Work-in-progress / early research phase

The goal of this project is to extend pyelftools so it can inspect CUDA and HIP/ROCm fat ELF binaries and detect which GPU devices (architectures) are supported by a given shared library or executable.

See REFERENCES.md for binary format documentation, tools, and research links.

$ python3 -m elfgpu libtorch_cuda.so 
sm_100
sm_100a
sm_103a
sm_120
sm_120a
sm_121a
sm_75
sm_80
sm_86
sm_89
sm_90
sm_90a

$ python3 -m elfgpu libtorch_hip.so 
gfx1030
gfx1100
gfx1101
gfx1102
gfx1150
gfx1151
gfx1200
gfx1201
gfx900
gfx906
gfx908
gfx90a
gfx942
gfx950