Follow @Openwall on Twitter for new release announcements and other news
[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Date: Wed, 02 May 2012 22:59:26 +0200
From: magnum <john.magnum@...hmail.com>
To: john-dev@...ts.openwall.com
Subject: Re: Nvidia compiler verbosity

On 05/02/2012 07:05 PM, Claudio André wrote:
> Hi, now it is possible to enabled verbose on NVIDIA compiler. No side
> effects.
> 
> Register, local, shared, and constant memory usages are reported by the
> compiler when compiling with the -cl-nv-verbose build option.

This never worked for me. Maybe I should try again.

> Let me know if you guys would like to have this on default version (only
> if REPORT_OPENCL_WARNINGS is defined).

Yes, I think it should default to be activated when
REPORT_OPENCL_WARNINGS is defined. Perhaps we should undefine that when
releasing Jumbo (the pragma unroll output is mostly annoying once you
know it works)

>     sprintf(include, "-I %s %s %s%d %s %s", path_expand(pathname),
>             get_device_type(dev_id) == CL_DEVICE_TYPE_CPU ?
>             "-DDEVICE_IS_CPU" : "",
>                 "-DDEVICE_INFO=", device_info,
> + =>                gpu_nvidia(device_info) ? "-cl-nv-verbose" : "",
>             "-cl-strict-aliasing -cl-mad-enable");

When I tried this earlier I never ever got any output. Some googling
indicated there might be some kind of PTX caching so it will only show
up with new code, but some rewriting just to avoid a cache hit did not
do the trick. Do you always get the output, even if running same kernel
two times in a row?

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.