19:12 <d1b2> <davoid> quick one, remember I saw a speed test where u achieved close to 30MB over bulk (cdc?) via libusb+python, correct or dream?

19:17 <d1b2> <Qyriad> Idk about CDC specifically, but we've managed to get around 42 MiB/s in Python with pretty tuned code

19:17 <d1b2> <Qyriad> This will depend a lot on your host controller too though

19:18 <d1b2> <davoid> tuned linux prios you mean? pr even rtpatched?

19:19 <d1b2> <Qyriad> Tuned Python code handing the transfers

19:19 <d1b2> <davoid> stock libusb?

19:19 <d1b2> <Qyriad> Yep

19:19 <d1b2> <davoid> on the transport/LL side, did you have to do something else than just one cdc endpoint to achieve that speed?

19:20 <d1b2> <davoid> i've seen problems with libusb being able to pack all microframes in an optimal way

19:20 <d1b2> <Qyriad> Before we optimized it we got it to ~30 MiB/s on some host controllers, but other host controllers it was less than half that

19:21 <d1b2> <davoid> to follow the opt history, where should I look?

19:21 <d1b2> <Qyriad> Not that I remember, but it wasn't CDC: vendor-specific bulk

19:22 <d1b2> <Qyriad> Look at the discussions and branches listed in this issue: https://github.com/greatscottgadgets/greatfet/issues/286

19:24 <d1b2> <Qyriad> Iirc the tl;dr was use python-libusb1 with async transfers instead of PyUSB, and benchmark to make sure any Python data-parsing code isn't getting in the way of the USB transfers

19:29 <d1b2> <davoid> thx

19:30 <d1b2> <davoid> because cdc was slower? or just to get a separate descriptor as you own both sides?

19:31 <d1b2> <Qyriad> idk, we weren't there for when it was implemented, but I'd assume it was just because it was simpler

19:32 <d1b2> <Qyriad> Didn't have to go and correctly implement a proper USB class; they just made a bulk endpoint that sporked raw data any time it was asked

19:33 <d1b2> <Qyriad> I don't know enough about CDC to know what overhead it has, so I don't know how helpful these results will be for you, but if you're having performance issues then switching to python-libusb1 and async transfers is a good bet

19:41 <d1b2> <davoid> will try and mimic that implementation

19:42 <d1b2> <davoid> i was mainly interested if several bulk endpoint was used to trick the OS of filling the outgoing phy buffers in a more efficient way

20:20 <d1b2> <Holonium> Any chance that a Glasgow ordered June 2021 will arrive before August this year under current conditions?