Does the JIT compiler perform device link-time optimization?

[phw89]

Possibly related?

Using device link-time optimization results in much larger fatbinaries

Let’s say I compile/link an application using device link-time optimization (available from CUDA 11.2+) using the following options: Compile -gencode=arch=compute_52,code=[compute_52,lto_52,lto_61] Link -gencode=arch=compute_52,code=[sm_52,sm_61] -dlto As expected, this will create a fatbinary containing PTX for sm_52, LTO intermediaries for sm_52 and sm_61, and link-time optimized SASS for sm_52 and sm_61. However, according to cuobjdump -all (output: cuobjdump.txt (4.7 KB)), the fatbinar…