Tony Reix

When rebuilding all (libopenblas_power5p-r0.3.10.a) after complete clean, by scripts, with -O0 -g, or -O1 -g, or -O2 -g, there is no more bug. Before, it was built with -Ofast and...

> First rebuild was probably without setting TARGET... Yes. I forgot this. Now using BINARY and TARGET. Not perfectly awaken this morning... Now using: ``` # cat MAKE export OBJECT_MODE=32...

By the way, the number of cores to be used on AIX is not computed by OpenBlas, thus leading to NUM_CORES=2 . The attached patch does make use of the...

> Is there an advantage to using SC_NPROCESSORS_ONLN rather than _CONF ? The man of sysconf on both AIX and Fedora say: AIX : _SC_NPROCESSORS_ONLN Number of processors online. Fedora:...

> Just asking as we used to see the opposite effect (on armv8 however)... On AIX, the number of CPUs available for a VM is far more complicated than on...

About this issue, after testing with the appropriate options, it looks like there are issues with GCC and gdb, in 32bit only and Power8. - gdb is unable to show...

> returning tp the original topic, what is a bit suspicious to me is that the Power8 GEMM kernels have a surprisingly tiny STACKSIZE defined in 32-bit mode (400 bytes...

I do no master Boost. I used %ul because my error message on AIX is talking about "unsigned long". Anyway, using %16ld, I see: ``` 1) TONY: BEG f.min_v_: 0...

OK. I'm not in a hurry. Anyway, I think that there is an issue here. Maybe it has been fixed within last versions, I don't know, but I'd like to...

I'm now using v1.73.0. And I think I still see the same behavior. A) Random intermediate values. 1rst launch of test ``` # ../bin.v2/libs/gil/test/core/channel/algorithm_channel_arithmetic.test/gcc-9/debug/threading-multi/visibility-hidden/algorithm_channel_arithmetic 2>&1 | grep BEG TONY: BEG...