Re: Poor performance on software rendering generator
Re: Poor performance on software rendering generator
- Subject: Re: Poor performance on software rendering generator
- From: Pierre Jasmin <email@hidden>
- Date: Sat, 30 Jul 2011 18:23:02 -0700
- Organization: RE:Vision Effects
On 7/30/2011 3:15 AM, Martin Baker wrote:
We're seeing that software rendering for a generator is extremely slow. All these are on a Motion 5 project at 1920x1080, 29.97fps, progressive at full res.
step 1: checking timing on our plug-in
Drawing takes 10-11ms
RGBA->ARGB conversion takes 20-28ms
Motion playback rate is 10-11fps
step 2: removing RGBA->ARGB code
Motion playback rate is still 10-11fps
step 3: removing ALL code from renderOutput method
Motion playback rate is 12fps
step 4: changed from software to hardware rendering, still with empty renderOutput method
Motion playback rate is 30fps
So even before we add any custom code we're hampered by using software rendering, and this is just a simple generator with an output only not a filter with input/output.
Looks like you are measuring basically the speed of moving RAM to
GPU after you and that ends up in the order of 83 ms (1/12) versus
33 ms (1/30) on your system (a delta of 50 ms per frame - and it's
one way as it's an output only/generator), even if you render to a
software file format like .png and you would expect nothing to ever
need to go through graphics memory...
Let me help you file your bug report while you are there...
How does Motion handles canDoHardware, canDoSoftware
// graphics hardware assisted rendering versus cpu only rendering
For example if one had a motion template with
effect 1 - canDoHardware=NO, canDoSoftware=YES
(your case 3)
effect 2 - canDoHardware=YES, canDoSoftware=YES
<does motion go GPU here and move data to GPU and back out?>
effect 3 - canDoHardware=NO, canDoSoftware=YES
With empty renderOutput methods, are you in a scenario like that
then regressing to something like 6 FPS playback with empty methods?
(less/more)
I would be curious of your render times: 1 versus 1,3 versus 1,2,3
(with many cores Sandy Bridge and AVX2 and what not etc it's
actually becoming hard to make a software effect that takes longer
then GPU->CPU->GPU memory transfers and I am pretty sure the
internal processing engine is probably pretty agnostic performance
wise, particularly for empty methods :) )...
Pierre
Thoughts?
Martin Baker
www.digital-heaven.co.uk
www.twitter.com/digitalheaven
www.videospaceonline.com
www.finalcutters.com
------------------------------------------------------
Digital Heaven Ltd is registered in England and Wales
Company registration number: 3628396
Registered address: 55 Lynwood Drive, Worcester Park, Surrey KT4 7AE
VAT registration number: GB 736 0989 03
_______________________________________________
Do not post admin requests to the list. They will be ignored.
Pro-apps-dev mailing list (email@hidden)
Help/Unsubscribe/Update your Subscription:
This email sent to email@hidden
|
_______________________________________________
Do not post admin requests to the list. They will be ignored.
Pro-apps-dev mailing list (email@hidden)
Help/Unsubscribe/Update your Subscription:
This email sent to email@hidden