Further reading of open and closed PRs:
https://github.com/xianyi/OpenBLAS/issues/3139
suggests that TARGET=GENERIC is required, as the result is somehow
different than TARGET=native. The compile host should generate
old-cpu compatible code but doesn't quite. GCC vs OpenBLAS issues
afaik.
On 2021-12-21 00:45, Stuart Henderson wrote:
> On 2021/12/21 11:31, Kevin Lo wrote:
>> My intention was not replace blas and lapack. If porters/users would
>> like
>> to use openblas instead of blas and lapack, feel free to use it, no?
>
> If openblas does not replace existing libraries, then we need to be
> sure
> that other ports don't pick up openblas if installed when they're
> built.
Some codes and build systems are quite aggressive in hunting down
the necessary prerequisites. (Hello cmake.) Is there a way to
crash a build if it links to the wrong lib? Something like a
"poisoned" openblas?
A single-threaded build that covers old-enough CPUs as well as
avx512 should work OK.
The Makefiles for openblas are an interesting walk down memory
lane...I haven't thought about katmai or shanghai CPUs for a long
time.
John
No comments:
Post a Comment