I am using libNVVM 10.1 to generate PTX code from LLVM IR. I would like to generate system-wide atomic instructions, like for instance:
atom.global.sys.add.u32 d, [a], b;
Intrinsics that support this have been added to the main LLVM code line but it seems that libNVVM does not recognize them.
I would expect that the following intrinsic function would be translated into the instruction above. But it is actually not recognized.
Is there a way to generate system-wide atomic instructions with NVVM?