|
Message-ID: <542156d4f50e111f78c0e07202270af4@smtp.hushmail.com> Date: Sat, 19 Jan 2013 14:30:06 +0100 From: magnum <john.magnum@...hmail.com> To: john-dev@...ts.openwall.com Subject: Re: New shared find_best_gws On 16 Jan, 2013, at 3:38 , Claudio André <claudioandre.br@...il.com> wrote: > Hi, i created a new shared find_best_GWS. > > It uses: > - format->set_salt, format->set_key, format->crypt_all, etc. > - requires create_clobj and release_clobj (you have to give pointers to them so find_best_GWS can (m)alloc necessary memory). > - (maybe) a new crypt_all_benchmark function (with unique events for all OpenCL commands). > > It can handle a set of formats for sure. Anyway, it will be nice if new formats authors try it (at least to show its problems). > > As usual, you can dance naked using legacy code and it still works. Good stuff. I see you also added a opencl_find_best_lws() function. Is that a new version of opencl_find_best_workgroup() with more flexibility (eg. for split-kernel formats)? I think we might want to adopt core JtR style and move/add comments to common-opencl.h. For example, this comment in the .c file might be better put in the .h file: /* * NOTE: Requirements for using this function: * * - Your kernel (or main kernel) should be crypt_kernel. * - Use profilingEvent in your crypt_all() when enqueueing crypt_kernel. * - Do not use profilingEvent for transfers or other subkernels. * - For split kernels, use firstEvent and lastEvent instead. */ And similar how-to comments should be added for some/many of the other functions. For deprecated functions (which we might end up making static or rationalize away eventually) we could point to the newer/better alternative. Obviously we shouldn't bother with this for unstable, only bleeding. This would be of great help for developers trying to catch up, including myself! 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.