Thanks for the responses! I'm not a full-time comper so am unaware of some of the tricks out there, like using ZDefocus. And I should have added that large kernels / images are involved so the standard convolve node falls behind at the resolution we are working with.
Magno, I wasn't aware of your blinkscript. If I have some time I'll peek at the code. Someone should put together a commercial implementation for Nuke some day, sounds like there is demand out there for it. -Jon On Mon, Mar 20, 2017 at 7:18 AM, Magno Borgo <li...@borgo.tv> wrote: > I coded a naive Blinkscript DCT (and inverse) implementation a while back, > you can check the code and use as a start for a better implementation. > > http://www.nukepedia.com/blink/other/dct-discrete-cosine-transform > > Magno. > > > > > > On Mon, 20 Mar 2017 03:39:40 -0400, Mads Lund <madshl...@gmail.com> wrote: > > we did some CPU vs FFT tests aswell, but I can't remember the exact kernel > size where FFT started to be more efficient, but it was rather low. On the > flip side Convolve on the GPU seem to have some memory problems and (at > least for us) cause some random out of memory issues, even on beefy cards. > > We found that using gaming techniques: up-scaling low frequency areas (then > convolving) and only full convolving of high frequency areas to be to be > quite efficient for the vast majority of our renders where the plate is > gigantic or a large kernel size is needed. > The result is perceptually indistinguishable. > > > > > man. 20. mar. 2017 kl. 06.14 skrev Deke Kincaid <dekekinc...@gmail.com>: >> >> The convolve node was fixed up quite a while a go(4-5 years ago, it is >> what the zdefocus is based on). In our tests though we still find the FFT >> nodes faster on the farm vs convolve in CPU mode. If you have a gpu farm >> then convolve is faster. >> >> >> On Sun, Mar 19, 2017 at 6:23 PM Michael Habenicht <m...@tinitron.de> >> wrote: >>> >>> The latest version of the convolve node has gpu support. >>> >>> Am 20.03.2017 um 01:20 schrieb jon parker: >>> > Greetings Nuke users, >>> > >>> > I'm just wondering if there are any faster, more robust FFT tools >>> > available for Nuke besides the (hidden) built-in nodes? >>> > >>> > The built-ins do the job, but they are pretty slow and definitely >>> > prone to crashing fairly often. >>> > >>> > Or, alternatively, something that does fast image convolution, some >>> > other way, under the hood could work too. >>> > >>> > Cheers, >>> > Jon >>> > _______________________________________________ >>> > Nuke-users mailing list >>> > Nuke-users@support.thefoundry.co.uk, http://forums.thefoundry.co.uk/ >>> > http://support.thefoundry.co.uk/cgi-bin/mailman/listinfo/nuke-users >>> > >>> _______________________________________________ >>> Nuke-users mailing list >>> Nuke-users@support.thefoundry.co.uk, http://forums.thefoundry.co.uk/ >>> http://support.thefoundry.co.uk/cgi-bin/mailman/listinfo/nuke-users >> >> _______________________________________________ >> Nuke-users mailing list >> Nuke-users@support.thefoundry.co.uk, http://forums.thefoundry.co.uk/ >> http://support.thefoundry.co.uk/cgi-bin/mailman/listinfo/nuke-users > > -- > Best regards. Mads Hagbarth Lund > > > > > -- > Using Opera's mail client: http://www.opera.com/mail/ > > _______________________________________________ > Nuke-users mailing list > Nuke-users@support.thefoundry.co.uk, http://forums.thefoundry.co.uk/ > http://support.thefoundry.co.uk/cgi-bin/mailman/listinfo/nuke-users _______________________________________________ Nuke-users mailing list Nuke-users@support.thefoundry.co.uk, http://forums.thefoundry.co.uk/ http://support.thefoundry.co.uk/cgi-bin/mailman/listinfo/nuke-users
