[poppler] Migrating poppler to liblcms2 from liblcms1

Sun Jan 29 09:27:43 PST 2012

Great, I would suggest just one test more. Try to convert a PDF with an huge raster image > 400Mb. The xput gain of lcms2 should be evident in this case.
Regards.
Marti

-----Original Message-----
From: Albert Astals Cid [mailto:tsdgeos at gmail.com] On Behalf Of Albert Astals Cid
Sent: sábado, 28 de enero de 2012 16:48
To: poppler at lists.freedesktop.org
Cc: Maria, Marti; seb128 at ubuntu.com; Till Kamppeter; Koji Otani
Subject: Re: [poppler] Migrating poppler to liblcms2 from liblcms1

El Divendres, 27 de gener de 2012, a les 18:23:50, Albert Astals Cid va
escriure:
> El Divendres, 27 de gener de 2012, a les 10:39:18, Maria, Marti va escriure:
> > Bon dia Albert,
> 
> Hola
> 
> > That makes a lot of sense. Lcms2 spends more time on optimizing the 
> > transform in order to make it faster. That means create transform is 
> > currently slower that lcms1 but applying transforms  is way faster. 
> > This was indeed a design decision, as most of time creation of 
> > transforms happens only once at the very beginning and then this 
> > transform is applied to huge amounts of raster. The performance 
> > tests I send didn't take creation time into account.
> > 
> > I will take this action item to review creation time in the next 
> > release.
> > Otherwise, you may consider reusing transforms.
> 
> We reuse the transforms, it is just that this file has lots of them.
> 
> > That would speed up both
> > lcms1 and lcms2.  If this is not feasible, there is a flags 
> > cmsFLAGS_NOOPTIMIZE, that makes transform creation fast, at the 
> > expense of slower dotransfom operation. This is not recommended at 
> > all because you disable all optimizations, but may solve the issue 
> > when creating transforms that are applied only to few bytes.
> 
> Not sure if i can now to how many bytes it'll be applied but i'll do 
> some testing and maybe just enable it for the time being so at least 
> we have the same behaviour than when using lcms1, might not be faster 
> but we not slower either.

Using cmsFLAGS_NOOPTIMIZE brings back altona_technical_1v2_x3.pdf to an acceptable speed again.

Koji can you please update your patch so it uses the cmsFLAGS_NOOPTIMIZE when compiling with lmcs2?

Albert

> 
> Thanks,
>   Albert
> 
> > Again, thanks for reporting. I'm putting the transform creation time 
> > issue high in my priority list.
> > 
> > Best regards
> > 
> > Marti Maria,  Color Engineer
> > HP Large Format Printers Division - Barcelona marti.maria at hp.com
> > 
> > 
> > 
> > -----Original Message-----
> > From: Albert Astals Cid [mailto:tsdgeos at gmail.com] On Behalf Of 
> > Albert Astals Cid Sent: jueves, 26 de enero de 2012 19:43
> > To: poppler at lists.freedesktop.org
> > Cc: Maria, Marti; Till Kamppeter; seb128 at ubuntu.com
> > Subject: Re: [poppler] Migrating poppler to liblcms2 from liblcms1
> > 
> > El Dilluns, 23 de gener de 2012, a les 11:16:03, Maria, Marti va escriure:
> > > Hi,
> > > In my testings lcms2 is significantly faster that lcms1 except 
> > > when using floating point, then lcms2 favors precision.
> > 
> > Is this precision is so important to justify a 10x slowliness? Maybe 
> > it'd be a good idea to have a fallback fast mode for those not 
> > willing to take that 10x hit?
> > 
> > > http://littlecms2.blogspot.com/2010/03/xput-comparative-119-vs-20.
> > > ht
> > > ml
> > > 
> > > Altona test does many different color conversions it would be 
> > > interesting to know which transformation is going slow.
> > 
> > Attached the callgrind output of running poppler there, you can see 
> > the time is almost totally spent on creating the transforms.
> > 
> > Wild guess: I'm wondering if the thing is that not transformations 
> > are all precalculated (meaning a faster overall if you do lots of 
> > access) and previously they were done each time and thus if we end 
> > up doing "few" transformations it was just much faster?
> > 
> > Cheers,
> > 
> >   Albert
> >   
> > > Regards
> > > Marti
> > > 
> > > ________________________________________
> > > From: Till Kamppeter [till.kamppeter at gmail.com]
> > > Sent: Monday, January 23, 2012 8:58 AM
> > > To: Albert Astals Cid
> > > Cc: poppler at lists.freedesktop.org; Maria, Marti; Sebastien Bacher; 
> > > Koji Otani Subject: Re: [poppler] Migrating poppler to liblcms2 
> > > from
> > > 
> > > liblcms1 On 01/22/2012 11:18 PM, Albert Astals Cid wrote:
> > > > Same thing, altona_technical_1v2_x3.pdf still goes up from x 
> > > > seconds to 10x seconds when going from lcms1 to lcms2.
> > > > 
> > > > That is in my opinion not acceptable (of course i'm not ruling 
> > > > out that we in poppler are using lcms2 wrong). Till, does your 
> > > > ubuntu work cover fixing the regressions of moving from lcms1 to 
> > > > lcms2 or all your work covers is making stuff compile?
> > > 
> > > I downloaded altona_technical_1v2_x3.pdf from 
> > > http://www.eci.org/doku.php?id=es:downloads and tried pdftops on 
> > > Ubuntu Oneiric (11.10, lcms1) and on Precise (12.04, lcms2, Koji 
> > > Otani's patch) and with lcms1 it takes 3.5 seconds and with lcms2
> > > 17.5
> > > seconds.
> > > 
> > > The same problem occurs with Ghostscript: There the same 
> > > transition happened from Oneiric to Precise and here Oneiric is 
> > > much faster with this file, too. In Precise PDF rendering for 
> > > printing is done by Ghostscript and not by Poppler, so lcms2 is 
> > > already used for printing in Precise for some time. As there are 
> > > no performance complaints by users yet, it seems that for everyday 
> > > files there is not such a significant performance loss. and this
> > > 
> > > Marti, can you check what is the problem with lcms2 and this PDF 
> > > file and also why most files are somewhat slower with lcms2?
> > > 
> > >     Till
> > > 
> > > _______________________________________________
> > > poppler mailing list
> > > poppler at lists.freedesktop.org
> > > http://lists.freedesktop.org/mailman/listinfo/poppler
> 
> _______________________________________________
> poppler mailing list
> poppler at lists.freedesktop.org
> http://lists.freedesktop.org/mailman/listinfo/poppler