3

The CUDA 11 features announcement, it's said that there are now:

New link time optimization capabilities

what link-time optimizations does nvcc actually employ (e.g. relative to the LTO capabilities in host-side code with g++ or clang++)?

Also - is there something one needs to do to get LTO enabled, or does it always occur (unlike with host-side code where you need to compile with an -flto switch?

einpoklum
  • 118,144
  • 57
  • 340
  • 684
  • 1
    I haven't explored that new feature yet, but would *assume* that function inlining across compilation units is one of those capabilities. Should be easy enough to confirm or refute with a simple experiment. – njuffa Feb 24 '21 at 21:57

2 Answers2

2

Partial answer:

To enable link-time optimization, use --dlink-time-opt (or dlto) when invoking the NVCC compiler, both for compilation and for device-side code linking. No (link-time) optimization will be applied if the compiler can't find the relevant intermediate information.

einpoklum
  • 118,144
  • 57
  • 340
  • 684
0

my guess is that -dlto has to be with compile time and link time, if you link your program using non-nvcc, such as gcc or g++, then you may not get the best performance

Yuxiang Lin
  • 1