I am reading the code of CUPTI concurrent_profiling sample, and I get some confusions. The function ProfileKernels takes &deviceData as one of the parameters, and RUNTIME_API_CALL(cudaSetDevice(deviceData.deviceID)) is called at the beginning of this function.
My question is that, is deviceData.deviceID (instead of deviceData.config.deviceID, which is properly initialized) actually initialized to zero by the constructor of vector<PerDeviceData> deviceData(numDevices)? I cannot find any further modification of deviceData[i].deviceID; if so, does that mean the test on multiple devices will unexpectedly run on the device 0 only? I appreciate any help.