[Developers] inv problems.

John Sibert sibert at hawaii.edu
Sat May 21 18:34:12 PDT 2011


I'm not sure how to interpret the display with respect to the number of 
calls to
cltudecomp_for_adjoint::ludecomp_pivot_for_adjoint_1()
(if that is what you are talking about),  but I note that number of 
calls to new exceeds the calls to delete by 14 calls. Is there a memory 
leak?
John

On 05/21/2011 06:17 AM, Derek Seiple wrote:
> I looked through the file Dave gave me with the inv code in it. I
> didn't look very different from what is in version 10. I swapped it
> out any way to see if it made any difference in the 'vol' example, but
> it ran at the same speed. The profile I have of the 'vol' example
> (attached: open with kcachegrind) suggests that a lot of time is being
> spent in the headers for the lu-decomposition. Specifically the
> decomp_for_adjoint.
>
> Derek
>
> On Thu, May 19, 2011 at 4:36 PM, Derek Seiple<dseiple84 at gmail.com>  wrote:
>> Hi All,
>> Yes, I have made some improvements, but it is not yet back to the
>> speed that 9.1 had. I am going to make one more change to the code and
>> then see where we are. Once I've made the change and tested it I'll
>> report back.
>>
>> Derek
>>
>> On Thu, May 19, 2011 at 3:13 PM, John Sibert<sibert at hawaii.edu>  wrote:
>>> I seem to recall a recent message from Derek implying that he had fixed the
>>> speed issue.
>>> John
>>>
>>> On 05/19/2011 08:02 AM, Arni Magnusson wrote:
>>>> Hi Dave,
>>>>
>>>> I notice that I'm the only recipient of this important question. I
>>>> appreciate your faith in me, but efficient linear algebra with AD objects is
>>>> really outside my area of expertise.
>>>>
>>>> I haven't looked at the inverse or solve code (old, new, or proposed), but
>>>> posted the GCC profile diagnostics to help the development team to improve
>>>> the performance of the 'vol' example, which became 3 times slower between
>>>> 9.1 and 10.0. The profile diagnostics point at some culprit functions, and
>>>> Derek Seiple told me (25 Apr email) that he is currently following those
>>>> leads.
>>>>
>>>> This is definitely a high-priority issue. Given the nature of the problem
>>>> (we have 9.1 code that works fast) and the upcoming developer workshop, I'm
>>>> optimistic it will be solved in version 10.2 or 11.0.
>>>>
>>>> Cheers,
>>>>
>>>> Arni
>>>>
>>>>
>>>>
>>>> On Thu, 19 May 2011, dave fournier wrote:
>>>>
>>>>> Hi,
>>>>>
>>>>> I'm a bit woried about these problmes with inv and maybe solve taking so
>>>>> long.
>>>>>
>>>>> It kind of breaks the entire software. That is what I alway expected to
>>>>> happen eventually.  The idea is that it is supposed to be fast. I produced a
>>>>> sketch of the code proper working code but it seems to have been screwed up.
>>>>> Any thoughts?
>>>>>
>>>>> Dave
>>>>>
>>>> _______________________________________________
>>>> Developers mailing list
>>>> Developers at admb-project.org
>>>> http://lists.admb-project.org/mailman/listinfo/developers
>>>>
>>> --
>>> John Sibert
>>> Emeritus Researcher, SOEST
>>> University of Hawaii at Manoa
>>>
>>> Visit the ADMB project http://admb-project.org/
>>>
>>>

-- 
John Sibert
Emeritus Researcher, SOEST
University of Hawaii at Manoa

Visit the ADMB project http://admb-project.org/



More information about the Developers mailing list