[Polly] [WIP] [NFC] [GPUJit] call cuProfilerStop before destroying the context to flush profiler cache

[bollu]

@grosser, @singam-sanjay Do you know how I can add a test case for this?

[bollu]

Spoke with @grosser about this. We don't have a clean way to test this stuff, so I'm upstreaming this without a test for now.

[bollu]

• [NFC] whitespace nit.

tools/GPURuntime/GPUJIT.c

 static CuMemcpyHtoDFcnTy *CuMemcpyHtoDFcnPtr;
 
 typedef CUresult CUDAAPI CuMemFreeFcnTy(CUdeviceptr);
 static CuMemFreeFcnTy *CuMemFreeFcnPtr;
 
 typedef CUresult CUDAAPI CuModuleUnloadFcnTy(CUmodule);
 static CuModuleUnloadFcnTy *CuModuleUnloadFcnPtr;
 
+
+typedef CUresult CUDAAPI CuProfilerStopFcnTy();
+static CuProfilerStopFcnTy *CuProfilerStopFcnPtr;
+
 typedef CUresult CUDAAPI CuCtxDestroyFcnTy(CUcontext);
 static CuCtxDestroyFcnTy *CuCtxDestroyFcnPtr;
 
 typedef CUresult CUDAAPI CuInitFcnTy(unsigned int);
 static CuInitFcnTy *CuInitFcnPtr;
 
 typedef CUresult CUDAAPI CuDeviceGetCountFcnTy(int *);
 static CuDeviceGetCountFcnTy *CuDeviceGetCountFcnPtr;
CuMemcpyDtoHFcnPtr =
       (CuMemcpyDtoHFcnTy *)getAPIHandleCUDA(HandleCuda, "cuMemcpyDtoH_v2");
 
   CuMemcpyHtoDFcnPtr =
       (CuMemcpyHtoDFcnTy *)getAPIHandleCUDA(HandleCuda, "cuMemcpyHtoD_v2");
 
   CuModuleUnloadFcnPtr =
       (CuModuleUnloadFcnTy *)getAPIHandleCUDA(HandleCuda, "cuModuleUnload");
 
+  CuProfilerStopFcnPtr =
+      (CuProfilerStopFcnTy *)getAPIHandleCUDA(HandleCuda, "cuProfilerStop");
+
   CuCtxDestroyFcnPtr =
       (CuCtxDestroyFcnTy *)getAPIHandleCUDA(HandleCuda, "cuCtxDestroy");
 
   CuInitFcnPtr = (CuInitFcnTy *)getAPIHandleCUDA(HandleCuda, "cuInit");
 
   CuDeviceGetCountFcnPtr =
       (CuDeviceGetCountFcnTy *)getAPIHandleCUDA(HandleCuda, "cuDeviceGetCount");
 
static void *getDevicePtrCUDA(PollyGPUDevicePtr *Allocation) {
   return (void *)DevPtr->Cuda;
 }
 
 static void freeContextCUDA(PollyGPUContext *Context) {
   dump_function();
 
   CUDAContext *Ctx = (CUDAContext *)Context->Context;
   if (Ctx->Cuda) {
+    CuProfilerStopFcnPtr();
     CuCtxDestroyFcnPtr(Ctx->Cuda);
     free(Ctx);
     free(Context);
   }
 
   dlclose(HandleCuda);
   dlclose(HandleCudaRT);
 }