When I reproduce the benchmark results on the following url, I got well performance on Arm-cpu, but the results on Mali-gpu is terrible.
https://github.com/dmlc/tvm/wiki/Benchmark#mobile-gpu
Example results on Mali-T860:
Which script are you using for benchmarking? Also, are you using the GPU for other tasks, such as running the desktop environment?
I run the following code:
~/tvm/apps/benchmark$ python3 mobile_gpu_imagenet_bench.py --model rk3399 --rpc-key rk3399
I close the other tasks and stop the destop of rk3399 by
sudo /etc/init.d/lightdm stop
sudo -i
echo performance > /sys/class/misc/mali0/device/devfreq/ff9a0000.gpu/governor
and rerun the above codes, no major changes in the results.
Thanks for the detailed information; I will check on my RK3399 boards.
I use another mali gpu.
when I am tuning the model, some layers show no Latency and Speed info.
The tuing result is verry terrible.
The tvm_runtime was built by cross compiling not on the device platform.
Could you give me some ideas about this problem?