I would like to know the approximate time it takes to implement super resolution in RK3588

[Zhang-Senhua]

Do you have test documentation? I would like to know the approximate time it takes to implement super resolution in RK3588. thank you

[marty1885]

Sorry for the late reply. I was very busy this week. I did some benchmark. My example code took 9.75s to upscale an 1920x1200 image using an uninitialized model (ie, running in fp16 mode, did not set stddev and mean). This includes the image decoding and exporting time. I suspect 1 or 2 seconds are on the CPU. So about 8seconds running the actual model (across the entire image).

[Zhang-Senhua]

Sorry for the late reply. I was very busy this week. I did some benchmark. My example code took 9.75s to upscale an 1920x1200 image using an uninitialized model (ie, running in fp16 mode, did not set stddev and mean). This includes the image decoding and exporting time. I suspect 1 or 2 seconds are on the CPU. So about 8seconds running the actual model (across the entire image).

Thanks for replying, Also, I would like to know if you have tried calling NPU on rk3588 to accelerate the inference process?

[marty1885]

Ah, I had to correct my statements. I re-benchmark and find my previous result wrong.

The following table shows how long it takes to upscale an entire 1920x1200 image. A total of 400 inference calls are made.

RKNN (NPU)	ONNXRuntime (CPU / 4 threads)
15.4s	73.608s

RKNN is about 5x faster then ONNX by default. The infer.py in this repo is running on NPU.

[Zhang-Senhua]

Got it ! Thanks a lot!