Starting a TotalView CUDA Session
A standard TotalView installation supports debugging CUDA applications running on both the host and GPU processors. TotalView dynamically detects a CUDA install on your system. To start the TotalView GUI or CLI, provide the name of your CUDA host executable to the totalview or totalviewcli command. For example, to start the TotalView GUI on the sample program, use the following command:
% totalview tx_cuda_matmul
If TotalView successfully loads the CUDA debugging library, it prints the current CUDA debugger API version and the NVIDIA driver version:
CUDA library loaded: Current DLL API version is “8.0.128”; NVIDIA driver version 384.125
...
After reading the symbol table information for the CUDA host executable, TotalView opens the initial process window focused on main in the host code, as shown in Figure 268.
 
Figure 268, Initial process window opened on CUDA host code
You can debug the CUDA host code using the normal TotalView commands and procedures.