Date: Tue, 25 Aug 2015 08:50:58 +0300 From: Solar Designer <solar@...nwall.com> To: john-dev@...ts.openwall.com Subject: Re: md5crypt-opencl On Mon, Aug 24, 2015 at 07:57:20PM +0200, magnum wrote: > On 2015-08-24 11:38, Solar Designer wrote: > >With the inline ptx asm enabled (need proper compile-time check for it): > > > >Raw: 1664K c/s real, 1664K c/s virtual > > I added a proper check in 7531c0b, although it's starting to go south - > I think we should actually export the major/minor numbers to macros > instead of defining functions for specific sm_xx combinations. But > anyway it works correctly right now. Confirmed, thanks! Testing this further, I notice that our LWS/GWS tuning which tries powers of 2 only is highly non-optimal for this card. Simply adjusting LWS and GWS, I get much better results than the above: [solar@...er run]$ LWS=96 GWS=129024 ./john -test -form=md5crypt-opencl -dev=5 Device 5: GeForce GTX TITAN Benchmarking: md5crypt-opencl, crypt(3) $1$ [MD5 OpenCL]... DONE Raw: 1984K c/s real, 1984K c/s virtual and actual cracking: [solar@...er run]$ LWS=96 GWS=129024 ./john -form=md5crypt-opencl 6.md5crypt.pw -dev=5 -mask='?l?l?l?l?l?l?l?l' [...] 0g 0:00:02:03 0g/s 0p/s 2395Kc/s 2395KC/s GPU:76б╟C util:99% fan:50% aaaaaaaa..lwihaaaa (was ~1850K before) Also, GWS=129024 is actually much lower than what was auto-tuned (512K), which is helpful for small wordlist runs and to reduce the amount of work lost on interrupt/restore. In fact, seeing how low we can get: [solar@...er run]$ LWS=96 GWS=16128 ./john -test -form=md5crypt-opencl -dev=5 Device 5: GeForce GTX TITAN Benchmarking: md5crypt-opencl, crypt(3) $1$ [MD5 OpenCL]... DONE Raw: 1756K c/s real, 1774K c/s virtual This is suboptimal, but still better than our auto-tuning - and that's ~32.5 times lower than the auto-tuned GWS, so has much better usability for quick runs. Alexander
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.