Chapter 25 About the TotalView CUDA Debugger
The TotalView CUDA debugger is an integrated debugging tool capable of simultaneously debugging CUDA code that is running on the host Linux-x86_64 and the NVIDIA® GPU. CUDA support is an extension to the standard version of Linux-x86_64 TotalView, and is capable of debugging 64-bit CUDA programs on Linux-x86_64. Debugging 32-bit CUDA programs is currently not supported.
Supported major features:
• Debug CUDA application running directly on GPU hardware
• Set breakpoints, pause execution, and single step in GPU code
• View GPU variables in PTX registers, local, parameter, global, or shared memory
• Access runtime variables, such as threadIdx, blockIdx, blockDim, etc.
• Debug multiple GPU devices per process
• Support for the CUDA MemoryChecker
• Debug remote, distributed and clustered systems
• Support for directive-based programming languages
• Support for Linux-x86_64 host debugging features
Requirements:
• CUDA SDK 7.0, 7.5, or 8.0
— With SDK 7.0, TotalView 8.15.4 and 8.15.7
— With SDK 7.5, TotalView 8.15.10 through TotalView for HPC 2016.06
— With SDK 8.0, TotalView for HPC 2016.07
• Tesla, Fermi or Kepler hardware supported by NVIDIA
• A linux-x86_64 distribution supported by NVIDIA