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

profiler/include/profile_gemm_bilinear_impl.hpp → profiler/include/profiler/profile_gemm_bilinear_impl.hpp

@@ -16,6 +16,7 @@
 #include "ck/library/utility/device_memory.hpp"
 #include "ck/library/utility/host_tensor.hpp"
 #include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
 #include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
 
 namespace ck {
@@ -46,15 +47,15 @@ bool profile_gemm_bilinear_impl(int do_verification,
 {
     auto f_host_tensor_descriptor =
         [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
             if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
             {
-                return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
-                                            std::vector<std::size_t>({stride, 1}));
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
             }
             else
             {
-                return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
-                                            std::vector<std::size_t>({1, stride}));
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
             }
         };
 
@@ -116,8 +117,7 @@ bool profile_gemm_bilinear_impl(int do_verification,
     // run reference
     if(do_verification)
     {
-        Tensor<AccDataType> c_m_n(HostTensorDescriptor(
-            std::vector<std::size_t>{static_cast<std::size_t>(M), static_cast<std::size_t>(N)}));
+        Tensor<AccDataType> c_m_n({M, N});
 
         using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
                                                                                 BDataType,
@@ -215,8 +215,7 @@ bool profile_gemm_bilinear_impl(int do_verification,
             {
                 e_device_buf.FromDevice(e_m_n_device_result.mData.data());
 
-                pass = pass &&
-                       ck::utils::check_err(e_m_n_device_result.mData, e_m_n_host_result.mData);
+                pass = pass && ck::utils::check_err(e_m_n_device_result, e_m_n_host_result);
             }
         }
         else

profiler/include/profile_gemm_impl.hpp → profiler/include/profiler/profile_gemm_impl.hpp

@@ -18,6 +18,7 @@
 #include "ck/library/utility/device_memory.hpp"
 #include "ck/library/utility/host_tensor.hpp"
 #include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
 #include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
 
 namespace ck {
@@ -45,15 +46,15 @@ int profile_gemm_impl(int do_verification,
 
     auto f_host_tensor_descriptor =
         [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
             if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
             {
-                return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
-                                            std::vector<std::size_t>({stride, 1}));
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
             }
             else
             {
-                return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
-                                            std::vector<std::size_t>({1, stride}));
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
             }
         };
 
@@ -187,8 +188,7 @@ int profile_gemm_impl(int do_verification,
             {
                 c_device_buf.FromDevice(c_m_n_device_result.mData.data());
 
-                pass =
-                    pass & ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData);
+                pass = pass & ck::utils::check_err(c_m_n_device_result, c_m_n_host_result);
 
                 if(do_log)
                 {

profiler/include/profile_gemm_reduce_impl.hpp → profiler/include/profiler/profile_gemm_reduce_impl.hpp

@@ -14,6 +14,7 @@
 #include "ck/library/utility/device_memory.hpp"
 #include "ck/library/utility/host_tensor.hpp"
 #include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
 #include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
 
 namespace ck {
@@ -75,15 +76,15 @@ bool profile_gemm_reduce_impl(int do_verification,
 
     auto f_host_tensor_descriptor =
         [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
             if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
             {
-                return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
-                                            std::vector<std::size_t>({stride, 1}));
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
             }
             else
             {
-                return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
-                                            std::vector<std::size_t>({1, stride}));
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
             }
         };
 
@@ -91,16 +92,12 @@ bool profile_gemm_reduce_impl(int do_verification,
     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<ReduceDataType> reduce0_m_host_result(
-        HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(M)})));
-    Tensor<ReduceDataType> reduce1_m_host_result(
-        HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(M)})));
+    Tensor<ReduceDataType> reduce0_m_host_result({M});
+    Tensor<ReduceDataType> reduce1_m_host_result({M});
 
     Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
-    Tensor<ReduceDataType> reduce0_m_device_result(
-        HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(M)})));
-    Tensor<ReduceDataType> reduce1_m_device_result(
-        HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(M)})));
+    Tensor<ReduceDataType> reduce0_m_device_result({M});
+    Tensor<ReduceDataType> reduce1_m_device_result({M});
 
     std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
     std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
@@ -313,9 +310,9 @@ bool profile_gemm_reduce_impl(int do_verification,
                 reduce0_device_buf.FromDevice(reduce0_m_device_result.mData.data());
                 reduce1_device_buf.FromDevice(reduce1_m_device_result.mData.data());
 
-                ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData);
-                ck::utils::check_err(reduce0_m_device_result.mData, reduce0_m_host_result.mData);
-                ck::utils::check_err(reduce1_m_device_result.mData, reduce1_m_host_result.mData);
+                ck::utils::check_err(c_m_n_device_result, c_m_n_host_result);
+                ck::utils::check_err(reduce0_m_device_result, reduce0_m_host_result);
+                ck::utils::check_err(reduce1_m_device_result, reduce1_m_host_result);
 
                 if(do_log)
                 {

profiler/include/profile_gemm_splitk_impl.hpp → profiler/include/profiler/profile_gemm_splitk_impl.hpp

@@ -18,6 +18,7 @@
 #include "ck/library/utility/device_memory.hpp"
 #include "ck/library/utility/host_tensor.hpp"
 #include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
 #include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
 
 namespace ck {
@@ -46,15 +47,15 @@ bool profile_gemm_splitk_impl(int do_verification,
 
     auto f_host_tensor_descriptor =
         [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
             if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
             {
-                return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
-                                            std::vector<std::size_t>({stride, 1}));
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
             }
             else
             {
-                return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
-                                            std::vector<std::size_t>({1, stride}));
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
             }
         };
 
@@ -190,8 +191,7 @@ bool profile_gemm_splitk_impl(int do_verification,
             {
                 c_device_buf.FromDevice(c_m_n_device_result.mData.data());
 
-                pass =
-                    pass & ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData);
+                pass = pass & ck::utils::check_err(c_m_n_device_result, c_m_n_host_result);
 
                 if(do_log)
                 {

profiler/include/profile_conv_bwd_weight_impl.hpp → profiler/include/profiler/profile_grouped_conv_bwd_weight_impl.hpp

@@ -3,9 +3,10 @@
 
 #pragma once
 
-#include "ck/ck.hpp"
+#include <algorithm>
 #include <iomanip>
 #include <iostream>
+#include <iterator>
 #include <typeinfo>
 
 #include "ck/ck.hpp"
@@ -13,7 +14,7 @@
 #include "ck/tensor_operation/gpu/device/device_conv_fwd.hpp"
 #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
 
-#include "ck/library/tensor_operation_instance/gpu/convolution_backward_weight.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_weight.hpp"
 
 #include "ck/library/utility/check_err.hpp"
 #include "ck/library/utility/device_memory.hpp"
@@ -26,32 +27,6 @@
 namespace ck {
 namespace profiler {
 
-template <typename DataType>
-void show_data_nhwc_layout(Tensor<DataType>& nhwc)
-{
-    std::cout << "[";
-    for(int n = 0; n < ck::type_convert<int>(nhwc.mDesc.GetLengths()[0]); n++)
-    {
-        std::cout << "[";
-        for(int hi = 0; hi < ck::type_convert<int>(nhwc.mDesc.GetLengths()[2]); hi++)
-        {
-            std::cout << "[";
-            for(int wi = 0; wi < ck::type_convert<int>(nhwc.mDesc.GetLengths()[3]); wi++)
-            {
-                std::cout << "[";
-                for(int c = 0; c < ck::type_convert<int>(nhwc.mDesc.GetLengths()[1]); c++)
-                {
-                    std::cout << static_cast<float>(nhwc(n, c, hi, wi)) << "  ";
-                }
-                std::cout << "]";
-            }
-            std::cout << "]";
-        }
-        std::cout << "]";
-    }
-    std::cout << "]";
-}
-
 template <ck::index_t NDimSpatial,
           typename InLayout,
           typename WeiLayout,
@@ -59,12 +34,12 @@ template <ck::index_t NDimSpatial,
           typename InDataType,
           typename WeiDataType,
           typename OutDataType>
-bool profile_conv_bwd_weight_impl(int do_verification,
-                                  int init_method,
-                                  bool do_log,
-                                  bool time_kernel,
-                                  const ck::utils::conv::ConvParam& conv_param,
-                                  ck::index_t split_k)
+bool profile_grouped_conv_bwd_weight_impl(int do_verification,
+                                          int init_method,
+                                          bool do_log,
+                                          bool time_kernel,
+                                          const ck::utils::conv::ConvParam& conv_param,
+                                          ck::index_t split_k)
 {
     using InElementOp  = ck::tensor_operation::element_wise::PassThrough;
     using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
@@ -114,16 +89,14 @@ bool profile_conv_bwd_weight_impl(int do_verification,
 
     if(do_verification)
     {
-        auto ref_conv = ck::tensor_operation::host::ReferenceConvBwdWeight<NDimSpatial,
+        auto ref_conv     = ck::tensor_operation::host::ReferenceConvBwdWeight<NDimSpatial,
                                                                            InDataType,
                                                                            WeiDataType,
                                                                            OutDataType,
                                                                            InElementOp,
                                                                            WeiElementOp,
                                                                            OutElementOp>{};
-
-        auto ref_invoker = ref_conv.MakeInvoker();
-
+        auto ref_invoker  = ref_conv.MakeInvoker();
         auto ref_argument = ref_conv.MakeArgument(input,
                                                   weight_host_result,
                                                   output,
@@ -138,16 +111,16 @@ bool profile_conv_bwd_weight_impl(int do_verification,
         ref_invoker.Run(ref_argument);
     }
 
-    using DeviceOp = ck::tensor_operation::device::DeviceConvBwdWeight<NDimSpatial,
-                                                                       InLayout,
-                                                                       WeiLayout,
-                                                                       OutLayout,
-                                                                       InDataType,
-                                                                       WeiDataType,
-                                                                       OutDataType,
-                                                                       InElementOp,
-                                                                       WeiElementOp,
-                                                                       OutElementOp>;
+    using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvBwdWeight<NDimSpatial,
+                                                                              InLayout,
+                                                                              WeiLayout,
+                                                                              OutLayout,
+                                                                              InDataType,
+                                                                              WeiDataType,
+                                                                              OutDataType,
+                                                                              InElementOp,
+                                                                              WeiElementOp,
+                                                                              OutElementOp>;
 
     // get device op instances
     const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
@@ -163,22 +136,41 @@ bool profile_conv_bwd_weight_impl(int do_verification,
     // profile device Conv instances
     bool all_pass = true;
 
+    std::array<ck::index_t, NDimSpatial> input_spatial_lengths{};
+    std::array<ck::index_t, NDimSpatial> filter_spatial_lengths{};
+    std::array<ck::index_t, NDimSpatial> output_spatial_lengths{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_dilations{};
+    std::array<ck::index_t, NDimSpatial> input_left_pads{};
+    std::array<ck::index_t, NDimSpatial> input_right_pads{};
+
+    auto range_copy = [](const auto& from, auto to) { std::copy(begin(from), end(from), to); };
+
+    range_copy(conv_param.input_spatial_lengths_, begin(input_spatial_lengths));
+    range_copy(conv_param.filter_spatial_lengths_, begin(filter_spatial_lengths));
+    range_copy(conv_param.output_spatial_lengths_, begin(output_spatial_lengths));
+    range_copy(conv_param.conv_filter_strides_, begin(conv_filter_strides));
+    range_copy(conv_param.conv_filter_dilations_, begin(conv_filter_dilations));
+    range_copy(conv_param.input_left_pads_, begin(input_left_pads));
+    range_copy(conv_param.input_right_pads_, begin(input_right_pads));
+
     for(auto& op_ptr : op_ptrs)
     {
         auto argument_ptr =
             op_ptr->MakeArgumentPointer(static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
                                         static_cast<WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
                                         static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
+                                        conv_param.G_,
                                         conv_param.N_,
                                         conv_param.K_,
                                         conv_param.C_,
-                                        conv_param.input_spatial_lengths_,
-                                        conv_param.filter_spatial_lengths_,
-                                        conv_param.output_spatial_lengths_,
-                                        conv_param.conv_filter_strides_,
-                                        conv_param.conv_filter_dilations_,
-                                        conv_param.input_left_pads_,
-                                        conv_param.input_right_pads_,
+                                        input_spatial_lengths,
+                                        filter_spatial_lengths,
+                                        output_spatial_lengths,
+                                        conv_filter_strides,
+                                        conv_filter_dilations,
+                                        input_left_pads,
+                                        input_right_pads,
                                         in_element_op,
                                         wei_element_op,
                                         out_element_op,
@@ -217,33 +209,29 @@ bool profile_conv_bwd_weight_impl(int do_verification,
             {
                 wei_device_buf.FromDevice(weight_device_result.mData.data());
 
-                bool pass =
-                    ck::utils::check_err(weight_host_result.mData, weight_device_result.mData);
+                bool pass = ck::utils::check_err(weight_device_result, weight_host_result);
 
                 if(!pass)
                 {
-                    std::cout << "Fail info:" << op_ptr->GetTypeString() << std::endl;
+                    std::cout << "Fail info: " << op_ptr->GetTypeString() << std::endl;
                 }
 
                 all_pass &= pass;
 
                 if(do_log)
                 {
-                    std::cout << "in : ";
-                    show_data_nhwc_layout(output);
-                    std::cout << std::endl;
-
-                    std::cout << "wei: ";
-                    show_data_nhwc_layout(weight_host_result);
-                    std::cout << std::endl;
-
-                    std::cout << "out  : ";
-                    show_data_nhwc_layout(input);
-                    std::cout << std::endl;
-
-                    std::cout << "wei_device: ";
-                    show_data_nhwc_layout(weight_device_result);
-                    std::cout << std::endl;
+                    LogRangeAsType<float>(std::cout << "output : ", output.mData, ",") << std::endl;
+                    ;
+                    LogRangeAsType<float>(
+                        std::cout << "weight (device): ", weight_device_result.mData, ",")
+                        << std::endl;
+                    ;
+                    LogRangeAsType<float>(
+                        std::cout << "weight (host): ", weight_host_result.mData, ",")
+                        << std::endl;
+                    ;
+                    LogRangeAsType<float>(std::cout << "input: ", input.mData, ",") << std::endl;
+                    ;
                 }
             }
         }

profiler/include/profile_grouped_conv_fwd_impl.hpp → profiler/include/profiler/profile_grouped_conv_fwd_impl.hpp

@@ -9,14 +9,11 @@
 
 #include "ck/ck.hpp"
 #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
-#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_d.hpp"
 #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
 
 #include "ck/library/tensor_operation_instance/gpu/grouped_convolution_forward.hpp"
 
-#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd.hpp"
-#include "ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_dl.hpp"
-
+#include "ck/library/utility/algorithm.hpp"
 #include "ck/library/utility/check_err.hpp"
 #include "ck/library/utility/device_memory.hpp"
 #include "ck/library/utility/host_tensor.hpp"
@@ -69,7 +66,7 @@ bool profile_grouped_conv_fwd_impl(int do_verification,
     std::array<ck::index_t, NDimSpatial> input_left_pads{};
     std::array<ck::index_t, NDimSpatial> input_right_pads{};
 
-    auto copy = [](auto& x, auto& y) { std::copy(x.begin(), x.end(), y.begin()); };
+    auto copy = [](const auto& x, auto& y) { ck::ranges::copy(x, y.begin()); };
 
     copy(in_g_n_c_wis_desc.GetLengths(), a_g_n_c_wis_lengths);
     copy(in_g_n_c_wis_desc.GetStrides(), a_g_n_c_wis_strides);
@@ -182,7 +179,7 @@ bool profile_grouped_conv_fwd_impl(int do_verification,
             {
                 out_device_buf.FromDevice(device_output.mData.data());
 
-                pass = pass & ck::utils::check_err(device_output.mData, host_output.mData);
+                pass = pass & ck::utils::check_err(device_output, host_output);
 
                 if(do_log)
                 {
@@ -201,92 +198,48 @@ bool profile_grouped_conv_fwd_impl(int do_verification,
         }
     };
 
-    // xdl
-    {
-        using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<NDimSpatial,
-                                                                                     InLayout,
-                                                                                     WeiLayout,
-                                                                                     ck::Tuple<>,
-                                                                                     OutLayout,
-                                                                                     InDataType,
-                                                                                     WeiDataType,
-                                                                                     ck::Tuple<>,
-                                                                                     OutDataType,
-                                                                                     InElementOp,
-                                                                                     WeiElementOp,
-                                                                                     OutElementOp>;
-
-        // get device op instances
-        const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
-            DeviceOp>::GetInstances();
-
-        std::cout << "xdl found " << op_ptrs.size() << " instances" << std::endl;
-
-        for(auto& op_ptr : op_ptrs)
-        {
-            auto argument_ptr = op_ptr->MakeArgumentPointer(
-                in_device_buf.GetDeviceBuffer(),
-                wei_device_buf.GetDeviceBuffer(),
-                std::array<const void*, 0>{},
-                out_device_buf.GetDeviceBuffer(),
-                a_g_n_c_wis_lengths,
-                a_g_n_c_wis_strides,
-                b_g_k_c_xs_lengths,
-                b_g_k_c_xs_strides,
-                std::array<std::array<ck::index_t, NDimSpatial + 3>, 0>{{}},
-                std::array<std::array<ck::index_t, NDimSpatial + 3>, 0>{{}},
-                e_g_n_k_wos_lengths,
-                e_g_n_k_wos_strides,
-                conv_filter_strides,
-                conv_filter_dilations,
-                input_left_pads,
-                input_right_pads,
-                in_element_op,
-                wei_element_op,
-                out_element_op);
-
-            run_impl(op_ptr, argument_ptr);
-        }
-    }
-
-    // dl
+    using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<NDimSpatial,
+                                                                                 InLayout,
+                                                                                 WeiLayout,
+                                                                                 ck::Tuple<>,
+                                                                                 OutLayout,
+                                                                                 InDataType,
+                                                                                 WeiDataType,
+                                                                                 ck::Tuple<>,
+                                                                                 OutDataType,
+                                                                                 InElementOp,
+                                                                                 WeiElementOp,
+                                                                                 OutElementOp>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "xdl found " << op_ptrs.size() << " instances" << std::endl;
+
+    for(auto& op_ptr : op_ptrs)
     {
-        using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwd<NDimSpatial,
-                                                                            InLayout,
-                                                                            WeiLayout,
-                                                                            OutLayout,
-                                                                            InDataType,
-                                                                            WeiDataType,
-                                                                            OutDataType,
-                                                                            InElementOp,
-                                                                            WeiElementOp,
-                                                                            OutElementOp>;
-
-        const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
-            DeviceOp>::GetInstances();
-        std::cout << "dl found " << op_ptrs.size() << " instances" << std::endl;
-
-        for(auto& op_ptr : op_ptrs)
-        {
-            auto argument_ptr = op_ptr->MakeArgumentPointer(in_device_buf.GetDeviceBuffer(),
-                                                            wei_device_buf.GetDeviceBuffer(),
-                                                            out_device_buf.GetDeviceBuffer(),
-                                                            a_g_n_c_wis_lengths,
-                                                            a_g_n_c_wis_strides,
-                                                            b_g_k_c_xs_lengths,
-                                                            b_g_k_c_xs_strides,
-                                                            e_g_n_k_wos_lengths,
-                                                            e_g_n_k_wos_strides,
-                                                            conv_filter_strides,
-                                                            conv_filter_dilations,
-                                                            input_left_pads,
-                                                            input_right_pads,
-                                                            in_element_op,
-                                                            wei_element_op,
-                                                            out_element_op);
-
-            run_impl(op_ptr, argument_ptr);
-        }
+        auto argument_ptr = op_ptr->MakeArgumentPointer(in_device_buf.GetDeviceBuffer(),
+                                                        wei_device_buf.GetDeviceBuffer(),
+                                                        {},
+                                                        out_device_buf.GetDeviceBuffer(),
+                                                        a_g_n_c_wis_lengths,
+                                                        a_g_n_c_wis_strides,
+                                                        b_g_k_c_xs_lengths,
+                                                        b_g_k_c_xs_strides,
+                                                        {},
+                                                        {},
+                                                        e_g_n_k_wos_lengths,
+                                                        e_g_n_k_wos_strides,
+                                                        conv_filter_strides,
+                                                        conv_filter_dilations,
+                                                        input_left_pads,
+                                                        input_right_pads,
+                                                        in_element_op,
+                                                        wei_element_op,
+                                                        out_element_op);
+
+        run_impl(op_ptr, argument_ptr);
     }
 
     std::cout << "Best configuration parameters:"

profiler/include/profile_grouped_gemm_impl.hpp → profiler/include/profiler/profile_grouped_gemm_impl.hpp

@@ -17,6 +17,7 @@
 #include "ck/library/utility/device_memory.hpp"
 #include "ck/library/utility/host_tensor.hpp"
 #include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
 #include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
 
 namespace ck {
@@ -45,15 +46,15 @@ bool profile_grouped_gemm_impl(int do_verification,
 
     auto f_host_tensor_descriptor =
         [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
             if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
             {
-                return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
-                                            std::vector<std::size_t>({stride, 1}));
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
             }
             else
             {
-                return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
-                                            std::vector<std::size_t>({1, stride}));
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
             }
         };
 
@@ -257,8 +258,7 @@ bool profile_grouped_gemm_impl(int do_verification,
                                                               c_element_op);
 
                     ref_invoker.Run(ref_argument);
-                    pass = pass && ck::utils::check_err(c_m_n_device_results[i].mData,
-                                                        c_m_n_host_result.mData);
+                    pass = pass && ck::utils::check_err(c_m_n_device_results[i], c_m_n_host_result);
 
                     if(do_log)
                     {

profiler/include/profile_groupnorm_impl.hpp → profiler/include/profiler/profile_groupnorm_impl.hpp

@@ -165,8 +165,7 @@ bool profile_groupnorm_impl(int do_verification,
         {
             y_dev.FromDevice(y.mData.data());
 
-            bool pass =
-                ck::utils::check_err(y.mData, host_y.mData, "Error: Incorrect results", 1e-3, 1e-3);
+            bool pass = ck::utils::check_err(y, host_y, "Error: Incorrect results", 1e-3, 1e-3);
 
             if(do_log)
             {

profiler/include/profile_reduce_impl.hpp → profiler/include/profiler/profile_reduce_impl.hpp

@@ -6,8 +6,9 @@
 #include "ck/utility/reduction_enums.hpp"
 #include "ck/tensor_operation/gpu/device/device_reduce.hpp"
 
-#include "ck/library/utility/check_err.hpp"
 #include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance.hpp"
+#include "ck/library/utility/algorithm.hpp"
+#include "ck/library/utility/check_err.hpp"
 #include "ck/library/utility/device_memory.hpp"
 #include "ck/library/utility/host_reduction.hpp"
 #include "ck/library/utility/host_common_util.hpp"
@@ -359,10 +360,10 @@ bool profile_reduce_impl_impl(bool do_verification,
         std::array<index_t, NumOutDim> arrOutLengths;
         std::array<index_t, NumOutDim> arrOutStrides;
 
-        std::copy(inLengths.begin(), inLengths.end(), arrInLengths.begin());
-        std::copy(inStrides.begin(), inStrides.end(), arrInStrides.begin());
-        std::copy(outLengths.begin(), outLengths.end(), arrOutLengths.begin());
-        std::copy(outStrides.begin(), outStrides.end(), arrOutStrides.begin());
+        ck::ranges::copy(inLengths, arrInLengths.begin());
+        ck::ranges::copy(inStrides, arrInStrides.begin());
+        ck::ranges::copy(outLengths, arrOutLengths.begin());
+        ck::ranges::copy(outStrides, arrOutStrides.begin());
 
         for(auto& reduce_ptr : reduce_ptrs)
         {
@@ -411,13 +412,12 @@ bool profile_reduce_impl_impl(bool do_verification,
                 bool single_pass;
 
                 out_dev.FromDevice(out.mData.data());
-                single_pass = ck::utils::check_err(out.mData, out_ref.mData);
+                single_pass = ck::utils::check_err(out, out_ref);
 
                 if(OutputIndex)
                 {
                     out_indices_dev.FromDevice(out_indices.mData.data());
-                    single_pass = single_pass &&
-                                  ck::utils::check_err(out_indices.mData, out_indices_ref.mData);
+                    single_pass = single_pass && ck::utils::check_err(out_indices, out_indices_ref);
                 };
 
                 if(!single_pass)
@@ -492,7 +492,7 @@ bool profile_reduce_impl(bool do_verification,
 
         std::array<ck::index_t, descType::NumReduceDim_> arrReduceDims;
 
-        std::copy(reduceDims.begin(), reduceDims.end(), arrReduceDims.begin());
+        ck::ranges::copy(reduceDims, arrReduceDims.begin());
 
         pass = pass && profile_reduce_impl_impl<InDataType,
                                                 AccDataType,

profiler/src/CMakeLists.txt

+# ckProfiler
+set(PROFILER_SOURCES
+    profiler.cpp
+    profile_gemm.cpp
+    profile_gemm_splitk.cpp
+    profile_gemm_bilinear.cpp
+    profile_gemm_bias_add_reduce.cpp
+    profile_gemm_add_add_fastgelu.cpp
+    profile_gemm_reduce.cpp
+    profile_batched_gemm.cpp
+    profile_batched_gemm_gemm.cpp
+    profile_batched_gemm_add_relu_gemm_add.cpp
+    profile_batched_gemm_reduce.cpp
+    profile_grouped_gemm.cpp
+    profile_conv_fwd.cpp
+    profile_conv_fwd_bias_relu.cpp
+    profile_conv_fwd_bias_relu_add.cpp
+    profile_conv_bwd_data.cpp
+    profile_grouped_conv_fwd.cpp
+    profile_grouped_conv_bwd_weight.cpp
+    profile_reduce.cpp
+    profile_groupnorm.cpp
+    profile_layernorm.cpp
+    profile_softmax.cpp
+    profile_batchnorm_fwd.cpp
+    profile_batchnorm_bwd.cpp
+)
+
+set(PROFILER_EXECUTABLE ckProfiler)
+
+add_executable(${PROFILER_EXECUTABLE} ${PROFILER_SOURCES})
+target_compile_options(${PROFILER_EXECUTABLE} PRIVATE -Wno-global-constructors)
+
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE utility)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_splitk_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_bilinear_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_add_fastgelu_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_reduce_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_bias_add_reduce_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_gemm_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_add_relu_gemm_add_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_reduce_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv1d_fwd_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_fwd_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_fwd_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv1d_bwd_data_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_bwd_data_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv3d_bwd_data_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv1d_bwd_weight_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_bwd_weight_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_bwd_weight_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_bias_relu_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_conv2d_fwd_bias_relu_add_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_normalization_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_softmax_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_reduce_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batchnorm_instance)
+
+rocm_install(TARGETS ${PROFILER_EXECUTABLE} COMPONENT profiler)

profiler/src/profile_batched_gemm.cpp

@@ -7,7 +7,8 @@
 #include <initializer_list>
 #include <cstdlib>
 
-#include "profiler/include/profile_batched_gemm_impl.hpp"
+#include "profiler/profile_batched_gemm_impl.hpp"
+#include "profiler_operation_registry.hpp"
 
 enum struct GemmMatrixLayout
 {
@@ -25,12 +26,15 @@ enum struct GemmDataType
     INT8_INT8_INT8, // 3
 };
 
+#define OP_NAME "batched_gemm"
+#define OP_DESC "Batched GEMM"
+
 int profile_batched_gemm(int argc, char* argv[])
 {
     if(argc != 18)
     {
         // clang-format off
-        printf("arg1: tensor operation (batched_gemm: Batched GEMM)\n");
+        printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
         printf("arg2: data type (0: fp32; 1: fp16, 2: bf16, 3: int8)\n");
         printf("arg3: matrix layout (0: A[g, m, k] * B[g, k, n] = C[g, m, n];\n");
         printf("                     1: A[g, m, k] * B[g, n, k] = C[g, m, n];\n");
@@ -195,3 +199,5 @@ int profile_batched_gemm(int argc, char* argv[])
         return 1;
     }
 }
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_batched_gemm);

profiler/src/profile_batched_gemm_add_relu_gemm_add.cpp

@@ -6,7 +6,8 @@
 #include <initializer_list>
 #include <cstdlib>
 
-#include "profiler/include/profile_batched_gemm_add_relu_gemm_add_impl.hpp"
+#include "profiler/profile_batched_gemm_add_relu_gemm_add_impl.hpp"
+#include "profiler_operation_registry.hpp"
 
 using F16 = ck::half_t;
 using F32 = float;
@@ -14,6 +15,9 @@ using F32 = float;
 using Row = ck::tensor_layout::gemm::RowMajor;
 using Col = ck::tensor_layout::gemm::ColumnMajor;
 
+#define OP_NAME "batched_gemm_add_relu_gemm_add"
+#define OP_DESC "Batched GEMM+Add+Relu+GEMM+Add"
+
 int profile_batched_gemm_add_relu_gemm_add(int argc, char* argv[])
 {
     enum struct GemmMatrixLayout
@@ -109,8 +113,7 @@ int profile_batched_gemm_add_relu_gemm_add(int argc, char* argv[])
     }
     else
     {
-        printf("arg1: tensor operation (batched_gemm_add_relu_gemm_add: "
-               "Batched_GEMM+Add+Relu+Gemm+Add)\n");
+        printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
         printf("arg2: data type (1: fp16)\n");
         printf("arg3: matrix layout (0: Relu(A0[m, k] * B0[n, k] + D0[m, n]) * B1[n, o] + D1[m, o] "
                "= E1[m, o]; 1: Relu(A0[m, k] * B0[n, k] + D0[m, n]) * B1[o, n] + D1[m, o] = "
@@ -207,3 +210,5 @@ int profile_batched_gemm_add_relu_gemm_add(int argc, char* argv[])
 
     return 0;
 }
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_batched_gemm_add_relu_gemm_add);

profiler/src/profile_batched_gemm_gemm.cpp

@@ -6,7 +6,8 @@
 #include <initializer_list>
 #include <cstdlib>
 
-#include "profiler/include/profile_batched_gemm_gemm_impl.hpp"
+#include "profiler/profile_batched_gemm_gemm_impl.hpp"
+#include "profiler_operation_registry.hpp"
 
 using F16 = ck::half_t;
 using F32 = float;
@@ -14,6 +15,9 @@ using F32 = float;
 using Row = ck::tensor_layout::gemm::RowMajor;
 using Col = ck::tensor_layout::gemm::ColumnMajor;
 
+#define OP_NAME "batched_gemm_gemm"
+#define OP_DESC "Batched GEMM+GEMM"
+
 int profile_batched_gemm_gemm(int argc, char* argv[])
 {
     enum struct GemmMatrixLayout
@@ -101,7 +105,7 @@ int profile_batched_gemm_gemm(int argc, char* argv[])
     }
     else
     {
-        printf("arg1: tensor operation (batched_gemm_gemm: Batched_GEMM+Gemm)\n");
+        printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
         printf("arg2: data type (1: fp16)\n");
         printf("arg3: matrix layout (0: Relu(A0[m, k] * B0[n, k] + D0[m, n]) * B1[n, o] + D1[m, o] "
                "= E1[m, o];  1: Relu(A0[m, k] * B0[n, k] + D0[m, n]) * B1[o, n] + D1[m, o] = E1[m, "
@@ -179,3 +183,5 @@ int profile_batched_gemm_gemm(int argc, char* argv[])
 
     return 0;
 }
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_batched_gemm_gemm);

profiler/src/profile_batched_gemm_reduce.cpp

@@ -6,7 +6,11 @@
 #include <initializer_list>
 #include <cstdlib>
 
-#include "profiler/include/profile_batched_gemm_reduce_impl.hpp"
+#include "profiler/profile_batched_gemm_reduce_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+#define OP_NAME "batched_gemm_reduce"
+#define OP_DESC "Batched GEMM+Reduce"
 
 int profile_batched_gemm_reduce(int argc, char* argv[])
 {
@@ -26,7 +30,7 @@ int profile_batched_gemm_reduce(int argc, char* argv[])
 
     if(argc != 15)
     {
-        printf("arg1: tensor operation (batched_gemm_reduce: BatchedGEMM+Reduce)\n");
+        printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
         printf("arg2: data type (0: fp32; 1: fp16)\n");
         printf("arg3: matrix layout (0: A[m, k] * B[k, n] = C[m, n];\n");
         printf("                     1: A[m, k] * B[n, k] = C[m, n];\n");
@@ -151,3 +155,5 @@ int profile_batched_gemm_reduce(int argc, char* argv[])
 
     return 0;
 }
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_batched_gemm_reduce);

profiler/src/profile_batchnorm_bwd.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <vector>
+#include <getopt.h>
+
+#include "ck/library/utility/host_common_util.hpp"
+#include "profiler/profile_batchnorm_backward_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+using ck::index_t;
+
+using namespace std;
+
+static const struct option long_options[] = {{"inOutLengths", required_argument, nullptr, 'D'},
+                                             {"reduceDims", required_argument, nullptr, 'R'},
+                                             {"dumpout", required_argument, nullptr, 'o'},
+                                             {"verify", required_argument, nullptr, 'v'},
+                                             {"help", no_argument, nullptr, '?'},
+                                             {nullptr, 0, nullptr, 0}};
+
+class BatchnormBwdArgParser
+{
+    private:
+    int option_index = 0;
+
+    public:
+    std::vector<size_t> inLengths;
+    std::vector<int> reduceDims;
+
+    bool do_verification = false;
+    bool do_dumpout      = false;
+
+    bool haveSavedMeanInvVar;
+
+    int data_type    = 0;
+    int init_method  = 2;
+    bool time_kernel = false;
+
+    BatchnormBwdArgParser()  = default;
+    ~BatchnormBwdArgParser() = default;
+
+    void show_usage(const char* cmd)
+    {
+        // clang-format off
+        std::cout << "Usage of " << cmd << std::endl;
+        std::cout << "--inOutLengths or -D, comma separated list of input tensor dimension lengths, must have 4 integers for nhwc" << std::endl;
+        std::cout << "--reduceDims or -R, comma separated list of dimensions to reduce on" << std::endl;  
+        std::cout << "--verify or -v, 1/0 to indicate whether to verify the result by comparing with the host-based batch-normalization" << std::endl;
+        std::cout << "Arg1: data type (0: fp16, 1: fp32, 5: bp16, 6: fp64)" << std::endl;
+        std::cout << "Arg2 -- 1/0 to indicate whether to use saved mean and invVariance" << std::endl;
+        std::cout << "Arg3 -- init method used for dy and bnScale (0=no init, 1=single integer value, 2=scope integer value, 3=decimal value)" << std::endl;
+        std::cout << "Arg4 -- time kernel (0=no, 1=yes)" << std::endl;
+        // clang-format on
+    };
+
+    int operator()(int argc, char* argv[])
+    {
+        using ck::host_common::getTypeValuesFromString;
+
+        int ch;
+
+        optind++; // to skip the module name
+
+        while(1)
+        {
+            ch = getopt_long(argc, argv, "D:R:v:o:", long_options, &option_index);
+            if(ch == -1)
+                break;
+            switch(ch)
+            {
+            case 'D':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                inLengths = getTypeValuesFromString<size_t>(optarg);
+                break;
+            case 'R':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                reduceDims = getTypeValuesFromString<int>(optarg);
+                break;
+            case 'v':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                do_verification = static_cast<bool>(std::atoi(optarg));
+                break;
+            case 'o':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                do_dumpout = static_cast<bool>(std::atoi(optarg));
+                break;
+            case '?':
+                if(std::string(long_options[option_index].name) == "help")
+                {
+                    show_usage(argv[0]);
+                    return -1;
+                };
+                break;
+
+            default:
+                show_usage(argv[0]);
+                std::cerr << "Invalid cmd-line options!" << std::endl;
+                return -1;
+            };
+        };
+
+        if(optind + 4 > argc)
+            throw std::runtime_error("Invalid cmd-line arguments, more argumetns are needed!");
+
+        data_type           = std::atoi(argv[optind++]);
+        haveSavedMeanInvVar = std::atoi(argv[optind++]);
+        init_method         = std::atoi(argv[optind++]);
+        time_kernel         = static_cast<bool>(std::atoi(argv[optind++]));
+
+        if(data_type != 0 && data_type != 1 && data_type != 3 && data_type != 5 && data_type != 6)
+            return -1;
+
+        return 0;
+    };
+}; // end of class AppArgs
+
+static const double epsilon = std::numeric_limits<float>::epsilon();
+
+int profile_batchnorm_backward(int argc, char* argv[])
+{
+    using ck::profiler::profile_batchnorm_backward_impl;
+
+    BatchnormBwdArgParser arg_parser;
+
+    if(arg_parser(argc, argv) != 0)
+        return -1;
+
+    using F16  = ck::half_t;
+    using F32  = float;
+    using BF16 = ck::bhalf_t;
+    using F64  = double;
+
+    if(arg_parser.data_type == 0)
+    {
+        if(arg_parser.inLengths.size() == 4 && arg_parser.reduceDims.size() == 3)
+        {
+            profile_batchnorm_backward_impl<F16, F32, F32, F32, F16, F32, F32, 4, 3>(
+                arg_parser.do_verification,
+                arg_parser.init_method,
+                arg_parser.do_dumpout,
+                arg_parser.time_kernel,
+                arg_parser.inLengths,
+                arg_parser.reduceDims,
+                arg_parser.haveSavedMeanInvVar,
+                epsilon);
+        };
+    }
+    else if(arg_parser.data_type == 1)
+    {
+        if(arg_parser.inLengths.size() == 4 && arg_parser.reduceDims.size() == 3)
+        {
+            profile_batchnorm_backward_impl<F32, F32, F32, F32, F32, F32, F32, 4, 3>(
+                arg_parser.do_verification,
+                arg_parser.init_method,
+                arg_parser.do_dumpout,
+                arg_parser.time_kernel,
+                arg_parser.inLengths,
+                arg_parser.reduceDims,
+                arg_parser.haveSavedMeanInvVar,
+                epsilon);
+        };
+    }
+    else if(arg_parser.data_type == 5)
+    {
+        if(arg_parser.inLengths.size() == 4 && arg_parser.reduceDims.size() == 3)
+        {
+            profile_batchnorm_backward_impl<BF16, F32, F32, F32, BF16, F32, F32, 4, 3>(
+                arg_parser.do_verification,
+                arg_parser.init_method,
+                arg_parser.do_dumpout,
+                arg_parser.time_kernel,
+                arg_parser.inLengths,
+                arg_parser.reduceDims,
+                arg_parser.haveSavedMeanInvVar,
+                epsilon);
+        };
+    }
+    else if(arg_parser.data_type == 6)
+    {
+        if(arg_parser.inLengths.size() == 4 && arg_parser.reduceDims.size() == 3)
+        {
+            profile_batchnorm_backward_impl<F64, F64, F64, F64, F64, F64, F64, 4, 3>(
+                arg_parser.do_verification,
+                arg_parser.init_method,
+                arg_parser.do_dumpout,
+                arg_parser.time_kernel,
+                arg_parser.inLengths,
+                arg_parser.reduceDims,
+                arg_parser.haveSavedMeanInvVar,
+                epsilon);
+        };
+    }
+
+    return 0;
+}
+
+REGISTER_PROFILER_OPERATION("bnorm_bwd", "Batchnorm backward", profile_batchnorm_backward);

profiler/src/profile_batchnorm_fwd.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <vector>
+#include <getopt.h>
+
+#include "ck/library/utility/host_common_util.hpp"
+#include "profiler/profile_batchnorm_forward_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+using ck::index_t;
+
+using namespace std;
+
+static const struct option long_options[] = {{"inOutLengths", required_argument, nullptr, 'D'},
+                                             {"reduceDims", required_argument, nullptr, 'R'},
+                                             {"dumpout", required_argument, nullptr, 'o'},
+                                             {"verify", required_argument, nullptr, 'v'},
+                                             {"help", no_argument, nullptr, '?'},
+                                             {nullptr, 0, nullptr, 0}};
+
+class BatchnormFwdArgParser
+{
+    private:
+    int option_index = 0;
+
+    public:
+    std::vector<size_t> inLengths;
+    std::vector<int> reduceDims;
+
+    bool do_verification = false;
+    bool do_dumpout      = false;
+
+    bool updateMovingAverage;
+    bool saveMeanAndInvVariance;
+
+    int data_type    = 0;
+    int init_method  = 2;
+    bool time_kernel = false;
+
+    BatchnormFwdArgParser()  = default;
+    ~BatchnormFwdArgParser() = default;
+
+    void show_usage(const char* cmd)
+    {
+        // clang-format off
+        std::cout << "Usage of " << cmd << std::endl;
+        std::cout << "--inOutLengths or -D, comma separated list of input tensor dimension lengths, must have 4 integers for nhwc" << std::endl;
+        std::cout << "--reduceDims or -R, comma separated list of dimensions to reduce on" << std::endl;  
+        std::cout << "--verify or -v, 1/0 to indicate whether to verify the result by comparing with the host-based batch-normalization" << std::endl;
+        std::cout << "Arg1: data type (0: fp16, 1: fp32, 5: bp16, 6: fp64)" << std::endl;
+        std::cout << "Arg2: 1/0 to indicate whether to update the moving average and variance (0=no, 1=yes)" << std::endl;
+        std::cout << "Arg3: 1/0 to indicate whether to save the calculated mean and invVariance (0=no, 1=yes)" << std::endl;
+        std::cout << "Arg4: init method used for bnScale and bnBias (0=no init, 1=single integer value, 2=scope integer value, 3=decimal value)" << std::endl;
+        std::cout << "Arg5: time kernel (0=no, 1=yes)" << std::endl;
+        // clang-format on
+    };
+
+    int operator()(int argc, char* argv[])
+    {
+        using ck::host_common::getTypeValuesFromString;
+
+        int ch;
+
+        optind++; // to skip the module name
+
+        while(1)
+        {
+            ch = getopt_long(argc, argv, "D:R:v:o:", long_options, &option_index);
+            if(ch == -1)
+                break;
+            switch(ch)
+            {
+            case 'D':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                inLengths = getTypeValuesFromString<size_t>(optarg);
+                break;
+            case 'R':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                reduceDims = getTypeValuesFromString<int>(optarg);
+                break;
+            case 'v':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                do_verification = static_cast<bool>(std::atoi(optarg));
+                break;
+            case 'o':
+                if(!optarg)
+                    throw std::runtime_error("Invalid option format!");
+
+                do_dumpout = static_cast<bool>(std::atoi(optarg));
+                break;
+            case '?':
+                if(std::string(long_options[option_index].name) == "help")
+                {
+                    show_usage(argv[0]);
+                    return -1;
+                };
+                break;
+
+            default:
+                show_usage(argv[0]);
+                std::cerr << "Invalid cmd-line options!" << std::endl;
+                return -1;
+            };
+        };
+
+        if(optind + 5 > argc)
+            throw std::runtime_error("Invalid cmd-line arguments, more argumetns are needed!");
+
+        data_type              = std::atoi(argv[optind++]);
+        updateMovingAverage    = std::atoi(argv[optind++]);
+        saveMeanAndInvVariance = std::atoi(argv[optind++]);
+        init_method            = std::atoi(argv[optind++]);
+        time_kernel            = static_cast<bool>(std::atoi(argv[optind++]));
+
+        if(data_type != 0 && data_type != 1 && data_type != 3 && data_type != 5 && data_type != 6)
+            return -1;
+
+        return 0;
+    };
+}; // end of class AppArgs
+
+static const double epsilon       = std::numeric_limits<float>::epsilon();
+static const double averageFactor = 0.1;
+
+int profile_batchnorm_forward(int argc, char* argv[])
+{
+    using ck::profiler::profile_batchnorm_forward_impl;
+
+    BatchnormFwdArgParser arg_parser;
+
+    if(arg_parser(argc, argv) != 0)
+        return -1;
+
+    using F16  = ck::half_t;
+    using F32  = float;
+    using BF16 = ck::bhalf_t;
+    using F64  = double;
+
+    if(arg_parser.data_type == 0)
+    {
+        if(arg_parser.inLengths.size() == 4 && arg_parser.reduceDims.size() == 3)
+        {
+            profile_batchnorm_forward_impl<F16, F16, F32, F16, F16, F16, 4, 3>(
+                arg_parser.do_verification,
+                arg_parser.init_method,
+                arg_parser.do_dumpout,
+                arg_parser.time_kernel,
+                arg_parser.inLengths,
+                arg_parser.reduceDims,
+                arg_parser.updateMovingAverage,
+                arg_parser.saveMeanAndInvVariance,
+                epsilon,
+                averageFactor);
+        };
+    }
+    else if(arg_parser.data_type == 1)
+    {
+        if(arg_parser.inLengths.size() == 4 && arg_parser.reduceDims.size() == 3)
+        {
+            profile_batchnorm_forward_impl<F32, F32, F32, F32, F32, F32, 4, 3>(
+                arg_parser.do_verification,
+                arg_parser.init_method,
+                arg_parser.do_dumpout,
+                arg_parser.time_kernel,
+                arg_parser.inLengths,
+                arg_parser.reduceDims,
+                arg_parser.updateMovingAverage,
+                arg_parser.saveMeanAndInvVariance,
+                epsilon,
+                averageFactor);
+        };
+    }
+    else if(arg_parser.data_type == 5)
+    {
+        if(arg_parser.inLengths.size() == 4 && arg_parser.reduceDims.size() == 3)
+        {
+            profile_batchnorm_forward_impl<BF16, BF16, F32, BF16, BF16, F32, 4, 3>(
+                arg_parser.do_verification,
+                arg_parser.init_method,
+                arg_parser.do_dumpout,
+                arg_parser.time_kernel,
+                arg_parser.inLengths,
+                arg_parser.reduceDims,
+                arg_parser.updateMovingAverage,
+                arg_parser.saveMeanAndInvVariance,
+                epsilon,
+                averageFactor);
+        };
+    }
+    else if(arg_parser.data_type == 6)
+    {
+        if(arg_parser.inLengths.size() == 4 && arg_parser.reduceDims.size() == 3)
+        {
+            profile_batchnorm_forward_impl<F64, F64, F64, F64, F64, F64, 4, 3>(
+                arg_parser.do_verification,
+                arg_parser.init_method,
+                arg_parser.do_dumpout,
+                arg_parser.time_kernel,
+                arg_parser.inLengths,
+                arg_parser.reduceDims,
+                arg_parser.updateMovingAverage,
+                arg_parser.saveMeanAndInvVariance,
+                epsilon,
+                averageFactor);
+        };
+    }
+
+    return 0;
+}
+
+REGISTER_PROFILER_OPERATION("bnorm_fwd", "Batchnorm forward", profile_batchnorm_forward);

profiler/src/profile_conv_bwd_data.cpp

@@ -6,7 +6,8 @@
 #include <initializer_list>
 #include <cstdlib>
 
-#include "profiler/include/profile_conv_bwd_data_impl.hpp"
+#include "profiler/profile_conv_bwd_data_impl.hpp"
+#include "profiler_operation_registry.hpp"
 
 namespace {
 
@@ -24,10 +25,13 @@ enum struct ConvDataType
     INT8_INT8_INT8, // 3
 };
 
+#define OP_NAME "conv_bwd_data"
+#define OP_DESC "Convolution Backward Data"
+
 static void print_helper_msg()
 {
     std::cout
-        << "arg1: tensor operation (conv_bwd_data: Convolution Backward Data)\n"
+        << "arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n"
         << "arg2: data type (0: Input fp32, Weight fp32, Output fp32\n"
         << "                 1: Input fp16, Weight fp16, Output fp16\n"
         << "                 2: Input bf16, Weight bf16, Output bf16\n"
@@ -182,3 +186,5 @@ int profile_conv_bwd_data(int argc, char* argv[])
 
     return 1;
 }
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_conv_bwd_data);

profiler/src/profile_conv_fwd.cpp

@@ -6,7 +6,8 @@
 #include <initializer_list>
 #include <cstdlib>
 
-#include "profiler/include/profile_conv_fwd_impl.hpp"
+#include "profiler/profile_conv_fwd_impl.hpp"
+#include "profiler_operation_registry.hpp"
 
 namespace {
 
@@ -24,11 +25,14 @@ enum struct ConvDataType
     INT8_INT8_INT8, // 3
 };
 
+#define OP_NAME "conv_fwd"
+#define OP_DESC "Convolution Forward"
+
 static void print_helper_msg()
 {
     std::cout
         // clang-format-off
-        << "arg1: tensor operation (conv_fwd: Convolution Forward)\n"
+        << "arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n"
         << "arg2: data type (0: Input fp32, Weight fp32, Output fp32\n"
         << "                 1: Input fp16, Weight fp16, Output fp16\n"
         << "                 2: Input bf16, Weight bf16, Output bf16\n"
@@ -184,3 +188,5 @@ int profile_conv_fwd(int argc, char* argv[])
 
     return 1;
 }
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_conv_fwd);