Date: Thu, 03 May 2012 11:25:49 -0300 From: Claudio André <claudioandre.br@...il.com> To: john-dev@...ts.openwall.com Subject: Re: cryptsha512-opencl Em 03-05-2012 07:59, magnum escreveu: > Claudio, > > Do you think you can get it working on AMD soon too? I saw a really crazy thing in 7970. For example: The code: output->v = (i < rounds); output->v = (i >= rounds); output->v = rounds; output->v = i; The output (after o loop, so "i" should be 4999 [decimal]): Tahiti: 0000000000000000 0000000000000001 0000000000001388 0000000000000001 <==> (1 >= 5000)??????? CPU: 0000000000000001 0000000000000000 0000000000001388 0000000000001387 Trying some other tests, i found a condition where: (i < rounds) == false AND (i >= rounds) == false My understanding is that ((int + uint) + (maybe) local memory access) do the trick. Not sure. I don't want to create a kernel only to succeed on this crazy environment. Trying to find a pattern. > We might try to > wrap up a Jumbo release soon. Also (or at the same time but this is > lower prio) you might want to get rid of byte stores. The AMD will honor > this setting: > > #ifdef cl_khr_byte_addressable_store > #pragma OPENCL EXTENSION cl_khr_byte_addressable_store : disable > #endif > > This will make your format fail until you have replaced all byte pointer > writes to macros like the PUTCHAR one in rar (it's normally as simple as > that). Note that besides making your format work on more devices, this > will probably also speed things up. > > BTW, that pragma disable *should* work on nvidia too but they just > silently ignore it. > > magnum Will do it.
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.