CPU feature detection fails with NATIVE=off about ntl HOT 5 CLOSED

from ntl.

I know that it would be annoying to implement, but for the sake of argument, a perfect solution for us would be to have build flags that allow the user to override the detection of each CPU feature. The correct flag values could then be determined by looking at the CPUID, as in https://github.com/projg2/cpuid2cpuflags/, and specified by the user -- relieving NTL of the responsibility to detect them.

So, for example, a new flag:

FMA=on|off|auto (default: auto)

If I don't specify it, the default (auto) would continue to work as it does now. But if I know that my CPU has FMA support and if NTL can't detect it (due to missing CXXFLAGS=-march or some other reason), then I could pass FMA=on to DoConfig.

I think that would work in most cases, but some refactoring would be required in any tests that check not only for the CPU feature, but also some additional stuff. Here I have AVX2 in mind, which checks not only for AVX2 support in the CPU, but also the size of int/long and the FP precision.

from ntl.

from ntl.

You're right. I thought there were fallbacks for the builtins but it's not for all of them. My CPUs are too old to test most but e.g. PCLMUL does fail without the right compiler flags.

I'll close this now since the description is misleading. But in light of that, my one remaining suggestion would be to have NTL_ENABLE_AVX_FFT=on ignored when the corresponding CPU features are unavailable. Right now, the build fails if NTL_ENABLE_AVX_FFT=on but NTL doesn't like the CPU, making it hard for us to determine if it's safe to turn such a feature on.

Thanks for clarifying.

from ntl.

from ntl.