[DevTSAN] Do early return if kernel has no data race reports (#19133)

Sync runtime data from device to host will cost a lot of time, so we'd
better check if kernel has reports first to avoid unnecessary data sync.

Author: zhaomaosu

unified-runtime/source/loader/layers/sanitizer/tsan/tsan_interceptor.cpp

@@ -61,6 +61,17 @@ ur_result_t TsanRuntimeDataWrapper::syncToDevice(ur_queue_handle_t Queue) {
   return UR_RESULT_SUCCESS;
 }
 
+bool TsanRuntimeDataWrapper::hasReport(ur_queue_handle_t Queue) {
+  ur_result_t URes = getContext()->urDdiTable.Enqueue.pfnUSMMemcpy(
+      Queue, true, ur_cast<void *>(&Host), getDevicePtr(),
+      sizeof(TsanRuntimeData::RecordedReportCount), 0, nullptr, nullptr);
+  if (URes != UR_RESULT_SUCCESS) {
+    UR_LOG(ERR, "Failed to sync runtime data to host: {}", URes);
+    return false;
+  }
+  return Host.RecordedReportCount != 0;
+}
+
 ur_result_t TsanRuntimeDataWrapper::importLocalArgsInfo(
     ur_queue_handle_t Queue, const std::vector<TsanLocalArgsInfo> &LocalArgs) {
   assert(!LocalArgs.empty());
@@ -310,6 +321,9 @@ ur_result_t TsanInterceptor::postLaunchKernel(ur_kernel_handle_t Kernel,
   // urEventSetCallback
   UR_CALL(getContext()->urDdiTable.Queue.pfnFinish(Queue));
 
+  if (!LaunchInfo.Data.hasReport(Queue))
+    return UR_RESULT_SUCCESS;
+
   UR_CALL(LaunchInfo.Data.syncFromDevice(Queue));
 
   for (uptr ReportIndex = 0;

unified-runtime/source/loader/layers/sanitizer/tsan/tsan_interceptor.hpp

@@ -133,6 +133,8 @@ struct TsanRuntimeDataWrapper {
 
   ur_result_t syncToDevice(ur_queue_handle_t Queue);
 
+  bool hasReport(ur_queue_handle_t Queue);
+
   ur_result_t
   importLocalArgsInfo(ur_queue_handle_t Queue,
                       const std::vector<TsanLocalArgsInfo> &LocalArgs);

unified-runtime/source/loader/layers/sanitizer/tsan/tsan_libdevice.hpp

@@ -82,6 +82,8 @@ struct TsanLocalArgsInfo {
 constexpr uint64_t TSAN_MAX_NUM_REPORTS = 128;
 
 struct TsanRuntimeData {
+  uint32_t RecordedReportCount = 0;
+
   uintptr_t GlobalShadowOffset = 0;
 
   uintptr_t GlobalShadowOffsetEnd = 0;
@@ -103,8 +105,6 @@ struct TsanRuntimeData {
 
   int Lock = 0;
 
-  uint32_t RecordedReportCount = 0;
-
   TsanErrorReport Report[TSAN_MAX_NUM_REPORTS];
 };
 

unified-runtime/source/loader/layers/sanitizer/tsan/tsan_shadow.cpp

@@ -169,19 +169,12 @@ ur_result_t ShadowMemoryGPU::CleanShadow(ur_queue_handle_t Queue, uptr Ptr,
         UR_LOG_L(getContext()->logger, DEBUG, "urVirtualMemMap: {} ~ {}",
                  (void *)MappedPtr, (void *)(MappedPtr + PageSize - 1));
 
-        // Initialize to zero
-        URes = EnqueueUSMSet(Queue, (void *)MappedPtr, (char)0, PageSize);
-        if (URes != UR_RESULT_SUCCESS) {
-          UR_LOG_L(getContext()->logger, ERR, "EnqueueUSMBlockingSet(): {}",
-                   URes);
-          return URes;
-        }
-
         VirtualMemMaps[MappedPtr] = PhysicalMem;
       }
     }
   }
 
+  // Initialize to zero
   auto URes = EnqueueUSMSet(Queue, (void *)Begin, (char)0,
                             Size / kShadowCell * kShadowCnt * kShadowSize);
   if (URes != UR_RESULT_SUCCESS) {