Date: Thu, 19 Jul 2012 22:26:32 +0400 From: Solar Designer <solar@...nwall.com> To: john-dev@...ts.openwall.com Subject: Re: Keychain format OpenCL implementation, please review Dhiru - On Tue, Jul 17, 2012 at 10:16:48PM +0400, Solar Designer wrote: > On Tue, Jul 17, 2012 at 10:02:05PM +0530, Dhiru Kholia wrote: > > Speed is 132K c/s on ATI 6970. It is roughly 300x faster than > > single-core CPU version (AMD X3 720). Hopefully, I haven't screwed > > things up ;) > > I guess this means that your CPU code is a lot slower than it can/should > be. ;-) And your 140K c/s figure for 7970 (mentioned elsewhere) > suggests that you might still be spending much time in CPU code, which > is why changing the GPU makes little difference. FYI, I am seeing 49% GPU usage when running your keychain-opencl code currently in magnum-jumbo on FX-8120 and HD 7970. This is a non-OpenMP build. So a 2x speedup might be possible with the same OpenCL kernel by optimizing the CPU code and/or running it simultaneously with the GPU code. To better support the latter, I am going to introduce support for "block IDs" into the formats interface, so that you could interleave two such blocks. 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.