Pull request opened here https://github.com/JuliaLang/julia/pull/7476

If you get into trouble compiling the Haswell kernels due to "no such 
instruction: vpermpd", give NO_AVX a try. Or upgrade binutils, up to you.


On Monday, June 30, 2014 3:50:58 PM UTC-7, Florian Oswald wrote:
>
> thanks tony, just giving this a try from scratch. to be honest, I dont' 
> know the exact age of those processors. Let me see how far I get without 
> those newer binutils.
>
> On Monday, 30 June 2014 23:17:00 UTC+1, Tony Kelman wrote:
>>
>> That arrayops.jl test failure is because you're missing an LLVM patch, 
>> see https://github.com/JuliaLang/julia/issues/7197#issuecomment-46688404
>>
>> A lot of these build customization options should be better-documented, 
>> especially for the heterogenous-processor case. Don't copy Make.inc to 
>> Make.user, rather create a new Make.user from scratch where you set only 
>> the options you need to change from their defaults.
>>
>> If you have both Intel and AMD processors, you will definitely need to 
>> have OPENBLAS_DYNAMIC_ARCH enabled. Are any of your processors new enough 
>> to support AVX instructions? If so, you'll need a newer version of binutils 
>> than provided by Red Hat 6. See 
>> https://github.com/JuliaLang/julia/issues/7363#issuecomment-46927356
>>
>> I'll open a pull request shortly to add another OpenBLAS build option to 
>> disable AVX instructions for when you need DYNAMIC_ARCH but installing 
>> newer binutils is not possible.
>>
>>
>> On Monday, June 30, 2014 2:56:48 PM UTC-7, Florian Oswald wrote:
>>>
>>> additional to that, `make testall` errors with this:
>>>
>>> while loading arrayops.jl, in expression starting on line 895
>>>
>>> ERROR: test failed: i7197() == (2,2)
>>>
>>>
>>> On 30 June 2014 22:50, Florian Oswald <[email protected]> wrote:
>>>
>>>> Hi all,
>>>>
>>>> I've got some very basic questions about building Julia from source on 
>>>> an hpc cluster. I got it to build fine a month ago, but now the makefile 
>>>> changed. I was partially  succesful this time as well (my test runs), but 
>>>> some strange things happen and there's a lot of stuff in that makefile I 
>>>> dont' understand.
>>>>
>>>>    - I can't get Make.user to work. My understanding was that i `cp 
>>>>    Make.inc Make.user`, then change the options I want changed, then 
>>>> `make`? 
>>>>    Nothing happens. Do I have to delete Make.inc after having made my 
>>>> copy? 
>>>>    - I needed to set `JULIA_CPU_TARGET=core2` when I last got it to 
>>>>    build. The cluster is built out of several different CPU types, so not 
>>>>    setting this resulted in target mismatch errors. I can't find that line 
>>>>    anymore in the current Make.inc, so I changed 
>>>>    https://github.com/JuliaLang/julia/blob/master/Make.inc#L345 from 
>>>>
>>>>          JULIA_CPU_TARGET ?= native
>>>> to
>>>>          JULIA_CPU_TARGET = core2
>>>>
>>>>    - Is this still recommended practice or should I not touch this at 
>>>>    all?
>>>>    - the openblas default build failed. I set this: 
>>>>    - OPENBLAS_DYNAMIC_ARCH=0
>>>>    - OPENBLAS_USE_THREAD=0
>>>>    - OPENBLAS_TARGET_ARCH=BARCELONA 
>>>>    - It says in the trouble-shooting section on 
>>>>    https://github.com/JuliaLang/julia/readme.md that a possibel 
>>>>    solution to problems is to set the last variable to BARCELONA for AMD 
>>>> and 
>>>>    NEHALEM for Intel. What if I have both CPUs? Does it matter? 
>>>>    - With those settings it compiles, and my little test script runs. 
>>>>    However, I get this error/warning from all workers:
>>>>    - ./julia: /lib64/libz.so.1: no version information available 
>>>>    (required by /data/uctpfos/git/julia/usr/bin/../lib/libjulia.so) 
>>>>    - In case this is helpful, my `cat /proc/version` is:
>>>>    - Linux version 2.6.32-279.19.1.el6.x86_64 ([email protected]) 
>>>>    (gcc version 4.4.6 20120305 (Red Hat 4.4.6-4) (GCC) ) #1 SMP Tue Dec 18 
>>>>    17:22:54 CST 2012 
>>>>
>>>>
>>>>
>>>

Reply via email to