Failed build on FreeBSD #829
Comments
|
Seems like our patch for dSFMT.h is not being applied properly. Is there a functional difference between the |
|
There are both patch and gpatch on FreeBSD, I just discovered prompted by your question. Analogous to make and gmake, I guess. So, probably, yes. I am not sure the best way to set this up -- I don't want to link patch to gpatch in the file system as root (it will probably mess a pile of stuff up), but I do want to make it work for the entire build of Julia. Ideas? |
|
Can you replace all calls to |
|
Changed as recommended, get the following errors: julia/ (j=0,r=0)$ gmake |
|
can you try |
|
Cranking away. Summary of changes: (1) patch to gpatch. (2) copied install-name. (3) changed the latter to /usr/local/bin/bash. (4) got rid of Microsoft Windows newlines in that file. Thanks! |
|
Whoopsie! Here is the next error. In file included from ../common.h:326, |
|
This is an upstream issue. Just filed it here: OpenMathLib/OpenBLAS#106. For now, can you replace |
|
Installed wget, restarted gmake, and got this error w/r/t lapack. julia/ (j=2,r=0)$ gmake 100%[===============================================================================================>] 6,147,915 378K/s in 16s 2012-05-12 15:50:54 (367 KB/s) - `lapack-3.4.1.tgz' saved [6147915/6147915] md5sum: not found |
|
Seems like OpenBlas is trying to verify the md5 checksum of the downloaded package. Is there anyway to compute the MD5 checksum of a file on FreeBSD? |
|
http://lists.freebsd.org/pipermail/freebsd-questions/2004-March/038150.html Seems like "md5" is what we want. Never used it, so I don't know how it works exactly. |
|
grepped and changed md5sum to md5 in the makefile, restarted make, got the following: julia/ (j=2,r=0)$ gmake |
|
What does manually downloading http://www.netlib.org/lapack/lapack-3.4.1.tgz and using |
|
Yes, same md5. Ran your comand with "gmake", compiling now. will report. Fails at some point with a missing "md5sum", probably because a Makefile gets downloaded after I go through and change it to only use "md5". I just symlinked md5sum to md5, and we will see if that makes it work. |
|
This has been fixed upstream. Can you download https://github.com/xianyi/OpenBLAS/zipball/develop and extract it into the deps/openblas-v0.1.1 directory to verify that it works as expected? |
|
I re-cloned and am building now. I crapped out with a bad md5sum, but I couldn't figure out where. Any pointers on how to disable md5 checking, just in case? |
|
I re-cloned, it failed on a bad md5sum. The check seems to be on line number 243 of openblas' Makefile -- all that awk stuff looks pretty tricky; it also might be relying on gnu awk, so I edited to "gawk". I downloaded the zip you specificied manually and it worked, but I didn't try to use it in the build. How do we disable md5sum checking, for now? (Thanks for all your help! I might take a break for tonight soon, but I should be able to re-attack in the morning. Once we get it running, I will try to write a summary.) double checked lapacks md5: MD5 (lapack-3.4.1.tgz) = 44c3869c38c8335c2b9c2a8bb276eb55 Failed on this: [pao: formatting] |
|
I have gotten much farther with my freebsd build (will report later -- basically by using my PATH variable and a directory of symlinks to point to GNU tools). Now I am erroring out when it calls ld with gfortran. I just exported LD_RUN_PATH to include the symlink directory. I will report further when I have made more progress. |
|
Thanks for working on this. Would be great to officially support FreeBSD too! |
|
This used to work a little while ago, but with nobody (that I known of) in the core development group using FreeBSD, it's really easy to lose track. Another reason we should get around to setting up some sort of CI sometime ;). |
|
closer, but not finished (see below). Any ideas on how to get ld to find gfortran? gfortran46 is installed by FreeBSD ports, but I symlinked gfortran to it, and also set envars before compiling, to no avail. (Oh -- I am glad to work on -- FreeBSD awesome, and I desperately want a new math language to use ;)) END OF TESTS julia/ (j=0,r=0)$ echo $LD_LIBRARY_PATH |
|
It's looking for libgfortran, not the gfortran executable. |
|
Should I symlink the lib? Have to run, will try tonight, but I am still interested in any thoughts... |
|
Created symlinks to libgfortran in /usr/lib (which is not a good thing to do on a FreeBSD machine). Was unable to figure out how to make export LDFLAGS+='-L/usr/local/lib/gcc46' actually work. It seemed like the makefiles couldn't see it, even though I marked it with export. Now I have what looks like a bug: /usr/bin/ld: cannot find -l/usr/local/lib Any pointers on how to diagnose which Makefile could be fixed to deal with these problems? Does julia really build on arbitrary Linux machines? I am sure it builds on the devs machines, but I would be comforted (and slightly surprised) to hear that it can be cloned and built on a random Ubuntu box. This is hitting the point at which I don't really want to bother anymore, as I fail to rebuild dependencies that I know I have installed locally, etc, etc, etc. I think I am going to try one or two more times, but I might pitch it and come back in a few months... |
|
I don't know about "random Ubuntu box," but "typical Ubuntu box" works fine. Other devs use Fedora and OSX so we tend to notice when any of those break. |
|
Oh well. I can't really spend any more time on it; I don't really know enough to know how to determine what the problems are at this point. EDIT: I think I finally got the julia core to build, but failed on too many of the dependencies to make it worth continuing. Some of these dependencies are notoriously hard to build even on a Linux system, where they assume GNU tools, and when you try it on a FreeBSD system with multiple small differences (md5 != md5sum, make != gmake, etc, etc) it is a real pain, probably even for someone with better experience porting software than I have. I also found the build process really hard to diagnose -- I crapped out with errors, but even with gmake -v, I couldn't figure out even what dependency package I was failing on. On a FreeBSD system, it would probably be best to rely on pre-installed dependencies rather than building them, since these deps have been ported already with the ports system. If I can help in any way, let me know, but I don't think I can work on this anymore, like I say above. |
|
For relying on system libraries, look in Make.inc for the |
|
Thanks for the pointer. I set everything to be 1, and made sure I installed it all, and then ran gmake, with the output below. [pao: formatting] |
|
SuiteSparse is special; take a look at #785 (comment). |
|
I set the flag in Make.inc to download and build suitesparse and it seems to get past that. Now the problem is that nothing seems to be looking in /usr/local/(lib|include|whatever). When I symlink a library into /usr/lib then I make some progress, but have to symlink the next thing, etc... EDIT: What i mean is "is there a place to set the search path for include and lib to include both /usr and /usr/local varieties, and have it propagate down? Like Make.inc?" |
|
Julia should now build on FreeBSD 9.0, if you do the following:
|
|
@nolta that was some awesome work! I presume this can be closed once @forkandwait confirms. |
|
@nolta thanks! I will try tonight. I have a few questions though:
All that being said... super thanks to all for working with me, Julia is is the most excited I have been about a language in a while, and everyone seems great! |
|
@forkandwait: you can force the download and compile of OpenBLAS immediately by doing We definitely shouldn't have any files committed anywhere with MS DOS new lines. Which files are you seeing this for? I'll try to fix it. |
|
I see newlines whenever I open any files in vi (emacs hides the conversions). little "^M" at the end of the lines. I might be misconfigured ("misconfigured from birth"... hehe) |
|
Somewhat ironically, the only file that contains the |
|
I'm not sure why that's happening, but I don't see any '\r' characters anywhere so it must be a misconfiguration on your end, no matter how bizarre that may seem. |
|
Hmm. I am definitely seeing them on a new clone. I will try to figure it out -- maybe I am converting backwards. |
Yes, these patches have been accepted upstream.
My FreeBSD VM has the same |
|
@nolta Could you run patch --version and maybe look to see if it is linked? Have you done any other linking (like of libraries)? You say VM -- did you download an image from somewhere (someone might have made the image "nicer" for Linux users)? |
I'm guessing the
Nope, i used the standard iso: ftp://ftp.freebsd.org/pub/FreeBSD/releases/i386/i386/ISO-IMAGES/9.0/FreeBSD-9.0-RELEASE-i386-disc1.iso and here's the list of packages: |
|
@nolta Well, if I actually read the blurb about newlines, and set git to "input", seems to work pretty well. I still needed to override using my gnu tools directory (to fix md5sum). Hmm... I am currently bombing out when trying to connect to host for libunwind. So I have tried setting Make.inc to use it locally, but it still won't find my libs and includes (trying to compile again, so don't have the output to paste). Favor: could you try installing one of the dependencies (maybe readline?) and see if you get the same problem if you try to use it? If we need to set LDFLAGS and CPPFLAGS to get us /usr/local/lib, can we do it in Make.inc? Maybe that is a bad approach -- I am just a humble scripter, so I am learning a lot about real projects and Makefiles and such.... FYI: I am using ports, not packages, and I have a lot of them ;) Random thread about FreeBSD not checking /usr/local/* by default during compiles: http://forums.freebsd.org/showthread.php?t=19956 |
|
Are you still unable to download libunwind? |
|
Woot! I got it to build! md5 still doesn't work out of the box unless I tweak it like I said above. I wonder if it was binutils that gave you a linux style md5sum. I am going to let @nolta close it if he thinks its ready. We should open a new ticket to create an official freebsd port (which might take care of some of the weird path stuff, we can set it to use gmake, etc). I will start helping on that next week, after I play with the language a little bit. Thanks to everyone! |
|
Congrats to all involved in this effort. An official FreeBSD port would be incredible :-) |
|
Excellent, i'm glad it works for you @forkandwait. I think i'll leave this open until the next version of openblas is released. |
|
I think the only problem remaining is that my standard issue freebsd does not have md5sum, but rather md5. @nolta, could you run 'which md5sum' on your freebsd system? just curious, but maybe there could be a test in the Makefile. Also, I think md5sum (Linux) is called differently from md5 on FreeBSD. |
|
Doesn't md5 -r do the same thing as md5sum? -viral On 27-May-2012, at 10:10 AM, forkandwait wrote:
|
|
I'll double check tomorrow, but I thought this was fixed by the openblas patches. |
|
I got it to build by writing a single line md5sum script and putting it in my PATH. otherwise no changes were necessary. I think the problem happens in the lapack download, not openblas, I don't know. md5 -r does the same thing as md5sum, but I think the latter is hardcoded into one of the downloaded makefiles in lapack. |
|
The develop branch of openblas fixes this. We should have a release soon enough. -viral On 27-May-2012, at 6:31 PM, forkandwaitreply@reply.github.com wrote:
|
|
or you can just temporarily change the |
Hi all,
I followed the directions for building on FreeBSD, including installing libunwind, but I get an error. Any pointers would be appreciated. See below for some info.
My machine (more specs avail if important):
The most recent git commit in the log is e4ca3a9:
gmake output:
[pao: formatting]
The text was updated successfully, but these errors were encountered: