**** BEGIN LOGGING AT Wed Mar 11 03:02:27 2020 Mar 11 14:07:53 I didn't quite understand what you wanted to say here jkridner . Mar 11 14:10:57 I took a look at it, it was quite useful, so the data which EEPROM sends shouldn't be more than what goes into manifest. Mar 11 14:11:50 But, we would need a protocol to parse the platform data we receive from the EEPROM via I2C. Mar 11 14:14:58 What cape bus does is it reads an identifer from EEPROM, then reads data with that identifier from firmware. Mar 11 14:51:59 I must admit that it's just these superiority of FeatherCNN that has kept me stuck to it. Mar 11 14:53:46 * pradan[m] uploaded an image: PicsArt_03-11-08.23.19.png (109KB) < https://matrix.org/_matrix/media/r0/download/matrix.org/GLJHEzWgMcrsEnmmlKKLuSUh > Mar 11 14:55:36 Due to the scarce documentation and no response from the maintainers, I am finding it challenging to proceed. Still, I will be giving it some more time OR I'll have to revert to NNPACK Mar 11 14:55:36 @ds2 Mar 11 15:32:57 * hendersa waves to jkridner Mar 11 16:25:54 Hi everyone Mar 11 16:26:08 hi hendersa Abhishek_ ! Mar 11 16:29:03 we are getting engaged pretty late, but hopefully we can engage fully. Mar 11 16:30:31 I keep promising to start doing a weekly online Linux training via YouTube. I think I want to start with the release of the upcoming Debian images. Mar 11 16:30:42 jkridner[m]: I haven't been keeping up with the IRC traffic pre-GSoC like I was last year, but I'll review the logs and see what's been going on. Mar 11 16:30:44 * jkridner[m] looks for @rcn-ee Mar 11 16:31:18 not many students have been getting answers on http://bbb.io/gsocml, so starting there would help a lot. Mar 11 16:31:47 I've been encouraging students this year to put their proposals up on elinux.org using the template, so I'd suggest you encourage them to do the same. Mar 11 16:34:31 *in it becoming Mar 11 16:35:09 Haven’t seen interest on the parallel bus project so far, but I’m interested in becoming a vehicle to interface the Beagle to FPGAs with OSS toolchains. hendersa - how can we make it happen Mar 11 16:35:54 well, don't we have some good capes already out there? Mar 11 16:36:22 I think the bus isn't a project we should focus on this time around. We gave it a shot last time. Time to try something new. Mar 11 16:36:25 can we do anything related to the easy kernel driver integration with PRU? Mar 11 16:38:00 Abhishek_: I've been doing some PRU work with interfacing with CPLDs. Ironically, I've been using BeagleLogic as a template for that. Maybe a project to rework BeagleLogic into a more generic PRU interface framework. Mar 11 16:38:39 The remoteproc driver is lagging. IIRC last release was in 5.0 LTS kernel, tried checking with rcn-ee multiple times about it but no updates really. Mar 11 16:38:53 can we use rpmsg communication to FPGA? Mar 11 16:40:44 You might be able to use a PRU's R31/R30 GPIOs to talk an IP core-friendly protocol. Something like wishbone. Mar 11 16:41:25 Then you can tie your FPGA IP cores onto such a bus and route it into the BBB via the PRUs to get rpmsg capability. Mar 11 16:41:40 That's the best I can think of. Mar 11 16:42:28 If it help in someway streamlining the development of drivers, it seems useful. Mar 11 16:42:46 hendersa: Have something similar in mind Mar 11 16:42:46 Are there any active FPGA capes these days? Mar 11 16:42:57 does rpmsg have some size limit in the messages? Mar 11 16:43:02 There’s BeagleWire by mwelling Mar 11 16:43:26 I would like to design an ECP5 based cape for the AI Mar 11 16:43:59 is the AI truely usable thesedays? Mar 11 16:44:11 But is BeagleWire something that you can purchase right now? I remember seeing it as a group buy a while back. Mar 11 16:44:56 If you're interested in FPGA integration as a project, you'll need an already-available cape or configuration to target to narrow the scope for the student. Mar 11 16:46:25 * hendersa wonders where the students all are. Mar 11 16:46:28 https://www.newark.com/logi/logi-bone-2/silicon-manufacturer-xilinx/dp/38Y8591?st=beaglebone%20fpga hmm.... :-( Mar 11 16:46:32 didn't realize it went out of production. Mar 11 16:46:46 Yeah, I was shopping around a while back and came up dry. Mar 11 16:47:38 ds2: BeagleLogic porting to AI is pending. Calculated long back that it’s feasible but haven’t gotten around to it yet Mar 11 16:48:23 Abhishek_: I mean AI as a platform... heat issues, stuck at 400MHz, etc Mar 11 16:48:48 https://www.seeedstudio.com/KiwiSDR-Kit-p-2725.html Mar 11 16:56:16 jkridner[m]: Does bb.org have a set of development priorities that could help focus student efforts on things that benefit the platforms? Mar 11 16:58:36 hi Pac23 Mar 11 16:59:09 jkridner: I see, but price is on the higher side for KiwiSDR board. Mar 11 16:59:23 The mikroBus, Greybus, and BeagleBone AI cape support I'd say are the biggest BeagleBoard.org priorities.... Mar 11 16:59:54 the improvements to the AI support on BeagleBone AI is also important, but I think BeagleBone AI is most useful as just a nice update to BeagleBone Black. Mar 11 17:00:15 Abhishek_: for students, we can of course provide the boards. Mar 11 17:00:35 We might ask Seeed to provide a cheaper board to the community if there's interest. Mar 11 17:02:27 jkridner[m]: is the AI currently useable today (heat, clock limits, etc)? Mar 11 17:05:23 yeah, if you clock it down to 400MHz. Mar 11 17:05:32 otherwise, use one of the fans out there. Mar 11 17:06:43 is there plans (or prehaps a GSoC project?) to make the AI useable for > 400MHz? Mar 11 17:06:52 You can still do 1.5GHz bursts, but you need to slow it down before it gets too hot or it just won't cool back down. Mar 11 17:07:23 ds2: you mean add more clock gating code? Mar 11 17:08:22 jkridner[m]: no, figure a better work around for the heat issue... a 400MHz limit is making the AI look pretty bad Mar 11 17:10:54 even with the A15/A8 differences, the core on the AI is looking a lot like the Black's ARM Mar 11 17:11:35 SGX is faster (but no one really uses that right now) and EVE/DSP support is a bit limiting in TIDL Mar 11 17:13:45 ds2: does enablement need to be tied to fanless or can we assume fans? Mar 11 17:14:21 https://www.newark.com/element14/6100310/beaglebone-ai-fan-cape/dp/50AH3704 Mar 11 17:14:25 hi pdp7 Mar 11 17:14:40 * jkridner[m] has been distracted from this for a while by pdp7 Mar 11 17:14:47 jkridner[m]: IMO, fans are secondary as long as the students gets a kit with them if it is needed Mar 11 17:15:19 I can get them fans for development. Mar 11 17:15:50 but do we have images that will go faster w/o toppling over? Mar 11 17:16:03 I'm not sure we did a much better job of advertising cool projects this time, but we do have a number of potential students lurking here right now. Mar 11 17:16:06 i am finding the AI to be very fragile Mar 11 17:16:20 Over heat? Mar 11 17:16:36 anyone in here now planning to make a project proposal submission next week? Mar 11 17:16:49 ds2: can you be specific? Mar 11 17:16:52 yes unless zmatt's sledge hammering on the PRCM controls is applied Mar 11 17:16:56 ds2: when adding cape hardware? Mar 11 17:17:13 jkridner[m]: I am just talking about using the AI for AI Mar 11 17:17:29 The TIDL? Mar 11 17:17:32 but to be fair, I am not running a fan... Mar 11 17:17:37 ds2: hmmm.... we need to get zmatt's code into the default kernel... probably lots of simple house-keeping stuff like that to do. Mar 11 17:18:06 some of the TFLite conversion bits have landed in the repos.... Mar 11 17:18:25 integration of the latest OpenCV stuff seems to have fallen off. Mar 11 17:18:41 jkridner[m]: thought there was suppose to be a "proper fix", not his sledge hammer style thing (use /dev/mem and force bits on) Mar 11 17:19:03 TIDL is... uh... not very useful :D Mar 11 17:19:25 jkridner (@jkridner:matrix.org): I am planning to submit my proposal next week Mar 11 17:20:04 no trainning support; have to dealwith those .txt files for config; a bit C++ mess that pulls in stuff that isn't needed (IMO) Mar 11 17:20:17 I'm just waiting for some of my questions to get cleared.. So i can create a better proposal Mar 11 17:20:49 * jkridner[m] wishes Manisha was here. Mar 11 17:21:08 Manisha has done some presentation materials on how to train, but I don't see much public. Mar 11 17:22:03 training on the AM57x isn't supported by TIDL, right? Mar 11 17:22:13 it has to be done on x86 as I understand it Mar 11 17:22:52 pradan[m]: how's the app going? Mar 11 17:22:57 there is some work on RNNs and on-going training. Mar 11 17:25:06 there is always fall back to the x15 :D Mar 11 17:25:41 indeed. Mar 11 17:26:13 all: what would be most useful right now, I believe is engaging those students who have sumitted questions to https://bbb.io/gsocml Mar 11 17:26:52 has anyone else replied? :D Mar 11 17:31:28 ds2: meet manisha.agrawal !!! Mar 11 17:31:32 thank you so much manisha.agrawal Mar 11 17:32:05 ? Mar 11 17:32:07 ds2 is a long-time BeagleBoard.org community member/leader and is trying to instigate a project related to machine learning this year. Mar 11 17:32:30 I don't see anyone join beagle-* in the last 2 mins Mar 11 17:33:03 manisha.agrawal: I think you'd shown me a presentation at one point for training of networks for TIDL. Is that published? Mar 11 17:33:17 ds2: it might be a nutty matrix bridge thing. Mar 11 17:33:24 did you see what she typed? Mar 11 17:33:28 nope Mar 11 17:33:38 Me neither. It's a bridge problem. Mar 11 17:33:43 hmmmm.... I hope it isn't a voice thing. Mar 11 17:34:01 ah good, thought freenode is lagging badly Mar 11 17:34:10 manisha.agrawal: are you seeing what hendersa and ds2 are typing? Mar 11 17:35:23 note sure what is wrong with the bridge Mar 11 17:35:46 manisha.agrawalsaid: I can only see ds2 typing Mar 11 17:35:46 Please check this link http://software-dl.ti.com/processor-sdk-linux/esd/docs/latest/linux/Foundational_Components/Machine_Learning/tidl.html#neural-network-layers-supported-by-tidl Mar 11 17:35:46 It has list of all the operators supported by TIDL Mar 11 17:36:17 What is the original question on TIDL Mar 11 17:37:39 is there examples and docs on how to get TIDL going with something new? most of the tools want these .txt file inputs Mar 11 17:37:41 ds2: I don't think we'll have Manisha very long, so it would be good if you could share any questions that have been outstanding. Mar 11 17:38:03 ds2 said: is there examples and docs on how to get TIDL going with something new? most of the tools want these .txt file inputs Mar 11 17:38:43 and is that link applicable to the top of the git tree on git.ti? Mar 11 17:38:44 I can only see ds2 typing Mar 11 17:38:44 Please check this link http://software-dl.ti.com/processor-sdk-linux/esd/docs/latest/linux/Foundational_Components/Machine_Learning/tidl.html#neural-network-layers-supported-by-tidl Mar 11 17:38:44 It has list of all the operators supported by TIDL Mar 11 17:38:44 What's the original question on TIDL Mar 11 17:38:44 Wha's meant by TIDL going with something new? Mar 11 17:38:44 HI Jason, I can see messages from ds2 Mar 11 17:39:11 Yes, that link is applicable to git tree Mar 11 17:39:29 manishaagrawal[m: You're showing up for me in the IRC channel now, as well. Mar 11 17:39:37 Okay, great Mar 11 17:40:02 howdy nerdboy Mar 11 17:40:05 manishaagrawal[m: is there an example for using a converted model? (context - I am trying to figure a way to get YOLO going).. based on another post, it is understood some YOLO layers may need to be done on the ARM side Mar 11 17:40:54 That's correct. All the layers from YOLO model is not supported by TIDL Mar 11 17:41:05 There's couple of ways this can be achieved Mar 11 17:42:09 the big barriers I ran into are - the .txt files and the examples being buried in C++ using libraries that aren't needed (so it winds up being a distraction) - OpenCV is a big one. I think I hacked up an old version (~Nov time frame) to remove excess OpenCV to try to figure out what TIDL can do/does but it was a lot of work Mar 11 17:42:51 We are adding support for TFLite and Neo runtime inference engine on BBAI. We will have heterogenous compute feature added to these run time engines. All the operators supported by TIDL will run on the accelerators, rest will run on Arm core. Mar 11 17:43:35 will this include examples on doing that? Mar 11 17:43:47 Adding TIDL as backend in OpenCV DNN kit is also one of the option and I recommend that to be one of the google summer code project Mar 11 17:43:56 Yes Mar 11 17:44:45 I rather not have OpenCV involved... it is a big C++ mess Mar 11 17:45:05 I also suggest to have GSoC project on running models from model zoo using SageMaker Neo AI. That inlcude YOLO mode Mar 11 17:45:34 pardan[m]: is a potential student in this area Mar 11 17:45:48 Oh, I didn't know that. Yesterday I attended Gary's talk (OpenCV founder) and he was selling DNN kit in big way Mar 11 17:46:26 Which area? OpenCV or Neo? Mar 11 17:46:47 manishaagrawal[m: C++ is a pain when it comes to embedded stuff... I have doen projects where I have to debug C++ things Mar 11 17:46:48 I need to leave soon. Mar 11 17:47:12 manishaagrawal[m: the current thing is get YOLO running at a reasoanble rate but proposal is not written (nor have i seen it) Mar 11 17:47:20 I've pointed to the TFLite conversion stuff. Mar 11 17:47:25 thanks for dropping in... comments appreciated! Mar 11 17:48:04 I am not sure what specific pain point you are talking, My experience so far is good with C++ Mar 11 17:48:12 https://git.ti.com/cgit/tidl/tidl-utils/tree/src/importTool/modules/ti_dl/utils/tfImport Mar 11 17:48:34 tracing down memory corruption is very very painful Mar 11 17:48:46 C++ has too many implied things that just happens Mar 11 17:49:14 I believe there are also some Python bindings today: https://git.ti.com/cgit/tidl/tidl-api/tree/tidl_api/pybind11 Mar 11 17:49:16 It was nice chatting on TIDL stuffs. I gotta go. I would love to explore some projects on Neo and BBAI Mar 11 17:49:22 I'd expect Python support to get better. Mar 11 17:49:48 I've tried to connect the two of you by e-mail for follow-up. Mar 11 17:49:53 and C++ libraries means one HAS to pull in C++ stuff Mar 11 17:50:06 (which may have licensing issues dragged in) Mar 11 17:50:14 * jkridner[m] agrees that C would be nicer than C++ Mar 11 17:50:37 hi cwicks, you around? Mar 11 17:50:42 just another data point -- I have a Jetson nano next to me and it right now is so much easier to use for AI stuff Mar 11 18:09:15 i think the python bindings would only be as good as what they're wrapping... Mar 11 19:14:39 Sorry for not being present during the ACTUAL conversation (even if I just had to listen and not speak). Mar 11 19:38:30 pradan[m]: any updates? Mar 11 19:39:19 pradan[m]: started on an app yet? Mar 11 19:40:09 Sorry but what should I infer from app ? I didn't get it... Mar 11 19:40:23 pradan[m]: app as in application for GSoC Mar 11 19:42:47 Oh ... yeah I am done with the pre-requisites and made my page on elinux but still not completed the application. Actually, I wished to clarify WHAT was I actually gonna do. So, thought of devoting another day... Hopefully, I will draft the first version by tommorow evening (approx this exact time) Mar 11 22:24:26 pradan[m]: anyone else responding to you besides me? Mar 12 02:23:50 pradands2 are you looking for more input on the task? Ideally, I'd like to see ds2 really providing the guidance for what needs to be done on this one. **** ENDING LOGGING AT Thu Mar 12 02:59:57 2020