john-dev - md5crypt-opencl optimizations

Follow @Openwall on Twitter for new release announcements and other news

[<prev] [next>] [day] [month] [year] [list]

Message-ID: <f3e0cc72783339c12b7af0b69d36e3d6@smtp.hushmail.com>
Date: Thu, 10 Jan 2013 02:45:02 +0100
From: magnum <john.magnum@...hmail.com>
To: "john-dev@...ts.openwall.com" <john-dev@...ts.openwall.com>
Subject: md5crypt-opencl optimizations

Before:

   Benchmarking: md5crypt [CUDA]... DONE
   Raw:    732298 c/s real, 732298 c/s virtual

   Using device 0: GeForce GTX 570
   Benchmarking: md5crypt [OpenCL]... DONE
   Raw:    232672 c/s real, 234057 c/s virtual

   Using device 0: Tahiti
   Benchmarking: md5crypt [OpenCL]... DONE
   Raw:    241920 c/s real, 3518K c/s virtual

Basically I just got rid of global memory use in the inner loop and reduced register waste (context buffer was using ints for storing single chars). OpenCL now:

   Using device 0: GeForce GTX 570
   Benchmarking: md5crypt [OpenCL]... DONE
   Raw:    892405 c/s real, 892405 c/s virtual

   Using device 0: Tahiti
   Benchmarking: md5crypt [OpenCL]... DONE
   Raw:    607171 c/s real, 8601K c/s virtual

Still far from Hashcat, but a little less embarrasing now. I have no idea why Tahiti is so slow here. Maybe we should use local memory.

magnum

Confused about mailing lists and their use? Read about mailing lists on Wikipedia and check out these guidelines on proper formatting of your messages.