[CP2K-user] [CP2K:20421] Re: FFTW Wisdom File Error

[Frederick Stein]

Dear Santiago,
TLDR I finally fixed the culprit: The current master or the next release 
contain a proper fix of import/export of wisdom.
Long version: We missed the FFTW initialization for threaded environments 
such that the import was done with a non-threaded library and later 
threaded calls "turned" it to a threaded library such that the export 
occured with a threaded version (undefined behavior). Thus, the fftw-wisdom 
has to be called accordingly: without "-t <thread_count>", it is unthreaded 
(not compatible with CP2K), with this flag, it is threaded. Thus, creating 
wisdom with "fftw-wisdom" for CP2K requires you to set the number of 
threads. Incompatible wisdom files do not stop CP2K because with certain 
configurations we observed that the import/export is broken for unknown 
reasons.

Thank you for your cooperation and please excuse the inconveniences.
Best,
Frederick

Santiago Movilla schrieb am Freitag, 5. Juli 2024 um 11:54:45 UTC+2:

> Dear Frederick,
>
> The devel version has finally been installed in my computer center and 
> optimized for the architecture. I have tested the wisdom file generated via 
> fftw-wisdom and found it to be functioning without any issues (just for 
> estimate planning). For the time being, I believe this is the most 
> effective approach, as it offers a performance boost. Please let me know 
> the progress of the new release and about its compatibility with other 
> types of plans in the wisdom file.
>
> Thank you for the support, 
> Best,
>
> Santiago Movilla
>
> On Thursday, June 20, 2024 at 9:25:41 PM UTC+2 Frederick Stein wrote:
>
>> Dear Santiago,
>> Regarding the version: There will be a new version in 1-2 weeks if that 
>> is more suitable for your compute center.
>> I have just tried. It did work for me (at least with different planning 
>> types and a wisdom file for patient). I didn't check the number of threads. 
>> The details of the library are unknown to me. It is also difficult to check 
>> because there is no direct source code of FFTW3 available as it is the case 
>> with other libraries.
>> Tell me as soon as you have any news.
>> Best,
>> Frederick
>>
>> Santiago Movilla schrieb am Donnerstag, 20. Juni 2024 um 18:37:38 UTC+2:
>>
>>> Dear Frederick, 
>>>
>>> Yes, I am using version 2024.1. Right now it is impossible for me to 
>>> recompile the one from the Github repository (ITS must do it). Tomorrow as 
>>> soon as they can I will do it and let you know what happens. 
>>>
>>> On the other hand, if you create the wisdom with the default patient, 
>>> then the pre-planned routines cannot be used in CP2K in ESTIMATE type mode, 
>>> or is it possible? I thought FFTW was strict about the architecture, the 
>>> work plan and the number of threads.
>>>
>>> Thanks for the help today, when I have new tests I will let you know 
>>> here. 
>>>
>>> Best, 
>>>
>>> Santiago 
>>>
>>> On Thursday, June 20, 2024 at 6:27:53 PM UTC+2 Frederick Stein wrote:
>>>
>>>> Dear Santiago,
>>>> That is interesting. I used "./fftw-wisdom -v -c -o wisdom", i.e. I 
>>>> optimized in patient mode (default). I do not know the internals of FFTW3 
>>>> too well. Unfortunately, they do not provide any hint what the actual issue 
>>>> is.
>>>> According to your CP2K output, you are using the latest release 
>>>> (2024.1), not the development version? If you use an official release, is 
>>>> it possible for you to pull and compile the current master trunk of the 
>>>> CP2K Github repository? This one has the fix (which will hopefully work for 
>>>> you). Your version is too old for sure.
>>>> Best,
>>>> Frederick
>>>>
>>>> Santiago Movilla schrieb am Donnerstag, 20. Juni 2024 um 18:07:41 UTC+2:
>>>>
>>>>> Dear Frederick, 
>>>>>
>>>>> I have tried to do it now. I got the same error again (Error importing 
>>>>> wisdom from file...) I attach the file by which I create the wisdom file 
>>>>> (using fftw-wisdom), in case there is any irregularity in this one. I also 
>>>>> attach the CP2K output in case in the technical details printed by the 
>>>>> program, at the beginning of the execution, you can detect any error in the 
>>>>> installation.
>>>>>
>>>>> I don't know why it works for you and not for me. 
>>>>>
>>>>> Best, 
>>>>>
>>>>> Santiago Movilla
>>>>>
>>>>> On Thursday, June 20, 2024 at 5:49:29 PM UTC+2 Frederick Stein wrote:
>>>>>
>>>>>> Dear Santiago,
>>>>>> If you have compiled CP2K yourself, can you recompile with the 
>>>>>> current master trunk and run CP2K with a wisdom file optimized from the 
>>>>>> script fftw-wisdom (delivered with FFTW3). With my bugfix, I was not able 
>>>>>> to run CP2K with a wisdom file optimized from a CP2K run but with an 
>>>>>> FFTW-optimized one (I have no idea why).
>>>>>> Best,
>>>>>> Frederick
>>>>>>
>>>>>> Santiago Movilla schrieb am Donnerstag, 20. Juni 2024 um 17:43:39 
>>>>>> UTC+2:
>>>>>>
>>>>>>> Dear Frederick, 
>>>>>>> I would like to express my gratitude for all the assistance provided.
>>>>>>> I have conducted two trials with FFTW3 and FFTSG, and I have 
>>>>>>> attached the timings report. It appears that FFTSG results in superior 
>>>>>>> performance. 
>>>>>>> I have not yet tested with the wisdom file, as I have not yet 
>>>>>>> generated it for the estimate plan type.
>>>>>>>
>>>>>>> Best, 
>>>>>>>
>>>>>>> Santiago Movilla
>>>>>>>
>>>>>>> On Thursday, June 20, 2024 at 4:40:51 PM UTC+2 Frederick Stein wrote:
>>>>>>>
>>>>>>>> I will push a bugfix. Now, I can get it running, if a wisdom file 
>>>>>>>> created from fftw-wisdom is used, but not if it was created during a CP2K 
>>>>>>>> run (better than nothing).
>>>>>>>> Can you provide the timing report (last part of the output file) to 
>>>>>>>> check the performance of your setup?
>>>>>>>>
>>>>>>>> Santiago Movilla schrieb am Donnerstag, 20. Juni 2024 um 16:32:57 
>>>>>>>> UTC+2:
>>>>>>>>
>>>>>>>>> Sure!, I did. 
>>>>>>>>>
>>>>>>>>> Without this line the calculation works perfectly. It finishes 
>>>>>>>>> without error. The problem is when I want to read or write the wisdom file.
>>>>>>>>>
>>>>>>>>>
>>>>>>>>> On Thursday, June 20, 2024 at 4:24:39 PM UTC+2 Frederick Stein 
>>>>>>>>> wrote:
>>>>>>>>>
>>>>>>>>>> Can you try it without the line with FFTW_WISDOM_FILE_NAME?
>>>>>>>>>>
>>>>>>>>>> Santiago Movilla schrieb am Donnerstag, 20. Juni 2024 um 16:22:54 
>>>>>>>>>> UTC+2:
>>>>>>>>>>
>>>>>>>>>>> Dear Frederick, 
>>>>>>>>>>> Thank you very much for the answers. Even using the default 
>>>>>>>>>>> (ESTIMATE) I still get the same errors. I have decided to attach the input 
>>>>>>>>>>> of my calculation in case you can find any other error that may be related 
>>>>>>>>>>> or, taking advantage of your expertise, if you can suggest any other kind 
>>>>>>>>>>> of change that may speed up the calculation time. 
>>>>>>>>>>> Best,
>>>>>>>>>>> Santiago Movilla
>>>>>>>>>>>
>>>>>>>>>>>

-- 
You received this message because you are subscribed to the Google Groups "cp2k" group.
To unsubscribe from this group and stop receiving emails from it, send an email to cp2k+unsubscribe at googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/cp2k/faf2e7b4-df9b-41c3-9e6b-980adac99a3dn%40googlegroups.com.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.cp2k.org/archives/cp2k-user/attachments/20240705/eecf5a8c/attachment-0001.htm>