Use the new ROCR support to automatically load tool libraries

The ROCR now detects already loaded tool libraries and calls OnLoad/
OnUnload in the order specified with HSA_AMD_TOOL_ORDER.

It is no longer necessary to set the HSA_TOOLS_LIB environment variable
to load the roctracer API. The roctracer tool library should be
pre-loaded with LD_PRELOAD.

Change-Id: I6de1b1bd4f93caa08d3554aad2376d242c74fb7e

src/roctracer/exportmap

@@ -1,17 +1,9 @@
 ROCTRACER_4.0 {
-global: roctracer_flush_buf;
-        roctracer_load;
-        roctracer_unload;
-};
-
-ROCTRACER_4.1 {
 global: OnLoad;
         OnUnload;
         roctracer_activity_pop_external_correlation_id;
         roctracer_activity_push_external_correlation_id;
-        roctracer_close_pool;
         roctracer_close_pool_expl;
-        roctracer_default_pool;
         roctracer_default_pool_expl;
         roctracer_disable_activity;
         roctracer_disable_callback;
@@ -19,29 +11,38 @@ global: OnLoad;
         roctracer_disable_domain_callback;
         roctracer_disable_op_activity;
         roctracer_disable_op_callback;
-        roctracer_enable_activity;
         roctracer_enable_activity_expl;
         roctracer_enable_callback;
-        roctracer_enable_domain_activity;
         roctracer_enable_domain_activity_expl;
         roctracer_enable_domain_callback;
-        roctracer_enable_op_activity;
         roctracer_enable_op_activity_expl;
         roctracer_enable_op_callback;
         roctracer_error_string;
-        roctracer_flush_activity;
         roctracer_flush_activity_expl;
+        roctracer_flush_buf;
         roctracer_get_timestamp;
+        roctracer_load;
         roctracer_mark;
-        roctracer_next_record;
         roctracer_op_code;
-        roctracer_open_pool;
         roctracer_open_pool_expl;
         roctracer_op_string;
         roctracer_set_properties;
         roctracer_start;
         roctracer_stop;
+        roctracer_unload;
         roctracer_version_major;
         roctracer_version_minor;
 local:  *;
-} ROCTRACER_4.0;
+};
+
+ROCTRACER_4.1 {
+global: HSA_AMD_TOOL_PRIORITY;
+        roctracer_close_pool;
+        roctracer_default_pool;
+        roctracer_enable_activity;
+        roctracer_enable_domain_activity;
+        roctracer_enable_op_activity;
+        roctracer_flush_activity;
+        roctracer_next_record;
+        roctracer_open_pool;
+} ROCTRACER_4.0;

src/roctracer/roctracer.cpp

@@ -1238,12 +1238,19 @@ __attribute__((destructor)) void destructor() {
   ONLOAD_TRACE_END();
 }
 
+extern "C" {
+
+// The HSA_AMD_TOOL_PRIORITY variable must be a constant value type initialized by the loader
+// itself, not by code during _init. 'extern const' seems do that although that is not a guarantee.
+ROCTRACER_EXPORT extern const uint32_t HSA_AMD_TOOL_PRIORITY = 50;
+
 // HSA-runtime tool on-load method
-extern "C" ROCTRACER_EXPORT bool OnLoad(HsaApiTable* table, uint64_t runtime_version,
-                                        uint64_t failed_tool_count,
-                                        const char* const* failed_tool_names) {
+ROCTRACER_EXPORT bool OnLoad(HsaApiTable* table, uint64_t runtime_version,
+                             uint64_t failed_tool_count, const char* const* failed_tool_names) {
   hsa_support::SaveHsaApi(table);
   return true;
 }
 
-extern "C" ROCTRACER_EXPORT void OnUnload() {}
+ROCTRACER_EXPORT void OnUnload() {}
+
+}  // extern "C"

src/tracer_tool/exportmap

@@ -1 +1 @@
-{ global: OnLoad; OnUnload; local: *; };
+{ global: HSA_AMD_TOOL_PRIORITY; OnLoad; OnUnload; local:  *; };

src/tracer_tool/tracer_tool.cpp

@@ -795,10 +795,15 @@ void tool_load() {
   ONLOAD_TRACE_END();
 }
 
+extern "C" {
+
+// The HSA_AMD_TOOL_PRIORITY variable must be a constant value type initialized by the loader
+// itself, not by code during _init. 'extern const' seems do that although that is not a guarantee.
+ROCTRACER_EXPORT extern const uint32_t HSA_AMD_TOOL_PRIORITY = 1050;
+
 // HSA-runtime tool on-load method
-extern "C" ROCTRACER_EXPORT bool OnLoad(HsaApiTable* table, uint64_t runtime_version,
-                                        uint64_t failed_tool_count,
-                                        const char* const* failed_tool_names) {
+ROCTRACER_EXPORT bool OnLoad(HsaApiTable* table, uint64_t runtime_version,
+                             uint64_t failed_tool_count, const char* const* failed_tool_names) {
   ONLOAD_TRACE_BEG();
 
   roctracer::hip_support::HIP_depth_max = 0;
@@ -927,8 +932,10 @@ extern "C" ROCTRACER_EXPORT bool OnLoad(HsaApiTable* table, uint64_t runtime_ver
 }
 
 // HSA-runtime on-unload method
-extern "C" ROCTRACER_EXPORT void OnUnload() {
+ROCTRACER_EXPORT void OnUnload() {
   ONLOAD_TRACE_BEG();
   tool_unload();
   ONLOAD_TRACE_END();
 }
+
+}  // extern "C"

test/run.sh

@@ -111,15 +111,15 @@ eval_test "MatrixTranspose dry run" ./test/MatrixTranspose MatrixTranspose_dryru
 eval_test "copy dry run" ./test/copy copy_dryrun_trace
 
 # Standalone test
-# rocTrecer is used explicitely by test
+# ROCtracer is used explicitely by test
 eval_test "standalone C test" "./test/MatrixTranspose_ctest" MatrixTranspose_ctest_trace
 eval_test "standalone HIP test" "./test/MatrixTranspose_test" MatrixTranspose_test_trace
 eval_test "standalone HIP hipaact test" "./test/MatrixTranspose_hipaact_test" MatrixTranspose_hipaact_test_trace
 eval_test "standalone HIP MGPU test" "./test/MatrixTranspose_mgpu" MatrixTranspose_mgpu_trace
 
 # Tool test
-# rocTracer/tool is loaded by HSA runtime
-export HSA_TOOLS_LIB="$ROCTRACER_LIB_PATH/libroctracer64.so $ROCTRACER_TOOL_PATH/libroctracer_tool.so"
+# ROCtracer/tool is loaded by HSA runtime
+export LD_PRELOAD="$ROCTRACER_TOOL_PATH/libroctracer_tool.so"
 
 # ROCTX test
 export ROCTRACER_DOMAIN="roctx"
@@ -166,11 +166,11 @@ unset ROCP_INPUT
 # Check that the tracer tool can be unloaded and then reloaded.
 eval_test "Load/Unload/Reload the tracer tool" ./test/load_unload_reload_test load_unload_reload_trace
 
-export HSA_TOOLS_LIB="$ROCTRACER_LIB_PATH/libroctracer64.so ./test/libhsaco_test.so"
+export HSA_TOOLS_LIB="./test/libhsaco_test.so"
 eval_test "tool HSA codeobj" ./test/MatrixTranspose hsa_co_trace
 
 export ROCP_TOOL_LIB=./test/libcodeobj_test.so
-export HSA_TOOLS_LIB="$ROCTRACER_LIB_PATH/libroctracer64.so librocprofiler64.so"
+export HSA_TOOLS_LIB="librocprofiler64.so"
 eval_test "tool tracer codeobj" ./test/MatrixTranspose code_obj_trace
 
 #valgrind --leak-check=full $tbin