support for hip multi kernels

Change-Id: Idc8945f924ceb35081e153921181b96ee5f653b6


[ROCm/roctracer commit: 64ee8f29ef]

projects/roctracer/src/core/loader.h

@@ -115,6 +115,7 @@ class HipApi {
   typedef decltype(hipRemoveActivityCallback) RemoveActivityCallback_t;
   typedef decltype(hipKernelNameRef) KernelNameRef_t;
   typedef decltype(hipKernelNameRefByPtr) KernelNameRefByPtr_t;
+  typedef decltype(hipGetStreamDeviceId) GetStreamDeviceId_t;
   typedef decltype(hipApiName) ApiName_t;
 
   RegisterApiCallback_t* RegisterApiCallback;
@@ -123,6 +124,7 @@ class HipApi {
   RemoveActivityCallback_t* RemoveActivityCallback;
   KernelNameRef_t* KernelNameRef;
   KernelNameRefByPtr_t* KernelNameRefByPtr;
+  GetStreamDeviceId_t* GetStreamDeviceId;
   ApiName_t* ApiName;
 
   protected:
@@ -133,6 +135,7 @@ class HipApi {
     RemoveActivityCallback = loader->GetFun<RemoveActivityCallback_t>("hipRemoveActivityCallback");
     KernelNameRef = loader->GetFun<KernelNameRef_t>("hipKernelNameRef");
     KernelNameRefByPtr = loader->GetFun<KernelNameRefByPtr_t>("hipKernelNameRefByPtr");
+    GetStreamDeviceId =  loader->GetFun<GetStreamDeviceId_t>("hipGetStreamDeviceId");
     ApiName = loader->GetFun<ApiName_t>("hipApiName");
   }
 };

projects/roctracer/test/tool/tracer_tool.cpp

@@ -348,12 +348,10 @@ roctracer::TraceBuffer<hip_api_trace_entry_t> hip_api_trace_buffer("HIP", 0x2000
 
 static inline bool is_hip_kernel_launch_api(const uint32_t& cid) {
   bool ret =
-#if 0 // HIP_VDI
-    (cid == HIP_API_ID_hipExtLaunchMultiKernelMultiDevice) ||
+    (cid == HIP_API_ID_hipLaunchKernel) ||
     (cid == HIP_API_ID_hipLaunchCooperativeKernel) ||
     (cid == HIP_API_ID_hipLaunchCooperativeKernelMultiDevice) ||
-#endif
-    (cid == HIP_API_ID_hipLaunchKernel) ||
+    (cid == HIP_API_ID_hipExtLaunchMultiKernelMultiDevice) ||
     (cid == HIP_API_ID_hipModuleLaunchKernel) ||
     (cid == HIP_API_ID_hipExtModuleLaunchKernel) ||
     (cid == HIP_API_ID_hipHccModuleLaunchKernel);
@@ -389,13 +387,36 @@ void hip_api_callback(
     if (cid == HIP_API_ID_hipMalloc) {
       entry->ptr = *(data->args.hipMalloc.ptr);
     } else if (is_hip_kernel_launch_api(cid)) {
-      if (cid == HIP_API_ID_hipLaunchKernel) {
-        const void* f = data->args.hipLaunchKernel.function_address;
-        hipStream_t stream = data->args.hipLaunchKernel.stream;
-        if (f != NULL) entry->name = strdup(roctracer::HipLoader::Instance().KernelNameRefByPtr(f, stream));
-      } else {
-        const hipFunction_t f = data->args.hipModuleLaunchKernel.f;
-        if (f != NULL) entry->name = strdup(roctracer::HipLoader::Instance().KernelNameRef(f));
+      switch(cid) {
+        case HIP_API_ID_hipExtLaunchMultiKernelMultiDevice:
+        case HIP_API_ID_hipLaunchCooperativeKernelMultiDevice:
+        {
+          const hipLaunchParams* listKernels = data->args.hipLaunchCooperativeKernelMultiDevice.launchParamsList;
+          std::string name_str = "";
+          for (int i = 0; i < data->args.hipLaunchCooperativeKernelMultiDevice.numDevices; ++i) {
+            const hipLaunchParams& lp = listKernels[i];
+            if (lp.func != NULL) {
+              const char* kernel_name = roctracer::HipLoader::Instance().KernelNameRefByPtr(lp.func, lp.stream);
+              const int device_id = roctracer::HipLoader::Instance().GetStreamDeviceId(lp.stream);
+              name_str += std::string(kernel_name) + ":" + std::to_string(device_id) + ";";
+            }
+          }
+          entry->name = strdup(name_str.c_str());
+          break;
+        }
+        case HIP_API_ID_hipLaunchKernel:
+        case HIP_API_ID_hipLaunchCooperativeKernel:
+        {
+          const void* f = data->args.hipLaunchKernel.function_address;
+          hipStream_t stream = data->args.hipLaunchKernel.stream;
+          if (f != NULL) entry->name = strdup(roctracer::HipLoader::Instance().KernelNameRefByPtr(f, stream));
+          break;
+        }
+        default:
+        {
+          const hipFunction_t f = data->args.hipModuleLaunchKernel.f;
+          if (f != NULL) entry->name = strdup(roctracer::HipLoader::Instance().KernelNameRef(f));
+        }
       }
     }
   }