Hey guys!
What is the most efficient way that I can go about playing video in starling? I need to ideally layer behind my starling objects, so I can't use StageVideo (gah!)
Presently, I'm sampling bitmapdata from my video object and writing a Texture to an Image that's on the stage, however this doesn't really feel like the best way of doing things...
I understand that genome2d has a video player... I would love to hear how sHTiF is doing it!
Pretty much as you said with couple of options for downsampling etc. The things you need to do are:
- work with a single texture allocation and just reupload the new data, DON'T ALLOCATE NEW TEXTURE each update
- use blendmode none
- don't use alpha if you don't need to, and I really mean NEED on the video quad as it will be significatnly faster on low level devices as mul fix can kick in
- downsampling is handy here, especially custom, for example if you try to play full hd 1920x1080 you will need to allocate 2048x2048 texture with downsampling to 1820x1024 it will allocate only 2048x1024 texture which will be significantly faster due to fillrate and not that far from the real resolution
Thanks for those hints, sHTiF. I wrote a karaoke MP3+G renderer in AS3 a while back (basically, a very simple video format), and I've been trying to figure out how to make it run well in Starling. It needs to manipulate BitmapData every frame, and I wasn't sure how to get it onto the GPU efficiently.
The part about reuploading the BitmapData without creating a new texture, in particular, is probably what helped me the most. I didn't know that was possible (although I suspected as much since the same code runs well in renderMode gpu without Stage 3D), and that got me searching for a way to do it with Starling.
For anyone else who might be interested in a similar technique for uploading BitmapData often and efficiently, here's the call that worked for me:
flash.display3D.textures.Texture(starlingTexture.base).uploadFromBitmapData(bitmapData)
I'm getting 60fps on an iPhone 3GS. Although, it's worth noting that my "video" format is a tiny 300x216 (but stretched to fit the full width and height of the device).
Most excellent. Thanks shtif/Josh!
(testing on desktop in adl)
I seem to be getting really crappy performance doing:
flash.display3D.textures.Texture(_canvas.texture.base).uploadFromBitmapData(_cache[_counter]);
I'm looping through my video first and caching each frames' bitmapdata into a vector, then looping through that.
blendmode is set to none, transparency zero, hmmm :-/
Ok, so I changed my texture to a rendertexture and performance has gone from 10fps to 44fps on desktop. 
Ok and downsampling to half res has obviously kicked it right up to 60. Now to try it on my iPhone 
@tomph Do you mean that you simply set optimizeForRenderTexture to true when you initialially created the texture, and then continued using uploadFromBitmapData() like I suggested? Or are you using the RenderTexture class in some way?
I used RenderTexture instead of Texture (and then used uploadFromBitmapData).
It works in the exact same way, but gave me a huge performance increase.
Are the performance on device good?
I used RenderTexture to draw a dynamic shape every enter frame loop: on desktop I get 51/60 fps, on my iPhone I get 5/60 fps :(.
I'm not seeing any difference when I switch to using the RenderTexture class in my project. A tried a number of combinations, and I always had the same frame rate. However, my relatively small BitmapData may not be enough to overpower the GPU the way a larger video frame might.
Edit: I see that you're testing on desktop too. It must be very large video to overpower that.
Edit 2: Out of curiousity, I tried on my laptop with a 2048x2048 texture of generated noise. I saw no fps difference. Both a RenderTexture and a regular one gave me the same fps. I wonder what's different about our environments (besides, perhaps, the fact that my laptop is almost brand new).
Hmmm, interesting! I'll run a few more tests today, as it could (possibly) be something else that I did, that improved performance.
Just a side note,
Guys, remember that uploading textures to the GPU is a slow, expensive operation on the hardware pipeline and should ideally be minimized and relegated to initial stages. (I think it was Tom Krcha or Mike Jones from Adobe who told me that). Doing it on every frame may be choking your pipeline. That's why it's optimal for Starling to just manipulate textures that are already on the GPU.
In this case, perhaps sacrificing resolution for FPS speed is best...
I could be wrong on the following but:
2048 x 2048 x 32bits = 134217728 bits = 67108864 bytes = 65536 kb = 64mb for that size texture...
For a mobile device with limited system and GPU memory... ouch?
Hmm... can you try a test by playing the video the old way on the flash.display DisplayList stage that floats above Stage3D and let us know framerates?
You're not supposed to mix and match DisplayList and Stage3D because of the performance hit, but in this case, maybe you'll get better performance than redrawing the video on a texture for every frame? Then after you're done with the video, just removeChild the sprite overlay video container from the DisplayList.
You're not supposed to mix and match DisplayList and Stage3D because of the performance hit
This isn't necessarily true anymore. The display list is now synchronized to Stage 3D draw calls. Thibault Imbert confirmed it a while back.
object404
2048 x 2048 x 32 bits = 134217728 bits = 16777216 bytes = 16384 kB = 16MB
You must log in to post.
