Tried on armbian jammy with icecream95’s patchset. Both hevc_rkmpp(8&10b) and h264_rkmpp works fine for me. I don’t connect rock5b to a monitor, so I use ffmpeg transcode to verify this.
RGA performs a pixel format conversion (nv12|na12->yuv420p) and copy back it to memory.
Ahh, I tested it on kernel 5.10.66 and it worked. Kernel 5.10.110 does not. I will try it with linux-5.10-gen-rkr3.4 which has an updated mpp driver. Thanks.
Quick question, is rkmpp hardware decoding supposed to be working out of the box on the debian bullseye radxa distribution ? If not has anyone achieved it with a custom kernel/some custom package ?
Essentially the Rockchip BSP is a custom kernel with some custom packages where Devs here are trying various options and if you read you will find info above.
Also in the discord server devs and community are sharing images where much has been added. The Radxa images are more conservative with bleeding edge additions, but prob best to chat and ask in there.
If You are in USA then ameridroid seeme to be better option but both are ok. I ordered this month from both of them and i was about 3 working days for ameridroid to ship to Nevada and about 5 for allnet to ship to Europe.
This error occurs when the python compiler gets confused between function name and module name and try to run a module name as a function. This error statement TypeError: ‘module’ object is not callable is raised as you are being confused about the Class name and Module name. The problem is in the import line . You are importing a module, not a class. This happend because the module name and class name have the same name .
If you have a class MyClass in a file called MyClass.py , then you should write:
from MyClass import MyClass
In Python , a script is a module, whose name is determined by the filename . So when you start out your file MyClass.py with import MyClass you are creating a loop in the module structure. Also, note that this usually happens when you import any kind of module as a function.
@tonygruz
Thank you for the explanation. It was solved somehow at that time, most likely the same way.
I am about to review some of my tests with SDL3 and the latest rockchip rt but would like to try with yolov8 if possible.
I have read somewhere (can’t find the info anymore) people have (or had) some trouble converting it to rknn.
Are you willing to try converting it and sharing the rknn models, if so, please share your thoughts?
See here:
Seems Onnx is the most compatible so either get a Onnx Yolo8 or convert to Onnx then convert 2 Rknn2.
Rknn2 looks relatively easy and quite good but its a bit of a steep curve and the documentation isn’t great.
I had a look as was really intrested to see if https://github.com/usefulsensors/openai-whisper as pytorch whisper is converted to onnx before finally becoming tflite, but boy you have to really have some will power
I was thinking whisper would be a mighty model to have on the rk3588 NPU as on the cpu the test JFK.wav inference is taking just over a second with the tiny model.
The Armv8.2 SDOT and UDOT instructions provide access to many multiply and accumulate operations every cycle so greatly speed up ML so it would be interesting to compare a more complex model like whisper against the npu.
This may be a way too noob question. I have an M2 to USB3 adapter. https://amzn.to/3TRkLhN I can’t seem to get it working. Is there some config setting I need to make it work?
My understanding is that this card is not a USB3 adapter but a PCIe adapter that uses USB3 connectors (which are electrically compatible). It’s something common among the cryptomining community to extend the bus to reach multiple GPUs. If you run lspci I suspect you’ll see 4 new PCI busses. I agree that the article’s description is deceptive given that it’s written USB3 at a few places, but in fact it’s always associated with PCIe 1x which clearly is the bus in question. Now the good thing is that if you buy 4 of the other boards that connect to these ports, you’ll have 4 PCIe 1 slots for your board, on which you’ll probably be able to connect boards providing up to 4 USB3 ports each (the peak bandwidth will remain limited to PCIe 3.0 1x =8 Gbps though).
