#asahi-dev · 2022-11-28 19:09
alyssai am your pony waifu2x#asahi-dev · 2022-11-28 19:09
alyssasuch trust and vigour for life#asahi-dev · 2022-11-28 19:09
alyssasoft eyes to look into#asahi-dev · 2022-11-28 19:09
alyssayou've hair soft as feathery down#asahi-dev · 2022-11-28 19:08
alyssapovik: i see your attempt at nerdsniping dcp audio#asahi-dev · 2022-11-28 19:06
alyssathank you Dr. Peter#asahi-dev · 2022-11-28 19:05
alyssa(...strictly i need to graduate first, but I have enough credits for a minor...)#asahi-dev · 2022-11-28 19:05
alyssamarcan: and I'm a GPU developer with a minor in math, of couse I have random ML knowledge to drop :p#asahi-dev · 2022-11-28 19:05
alyssamarcan: go to sleep#asahi-dev · 2022-11-28 19:03
alyssalots of people are being paid big bucks to not solve these (-:#asahi-dev · 2022-11-28 19:03
alyssayou're not the only one having these problems#asahi-dev · 2022-11-28 19:03
alyssaso yeah... nothing standardized at that point... do what's best in the kernel side (probably DRM driver), do your own thing in userspace (libane with your own cute little compiler?), plumb libane into tensorflow and pytorch or whatever? IDK#asahi-dev · 2022-11-28 19:02
alyssanot sure there'd be any value in using LLVM or NIR, they may just get in the way..#asahi-dev · 2022-11-28 19:01
alyssaright... so quite a bit simpler than Habana I suspect#asahi-dev · 2022-11-28 18:58
alyssaHabanalabs is using LLVM for their commpiler ... while I won't stop you from doing so, as a maintainer of a competing project I won't be able to help if you go that route ;-)#asahi-dev · 2022-11-28 18:57
alyssaSomething in the middle?#asahi-dev · 2022-11-28 18:57
alyssaIs this "it can run C code"?#asahi-dev · 2022-11-28 18:57
alyssaIs this "we have a couple of registers for network weights and the rest is fixed"?#asahi-dev · 2022-11-28 18:57
alyssaI do want to understand the complexity of the ANE instruction set#asahi-dev · 2022-11-28 18:53
alyssasven: NPUs are just GPUs that suck at rasterization....#asahi-dev · 2022-11-28 18:51
alyssabut every so often the ML folk pass by my virtual cubicle at work :-p#asahi-dev · 2022-11-28 18:51
alyssathis isn't my 'area', I'm not the expert here#asahi-dev · 2022-11-28 18:50
alyssa("ANE done wrong" would be the downstream vendor approach to get some cool pics for Twitter/Mastodon and never end up getting shipped...)#asahi-dev · 2022-11-28 18:49
alyssaas the closest prior art for "ANE done right"#asahi-dev · 2022-11-28 18:49
alyssaso it's worth looking at what they're doing#asahi-dev · 2022-11-28 18:49
alyssaand ruminations of an accel subsystem as mentioned above#asahi-dev · 2022-11-28 18:49
alyssaI think Habana Labs (only) are doing things "properly" with upstream/mainline approach now#asahi-dev · 2022-11-28 18:48
alyssaUsually that's downstream char/misc devices in the kernel with custom vendor ioctls and a downstream fork of tflite or whatever ...#asahi-dev · 2022-11-28 18:48
alyssa(Because there's not a standard API, ie a hypothetical Khronos OpenML, that this hardware would implement... In effect the tensorflow API would be that API in this case...)#asahi-dev · 2022-11-28 18:47
alyssaI think the current state-of-the-art is adding a tensorflow backend for your hardware and yeeting#asahi-dev · 2022-11-28 18:46
alyssabut we don't currently have a home for userspace drivers for hardware of that class (Mesa for GPUs, libcamera for camera processing, what for NPUs?)#asahi-dev · 2022-11-28 18:46
alyssaIt does not make sense to support that in Mesa#asahi-dev · 2022-11-28 18:46
alyssawhich is more sensible if you're going to have an ANE at all#asahi-dev · 2022-11-28 18:45
alyssaOn the other extreme you have NPUs that are fixed function neural network hardware#asahi-dev · 2022-11-28 18:45
alyssaIMHO that design doesn't make a lot of sense on SoCs that have competent CPUs and GPUs, either of which can run OpenCL excellently, why add NPU into the mix?#asahi-dev · 2022-11-28 18:44
alyssahence is supported by the existing etnaviv GPU stack, supporting (only) OpenCL in userspace via mesa#asahi-dev · 2022-11-28 18:44
alyssathe Vivante NPU is literally just a Vivante GPU with all the graphics support stripped out#asahi-dev · 2022-11-28 18:43
alyssaSome NPUs (e.g. Vivante) are general purpose compute accelerators#asahi-dev · 2022-11-28 18:43
alyssaI think I'd like to understand the capabilities of the ANE more before I could give any advice#asahi-dev · 2022-11-28 18:43
alyssaStuff isn't too standardized yet, as you're seeing#asahi-dev · 2022-11-28 18:42
alyssaNPUs in mainline are in a bit of a weird spot right now#asahi-dev · 2022-11-28 18:42
alyssaeiln (for when they're back)--#asahi-dev · 2022-11-20 21:49
alyssadunno#asahi-dev · 2022-11-20 21:47
alyssa(and have nothing plugged to the left port ...)#asahi-dev · 2022-11-20 21:47
alyssa(I get that issue unless I use the right port ....)#asahi-dev · 2022-11-20 21:46
alyssatry switching which usb port on the proxyclient machine is used#asahi-dev · 2022-11-20 21:46
alyssatry unplugging all usb devices from the proxyclient machine, except for the cable to the m1#asahi-dev · 2022-11-20 21:46
alyssatry again#asahi-dev · 2022-11-20 21:12
alyssa:-D#asahi-dev · 2022-11-20 21:11
alyssajannau: nice