Given the GPU model, Compute Capability and the CUDA version, how can I tell which PTX version is in use?
Does the PTX version change with CUDA releases ?
Is the PTX generated/written by you or by nvcc?
PTX is an intermediate language, sometimes only used by the developers, sometimes delivered with the software to be compiled on-the-fly on the customer PC.
I just want to know what version, and therefore which instructions, is available to me given my setup.
The available instructions depend on both the PTX version defined by the system compiling the PTX (either ptxas and therefore SDK version; or the installed driver version) and on the compute capability.
After each instruction in the PTX manual the needed compute capabilities are mentioned.
For your own setup, use the newest SDK and driver, so the available instructions only depend on your compute capability. This setup is also compatible with older PTX versions.