Context Updates (#624)

* Improve error checks related to context create/start/stop/is_valid

* Bump version to 0.2.1

* Track number of kernels associated with correlation id

- add atomic kernel counter variable to context::correlation_id

* Update lib/rocprofiler-sdk/hsa/queue.cpp

- apply the +/- kernel count

VERSION

@@ -1 +1 @@
-0.2.0
+0.2.1

source/lib/rocprofiler-sdk/context.cpp

@@ -41,6 +41,10 @@ extern "C" {
 rocprofiler_status_t
 rocprofiler_create_context(rocprofiler_context_id_t* context_id)
 {
+    // context already registered
+    if(rocprofiler::context::get_registered_context(*context_id))
+        return ROCPROFILER_STATUS_ERROR_CONTEXT_INVALID;
+
     // always set to none first
     *context_id = rocprofiler_context_none;
 
@@ -56,7 +60,8 @@ rocprofiler_create_context(rocprofiler_context_id_t* context_id)
 rocprofiler_status_t
 rocprofiler_start_context(rocprofiler_context_id_t context_id)
 {
-    if(context_id.handle == rocprofiler_context_none.handle)
+    if(context_id.handle == rocprofiler_context_none.handle ||
+       !rocprofiler::context::get_registered_context(context_id))
         return ROCPROFILER_STATUS_ERROR_CONTEXT_NOT_FOUND;
 
     // if currently finalizing or finalized, don't allow starting a context
@@ -69,7 +74,8 @@ rocprofiler_start_context(rocprofiler_context_id_t context_id)
 rocprofiler_status_t
 rocprofiler_stop_context(rocprofiler_context_id_t context_id)
 {
-    if(context_id.handle == rocprofiler_context_none.handle)
+    if(context_id.handle == rocprofiler_context_none.handle ||
+       !rocprofiler::context::get_registered_context(context_id))
         return ROCPROFILER_STATUS_ERROR_CONTEXT_NOT_FOUND;
 
     return rocprofiler::context::stop_context(context_id);
@@ -105,7 +111,8 @@ rocprofiler_context_is_valid(rocprofiler_context_id_t context_id, int* status)
 {
     *status = 0;
 
-    if(context_id.handle == rocprofiler_context_none.handle)
+    if(context_id.handle == rocprofiler_context_none.handle ||
+       !rocprofiler::context::get_registered_context(context_id))
         return ROCPROFILER_STATUS_ERROR_CONTEXT_NOT_FOUND;
 
     for(const auto& itr : rocprofiler::context::get_registered_contexts())
@@ -117,6 +124,7 @@ rocprofiler_context_is_valid(rocprofiler_context_id_t context_id, int* status)
             return _ret;
         }
     }
+
     return ROCPROFILER_STATUS_ERROR_CONTEXT_NOT_FOUND;
 }
 }

source/lib/rocprofiler-sdk/context/context.cpp

@@ -181,6 +181,18 @@ correlation_id::sub_ref_count()
     return _ret;
 }
 
+uint32_t
+correlation_id::add_kern_count()
+{
+    return m_kern_count.fetch_add(1);
+}
+
+uint32_t
+correlation_id::sub_kern_count()
+{
+    return m_kern_count.fetch_sub(1);
+}
+
 correlation_id*
 correlation_tracing_service::construct(uint32_t _init_ref_count)
 {

source/lib/rocprofiler-sdk/context/context.hpp

@@ -79,8 +79,13 @@ struct correlation_id
     uint32_t add_ref_count();
     uint32_t sub_ref_count();
 
+    uint32_t get_kern_count() const { return m_kern_count.load(); }
+    uint32_t add_kern_count();
+    uint32_t sub_kern_count();
+
 private:
-    std::atomic<uint32_t> m_ref_count = {};
+    std::atomic<uint32_t> m_kern_count = {0};
+    std::atomic<uint32_t> m_ref_count  = {0};
 };
 
 correlation_id*

source/lib/rocprofiler-sdk/hsa/queue.cpp

@@ -192,6 +192,7 @@ AsyncSignalHandler(hsa_signal_value_t /*signal_v*/, void* data)
             << "reference counter for correlation id " << _corr_id->internal << " from thread "
             << _corr_id->thread_idx << " has no reference count";
         _corr_id->sub_ref_count();
+        _corr_id->sub_kern_count();
     }
 
     queue_info_session.queue.async_complete();
@@ -380,7 +381,11 @@ WriteInterceptor(const void* packets,
         LOG_IF(FATAL, packet_type != HSA_PACKET_TYPE_KERNEL_DISPATCH)
             << "get_kernel_id below might need to be updated";
 
-        if(corr_id) corr_id->add_ref_count();
+        if(corr_id)
+        {
+            corr_id->add_ref_count();
+            corr_id->add_kern_count();
+        }
 
         // Enqueue the signal into the handler. Will call completed_cb when
         // signal completes.