[libomptarget][amdgpu] Destruct HSA queues

[JonChesterfield]

Summary

Store queues in unique_ptr so they are destroyed when the global DeviceInfo is. Currently they leak which raises an assert in debug builds of hsa.

[JonChesterfield]

openmp/libomptarget/plugins/amdgpu/src/rtl.cpp

856–857

Storing (most) state in a global variable that closes hsa in the destructor make it difficult to use raii classes to manage hsa resources.

We could change to using multiple calls to hsa_init/destroy, since it's internally reference counted, and that would give us the last one to destroy closes hsa.

Better is probably to nest the lifetime - would like hsa_init to occur before any other objects are constructed and hsa_shut_down after they've all been torn down. Will think about how best to represent that (separate to this patch)

[JonChesterfield]

D109512 rearranges rtl.cpp so that hsa is reliably constructed before other member variables in the big global object. If we land that, this signal pool can be refactored to free the signals in the destructor again. Still some care needed to ensure it doesn't try to create any without checking hsa is available, i.e. don't prepopulate the pool in the constructor, but at least cleanup can be implicit.

Likewise could put queue in a unique_ptr or similar to reliably cleanup.

[JonChesterfield]

• roll back most of the patch

[JonChesterfield]

• put queues in unique_ptr

[pdhaliwal]

Looks good.

openmp/libomptarget/plugins/amdgpu/dynamic_hsa/hsa.h

 
 hsa_status_t hsa_queue_create(hsa_agent_t agent, uint32_t size,
                               hsa_queue_type32_t type,
                               void (*callback)(hsa_status_t status,
                                                hsa_queue_t *source, void *data),
                               void *data, uint32_t private_segment_size,
                               uint32_t group_segment_size, hsa_queue_t **queue);
 
+hsa_status_t hsa_queue_destroy(hsa_queue_t *queue);
+
 uint64_t hsa_queue_load_read_index_scacquire(const hsa_queue_t *queue);
 
 uint64_t hsa_queue_add_write_index_relaxed(const hsa_queue_t *queue,
                                            uint64_t value);
 
 typedef enum {
   HSA_PACKET_TYPE_KERNEL_DISPATCH = 2,
 } hsa_packet_type_t;

openmp/libomptarget/plugins/amdgpu/dynamic_hsa/hsa.cpp

 DLWRAP(hsa_agent_get_info, 3);
 DLWRAP(hsa_iterate_agents, 2);
 DLWRAP(hsa_signal_create, 4);
 DLWRAP(hsa_signal_destroy, 1);
 DLWRAP(hsa_signal_store_relaxed, 2);
 DLWRAP(hsa_signal_store_screlease, 2);
 DLWRAP(hsa_signal_wait_scacquire, 5);
 DLWRAP(hsa_queue_create, 8);
+DLWRAP(hsa_queue_destroy, 1);
 DLWRAP(hsa_queue_load_read_index_scacquire, 1);
 DLWRAP(hsa_queue_add_write_index_relaxed, 2);
 DLWRAP(hsa_memory_copy, 3);
 DLWRAP(hsa_executable_create, 4);
 DLWRAP(hsa_executable_destroy, 1);
 DLWRAP(hsa_executable_freeze, 2);
 DLWRAP(hsa_executable_symbol_get_info, 3);
 DLWRAP(hsa_executable_iterate_symbols, 3);

openmp/libomptarget/plugins/amdgpu/src/rtl.cpp

struct HSALifetime {
   }
 };
 
 /// Class containing all the device information
 class RTLDeviceInfoTy {
   HSALifetime HSA; // First field => constructed first and destructed last
   std::vector<std::list<FuncOrGblEntryTy>> FuncGblEntries;
 
+  struct QueueDeleter {
+    void operator()(hsa_queue_t *Q) {
+      if (Q) {
+        hsa_status_t Err = hsa_queue_destroy(Q);
+        if (Err != HSA_STATUS_SUCCESS) {
+          DP("Error destroying hsa queue: %s\n", get_error_string(Err));
+        }
+      }
+    }
+  };
+
 public:
   // load binary populates symbol tables and mutates various global state
   // run uses those symbol tables
   std::shared_timed_mutex load_run_lock;
 
   int NumberOfDevices = 0;
 
   // GPU devices
   std::vector<hsa_agent_t> HSAAgents;
-  std::vector<hsa_queue_t *> HSAQueues; // one per gpu
+  std::vector<std::unique_ptr<hsa_queue_t, QueueDeleter>>
+      HSAQueues; // one per gpu
 
   // CPUs
   std::vector<hsa_agent_t> CPUAgents;
 
   // Device properties
   std::vector<int> ComputeUnits;
   std::vector<int> GroupsPerDevice;
   std::vector<int> ThreadsPerGroup;
RTLDeviceInfoTy() {
 
     err = setupMemoryPools();
     if (err != HSA_STATUS_SUCCESS) {
       DP("Error when setting up memory pools");
       return;
     }
 
     for (int i = 0; i < NumberOfDevices; i++) {
-      HSAQueues[i] = nullptr;
-    }
-
-    for (int i = 0; i < NumberOfDevices; i++) {
       uint32_t queue_size = 0;
       {
         hsa_status_t err = hsa_agent_get_info(
             HSAAgents[i], HSA_AGENT_INFO_QUEUE_MAX_SIZE, &queue_size);
         if (err != HSA_STATUS_SUCCESS) {
           DP("HSA query QUEUE_MAX_SIZE failed for agent %d\n", i);
           return;
         }
         enum { MaxQueueSize = 4096 };
         if (queue_size > MaxQueueSize) {
           queue_size = MaxQueueSize;
         }
       }
 
-      hsa_status_t rc = hsa_queue_create(
-          HSAAgents[i], queue_size, HSA_QUEUE_TYPE_MULTI, callbackQueue, NULL,
-          UINT32_MAX, UINT32_MAX, &HSAQueues[i]);
+      {
+        hsa_queue_t *Q = nullptr;
+        hsa_status_t rc =
+            hsa_queue_create(HSAAgents[i], queue_size, HSA_QUEUE_TYPE_MULTI,
+                             callbackQueue, NULL, UINT32_MAX, UINT32_MAX, &Q);
         if (rc != HSA_STATUS_SUCCESS) {
           DP("Failed to create HSA queue %d\n", i);
           return;
         }
+        HSAQueues[i].reset(Q);
+      }
 
       deviceStateStore[i] = {nullptr, 0};
     }
 
     for (int i = 0; i < NumberOfDevices; i++) {
       ThreadsPerGroup[i] = RTLDeviceInfoTy::Default_WG_Size;
       GroupsPerDevice[i] = RTLDeviceInfoTy::DefaultNumTeams;
       ComputeUnits[i] = 1;
~RTLDeviceInfoTy() {
 
     hsa_status_t Err;
     for (uint32_t I = 0; I < HSAExecutables.size(); I++) {
       Err = hsa_executable_destroy(HSAExecutables[I]);
       if (Err != HSA_STATUS_SUCCESS) {
         DP("[%s:%d] %s failed: %s\n", __FILE__, __LINE__,
            "Destroying executable", get_error_string(Err));
       }
     }
   }
 };
 
 pthread_mutex_t SignalPoolT::mutex = PTHREAD_MUTEX_INITIALIZER;
 
 // TODO: May need to drop the trailing to fields until deviceRTL is updated
 struct omptarget_device_environmentTy {
   int32_t debug_level; // gets value of envvar LIBOMPTARGET_DEVICE_RTL_DEBUG
                        // only useful for Debug build of deviceRTLs
fprintf(traceToStdout ? stdout : stderr,
             device_id, KernelInfo->ExecutionMode, KernelInfo->ConstWGSize,
             arg_num, num_groups, WorkgroupSize, num_teams, thread_limit,
             group_segment_size, sgpr_count, vgpr_count, sgpr_spill_count,
             vgpr_spill_count, loop_tripcount, KernelInfo->Name);
   }
 
   // Run on the device.
   {
-    hsa_queue_t *queue = DeviceInfo.HSAQueues[device_id];
+    hsa_queue_t *queue = DeviceInfo.HSAQueues[device_id].get();
     if (!queue) {
       return OFFLOAD_FAIL;
     }
     uint64_t packet_id = acquire_available_packet_id(queue);
 
     const uint32_t mask = queue->size - 1; // size is a power of 2
     hsa_kernel_dispatch_packet_t *packet =
         (hsa_kernel_dispatch_packet_t *)queue->base_address +