Date: Wed, 5 Jun 2013 05:08:18 +0400 From: Solar Designer <solar@...nwall.com> To: john-dev@...ts.openwall.com Subject: Re: bcrypt-opencl On Tue, Jun 04, 2013 at 09:56:26AM +0530, Sayantan Datta wrote: > There was a problem with loop unroll(automatically done by compiler). So I > unrolled it manually. This kind of problem is very common with AMD's > opencl. It generally happens with for loops with smaller loop count. > Compiler doesn't unroll properly when the unroll is a bit complicated. OK. Thank you for working around this problem. > Another concern is that the kernel is spilling registers, which it > shouldn't. This is lowering the performance too. Most likely the P box is > spilling. > > Device 1: Tahiti (AMD Radeon HD 7900 Series) > Optimal Global Work Size:1024 > Benchmarking: bcrypt-opencl ("$2a$05", 32 iterations) [Blowfish OpenCL]... > DONE > Raw: 3960 c/s real, 179200 c/s virtual Yeah. This is a slight performance regression compared to what we had with Catalyst 12.8 - IIRC, the speed was 4200+ c/s. It'd be nice if you manage to avoid the register spilling. Thanks! 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.