So I can recreate this issue within Quartz Composer directly. Load a large movie, say a 1080p30 Apple Intermediate Codec movie.
drag it in, connect to billboard, take a look at CPU usage and FPS. make an invert patch - the one straight from the CIKernel because why we have two invert patches named the same and one is a patch the other is a filter is beyond me, but just pick the CI Filter one. then connect any 3rd party plugin patch that uses a QCPluginOutputImageProvider.Ive used one of my v002 Blurs for good measure since I know its highly performant (or used to be :P) Notice a huge spike in CPU. Profile in instruments and see the same CPU side pixel format conversion happening. Chris, I think this is a bug, no? https://dl.dropboxusercontent.com/u/42612525/QC%20Perf%20wat.mov I can see that ARGB ABGR convert function within QC. Using a large movie helps because its a large read back / larger CPU slice per frame. Shall I file a bug? > On Feb 1, 2016, at 2:09 PM, George Toledo <gtole...@gmail.com> wrote: > > Core Image did get some sort of overhaul for El Cap. Syntax was updated for > sure. > > I also had several Core Image filters that use Edit Filter Function, or > Advanced Sampler Options break, maybe due to the new syntax, maybe to bugs. > > I have noticed insanely bad performance from some custom filters that still > technically work. The observations about the kinds of chains you created and > the performance, sounds right in line with what I have seen here. > > Some of the built in QC "Image Units" also now perform really badly because > of their use (and maybe mix?) of custom and built in CIKernels. > > Thanks for sharing the results of that troubleshooting. > > > > > On Mon, Feb 1, 2016 at 1:46 PM, Anton Marini <dokt...@mac.com > <mailto:dokt...@mac.com>> wrote: > Hi Chris > > Thanks so much for the prompt reply . > > Ive noticed some interesting things in terms of chaining multiple ‘QCImage’ > backing causing issues. > > My chaining terminology refers to : > > rendering a QCRenderer or providing an appropriate image type (CIImage > CVOpenGLTextureRef, etc), rendering said renderer at time, getting the > valueForOutputKey:key ofType:@“QCImage” and setting that to the > valueForInputKey: of the next QCRenderer in the chain. > > all renderers use the same GL context, save for one, which uses a shared GL > context / and an attached drawable to view the results. > > -- > > if I chain a series of QCCompositions/Renderers which use pure CIImage > filters, as provided by the OS, and no other processing I have no problem. > > if I chain a series of QCCompositions/Renderers which use plugins which vend > textures as QCOutputImageProviders in series, I have no problem. > > If I chain a mix of the above, I have a problem. > > if I load a *single* QC Composition/Renderer which has a custom plugin, and > within that composition that plugin is fed, or feeds, a custom CIKernel, I > have the problem. > > This is on > > MacBook Pro (Retina, 15-inch, Mid 2014) > > NVIDIA GeForce GT 750M 2048 MB > > 10.11.2 (15C50) > > About to use GFXCardStatus to see if the pixel format swizzling as you > pointed out is a GPU or something else issue. > > Thanks again for the quick reply. Right now my reproduction case is > admittedly complex - I’m going to see if I cant boil it down some more. > > > > > > On Feb 1, 2016, at 1:29 PM, Chris Wright <christopher_wri...@apple.com > > <mailto:christopher_wri...@apple.com>> wrote: > > > > Looks like that _might_ be due to a CoreImage change (note a few frames > > down, it’s doing a ABGR8 -> ARGB8 conversion, which possibly implies a > > request for a pixel format that isn’t what the GPU’s currently using). I > > only say CI because QC’s use of pixel formats hasn’t meaningfully changed > > in quite a while, so this would be a surprising behavior change for QC. > > > > Do you happen to have configuration info, or if this behaves similarly on > > other families of GPUs? It looks like you’re on some family of GeForce > > (nvidia), so Intel or AMD behavior might be helpful if it’s a quirk below > > QC. > > > > Can you please file a radar for this? It may well affect several other > > things negatively. > > > > > >> On Feb 1, 2016, at 9:38 AM, Anton Marini <dokt...@mac.com > >> <mailto:dokt...@mac.com>> wrote: > >> > >> Hi > >> > >> My application uses the QCRenderer API to input and output images in the > >> opaque QCImage format and send them through a chain of published input and > >> output images in QCCompositions/Renderers I control. > >> > >> Ive noticed in 10.11, performance seems worse. Looking into it, I’m > >> noticing that *certain* configurations of Core Image filters and custom QC > >> Plugins / output images vended as QCImages cause read back to occur on the > >> CPU. > >> > >> Ive tried to account for various different causes - like pixel format > >> issues, color space issues, etc, but I cannot get rid of the CPU read > >> back, which I am 99.999% positive didn’t happen in older Quartz Composer > >> framework/API versions as older hardware got better frame rates - likely > >> keeping things on the GPU where they belong™. > >> > >> Has anyone else seen regressions in performance with QC? > >> > >> Specifically, I am seeing stacks like the following (see attached image). > >> > >> https://www.dropbox.com/s/s7sqn56gdcogxsj/PastedGraphic-1.png?dl=0 > >> <https://www.dropbox.com/s/s7sqn56gdcogxsj/PastedGraphic-1.png?dl=0> > >> > >> Any hints as to why there might be rendering path changes in QC or Core > >> Image? > >> > >> Thank you. > > > > -- > > Chris Wright > > christopher_wri...@apple.com <mailto:christopher_wri...@apple.com> > > > > > > > > > _______________________________________________ > Do not post admin requests to the list. They will be ignored. > Quartzcomposer-dev mailing list (Quartzcomposer-dev@lists.apple.com > <mailto:Quartzcomposer-dev@lists.apple.com>) > Help/Unsubscribe/Update your Subscription: > https://lists.apple.com/mailman/options/quartzcomposer-dev/gtoledo3%40gmail.com > > <https://lists.apple.com/mailman/options/quartzcomposer-dev/gtoledo3%40gmail.com> > > This email sent to gtole...@gmail.com <mailto:gtole...@gmail.com> > _______________________________________________ > Do not post admin requests to the list. They will be ignored. > Quartzcomposer-dev mailing list (Quartzcomposer-dev@lists.apple.com > <mailto:Quartzcomposer-dev@lists.apple.com>) > Help/Unsubscribe/Update your Subscription: > https://lists.apple.com/mailman/options/quartzcomposer-dev/lists%40vade.info > <https://lists.apple.com/mailman/options/quartzcomposer-dev/lists%40vade.info> > > This email sent to li...@vade.info <mailto:li...@vade.info>
_______________________________________________ Do not post admin requests to the list. They will be ignored. Quartzcomposer-dev mailing list (Quartzcomposer-dev@lists.apple.com) Help/Unsubscribe/Update your Subscription: https://lists.apple.com/mailman/options/quartzcomposer-dev/archive%40mail-archive.com This email sent to arch...@mail-archive.com
