Incorrect results on Haswell Win64

[spalato]

Hi,

There were discussions to get access to my machine for debugging purposes by @wernsaar and @tkelman.

I've been using OpenBlas through julia in cygwin64 on windows 8. There are a few layers there, and I don't know what the best way to proceed: setting up a cygwin SSH server or using teamviewer through windows. I will be on another network next week (Monday-Wednesday). This is a laptop, so it moves around quite a bit. Please let me know about time needs so we can plan accordingly.

Please let me know what kind of access/resources is needed.

[wernsaar]

Hi,

for remote access I can use ssh or vnc tunneled by ssh.
I also can use teamviewer, which version ?

Best regards

Werner

[wernsaar]

Hi,

I published a new dynamic binary for Windows-64bit with 64bit Integer size on sourceforge. Download is available from http://sourceforge.net/projects/openblas/files/v0.2.10-pre/
The test results for piledriver and haswell are correct.

Best regards
Werner

[spalato]

Hi,

Test results:
Using libopenblas.dll in the julia directory: tests linalg2 and linalg3 now pass. linalg1 fails later than before. Results from linalg4 didn't change. See detailed results

Tests succeed when setting OPENBLAS_CORETYPE=nehalem, as before.

I will try to setup both ssh and teamviewer today. That may take some time, with me being a windows user.

[wernsaar]

Hi,

@spalato
I installed teamviewer 9 on my linux box and we can try it, if you are ready

Best regards
Werner

[tkelman]

I'd be happy to join in and give some clarification on Julia-OpenBLAS interface and build process, if that would be helpful.

[spalato]

Sorry for the silence. I've been quite busy.
I'm currently attending a conference. I don't know how paranoid the network is here. We can try tomorrow (Tuesday 24). After that, I won't be available until next Tuesday (30th).

[tkelman]

I can reproduce this on an i7-4702MQ, x86_64-w64-mingw32. Setting OPENBLAS_CORETYPE=ivybridge results in Julia's linear algebra test suite passing. I built a debug version from @wernsaar's latest develop branch, let me know if there are any debugging steps in gdb or otherwise you would like us to try.

[wernsaar]

The bugs were found and fixed. I now close this issue

[tkelman]

I can confirm that the Haswell problems appear to be fixed on @wernsaar's branch, thank you.