const std::vector<std::size_t>& shape,
+                        Functor functor)
+{
+    using ctype = ck::remove_reference_t<decltype(C(0, 0, 0, 0))>;
+
+    for(std::size_t n = 0; n < shape[0]; ++n)
+        for(std::size_t c = 0; c < shape[1]; ++c)
+            for(std::size_t h = 0; h < shape[2]; ++h)
+                for(std::size_t w = 0; w < shape[3]; ++w)
+                {
+                    ComputeDataType a_val = static_cast<ComputeDataType>(A(n, c, h, w));
+                    ComputeDataType b_val = static_cast<ComputeDataType>(B(n, c, h, w));
+                    ComputeDataType c_val = 0;
+                    functor(c_val, a_val, b_val);
+                    C(n, c, h, w) = static_cast<ctype>(c_val);
+                }
+}
+
+int main()
+{
+    bool do_verification = true;
+    bool time_kernel     = false;
+
+    std::vector<std::size_t> nchw = {4, 16, 32, 32};
+
+    Tensor<ABDataType> a(nchw);
+    Tensor<ABDataType> b(nchw);
+    Tensor<CDataType> c(nchw);
+
+    a.GenerateTensorValue(GeneratorTensor_3<ABDataType>{0.0, 1.0});
+    b.GenerateTensorValue(GeneratorTensor_3<ABDataType>{0.0, 1.0});
+
+    DeviceMem a_device_buf(sizeof(ABDataType) * a.mDesc.GetElementSpace());
+    DeviceMem b_device_buf(sizeof(ABDataType) * b.mDesc.GetElementSpace());
+    DeviceMem c_device_buf(sizeof(CDataType) * c.mDesc.GetElementSpace());
+
+    a_device_buf.ToDevice(a.mData.data());
+    b_device_buf.ToDevice(b.mData.data());
+
+    auto broadcastAdd = DeviceElementwiseAddInstance{};
+    auto argument     = broadcastAdd.MakeArgumentPointer(
+        a_device_buf.GetDeviceBuffer(),
+        b_device_buf.GetDeviceBuffer(),
+        c_device_buf.GetDeviceBuffer(),
+        std::vector<ck::index_t>{nchw.begin(), nchw.end()},
+        std::vector<ck::index_t>{a.mDesc.GetStrides().begin(), a.mDesc.GetStrides().end()},
+        std::vector<ck::index_t>{b.mDesc.GetStrides().begin(), b.mDesc.GetStrides().end()},
+        std::vector<ck::index_t>{c.mDesc.GetStrides().begin(), c.mDesc.GetStrides().end()},
+        Add{});
+
+    if(!broadcastAdd.IsSupportedArgument(argument.get()))
+    {
+        throw std::runtime_error("The runtime parameters seems not supported by the "
+                                 "DeviceBinaryElementwise_2D instance, exiting!");
+    };
+
+    auto broadcastAdd_invoker_ptr = broadcastAdd.MakeInvokerPointer();
+    float ave_time =
+        broadcastAdd_invoker_ptr->Run(argument.get(), StreamConfig{nullptr, time_kernel});
+
+    std::cout << "Perf: " << ave_time << " ms" << std::endl;
+
+    bool pass = true;
+    if(do_verification)
+    {
+        c_device_buf.FromDevice(c.mData.data());
+        Tensor<CDataType> host_c(nchw);
+
+        host_elementwise4D<Tensor<ABDataType>,
+                           Tensor<ABDataType>,
+                           Tensor<CDataType>,
+                           EltwiseComputeDataType,
+                           Add>(host_c, a, b, nchw, Add{});
+
+        pass &=
+            ck::utils::check_err(c.mData, host_c.mData, "Error: Incorrect results d1", 1e-3, 1e-3);
+    }
+
+    return pass ? 0 : 1;
+}

example/20_convnd_bwd_weight_xdl/CMakeLists.txt

+add_example_executable(example_convnd_bwd_weight_xdl convnd_bwd_weight_xdl.cpp)
+target_link_libraries(example_convnd_bwd_weight_xdl PRIVATE conv_util)
\ No newline at end of file

example/20_convnd_bwd_weight_xdl/convnd_bwd_weight_xdl.cpp

+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+#include <stdlib.h>
+#include <half.hpp>
+
+#include "check_err.hpp"
+#include "conv_util.hpp"
+#include "config.hpp"
+#include "print.hpp"
+#include "device.hpp"
+#include "host_tensor.hpp"
+#include "host_tensor_generator.hpp"
+#include "device_tensor.hpp"
+#include "tensor_layout.hpp"
+#include "element_wise_operation.hpp"
+#include "device_convnd_backward_weight_xdl_c_shuffle_nhwc_kyxc_nhwk.hpp"
+#include "reference_conv_backward_weight.hpp"
+
+using InDataType  = ck::half_t;
+using WeiDataType = ck::half_t;
+using OutDataType = ck::half_t;
+using AccDataType = float;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using InElementOp  = ck::tensor_operation::element_wise::PassThrough;
+using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
+using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvBwdWeightDefault =
+    ck::tensor_operation::device::ConvolutionBackwardWeightSpecialization::Default;
+
+using DeviceConvBwdWeightBasePtr =
+    ck::tensor_operation::device::DeviceConvBwdWeightPtr<InElementOp, WeiElementOp, OutElementOp>;
+
+// clang-format off
+template <ck::index_t NumDimSpatial>
+using DeviceConvndBwdWeightInstance = ck::tensor_operation::device::
+    DeviceConvndBwdWeightXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<
+        InDataType,                       // InDataType
+        WeiDataType,                      // WeiDataType
+        OutDataType,                      // OutDataType
+        AccDataType,                      // AccDataType
+        InElementOp,                      // InElementwiseOperation
+        WeiElementOp,                     // WeiElementwiseOperation
+        OutElementOp,                     // OutElementwiseOperation
+        ConvBwdWeightDefault,             // ConvolutionBackwardWeightSpecialization
+        NumDimSpatial,                    // NumDimSpatial
+        256,                              // BlockSize
+        128,                              // MPerBlock
+        128,                              // NPerBlock
+        4,                                // K0PerBlock
+        8,                                // K1
+        32,                               // MPerXdl
+        32,                               // NPerXdl
+        2,                                // MXdlPerWave
+        2,                                // NXdlPerWave
+        S<1, 4, 16, 4>,                   // ABlockTransferThreadClusterLengths_K0_M_K1
+        S<0, 3, 1, 2>,                    // ABlockTransferThreadClusterArrangeOrder
+        S<0, 2, 1, 3>,                    // ABlockTransferSrcAccessOrder
+        2,                                // ABlockTransferSrcVectorDim
+        8,                                // ABlockTransferSrcScalarPerVector
+        2,                                // ABlockTransferDstScalarPerVector_K1
+        true,                             // ABlockLdsAddExtraM
+        S<1, 4, 16, 4>,                   // BBlockTransferThreadClusterLengths_K0_N_K1
+        S<0, 3, 1, 2>,                    // BBlockTransferThreadClusterArrangeOrder
+        S<0, 2, 1, 3>,                    // BBlockTransferSrcAccessOrder
+        2,                                // BBlockTransferSrcVectorDim
+        8,                                // BBlockTransferSrcScalarPerVector
+        2,                                // BBlockTransferDstScalarPerVector_K1
+        true,                             // BBlockLdsAddExtraN
+        1,                                // CShuffleMXdlPerWavePerShuffle
+        1,                                // CShuffleNXdlPerWavePerShuffle
+        S<1, 32, 1, 4>,                   // CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+        8>;                               // CBlockTransferScalarPerVector_NWaveNPerXdl
+// clang-format on
+
+template <ck::index_t NumDimSpatial>
+using ReferenceConvBwdWeightInstance =
+    ck::tensor_operation::host::ReferenceConvBwdWeight<InDataType,
+                                                       WeiDataType,
+                                                       OutDataType,
+                                                       InElementOp,
+                                                       WeiElementOp,
+                                                       OutElementOp,
+                                                       NumDimSpatial>;
+
+void print_use_msg()
+{
+    std::cout << "arg1: verification (0=no, 1=yes)\n"
+              << "arg2: initialization (0=no init, 1=random value, 2= init to 1 )\n"
+              << "arg3: time kernel (0=n0, 1=yes)\n"
+              << "arg4: is show log (0=no, 1=yes)\n"
+              << "arg5: split-k \n"
+              << "arg6: N spatial dimensions (default 2)\n"
+              << "Following arguments (depending on number of spatial dims):\n"
+              << " N, K, C, \n"
+              << " <filter spatial dimensions>, (ie Y, X for 2D)\n"
+              << " <input image spatial dimensions>, (ie Hi, Wi for 2D)\n"
+              << " <strides>, (ie Sy, Sx for 2D)\n"
+              << " <dilations>, (ie Dy, Dx for 2D)\n"
+              << " <left padding>, (ie LeftPy, LeftPx for 2D)\n"
+              << " <right padding>, (ie RightPy, RightPx for 2D)\n"
+              << std::endl;
+}
+
+ck::utils::conv::ConvParams parse_conv_params(int num_dim_spatial, char* argv[])
+{
+    // (N, K, C) + num_dim_spatial * 6 (filter, input, strides, dilations, pad left, pad right)
+    ck::utils::conv::ConvParams params;
+    int arg_idx = 7;
+
+    params.num_dim_spatial_ = num_dim_spatial;
+    params.N_               = std::stoi(argv[arg_idx++]);
+    params.K_               = std::stoi(argv[arg_idx++]);
+    params.C_               = std::stoi(argv[arg_idx++]);
+
+    params.filter_spatial_lengths_.resize(num_dim_spatial);
+    for(int i = 0; i < num_dim_spatial; ++i)
+    {
+        params.filter_spatial_lengths_[i] = std::stoi(argv[arg_idx++]);
+    }
+    params.input_spatial_lengths_.resize(num_dim_spatial);
+    for(int i = 0; i < num_dim_spatial; ++i)
+    {
+        params.input_spatial_lengths_[i] = std::stoi(argv[arg_idx++]);
+    }
+    params.conv_filter_strides_.resize(num_dim_spatial);
+    for(int i = 0; i < num_dim_spatial; ++i)
+    {
+        params.conv_filter_strides_[i] = std::stoi(argv[arg_idx++]);
+    }
+    params.conv_filter_dilations_.resize(num_dim_spatial);
+    for(int i = 0; i < num_dim_spatial; ++i)
+    {
+        params.conv_filter_dilations_[i] = std::stoi(argv[arg_idx++]);
+    }
+    params.input_left_pads_.resize(num_dim_spatial);
+    for(int i = 0; i < num_dim_spatial; ++i)
+    {
+        params.input_left_pads_[i] = std::stoi(argv[arg_idx++]);
+    }
+    params.input_right_pads_.resize(num_dim_spatial);
+    for(int i = 0; i < num_dim_spatial; ++i)
+    {
+        params.input_right_pads_[i] = std::stoi(argv[arg_idx++]);
+    }
+
+    return params;
+}
+
+DeviceConvBwdWeightBasePtr get_conv_instance(int num_dim_spatial)
+{
+    switch(num_dim_spatial)
+    {
+    case 3: {
+        return std::make_unique<DeviceConvndBwdWeightInstance<3>>();
+    }
+    case 2: {
+        return std::make_unique<DeviceConvndBwdWeightInstance<2>>();
+    }
+    case 1: {
+        return std::make_unique<DeviceConvndBwdWeightInstance<1>>();
+    }
+    default: {
+        throw std::runtime_error("Unsupported number of spatial dimensions provided!");
+    }
+    }
+}
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+    int num_dim_spatial  = 2;
+    int do_log           = 0;
+    int split_k          = 1;
+
+    ck::utils::conv::ConvParams params;
+    params.C_ = 128;
+
+    if(argc == 6)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+        do_log          = std::stoi(argv[4]);
+        split_k         = std::stoi(argv[5]);
+    }
+    else if(argc > 6)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+        do_log          = std::stoi(argv[4]);
+        split_k         = std::stoi(argv[5]);
+        num_dim_spatial = std::stoi(argv[6]);
+        // check args number
+        int conv_args     = 3 + num_dim_spatial * 6;
+        int cmdline_nargs = conv_args + 7;
+        if(cmdline_nargs != argc)
+        {
+            print_use_msg();
+            exit(1);
+        }
+
+        params = parse_conv_params(num_dim_spatial, argv);
+    }
+    else if(argc != 1)
+    {
+        print_use_msg();
+        exit(1);
+    }
+
+    std::vector<std::size_t> input_dims{static_cast<std::size_t>(params.N_),
+                                        static_cast<std::size_t>(params.C_)};
+    input_dims.insert(std::end(input_dims),
+                      std::begin(params.input_spatial_lengths_),
+                      std::end(params.input_spatial_lengths_));
+
+    std::vector<std::size_t> filter_dims{static_cast<std::size_t>(params.K_),
+                                         static_cast<std::size_t>(params.C_)};
+    filter_dims.insert(std::end(filter_dims),
+                       std::begin(params.filter_spatial_lengths_),
+                       std::end(params.filter_spatial_lengths_));
+
+    const std::vector<ck::index_t>& output_spatial_lengths = params.GetOutputSpatialLengths();
+    std::vector<std::size_t> output_dims{static_cast<std::size_t>(params.N_),
+                                         static_cast<std::size_t>(params.K_)};
+    output_dims.insert(std::end(output_dims),
+                       std::begin(output_spatial_lengths),
+                       std::end(output_spatial_lengths));
+
+    Tensor<InDataType> in_n_c_hi_wi(
+        ck::utils::conv::get_input_host_tensor_descriptor(input_dims, num_dim_spatial));
+    Tensor<WeiDataType> wei_k_c_y_x_host_result(
+        ck::utils::conv::get_filters_host_tensor_descriptor(filter_dims, num_dim_spatial));
+    Tensor<WeiDataType> wei_k_c_y_x_device_result(
+        ck::utils::conv::get_filters_host_tensor_descriptor(filter_dims, num_dim_spatial));
+    Tensor<OutDataType> out_n_k_ho_wo(
+        ck::utils::conv::get_output_host_tensor_descriptor(output_dims, num_dim_spatial));
+
+    std::cout << "in_n_c_hi_wi: " << in_n_c_hi_wi.mDesc << std::endl;
+    std::cout << "wei_k_c_y_x: " << wei_k_c_y_x_device_result.mDesc << std::endl;
+    std::cout << "out_n_k_ho_wo: " << out_n_k_ho_wo.mDesc << std::endl;
+
+    std::cout << "in_n_c_hi_wi: " << in_n_c_hi_wi.mDesc << std::endl;
+    std::cout << "wei_k_c_y_x: " << wei_k_c_y_x_host_result.mDesc << std::endl;
+    std::cout << "out_n_k_ho_wo: " << out_n_k_ho_wo.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        out_n_k_ho_wo.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-2, 2});
+        in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-2, 2});
+        break;
+    default:
+        out_n_k_ho_wo.GenerateTensorValue(GeneratorTensor_1<OutDataType>{1});
+        in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_1<WeiDataType>{1});
+    }
+
+    DeviceMem in_device_buf(sizeof(InDataType) * in_n_c_hi_wi.mDesc.GetElementSpace());
+    DeviceMem wei_device_buf(sizeof(WeiDataType) *
+                             wei_k_c_y_x_device_result.mDesc.GetElementSpace());
+    DeviceMem out_device_buf(sizeof(OutDataType) * out_n_k_ho_wo.mDesc.GetElementSpace());
+
+    in_device_buf.ToDevice(in_n_c_hi_wi.mData.data());
+    out_device_buf.ToDevice(out_n_k_ho_wo.mData.data());
+    // reset input to zero
+    wei_device_buf.SetZero();
+
+    // do GEMM
+    auto conv    = get_conv_instance(num_dim_spatial);
+    auto invoker = conv->MakeInvokerPointer();
+    auto argument =
+        conv->MakeArgumentPointer(static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
+                                  static_cast<WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
+                                  static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
+                                  params.N_,
+                                  params.K_,
+                                  params.C_,
+                                  params.input_spatial_lengths_,
+                                  params.filter_spatial_lengths_,
+                                  output_spatial_lengths,
+                                  params.conv_filter_strides_,
+                                  params.conv_filter_dilations_,
+                                  params.input_left_pads_,
+                                  params.input_right_pads_,
+                                  InElementOp{},
+                                  WeiElementOp{},
+                                  OutElementOp{},
+                                  split_k);
+
+    if(!conv->IsSupportedArgument(argument.get()))
+    {
+        std::cout << "wrong! device_conv with the specified compilation parameters does "
+                     "not support this Conv problem"
+                  << std::endl;
+        return 1;
+    }
+
+    float ave_time = invoker->Run(argument.get(), StreamConfig{nullptr, time_kernel});
+
+    std::size_t flop = ck::utils::conv::get_flops(
+        params.N_, params.C_, params.K_, params.filter_spatial_lengths_, output_spatial_lengths);
+    std::size_t num_btype = ck::utils::conv::get_btype<InDataType, WeiDataType, OutDataType>(
+        params.N_,
+        params.C_,
+        params.K_,
+        params.input_spatial_lengths_,
+        params.filter_spatial_lengths_,
+        output_spatial_lengths);
+
+    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"
+              << std::endl;
+
+    if(do_verification)
+    {
+        auto verify_f = [&](const auto& ref_conv) {
+            auto ref_invoker = ref_conv.MakeInvoker();
+
+            auto ref_argument = ref_conv.MakeArgument(in_n_c_hi_wi,
+                                                      wei_k_c_y_x_host_result,
+                                                      out_n_k_ho_wo,
+                                                      params.conv_filter_strides_,
+                                                      params.conv_filter_dilations_,
+                                                      params.input_left_pads_,
+                                                      params.input_right_pads_,
+                                                      InElementOp{},
+                                                      WeiElementOp{},
+                                                      OutElementOp{});
+
+            ref_invoker.Run(ref_argument);
+
+            wei_device_buf.FromDevice(wei_k_c_y_x_device_result.mData.data());
+
+            if(do_log)
+            {
+                LogRangeAsType<float>(std::cout << "out: ", out_n_k_ho_wo.mData, ",") << std::endl;
+                LogRangeAsType<float>(std::cout << "in : ", in_n_c_hi_wi.mData, ",") << std::endl;
+                LogRangeAsType<float>(
+                    std::cout << "wei_device(after): ", wei_k_c_y_x_device_result.mData, ",")
+                    << std::endl;
+                LogRangeAsType<float>(
+                    std::cout << "wei_host  : ", wei_k_c_y_x_host_result.mData, ",")
+                    << std::endl;
+            }
+
+            return ck::utils::check_err(wei_k_c_y_x_device_result.mData,
+                                        wei_k_c_y_x_host_result.mData)
+                       ? 0
+                       : 1;
+        };
+
+        switch(num_dim_spatial)
+        {
+        case 3: {
+            auto ref_conv = ReferenceConvBwdWeightInstance<3>();
+            verify_f(ref_conv);
+            break;
+        }
+        case 2: {
+            auto ref_conv = ReferenceConvBwdWeightInstance<2>();
+            verify_f(ref_conv);
+            break;
+        }
+        case 1: {
+            auto ref_conv = ReferenceConvBwdWeightInstance<1>();
+            verify_f(ref_conv);
+            break;
+        }
+        default: {
+            throw std::runtime_error("Unsupported number of spatial dimensions provided!");
+        }
+        }
+    }
+    return 0;
+}

example/CMakeLists.txt

@@ -19,13 +19,22 @@ include_directories(BEFORE
 
 add_custom_target(examples)
 
-function(add_example_executable EXAMPLE_NAME)
+function(add_example_executable EXAMPLE_NAME FILE_NAME)
     message("adding example ${EXAMPLE_NAME}")
-    add_executable(${EXAMPLE_NAME} ${ARGN})
+    add_executable(${EXAMPLE_NAME} ${FILE_NAME})
     target_link_libraries(${EXAMPLE_NAME} PRIVATE host_tensor)
+    add_test(NAME ${EXAMPLE_NAME} COMMAND $<TARGET_FILE:${EXAMPLE_NAME}> ${ARGN})
     add_dependencies(examples ${EXAMPLE_NAME})
+    add_dependencies(check ${EXAMPLE_NAME})
 endfunction(add_example_executable EXAMPLE_NAME)
 
+function(add_example_executable_no_testing EXAMPLE_NAME FILE_NAME)
+    message("adding example ${EXAMPLE_NAME}")
+    add_executable(${EXAMPLE_NAME} ${FILE_NAME})
+    target_link_libraries(${EXAMPLE_NAME} PRIVATE host_tensor)
+    add_dependencies(examples ${EXAMPLE_NAME})
+endfunction(add_example_executable_no_testing EXAMPLE_NAME)
+
 add_subdirectory(01_gemm)
 add_subdirectory(02_gemm_alpha_beta)
 add_subdirectory(03_gemm_bias_relu)
@@ -42,3 +51,5 @@ add_subdirectory(17_convnd_bwd_data_xdl)
 add_subdirectory(15_grouped_gemm)
 add_subdirectory(16_gemm_reduce)
 add_subdirectory(18_batched_gemm_reduce)
+add_subdirectory(19_binary_elementwise)
+add_subdirectory(20_convnd_bwd_weight_xdl)

include/ck/config.hpp

@@ -76,6 +76,12 @@
 #define CK_USE_AMD_BUFFER_ATOMIC_ADD_FLOAT 0
 #endif
 
+#if defined(__gfx90a__) // for GPU code
+#define CK_USE_AMD_BUFFER_ATOMIC_MAX_FLOAT64 1
+#else
+#define CK_USE_AMD_BUFFER_ATOMIC_MAX_FLOAT64 0
+#endif
+
 // inline asm
 #define CK_USE_AMD_INLINE_ASM 1
 
@@ -91,10 +97,11 @@
 // experimental feature: static tensor descriptor
 #define CK_EXPERIMENTAL_STATIC_TENSOR_DESCRIPTOR 0
 
-// experimental feature: buffer load/store/atomic-add OOB trick
+// experimental feature: buffer load/store/atomic-add/ OOB trick
 #define CK_EXPERIMENTAL_USE_BUFFER_LOAD_OOB_CHECK_OFFSET_TRICK 0
 #define CK_EXPERIMENTAL_USE_BUFFER_STORE_OOB_CHECK_OFFSET_TRICK 1
 #define CK_EXPERIMENTAL_USE_BUFFER_ATOMIC_ADD_OOB_CHECK_OFFSET_TRICK 1
+#define CK_EXPERIMENTAL_USE_BUFFER_ATOMIC_MAX_OOB_CHECK_OFFSET_TRICK 1
 
 // experimental feature: in-regsiter sub-dword transpose
 #define CK_EXPERIMENTAL_USE_IN_REGISTER_SUB_DWORD_TRANSPOSE 1
@@ -142,9 +149,23 @@ enum struct InMemoryDataOperationEnum
 {
     Set,
     AtomicAdd,
+    AtomicMax,
     Add
 };
 
+template <InMemoryDataOperationEnum... Is>
+struct InMemoryDataOperationEnumSequence
+{
+    static constexpr int mSize = sizeof...(Is);
+
+    __host__ __device__ static constexpr InMemoryDataOperationEnum At(int I)
+    {
+        // the last dummy element is to prevent compiler complain about empty array, when mSize = 0
+        const InMemoryDataOperationEnum mData[mSize + 1] = {Is..., InMemoryDataOperationEnum::Set};
+        return mData[I];
+    }
+};
+
 // TODO: no longer needed, remove this
 enum struct ActivTypeEnum
 {

include/ck/hip_version.hpp.in

-#pragma once
-
-// "_PACKAGE_" to avoid name contentions: the macros like
-// HIP_VERSION_MAJOR are defined in HIP_VERSION.h.
-// clang-format off
-#define CK_HIP_PACKAGE_VERSION_MAJOR @CK_HIP_VERSION_MAJOR@
-#define CK_HIP_PACKAGE_VERSION_MINOR @CK_HIP_VERSION_MINOR@
-#define CK_HIP_PACKAGE_VERSION_PATCH @CK_HIP_VERSION_PATCH@
-// clang-format on
-
-#ifndef CK_HIP_PACKAGE_VERSION_MAJOR
-#define CK_HIP_PACKAGE_VERSION_MAJOR 0
-#endif
-#ifndef CK_HIP_PACKAGE_VERSION_MINOR
-#define CK_HIP_PACKAGE_VERSION_MINOR 0
-#endif
-#ifndef CK_HIP_PACKAGE_VERSION_PATCH
-#define CK_HIP_PACKAGE_VERSION_PATCH 0
-#endif
-// 3 decimal digits for major and minor, 6 digits for patch number.
-// Max number is 999,999,999999 == 0xE8,D4A5,0FFF that fits into 64-bit math.
-#if CK_HIP_PACKAGE_VERSION_MAJOR > 999 || CK_HIP_PACKAGE_VERSION_MAJOR > 999 || \
-    CK_HIP_PACKAGE_VERSION_PATCH > 999999
-#error "Too big HIP version number(s)"
-#endif
-#define CK_HIP_PACKAGE_VERSION_FLAT                                                      \
-    ((CK_HIP_PACKAGE_VERSION_MAJOR * 1000ULL + CK_HIP_PACKAGE_VERSION_MINOR) * 1000000 + \
-     CK_HIP_PACKAGE_VERSION_PATCH)