@warrenm Have you worked with compute much in Metal? I’m trying to apply a kernel on a texture after rendering to it and it is dog slow.
-
@warrenm Is there a large cost to switch from render to compute or should you in theory be able to mix them in your pipeline?
@jfroy In theory it should remain high-performance when you interleave rendering and compute.
-
-
@warrenm Stay tuned, I may be allocation a depth texture every frame ;) Side note: when the GPU debugger crashes, system trace can help.Thanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
@warrenm Side note 2: You can blame me when the GPU debugger fails, I built the core infra for it (well the original GL one, not Metal). -
@jfroy I owe you a beer or three. -
@warrenm And sure :) I don’t have a WWDC ticket, but I live in SF so I’d be happy to meet. Or anytime. We can talk Metal, Vulkan, DX12 ;) -
@jfroy I missed out on WWDC this year. Would love to meet up at some point next month, once we know more about the direction of gfx on iOS.
End of conversation
New conversation -
-
-
@warrenm Re: Metel compute. There is either a large mode switch cost or kernel fns run way slower than fragment fns. Like, 2x slower.Thanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
@warrenm I’m wondering if kernel fns use more accuracy by default or are restricted to certain USCs or input ports on each. -
@jfroy Haven’t heard of such a distinction. Regarding precision, you could try the -f(no-)fast-math flags in the Metal compiler options. -
@warrenm I did, no love. If you try their image processing demo, which is really simple, it runs at 30 fps ish and pegs the device at 100%. -
@jfroy I’m curious which device. Also, note that you may be able to increase throughput with a higher threadsPerThreadgroup count. -
@warrenm A7 (iPad mini 2).
End of conversation
New conversation -
Loading seems to be taking a while.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.