[Openmp-commits] [llvm] [openmp] [Offload] Allow to record kernel launch stack traces (PR #100472)

[Johannes Doerfert via Openmp-commits]

================
@@ -3240,6 +3251,26 @@ struct AMDGPUPluginTy final : public GenericPluginTy {
     uint32_t Node = -1;
     hsa_agent_get_info(Event->memory_fault.agent, HSA_AGENT_INFO_NODE, &Node);
 
+    AMDGPUPluginTy &Plugin = *reinterpret_cast<AMDGPUPluginTy *>(PluginPtr);
+    for (uint32_t I = 0, E = Plugin.getNumDevices();
+         Node != uint32_t(-1) && I < E; ++I) {
+      AMDGPUDeviceTy &AMDGPUDevice =
+          reinterpret_cast<AMDGPUDeviceTy &>(Plugin.getDevice(I));
+      auto KernelTraceInfoRecord =
+          AMDGPUDevice.KernelLaunchTraces.getExclusiveAccessor();
+
+      uint32_t DeviceNode = -1;
+      if (auto Err =
+              AMDGPUDevice.getDeviceAttr(HSA_AGENT_INFO_NODE, DeviceNode)) {
----------------
jdoerfert wrote:

We use it in the error handler just above this, so I matched it. If we assume one agent per device, we could also compare agents.

https://github.com/llvm/llvm-project/pull/100472