Handle OMNITRACE_ENABLED + minor updates (#78)

- handle OMNITRACE_ENABLED=OFF by disabling everything
- use set_data to get wrappee in pthread_create_gotcha
- clear roctracer_data storage if roctracer not initialized

[ROCm/rocprofiler-systems commit: f82845388a]

projects/rocprofiler-systems/source/lib/omnitrace/library/components/pthread_create_gotcha.cpp

@@ -29,10 +29,9 @@
 #include "library/runtime.hpp"
 #include "library/sampling.hpp"
 #include "library/thread_data.hpp"
-#include "timemory/components/timing/wall_clock.hpp"
 
-#include <bits/stdint-intn.h>
 #include <timemory/backends/threading.hpp>
+#include <timemory/components/timing/wall_clock.hpp>
 #include <timemory/sampling/allocator.hpp>
 #include <timemory/utility/types.hpp>
 
@@ -203,6 +202,7 @@ void
 pthread_create_gotcha::configure()
 {
     pthread_create_gotcha_t::get_initializer() = []() {
+        if(!tim::settings::enabled()) return;
         pthread_create_gotcha_t::template configure<
             0, int, pthread_t*, const pthread_attr_t*, void* (*) (void*), void*>(
             "pthread_create");
@@ -257,6 +257,12 @@ pthread_create_gotcha::shutdown(int64_t _tid)
     }
 }
 
+void
+pthread_create_gotcha::set_data(wrappee_t _v)
+{
+    m_wrappee = _v;
+}
+
 // pthread_create
 int
 pthread_create_gotcha::operator()(pthread_t* thread, const pthread_attr_t* attr,
@@ -282,8 +288,7 @@ pthread_create_gotcha::operator()(pthread_t* thread, const pthread_attr_t* attr,
     {
         auto* _obj = new wrapper(start_routine, arg, _enable_sampling, _tid, nullptr);
         // create the thread
-        auto _ret =
-            ::pthread_create(thread, attr, &wrapper::wrap, static_cast<void*>(_obj));
+        auto _ret = (*m_wrappee)(thread, attr, &wrapper::wrap, static_cast<void*>(_obj));
         return _ret;
     }
 
@@ -300,7 +305,7 @@ pthread_create_gotcha::operator()(pthread_t* thread, const pthread_attr_t* attr,
     auto* _wrap    = new wrapper(start_routine, arg, _enable_sampling, _tid, &_promise);
 
     // create the thread
-    auto _ret = ::pthread_create(thread, attr, &wrapper::wrap, static_cast<void*>(_wrap));
+    auto _ret = (*m_wrappee)(thread, attr, &wrapper::wrap, static_cast<void*>(_wrap));
 
     // wait for thread to set promise
     OMNITRACE_DEBUG("waiting for child to signal it is setup...\n");

projects/rocprofiler-systems/source/lib/omnitrace/library/components/pthread_create_gotcha.hpp

@@ -34,9 +34,11 @@ namespace omnitrace
 {
 struct pthread_create_gotcha : tim::component::base<pthread_create_gotcha, void>
 {
+    using routine_t = void* (*) (void*);
+    using wrappee_t = int (*)(pthread_t*, const pthread_attr_t*, routine_t, void*);
+
     struct wrapper
     {
-        using routine_t = void* (*) (void*);
         using promise_t = std::promise<void>;
 
         wrapper(routine_t _routine, void* _arg, bool, int64_t, promise_t*);
@@ -68,6 +70,11 @@ struct pthread_create_gotcha : tim::component::base<pthread_create_gotcha, void>
 
     static auto& get_execution_time(int64_t _tid = threading::get_id());
     static bool  is_valid_execution_time(int64_t _tid, uint64_t _ts);
+
+    void set_data(wrappee_t);
+
+private:
+    wrappee_t m_wrappee = &pthread_create;
 };
 
 inline auto&

projects/rocprofiler-systems/source/lib/omnitrace/library/components/roctracer.cpp

@@ -166,7 +166,12 @@ void
 roctracer::shutdown()
 {
     auto_lock_t _lk{ type_mutex<roctracer>() };
-    if(!roctracer_is_setup()) return;
+    if(!roctracer_is_setup())
+    {
+        if(!roctracer_is_init() && tim::storage<comp::roctracer_data>::instance())
+            tim::storage<comp::roctracer_data>::instance()->reset();
+        return;
+    }
     roctracer_is_setup() = false;
 
     OMNITRACE_VERBOSE_F(1, "shutting down roctracer...\n");

projects/rocprofiler-systems/source/lib/omnitrace/library/components/roctracer_callbacks.cpp

@@ -853,6 +853,13 @@ hip_activity_callback(const char* begin, const char* end, void*)
     }
 }
 
+bool&
+roctracer_is_init()
+{
+    static bool _v = tim::get_env("OMNITRACE_ROCTRACER_IS_INIT", false);
+    return _v;
+}
+
 bool&
 roctracer_is_setup()
 {
@@ -892,7 +899,9 @@ extern "C"
                 const char* const* failed_tool_names)
     {
         if(!tim::get_env("OMNITRACE_INIT_TOOLING", true)) return true;
+        if(!tim::settings::enabled()) return true;
 
+        roctracer_is_init() = true;
         pthread_gotcha::push_enable_sampling_on_child_threads(false);
         OMNITRACE_CONDITIONAL_BASIC_PRINT_F(get_debug_env() || get_verbose_env() > 0,
                                             "\n");

projects/rocprofiler-systems/source/lib/omnitrace/library/components/roctracer_callbacks.hpp

@@ -71,6 +71,9 @@ hip_api_callback(uint32_t domain, uint32_t cid, const void* callback_data, void*
 void
 hip_activity_callback(const char* begin, const char* end, void*);
 
+bool&
+roctracer_is_init();
+
 bool&
 roctracer_is_setup();
 

projects/rocprofiler-systems/source/lib/omnitrace/library/config.cpp

@@ -604,6 +604,22 @@ configure_mode_settings()
     // recycle all subsequent thread ids
     threading::recycle_ids() =
         tim::get_env<bool>("OMNITRACE_RECYCLE_TIDS", !get_use_sampling());
+
+    if(!get_config()->get_enabled())
+    {
+        _set("OMNITRACE_USE_PERFETTO", false);
+        _set("OMNITRACE_USE_TIMEMORY", false);
+        _set("OMNITRACE_USE_ROCM_SMI", false);
+        _set("OMNITRACE_USE_ROCTRACER", false);
+        _set("OMNITRACE_USE_KOKKOSP", false);
+        _set("OMNITRACE_USE_OMPT", false);
+        _set("OMNITRACE_USE_SAMPLING", false);
+        _set("OMNITRACE_USE_PROCESS_SAMPLING", false);
+        _set("OMNITRACE_USE_CODE_COVERAGE", false);
+        _set("OMNITRACE_CRITICAL_TRACE", false);
+        set_setting_value("OMNITRACE_TIMEMORY_COMPONENTS", std::string{});
+        set_setting_value("OMNITRACE_PAPI_EVENTS", std::string{});
+    }
 }
 
 void

projects/rocprofiler-systems/source/lib/omnitrace/library/ompt.cpp

@@ -61,6 +61,7 @@ bool _init_toolset_off = (trait::runtime_enabled<ompt_toolset_t>::set(false),
 void
 setup()
 {
+    if(!tim::settings::enabled()) return;
     trait::runtime_enabled<ompt_toolset_t>::set(true);
     trait::runtime_enabled<ompt_context_t>::set(true);
     comp::user_ompt_bundle::global_init();