|
View your shopping cart.
Recent topics
Recent Comments |
10.6.3 is out so lets see how many woes it solves As per previous threads I thought 10.6.2 was a bit of a disaster as far as QC was concerned. It seems that 10.6.3 is now public. Im expecting it to be good news in several areas but I dont know what new problems it may introduce or what may still be broken. My brief tests so far suggest that at least some of the problems with iterator performance has been fixed, perhaps completely, and openCL seems a bit more stable, as well as at least some of the mesh filter bugs having been squashed. Perhaps a few specific mesh filters still arent working but it could be that the openCL code in some of the filter compositions files is at fault and needs updating rather than the underlying openCL or QC openCL implementation still being faulty, I dont know, need more time.
More like this
|
Well it looks like the iterator performance issues which have been around since snow leopard first came out are still present. Im sure Im getting better performance with some compositions though, I just havent worked out what exactly is performing better.
Were there any additional iterator performance issues that started with 10.6.2 that may have been fixed? I got so upset with other stuff being borked in 10.6.2 that Ive let me knowledge lapse somewhat by having a lengthy QC absence, maybe its messed up my expectations.
First finding - framework version is now 156.13 next finding - nil improvement on Lighting Environment with Shadows enabled.
Second Finding - Mesh Iterator macro works faster.
Third - largish DAE load well.
Fourth - Mesh Noise and Mesh Bend are this updates bad boys for whatever reason.
Shall look into that more fully.
OK from looking back at previous threads it looks like there were some iterator performance issues introduced with snow leopard, and others that only started with 10.6.2. Can anybody who had a 10.6.2-specific slowdown say whether its better with 10.6.3?
My virtual macro Mesh Iterator by Structure renders faster than before on 10.6.2 - now alternating between 32 to 56 fps
Steve,
If you have a few minutes, can you whip up some sample compositions that demonstrate the iterator performance issues you're experiencing, and file a radar?
Not that I doubt the problem's presence (I reported it too several months ago), but I want some "in the wild" reference material to profile against to help find out what's taking so long.
Thanks! :)
To upgrade, or not to upgrade... that is the question.
a|x
Just upgraded and I see good things happening. At least in the compositions I used to test and play with. The best example is the famous carousel by Memo Akten.. it works fine again as it was with older QC frameworks but only if you close and reopen the viewer. here it is: http://bit.ly/cRm2b1
I posted some console messages to the dev-list.
juan
I would do but to be honest I dont use iterators much, the only reason I started going on about it is that I know it was quite a big 10.6 and 10.6.2 issue for some people, but as far as testing goes I was just trying to follow along by using the example compositions people had posted in the past to the dev list.
oh man i might upgrade. i don't really want to make a partition to test and see if this will kill QC for me or not, but QC on 10.6.2 just didn't work at all for me period. couldn't even open QC or any other app that relied on anything with built with QC. i'm having a hard time keeping up with xcode releases by staying on 10.6.1. can anyone with a MBP unibody confirm this upgrade working ?
There were some nvidia cards that had absolutely abysmal performance in 10.6.2 (a severe regression) that got fixed in 10.6.3 -- this alone might be worth the upgrade, in my opinion (one of my dev machines "enjoyed" this quirk -- less than fun)
Otherwise, not a lot changed QC-wise (mostly bug fixes, not much for performance). I think CL's working again? (if not, file a bugreport -- it'd be really nice to get CL working in QC since it's a "feature" and all ;)
CL has stopped crashing QC for me (my machine does not support CL on the GPU), but it does not ever do CL CPU side software fallback. I re-posted a bug that was requested closed/updated with pertinent info.
As per some chat on the apple qc dev list, it transpires that this is mostly down to a few of the mesh filters OpenCL Kernel code having bugs, rather than OpenCL CPU fallback failing for Vade. We can fix these ourselves by editing the mesh filter qc patches in question. Ive not got time to post the details now, nor am I the best person to be cleanly solving this!
I havent got to the bottom of why some of the Apple OpenCL QC samples from their website dont work, might be similar issues or changes to the way the DAEs get loaded with some of the comps.
Ive noticed that the snow example doesnt work well on CPU - it runs but the snow settles in the wrong place. Not sure if this is down to bugs or simply the CPU not having all of the features to run the part of the OpenCL Kernel code that looks at the image and does stuff to velocity accordingly.
Well I decided that 10.6.3 is good enough for me so yesterday a GTX 285 card arrived for my 2008 Mac Pro. Not had a chance to do much with it yet but will be spending about 8 hours with it today. If people have anything specific they want me to test the performance of then let me know. All I can say initially is that I can run my old EarFluid stuff with the image being processed at 640x480 and get somewhat acceptable framerates now, time to do some more work on that composition methinks.