Date: Tue, 11 Mar 2014 20:07:13 -0400 From: kzug <kzug10@...il.com> To: john-users@...ts.openwall.com Subject: Re: Help needed on CUDA / OpenCL on OS X 10.9 Thanks Magnum, I'll try all of that and report success or failure. > If you install a real gcc (which I do recommend for OSX), why not use a much newer version? As john.conf mentions a llvm-gcc 4.2, I stuck to 4.2 for the reason that I am very good at breaking things, and not so good for fixing them. On TuesdayMar 11, 2014, at 4:13 PM, magnum wrote: > On 2014-03-11 18:27, K Zug wrote: >> CUDA won't compile properly, and OpenCL, even though compiling will either >> hang, freeze, or return a "Warning: salt() returned misaligned pointer" or >> gives a "Abort trap: 6" when working on Jeff's DMG. >> Please note that when benchmarking, the dmg-opencl benchmark correctly. > > The misaligned pointer warning is more often than not just a red herring. And on x86 it's no problem anyway, since it can handle unaligned data. > > The hang/freeze problems are likely due to the fact this format does not auto-tune and the default GWS of 9216 will hang most GPU's weaker than state-of-the-art, and often takes the whole system down as collateral damage. Export GWS=512 (or so) or add a line in john.conf, section [Options:OpenCL] that says dmg_GWS = 512 (or so). > > >> I tried as far as I could, even installed GCC 4.2 via homebrew to no avail. > > If you install a real gcc (which I do recommend for OSX), why not use a much newer version? For best performance, use eg. 4.7 and the -native build target. See bottom section of doc/INSTALL. > > >> I am ready to sacrifice a rooster or wear a necklace of garlic if >> necessary. > > Only a last resort. > > >> $ make clean macosx-x86-64-cuda >> (...) >> /usr/bin/llvm-gcc-4.2: No such file or directory >> make: *** [cuda_common.o] Error 1 >> make: *** [macosx-x86-64-gpu] Error 2 > > Hm, so this was what you meant with gcc 4.2. Try this: > $ brew install llvm-gcc-28 > $ sudo ln -s /usr/local/Cellar/llvm-gcc28/2.8/bin/llvm-gcc /usr/bin/llvm-gcc-4.2 > > Works like a champ here with default Makefile. But I *only* use that bastard for CUDA. My normal "gcc" is a symlink in /usr/local/bin/gcc that points to (Homebrew) gcc-4.7 in the same directory - and /usr/local/bin is first in my path. And I build macosx-x86-64-native-gpu, which give me both CUDA and OpenCL support at once. > > >> $ make clean macosx-x86-64-opencl >> (...) >> ld: warning: directory not found for option '-L/usr/local/ssl/lib' > > This hints about the ancient native OpenSSL being used instad of a decent version. I don't think it matters for DMG but you can fix it like this: > > $ brew install openssl > $ ln -s /usr/local/opt/openssl/ /usr/local/ssl > > Phew. I should update that OSX section of doc/INSTALL. > > >> $ ./john --format=dmg-opencl --wordlist=/WList/run/output_single.txt >> /Users/kris/Dropbox/Public/jeff-sparsebundle-hash/hash >> Device 1: GeForce GT 750M >> Local worksize (LWS) 64, Global worksize (GWS) 9216 >> Loaded 1 password hash (dmg-opencl, Apple DMG [PBKDF2-SHA1 3DES/AES OpenCL]) >> Press 'q' or Ctrl-C to abort, almost any other key for status >> Abort trap: 6 > > Like I said, it's probably the GWS figure that's too high. try this: > > $ GWS=128 ./john --format=dmg-opencl --wordlist=/WList /run/output_single.txt /Users/kris/Dropbox/Public/jeff-sparsebundle-hash/hash > > Check for lag of response when pressing a key for status. If it's immediate, try doubling that GWS value. If response lags more than 5-10 seconds the GWS is too high. > > magnum >
Powered by blists - more mailing lists
Confused about mailing lists and their use? Read about mailing lists on Wikipedia and check out these guidelines on proper formatting of your messages.