Compiler does not gracefully handle multi-core builds

[alerque]

Re:

https://github.com/ROCm/rocFFT/blob/1680fbde578bbc65d1abfbaadf53e49dd5e0494c/library/src/rtc_compile.cpp

This is what CPU utilization looks like:

Somehow the usage is artificially capped at 50/70% of each core rather than just getting the job done.

[evetsso]

The build is launching a lot of individual processes to compile kernels for distribution with the library. This is known to not be particularly efficient, but the long-term solution would be in hipRTC and its underlying LLVM library to allow for usefully multithreading compilation within a process.

In the short term, this overhead can be mitigated by either selecting a subset of GPU architectures to build with the GPU_TARGETS CMake property, or to disable this part of the build completely by setting ROCFFT_KERNEL_CACHE_ENABLE to OFF.

Any kernels that are needed by the library that aren't built at the time of building rocFFT will be generated and compiled at runtime. So no library functionality is impacted when these CMake parameters are changed.

Closing this issue, as nothing will be fixed for it in the short term and there are workarounds.