Merge branch 'develop' into gemm_layernorm_welford

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

@@ -12,7 +12,6 @@
 #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
 
 #include "ck/library/tensor_operation_instance/gpu/grouped_convolution_forward.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"
@@ -199,93 +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)
     {
-        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(),
-                                                            {},
-                                                            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);
-        }
-    }
-
-    // dl
-    {
-        using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwd<NDimSpatial,
-                                                                            InLayout,
-                                                                            WeiLayout,
-                                                                            OutLayout,
-                                                                            InDataType,
-                                                                            WeiDataType,
-                                                                            OutDataType,
-                                                                            InElementOp,
-                                                                            WeiElementOp,
-                                                                            OutElementOp>;
-
-        // get device op instances
-        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/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

@@ -6,7 +6,8 @@
 #include <getopt.h>
 
 #include "ck/library/utility/host_common_util.hpp"
-#include "profiler/include/profile_batchnorm_backward_impl.hpp"
+#include "profiler/profile_batchnorm_backward_impl.hpp"
+#include "profiler_operation_registry.hpp"
 
 using ck::index_t;
 
@@ -202,3 +203,5 @@ int profile_batchnorm_backward(int argc, char* argv[])
 
     return 0;
 }
+
+REGISTER_PROFILER_OPERATION("bnorm_bwd", "Batchnorm backward", profile_batchnorm_backward);

profiler/src/profile_batchnorm_fwd.cpp

@@ -6,7 +6,8 @@
 #include <getopt.h>
 
 #include "ck/library/utility/host_common_util.hpp"
-#include "profiler/include/profile_batchnorm_forward_impl.hpp"
+#include "profiler/profile_batchnorm_forward_impl.hpp"
+#include "profiler_operation_registry.hpp"
 
 using ck::index_t;
 
@@ -214,3 +215,5 @@ int profile_batchnorm_forward(int argc, char* argv[])
 
     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);

profiler/src/profile_conv_fwd_bias_relu.cpp

@@ -6,7 +6,8 @@
 #include <initializer_list>
 #include <cstdlib>
 
-#include "profiler/include/profile_conv_fwd_bias_relu_impl.hpp"
+#include "profiler/profile_conv_fwd_bias_relu_impl.hpp"
+#include "profiler_operation_registry.hpp"
 
 enum struct ConvDataType
 {
@@ -32,11 +33,14 @@ enum struct ConvOutputLayout
     NHWK, // 1
 };
 
+#define OP_NAME "conv_fwd_bias_relu"
+#define OP_DESC "Convolution Forward+Bias+ReLU"
+
 int profile_conv_fwd_bias_relu(int argc, char* argv[])
 {
     if(argc != 25)
     {
-        printf("arg1: tensor operation (conv_fwd_bias_relu: ForwardConvolution+Bias+ReLu)\n");
+        printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
         printf("arg2: data type (0: fp32; 1: fp16)\n");
         printf("arg3: input tensor layout (0: NCHW; 1: NHWC)\n");
         printf("arg4: weight tensor layout (0: KCYX; 1: KYXC)\n");
@@ -114,3 +118,5 @@ int profile_conv_fwd_bias_relu(int argc, char* argv[])
 
     return 0;
 }
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_conv_fwd_bias_relu);

profiler/src/profile_conv_fwd_bias_relu_add.cpp

@@ -6,7 +6,8 @@
 #include <initializer_list>
 #include <cstdlib>
 
-#include "profiler/include/profile_conv_fwd_bias_relu_add_impl.hpp"
+#include "profiler/profile_conv_fwd_bias_relu_add_impl.hpp"
+#include "profiler_operation_registry.hpp"
 
 enum struct ConvDataType
 {
@@ -32,12 +33,14 @@ enum struct ConvOutputLayout
     NHWK, // 1
 };
 
+#define OP_NAME "conv_fwd_bias_relu_add"
+#define OP_DESC "Convolution Forward+Bias+ReLU+Add"
+
 int profile_conv_fwd_bias_relu_add(int argc, char* argv[])
 {
     if(argc != 25)
     {
-        printf(
-            "arg1: tensor operation (conv_fwd_bias_relu_add: ForwardConvolution+Bias+ReLu+Add)\n");
+        printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
         printf("arg2: data type (0: fp32; 1: fp16)\n");
         printf("arg3: input tensor layout (0: NCHW; 1: NHWC)\n");
         printf("arg4: weight tensor layout (0: KCYX; 1: KYXC)\n");
@@ -115,3 +118,5 @@ int profile_conv_fwd_bias_relu_add(int argc, char* argv[])
 
     return 0;
 }
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_conv_fwd_bias_relu_add);

profiler/src/profile_gemm.cpp

@@ -6,7 +6,8 @@
 #include <initializer_list>
 #include <cstdlib>
 
-#include "profiler/include/profile_gemm_impl.hpp"
+#include "profiler/profile_gemm_impl.hpp"
+#include "profiler_operation_registry.hpp"
 
 enum struct GemmMatrixLayout
 {
@@ -24,9 +25,12 @@ enum struct GemmDataType
     INT8_INT8_INT8, // 3
 };
 
+#define OP_NAME "gemm"
+#define OP_DESC "GEMM"
+
 static void print_helper_msg()
 {
-    std::cout << "arg1: tensor operation (gemm: GEMM)\n"
+    std::cout << "arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n"
               << "arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8)\n"
               << "arg3: matrix layout (0: A[m, k] * B[k, n] = C[m, n];\n"
               << "                     1: A[m, k] * B[n, k] = C[m, n];\n"
@@ -184,3 +188,5 @@ int profile_gemm(int argc, char* argv[])
         return 1;
     }
 }
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_gemm);