PTX Compression Method

Following this old thread regarding how PTX is compressed inside a fatbinary:

Which is the compression method employed by the NVCC toolchain?

This method seems to be employed starting from CUDA 5. Back to CUDA 3/4 the PTX was saved in plaintext.

Same question asked differently?

Actually I’m the same author of the question you mentioned, but the question was a bit different: it was related to the fatbinary organization.

Now the point is how PTX code is compressed.