CPU Blocks on Metal

[teddyMeng]

I found that the performance on Metal is quite worse than on GLES with the same code.
I have profiled on XCode，found that the “nextDrawable” (renderer_mtl.mm) call blocks the CPU for about 10ms per frame(just on my game). The API is called by renderThread's submit，while the renderSem has not been posted，so it also blocks the main thread. the CPU has been blocked almost half time per frame.
Is there any way to avoid this CPU block? such as double command queue , or any other good idea?

[bkaradzic]

Is this with vsync or without vsync? Which example you're testing on?

[teddyMeng]

Thanks for your reply. I am testing on IOS. It seems that metal is always vsync on IOS(which can be set on macOS). When I opened vsync on GLES version, the things went all the same.I didn't run your samples,my code is a bit large to share but i think it's actually bgfx's bug. Is there any other info u want? such as profile results?

[bkaradzic]

If you think it's bgfx bug, then you should be able to reproduce it in examples.

[teddyMeng]

I have solved this problem by using tripple buffer(add a new m_render on bgfx.cpp,and modify some sync code,but it may produce more latency to GPU).Then I found that bgfx's metal code also have GPU block problem(I add some vertice on your example 17, to let it drop to 20 fps,then use xcode metal performance tool).
You can see that on bgfx gles samples and Apple metal samples(https://developer.apple.com/metal/sample-code/, need modify some code to drop fps).VS can work together with FS at the same time,and vs can start work just behind the encoding ends. But on bgfx metal samples, the VS can not work together with fs,and the vs must wait for "nextDrawable" and flip(). so the frame time is VStime+FStime.But on the Apple samples.the frame time is max(VStime,FStime).
I have tested lots of samples by using xcode performance tools on iPhone7.I am actually a newbee on metal,is there any good choice to improve the performance? such as thriple buffer on CPU/multi encoder/use currentDrawable to replace nextDrawable?

above is the xcode performance on my game ,the first once is the metal CPU block .the second is the GPU time compare(left is metal,right is gles).

[frolaqi]

I have the same problem, [metalLayer nextDrawable] always block main thread over 5 second. Because nextDrawable use semaphore to keep render sequence.