LLVM offload compress
Hello,
It seems that LLVM 18 has introduced the --offload-compress option [1].
This allows LLVM to use zlib or zstd compression on the AMD GPU code
embedded within shared object libraries. They will be decompressed by
the HIP runtime when needed. As GPU code tends to be very repetitive,
it's likely that this will result in a space savings >90% for many of
the largest ROCm libraries. That would be a dramatic reduction for the
installed size on disk.
The ROCm 6.2 release is based upon LLVM 18, so this is probably
something we should explore adding to the build flags once we get the
HIP runtime updated.
Sincerely,
Cory Bloor
[1]:
https://clang.llvm.org/docs/ClangOffloadBundler.html#compression-and-decompression
Reply to: