Merge branch 'develop' into lwpck-471

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"
@@ -199,8 +198,6 @@ bool profile_grouped_conv_fwd_impl(int do_verification,
         }
     };
 
-    // xdl
-    {
     using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD<NDimSpatial,
                                                                                  InLayout,
                                                                                  WeiLayout,
@@ -244,49 +241,6 @@ bool profile_grouped_conv_fwd_impl(int do_verification,
 
         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);
-        }
-    }
 
     std::cout << "Best configuration parameters:"
               << "\nname: " << best_op_name << "\navg_time: " << best_avg_time

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_add_fastgelu.cpp
+    profile_gemm_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_add_fastgelu_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_fastgelu_instance)
+target_link_libraries(${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_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);