Use Mat3x4 for model and view transforms to save bandwidth and ALUs #107923
Conversation
|
Cant make the artifact work for my s24 ultra , after the wheel spin is over, it gets stuck( downloaded the editor apk), 4.5 beta works fine tho. |
I haven't done the mobile renderer implementation yet! |
ik but even selecting the mobile renderer gives the same result as with forward+ |
Oh alright, I'll make sure to test on Android before marking this as ready for review. |
…perations and ALUs
|
Tested now on MacOS, PopOS, and Android. On my M2 macbook, I can't measure a difference. On my Intel integrated GPU I get a consistent 5%ish improvement in the test scene from #68959. I get similar results on my Pixel 7 (Mali-G710) and pixel 4 (Adreno 640) On other scenes more broadly I suspect the average performance benefit will be lower than 5% since the bottlenecks are often fragment processing and this does little to help with that. But at any rate, this is a free improvement to performance (in some cases) and battery life in all cases |
|
Tagging this for 4.6 dev 1. It should be a pretty safe change, but it is a little risky for the Beta cycle |
This improves performance in situations that are vertex shader bound (i.e. high vertex count). Early tests show that this makes an improvement on my intel integrated GPU quite broadly, but not on my M2 MBP.
I want to test a bit more widely to get a sense of the broad impact.
Checking with the Mali Offline Compiler, this change appears to shave off a few L/S operations and ALUs (about 10%). So I don't expect it to make a huge difference (especially on desktop). But its a free performance boost.
Built on top of #107876 to avoid conflicts