[PowerPC] Bus error when using OpenBLAS with PPC970 kernel (on G5): address 0x5dc4, cause 'invalid alignment' (but no error with PPCG4 kernel)
#4376
Comments
|
failure on launch would almost certainly have to be compiler or flags used, as it wouldn't have had a reason to access one of the cpu-specific BLAS kernels yet. the only other cause I can imagine is that the default BUFFERSIZE is somehow too large for this hardware (but I guess that would not raise sigbus) |
|
@martin-frbg I will rebuild PPC970 now, first dropping P. S. It is a bit strange case; I am pretty sure that It would be nice if someone could test PPC970 kernel on G5 with Linux or BSD. (Just in case, FreeBSD 13+ has broken multicpu support for PowerPC, so 12– are needed.) |
|
@martin-frbg Okay, @iains Iain, could you please say, if For some reason here it looks like |
|
@martin-frbg So, Now, I have no idea whether this will apply to Linux and BSD cases. I am also not sure if |
|
Thanks for this update - I have no idea either, but with hardware that old it is probably safer to have it working at all, rather than try to get the last millisecond of speedup out of compiler flags |
|
from memory (I am tied up with other stuff at the moment)
|
|
@iains Thank you very much! This is with P. S. By extension, we should also avoid |
on G5 boxes it might produce surprising results indeed (but, again, I have not tested what a recent compiler version actually does with this - there have been some changes in the handling of this stuff in the GCC-14 development cycle). We'd need a better bug report than "it doesn't work" ;) |
|
A quick test on my G5 box confirms:
|
@martin-frbg We got some obscure issue with PPC970 kernel. It does not show up with every dependent of
OpenBLAS, but did show up in two cases. A notable one is withR-float, which links to OpenBLAS and fails on launch like this:Crash log:
There is no failure when PPCG4 kernel is used, both with and without optflags. (My initial thought was to blame
-mcpuor/and-maltivec, but with PPCG4 kernel they do not break anything.)Second case is
flexiblas, which fails to detect OpenBLAS as functional at configure when PPC970 kernel is used, but detects it correctly with PPCG4 kernel: mpimd-csc/flexiblas#40There are no problems with linking against either version of OpenBLAS with whatever dependents I tested so far.
Notice, that I build for
ppc(32-bit), but on G5 cpu, so PPC970 is supposed to be the right one.What in principle might go wrong:
I have to rebuild OpenBLAS again a couple of times to rule out some of there, hopefully. To be updated.
The text was updated successfully, but these errors were encountered: