support for python apps profiling

src/core/loader.h

@@ -1,9 +1,15 @@
 #ifndef SRC_CORE_LOADER_H_
 #define SRC_CORE_LOADER_H_
 
+#include <atomic>
 #include <mutex>
 #include <dlfcn.h>
 
+#define LOADER_INSTANTIATE() \
+  std::atomic<roctracer::HipLoader*> roctracer::HipLoader::instance_{}; \
+  std::atomic<roctracer::HccLoader*> roctracer::HccLoader::instance_{}; \
+  roctracer::Loader::mutex_t roctracer::Loader::mutex_;
+
 namespace roctracer {
 
 // Base runtime loader class
@@ -95,13 +101,13 @@ class HccLoader : protected Loader {
     return *obj;
   }
 
-  HccLoader() : Loader("libmcwamp.so") {
+  HccLoader() : Loader("libmcwamp_hsa.so") {
     // Kalmar::CLAMP::InitActivityCallback
-    InitActivityCallback = GetFun<InitActivityCallback_t>("_ZN6Kalmar5CLAMP20InitActivityCallbackEPvS1_S1_");
+    InitActivityCallback = GetFun<InitActivityCallback_t>("InitActivityCallbackImpl");
     // Kalmar::CLAMP::EnableActivityIdCallback
-    EnableActivityCallback = GetFun<EnableActivityCallback_t>("_ZN6Kalmar5CLAMP22EnableActivityCallbackEjb");
+    EnableActivityCallback = GetFun<EnableActivityCallback_t>("EnableActivityCallbackImpl");
     // Kalmar::CLAMP::GetCmdName
-    GetCmdName = GetFun<GetCmdName_t>("_ZN6Kalmar5CLAMP10GetCmdNameEj");
+    GetCmdName = GetFun<GetCmdName_t>("GetCmdNameImpl");
   }
 
   InitActivityCallback_t* InitActivityCallback;

src/core/roctracer.cpp

@@ -439,12 +439,10 @@ std::atomic<util::Logger*> util::Logger::instance_{};
 MemoryPool* memory_pool = NULL;
 typedef std::recursive_mutex memory_pool_mutex_t;
 memory_pool_mutex_t memory_pool_mutex;
-
-Loader::mutex_t Loader::mutex_;
-std::atomic<HipLoader*> HipLoader::instance_{};
-std::atomic<HccLoader*> HccLoader::instance_{};
 }
 
+LOADER_INSTANTIATE();
+
 std::atomic<proxy::Tracker*> proxy::Tracker::instance_{};
 proxy::Tracker::mutex_t proxy::Tracker::glob_mutex_;
 proxy::Tracker::counter_t proxy::Tracker::counter_ = 0;

test/run.sh

@@ -45,14 +45,14 @@ eval_test() {
 
 # Standalone test
 # rocTrecer is used explicitely by test
-eval_test "standalone HIP test" "LD_PRELOAD=$HCC_HOME/lib/libmcwamp_hsa.so ./test/MatrixTranspose_test"
+eval_test "standalone HIP test" ./test/MatrixTranspose_test
 
 # Tool test
 # rocTracer/tool is loaded by HSA runtime
-export HSA_TOOLS_LIB="test/libtracer_tool.so libroctracer64.so"
+export HSA_TOOLS_LIB="libroctracer64.so test/libtracer_tool.so"
 
 # HIP test
-eval_test "tool HIP test" "LD_PRELOAD='$HCC_HOME/lib/libmcwamp_hsa.so $HSA_TOOLS_LIB' ./test/MatrixTranspose"
+eval_test "tool HIP test" ./test/MatrixTranspose
 
 # HSA test
 export ROCTRACER_DOMAIN="hsa"
@@ -69,11 +69,11 @@ export ROCP_AGENTS=1
 # each thread creates a queue pre GPU agent
 export ROCP_THRS=1
 
-eval_test "tool HSA test" "LD_PRELOAD='$HSA_TOOLS_LIB' ./test/hsa/ctrl"
+eval_test "tool HSA test" ./test/hsa/ctrl
 
 echo "<trace name=\"HSA\"><parameters api=\"hsa_agent_get_info, hsa_amd_memory_pool_allocate\"></parameters></trace>" > input.xml
 export ROCP_INPUT=input.xml
-eval_test "tool HSA test input" "LD_PRELOAD='$HSA_TOOLS_LIB' ./test/hsa/ctrl"
+eval_test "tool HSA test input" ./test/hsa/ctrl
 
 #valgrind --leak-check=full $tbin
 #valgrind --tool=massif $tbin

test/tool/tracer_tool.cpp

@@ -31,6 +31,7 @@ THE SOFTWARE.
 #include <inc/roctracer_hip.h>
 #include <inc/roctracer_hcc.h>
 #include <inc/ext/hsa_rt_utils.hpp>
+#include <src/core/loader.h>
 #include <util/xml.h>
 
 #define PUBLIC_API __attribute__((visibility("default")))
@@ -54,6 +55,8 @@ thread_local timestamp_t hip_begin_timestamp = 0;
 bool trace_hsa = false;
 bool trace_hip = false;
 
+LOADER_INSTANTIATE();
+
 // Global output file handle
 FILE* hsa_api_file_handle = NULL;
 FILE* hsa_async_copy_file_handle = NULL;
@@ -116,33 +119,40 @@ void hip_api_callback(
     hsa_begin_timestamp = timer->timestamp_fn_ns();
   } else {
     const timestamp_t end_timestamp = timer->timestamp_fn_ns();
-    fprintf(hip_api_file_handle, "%lu:%lu %u:%u %s(", hsa_begin_timestamp, end_timestamp, GetPid(), GetTid(), roctracer_op_string(ACTIVITY_DOMAIN_HIP_API, cid, 0));
+    std::ostringstream oss;                                                                        \
+    oss << std::dec <<
+      hsa_begin_timestamp << ":" << end_timestamp << " " << GetPid() << ":" << GetTid() << " " << roctracer_op_string(ACTIVITY_DOMAIN_HIP_API, cid, 0);                                                       \
+
     switch (cid) {
       case HIP_API_ID_hipMemcpy:
-        fprintf(hip_api_file_handle, "dst(%p) src(%p) size(0x%x) kind(%u)",
+        fprintf(hip_api_file_handle, "%s(dst(%p) src(%p) size(0x%x) kind(%u))",
+          oss.str().c_str(),
           data->args.hipMemcpy.dst,
           data->args.hipMemcpy.src,
           (uint32_t)(data->args.hipMemcpy.sizeBytes),
           (uint32_t)(data->args.hipMemcpy.kind));
         break;
       case HIP_API_ID_hipMalloc:
-        fprintf(hip_api_file_handle, "ptr(0x%p) size(0x%x)",
+        fprintf(hip_api_file_handle, "%s(ptr(0x%p) size(0x%x))",
+          oss.str().c_str(),
           *(data->args.hipMalloc.ptr),
           (uint32_t)(data->args.hipMalloc.size));
         break;
       case HIP_API_ID_hipFree:
-        fprintf(hip_api_file_handle, "ptr(%p)",
+        fprintf(hip_api_file_handle, "%s(ptr(%p))",
+          oss.str().c_str(),
           data->args.hipFree.ptr);
         break;
       case HIP_API_ID_hipModuleLaunchKernel:
-        fprintf(hip_api_file_handle, "kernel(%s) stream(%p)",
-          hipKernelNameRef(data->args.hipModuleLaunchKernel.f),
+        fprintf(hip_api_file_handle, "%s(kernel(%s) stream(%p))",
+          oss.str().c_str(),
+          roctracer::HipLoader::Instance().KernelNameRef(data->args.hipModuleLaunchKernel.f),
           data->args.hipModuleLaunchKernel.stream);
         break;
       default:
         break;
     }
-    fprintf(hip_api_file_handle, ")\n"); fflush(hip_api_file_handle);
+    fflush(hip_api_file_handle);
   }
 }
 
@@ -305,7 +315,7 @@ extern "C" PUBLIC_API bool OnLoad(HsaApiTable* table, uint64_t runtime_version,
       NULL};
     roctracer_set_properties(ACTIVITY_DOMAIN_HSA_OPS, &ops_properties);
 
-    printf("    HSA-trace(");
+    fprintf(stdout, "    HSA-trace("); fflush(stdout);
     if (hsa_api_vec.size() != 0) {
       for (unsigned i = 0; i < hsa_api_vec.size(); ++i) {
         uint32_t cid = HSA_API_ID_NUMBER;
@@ -326,7 +336,7 @@ extern "C" PUBLIC_API bool OnLoad(HsaApiTable* table, uint64_t runtime_version,
     hip_api_file_handle = open_output_file(output_prefix, "hip_api_trace.txt");
     hcc_activity_file_handle = open_output_file(output_prefix, "hcc_ops_trace.txt");
 
-    printf("    HIP-trace()\n");
+    fprintf(stdout, "    HIP-trace()\n"); fflush(stdout);
     // Allocating tracing pool
     roctracer_properties_t properties{};
     properties.buffer_size = 0x1000;