[MacOS][M1]Metal much slower than OpenCL

Troubleshooting
#1

I tried to measure my model using metal which is recommended for macos, however the result is very bad , as shown below

image

This troubles me a lot, because the result on cpu is faster than it which is about 50 ms.

With no idea, I tried changing metal to opencl , the result is much normal

image

the only difference in my code is just the target setting

image
image724×112 25.8 KB

In my opinion, either metal needs extra settings or metal on m1 has bugs now.

Anyone know what causes this problem?

#2

I am trying with the M1 Pro chip, and I also found this issue with the metal, which @vincentily has mentioned.

Is there anyone who can figure out this problem?

#3

Have you tried AutoScheduler with metal target? The default scheduling is not good for this target, need to tune

#4

ok , I will have a try, thanks

#5

Hi，friend! Have you solved this problem? Is Metal faster than CPU after tuning?