merging with latest develop branch

example/16_gemm_reduce/gemm_reduce_xdl_mean_squaremean_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 "reduction_operator.hpp"
+#include "reference_gemm.hpp"
+#include "gemm_specialization.hpp"
+#include "reduction_operator.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using ADataType         = F16;
+using BDataType         = F16;
+using CDataType         = F16;
+using GemmAccDataType   = F32;
+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;
+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<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 UnaryDivElementOp =
+    ck::tensor_operation::element_wise::UnaryIdentic<ReduceAccDataType, ReduceAccDataType, true>;
+using UnarySquareElementOp =
+    ck::tensor_operation::element_wise::UnarySquare<ReduceAccDataType, ReduceAccDataType, false>;
+using DxsInElementOp  = ck::Tuple<UnaryIdenticElementOp, UnarySquareElementOp>;
+using DxsOutElementOp = ck::Tuple<UnaryDivElementOp, UnaryDivElementOp>;
+
+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|         Dxs|     DxsInEleOp|     DxsAccEleOp|             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::ReferenceGemm<ADataType,
+                                                                        BDataType,
+                                                                        CDataType,
+                                                                        GemmAccDataType,
+                                                                        AElementOp,
+                                                                        BElementOp,
+                                                                        CElementOp>;
+
+template <typename ADataType, typename BDataType, typename CDataType, typename DDataType>
+void DumpGemmLayerNormPerf(float gemm_reduce_time, int M, int N, int K)
+{
+    std::size_t gemm_flop     = std::size_t(2) * M * N * K;
+    std::size_t gemm_num_byte = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
+                                sizeof(CDataType) * M * N + sizeof(DDataType) * M +
+                                sizeof(DDataType) * M;
+
+    float tflops          = static_cast<float>(gemm_flop) / 1.E9 / gemm_reduce_time;
+    float gemm_gb_per_sec = gemm_num_byte / 1.E6 / gemm_reduce_time;
+
+    std::cout << "gemm + reduce_mean + reduce_mean_square Perf: " << gemm_reduce_time << " ms, "
+              << tflops << " TFlops, " << gemm_gb_per_sec << " GB/s, " << std::endl;
+}
+
+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: time kernel (0=n0, 1=yes)\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> d0_m_host_result(
+        HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(M)})));
+    Tensor<DDataType> d1_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> d0_m_device_result(
+        HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(M)})));
+    Tensor<DDataType> d1_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 << "d0_m: " << d0_m_host_result.mDesc << std::endl;
+    std::cout << "d1_m: " << d1_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 d0_device_buf(sizeof(DDataType) * d0_m_device_result.mDesc.GetElementSpace());
+    DeviceMem d1_device_buf(sizeof(DDataType) * d1_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 dxs_global   = ck::make_tuple(static_cast<DDataType*>(d0_device_buf.GetDeviceBuffer()),
+                                     static_cast<DDataType*>(d1_device_buf.GetDeviceBuffer()));
+
+    auto dxs_in_element_op  = DxsInElementOp{};
+    auto dxs_out_element_op = DxsOutElementOp{M, M};
+
+    // 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()),
+                                      dxs_global,
+                                      M,
+                                      N,
+                                      K,
+                                      StrideA,
+                                      StrideB,
+                                      StrideC,
+                                      a_element_op,
+                                      b_element_op,
+                                      c_element_op,
+                                      dxs_in_element_op,
+                                      dxs_out_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 DO, D1 to 0
+    d0_device_buf.SetZero();
+    d1_device_buf.SetZero();
+
+    // 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
+    invoker.Run(argument, StreamConfig{nullptr, false});
+    bool pass = true;
+
+    if(do_verification)
+    {
+        c_device_buf.FromDevice(c_m_n_device_result.mData.data());
+        d0_device_buf.FromDevice(d0_m_device_result.mData.data());
+        d1_device_buf.FromDevice(d1_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 d0_reduce_op = D0ReduceOp{};
+        auto d1_reduce_op = D1ReduceOp{};
+
+        for(int m = 0; m < M; ++m)
+        {
+            float d0_acc = d0_reduce_op.GetIdentityValue();
+            float d1_acc = d1_reduce_op.GetIdentityValue();
+
+            for(int n = 0; n < N; ++n)
+            {
+                float c_val  = ck::type_convert<float>(c_m_n_host_result(m, n));
+                float d0_val = 0;
+                float d1_val = 0;
+
+                dxs_in_element_op(ck::Number<0>{})(d0_val, c_val);
+                dxs_in_element_op(ck::Number<1>{})(d1_val, c_val);
+                d0_reduce_op(d0_acc, d0_val);
+                d1_reduce_op(d1_acc, d1_val);
+            }
+
+            dxs_out_element_op(ck::Number<0>{})(d0_acc, d0_acc);
+            dxs_out_element_op(ck::Number<1>{})(d1_acc, d1_acc);
+            d0_m_host_result(m) = ck::type_convert<DDataType>(d0_acc);
+            d1_m_host_result(m) = ck::type_convert<DDataType>(d1_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(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);
+    }
+
+    if(time_kernel)
+    {
+        float ave_time = invoker.Run(argument, StreamConfig{nullptr, true});
+
+        DumpGemmLayerNormPerf<ADataType, BDataType, CDataType, DDataType>(ave_time, M, N, K);
+    }
+
+    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);
@@ -322,29 +322,30 @@ 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)
         {
         case 3: {
             auto ref_conv = ReferenceConvBwdDataInstance<3>();
-            verify_f(ref_conv);
-            break;
+            return verify_f(ref_conv);
         }
         case 2: {
             auto ref_conv = ReferenceConvBwdDataInstance<2>();
-            verify_f(ref_conv);
-            break;
+            return verify_f(ref_conv);
         }
         case 1: {
             auto ref_conv = ReferenceConvBwdDataInstance<1>();
-            verify_f(ref_conv);
-            break;
+            return verify_f(ref_conv);
         }
         default: {
             throw std::runtime_error("Unsupported number of spatial dimensions provided!");
         }
         }
     }
+    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"
@@ -24,10 +25,12 @@ using F32 = float;
 using Row = ck::tensor_layout::gemm::RowMajor;
 using Col = ck::tensor_layout::gemm::ColumnMajor;
 
-using ADataType = F16;
-using BDataType = F16;
-using CDataType = F16;
-using DDataType = F32;
+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|     DxsAccEleOp|            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);
     }
@@ -169,12 +183,11 @@ int main(int argc, char* argv[])
     a_device_buf.ToDevice(a_g_m_k.mData.data());
     b_device_buf.ToDevice(b_g_k_n.mData.data());
 
-    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 a_element_op = AElementOp{};
+    auto b_element_op = BElementOp{};
+    auto c_element_op = CElementOp{};
+    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;
+    // init DO, D1 to 0
+    d0_device_buf.SetZero();
+    d1_device_buf.SetZero();
 
-        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());
@@ -255,19 +252,24 @@ 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)
             {
-                float d0_acc = d0_reduce_op.GetReductionZeroVal();
-                float d1_acc = d1_reduce_op.GetReductionZeroVal();
+                float d0_acc = d0_reduce_op.GetIdentityValue();
+                float d1_acc = d1_reduce_op.GetIdentityValue();
 
                 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_amn_bn broadcast_add_2d_amn_bn.cpp)
+add_example_executable(example_broadcast_add_3d_am_bmnk broadcast_add_3d_am_bmnk.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_amn_bn.cpp

+/*******************************************************************************
+ *
+ * MIT License
+ *
+ * Copyright (c) 2022 Advanced Micro Devices, Inc.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ *
+ *******************************************************************************/
+#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<EltwiseComputeDataType, EltwiseComputeDataType, EltwiseComputeDataType>;
+
+using DeviceElementwiseAddInstance =
+    ck::tensor_operation::device::DeviceBinaryElementwise<ABDataType,
+                                                          ABDataType,
+                                                          CDataType,
+                                                          EltwiseComputeDataType,
+                                                          Add,
+                                                          2,
+                                                          8,
+                                                          8,
+                                                          8,
+                                                          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 = ck::type_convert<ComputeDataType>(A(m, n));
+            ComputeDataType Cmn = 0;
+            if constexpr(broadcastDim == 0)
+            {
+                ComputeDataType Bn = ck::type_convert<ComputeDataType>(B(n));
+                functor(Cmn, Amn, Bn);
+            }
+            else
+            {
+                ComputeDataType Bm = ck::type_convert<ComputeDataType>(B(m));
+                functor(Cmn, Amn, Bm);
+            }
+            C(m, n) = ck::type_convert<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 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 c", 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/21_gemm_layernorm/CMakeLists.txt

+add_example_executable(example_gemm_layernorm_xdl_fp16 gemm_layernorm_xdl_fp16.cpp)

example/22_cgemm/CMakeLists.txt

+add_example_executable(example_cgemm_xdl_fp16 cgemm_xdl_fp16.cpp)

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)

include/ck/options.hpp

+#pragma once
+
+#define CK_TIME_KERNEL 1