Tim, I think you miss the point. Doing parallel computing in your garage is possible, but it won't help if your I/O is slow, which is the case if you plug desktops PC in a farm. You just have to open darktable -d perf -d opencl to understand that most of the time spent during an image export is lost in GPU/CPU data transfers, when you copy the RAM in the CPU/GPU registers back and forth. What you need to do is parallelize on "close" cores to reduce the memory copies and work in registers as much as possible.
The context of my question is I'm currently working to implement blind deconvolution for DT, and this process is not a filter but an equation solver that minimizes both the blur and the noise to find the minimal-energy image. This is something your CUDA GPU is not ready for, and needs several hours for large blurs on large images. Similarly, I'm investigating the possiblity to refocus images. We are talking about FFT, convolutions products, gradients computations, and so on with iterative processes, and I'm still not sure wether these functionnalities fit into a general-use software like DT. And if they do, you may not like the computing time. *Aurélien PIERRE* aurelienpierre.com <http://aurelienpierre.com> ------------------------------------------------------------------------ Le 2017-11-02 à 20:05, steve a écrit : > > Worth a look at this first: https://www.youtube.com/watch?v=xuuiUhMr-lQ > > > On 11/02/2017 08:01 PM, n61...@gmail.com wrote: >> Instead think of how many people have multiple computers at home. I >> have a desktop, a laptop, my wife has laptop, kids.... >> Most homes have more than one computer. For some of the more complex >> stuff it would be interesting to potentially off load to a local >> spare machine. >> I know audio systems do this for virtual instruments (brother is a >> hobby audiophiles) >> >> >> Tim >> >> >> ------ Original message------ >> *From: *Steven Adler >> *Date: *Thu, Nov 2, 2017 7:08 PM >> *To: *Aurélien PIERRE; >> *Cc: *darktable; >> *Subject:*Re: [darktable-dev] Darktable + Cloud-computing >> >> My PC is more than capable of processing photos and from what I've >> seen of v2.3, the new GPU support in Darktable makes processing much >> much faster. Cloud based GPU support would be great for advanced AI >> algorithms for things like auto background replacement. But still >> GPUs are advancing rapidly and I doubt we will need cloud GPU >> augmentation for some years. >> >> On Nov 2, 2017 6:17 PM, "Aurélien PIERRE" <rese...@aurelienpierre.com >> <mailto:rese...@aurelienpierre.com>> wrote: >> >> Hi, >> >> as top-notch image processing algorithms become more and more >> demanding in computing power, but often highly parallelizable, >> and pictures resolutions double almost every 5 years (now 52 Mpx >> for the Canon 5DS R, 45 Mpx for the Nikon D850), most computers >> become hardly enough to just open the pictures. Let alone apply >> complex filters on them… >> >> Serious amateurs and pro may want to buy expensive workstations >> but… Cloud-computing solutions like Amazon EC3 >> <https://aws.amazon.com/ec2/instance-types/?nc1=h_ls> gives you >> remote access to Linux instances with Nvidia GPUs for 0.76 US >> $/hour (g2.2 instances >> <https://aws.amazon.com/ec2/pricing/on-demand/>, 8 vCPU). The >> instances are scalable in size automatically and several Linux >> distros are provided (Ubuntu, Debian, Red Hat, CentOS). At this >> price, you get the price of your killer PC (5000 $) in more than >> 6500 hours of use. That's 5 years of working-time (assuming 48 >> weeks/year, 35 h/week, because I'm French). >> >> So… what do you think of having the heavy filters processed in >> Darktable through the servers of Amazon or anybody else instead >> of having to break the bank for a new (almost) disposable >> computer ? Possible or science-fiction ? How many of you don't >> have a 1MB/s or faster internet connection ? How difficult would >> it be to code ? >> >> Full disclaimer : I have no previous experience in cloud >> computing and no Amazon shares. >> >> -- >> >> *Aurélien PIERRE* >> aurelienpierre.com <http://aurelienpierre.com> >> >> ------------------------------------------------------------------------ >> >> >> ___________________________________________________________________________ >> darktable developer mailing list to unsubscribe send a mail to >> darktable-dev+unsubscr...@lists.darktable.org >> <mailto:darktable-dev%2bunsubscr...@lists.darktable.org> >> >> >> ___________________________________________________________________________ >> darktable developer mailing list to unsubscribe send a mail >> todarktable-dev+unsubscr...@lists.darktable.org >> <mailto:%20darktable-dev+unsubscr...@lists.darktable.org> > > > ___________________________________________________________________________ > darktable developer mailing list to unsubscribe send a mail to > darktable-dev+unsubscr...@lists.darktable.org ___________________________________________________________________________ darktable developer mailing list to unsubscribe send a mail to darktable-dev+unsubscr...@lists.darktable.org
