NVCC is composed of two parts: the compiler driver, which is close-source, and the Open64 compiler, which is open-source. I think most of the CUDA bugs are compiler bugs, which the NVCC team may check into the Open64 source tree quickly. If we can compile NVCC from the Open64, we will get the bug fix with no delay. We won’t have to wait until the next release.
I am familiar with compiling Open64. I fixed quite a few bugs in that compiler. So compiling it should not be too hard. I only need to get the right configuration for the Makefiles. But I have no idea if I can get the compiler driver to work with the “custom built” Open64.
Is this possible? Where can I get the nVidia-specific Makefiles for Open64? And how do I integrate the custom built Open64 with NVCC compiler driver?