Date: Fri, 20 Apr 2012 22:55:51 +0300 From: Milen Rangelov <gat3way@...il.com> To: john-dev@...ts.openwall.com Subject: Re: Pragma unroll I would rather avoid pragma unroll too unless you stick to SDK 2.6. APP SDK 2.5 has many problems with this. Though anyway, even if working correctly, it tends to be counterproductive in many cases. You need to do very careful profiling. Often less conditionals just can't compete with less occupancy due to GPR usage. And loop unrolling can easily increase GPR utilization beyond what you expected. On Fri, Apr 20, 2012 at 7:21 PM, Solar Designer <solar@...nwall.com> wrote: > magnum, Sayantan - > > On Fri, Apr 20, 2012 at 05:00:50PM +0200, magnum wrote: > > If anyone cares, here is the correct one: > > > > #ifdef cl_nv_pragma_unroll > > #pragma OPENCL EXTENSION cl_nv_pragma_unroll : enable > > #endif > > Thanks! > > Sayantan - can you make use of this in your kernel, please? IIRC, right > now you have the code unrolled manually, without even using #define's > for the unrolling (like we do in C source code). > > Alexander > Content of type "text/html" skipped
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.