#asahi-dev · 2021-12-15 14:39
arndDocumentation/process/email-clients.rst has a few things on Thunderbird, not sure any of those help#asahi-dev · 2021-12-15 14:31
arndit's the same thing that happened last time, I would blame thunderbird#asahi-dev · 2021-12-15 14:23
arndmarcan: ok, I see it. There are no extra linewraps, but I still see the odd extra spaces at the start of each line that has one or more spaces at the beginning.. Not a problem since the commit message comes from the tag anyway, but I wonder what causes this#asahi-dev · 2021-12-15 13:22
arndmarcan: and if you did need this as a dependency, you still wouldn't need to merge in -rc6, but just merge the fixes branch that contains this commit into your DT branch, that way you'd avoid backmerging all of the mainline changes from other people#asahi-dev · 2021-12-15 13:21
arndmarcan: ok, so no functional dependency, just a trivial conflict affecting context, don't worry about that then#asahi-dev · 2021-12-15 12:23
arndmarcan: which mainline patch is causing the conflict?#asahi-dev · 2021-12-15 12:23
arndmarcan: torvalds and I are both used to sorting out merge conflicts. If the new patches you have work without the earlier bugfix, I'd prefer to keep the git history clean and do the conflict resolution on my end#asahi-dev · 2021-11-11 12:31
arndkettenis: yes, that too. I was thinking of the PCIe binding proposal in https://www.devicetree.org/open-firmware/bindings/pci/pci-express.txt, which I think is in an even earlier stage than "draft", and is overall not helpful because it completely leaves out the interesting bits like MSI#asahi-dev · 2021-11-11 11:58
arndIIRC there was a proposed PCIe binding, but unlike the PCI one it was never ratified, and the three main companies shipping PCIe with OF (IBM, Apple, Sun) each had their own way of doing it, while the embedded systems did just enough to get linux booting#asahi-dev · 2021-11-11 11:56
arndthe sb600 is only used on Nemo, so this is likely worse than what ever they had on Electra. I remember the way the PCIe bringup went for the IBM PPC blades, with the kernel and firmware teams slowly converging on some DT representation that was halfway and working, but not according to spec#asahi-dev · 2021-11-11 11:46
arndI guess he doesn't really want to get involved either way ;-)#asahi-dev · 2021-11-11 11:46
arndI wonder if this is the same firmware that ojn used on the Electra#asahi-dev · 2021-11-11 11:45
arndok, makes sense. I do remember a number of problems with conflicting interpretations of the PCI interrupt properties#asahi-dev · 2021-11-11 11:34
arndmaz: if their DT is clearly broken, I think the easiest workaround would be to have some code in pas_init_IRQ() or pas_pci_init() that adjusts the respective properties. Did you figure out what exactly goes wrong?#asahi-dev · 2021-11-10 10:18
arndj_ey: We normally don't take new material after the start of the merge window. For .dts files, there isn't much need anyway, as this can easily be shipped outside of the kernel#asahi-dev · 2021-11-09 10:35
arndmarcan: for testing the read_b/read_w change, it shouldn't be hard to find the right person: I see that akashi worked on https://www.spinics.net/lists/linux-mmc/msg61479.html, I can ask him about it when I see him on IRC, and apparently the Ben Chuang from Genesys is active on the driver as well#asahi-dev · 2021-11-09 10:18
arndmarcan: is it reading the wrong byte (e.g. some byteswap gone wrong), or is it always 0xff for any address?#asahi-dev · 2021-11-09 10:04
arndmarcan: there are standard properties for inverted wp/cd signals, as documented in Documentation/devicetree/bindings/mmc/mmc-controller.yaml, and I think there is a helper function to parse those#asahi-dev · 2021-10-21 16:10
arndxhci-tegra and xhci-rcar call request_firmware()#asahi-dev · 2021-10-21 16:08
arndno, I don't think they do that, maybe some maskrom inside the dwc3 block though#asahi-dev · 2021-10-21 16:07
arndI haven't seen a confirmation for those, but I'm pretty sure that any XHCI needs to have firmware running on some MCU, and ARC is what Designware tends to use because they ownit#asahi-dev · 2021-10-21 16:06
arndah, so you are not even counting the ARC cores in the XHCI then#asahi-dev · 2021-10-21 13:12
arndmarcan: how many different instruction sets in those 15 coprocessors? The highest I'm aware of on that scale was Snapdragon, but I forget the exact number that I was told#asahi-dev · 2021-10-20 12:34
arndI'm thinking of applications that use cublas on nvidia GPUs today, not their higher level interfaces or the tensor cores#asahi-dev · 2021-10-20 12:33
arndalyssa: on a related note, do you think it would be possible to have an OpenCL based BLAS implementation on top of your gallium driver, and use that for machine learning on the GPU instead of the ANE?#asahi-dev · 2021-10-20 12:30
arndwhich in turn is what apple's Accelerate framework or Intel's OneAPI do as well, but I have not seen anyone do such an abstraction on the kernel to user boundary#asahi-dev · 2021-10-20 12:29
arndone could do something that looks like cudnn or cublas, and build on top of that#asahi-dev · 2021-10-20 12:29
arndTheLink: cuda is a user space interface, not kernel level, and their kernel interface is not an abstraction but hardware specific#asahi-dev · 2021-10-20 12:26
arndThe GPU programming model always feels like a layering violation to me, the same way that using DPDK to send network data does, or using a custom interface to a DSP doing video encoding instead of using v4l2. OTOH I have no idea what a good kernel abstraction for machine learning hardware would actually look like, so it's probably the best we can do here#asahi-dev · 2021-10-20 11:09
arndin this case, I don't think the question of open source user space is the main issue. As we won't be running MacOS user space on Linux for this, someone has to complete the reverse-engineering anyway in order to create a new compiler#asahi-dev · 2021-10-20 11:04
arndhttps://lwn.net/Articles/870418/ summarizes the discussion we had at the kernel summit about those devices. #asahi-dev · 2021-10-20 11:02
arndright, this does sound like an ioctl-type interface to send compiled code to the engine and run that, which is indeed similar to what GPUs would do, and also a bit like what we did a long time ago with spufs#asahi-dev · 2021-10-20 10:47
arndthe main question to me is whether it implements a particular set of high-level matrix operations (GEMM, convolution, ...) that can be abstracted using a kernel interface, or if this isx a fully programmable unit that relies on JIT-compiling your ML model into a custom ISA and run it autonomously#asahi-dev · 2021-10-20 10:45
arndah, right. What do we know about the hardware side, and about how applications access it in MacOS?#asahi-dev · 2021-10-20 10:44
arndwhat is ANE?#asahi-dev · 2021-10-13 12:51
arndI'd also include the way that Apple hooks up the timer to the FIQ in this, regardless of whether this is technically compliant or not#asahi-dev · 2021-10-13 12:49
arnd(I assume newer versions fixed it, this was exynos5 IIRC)#asahi-dev · 2021-10-13 12:49
arndalso, timer bugs are notoriously hard to find, because you can't rely on the timer to tell you when it goes wrong. One example is the Samsung Exynos, which has a timer that almost always has the correct values, but once in a blue moon produces nonsense#asahi-dev · 2021-10-13 12:47
arndDocumentation/devicetree/bindings/timer/arm,arch_timer.yaml in Linux lists a number of errata for the timer. In most cases this ends up meaning that we can't use it from user space because it's unreliable#asahi-dev · 2021-10-13 12:40
arndkettenis: The hardware timer is supposed to be there on ARMV7VE (Cortex-A7/A15) or later, but there are enough examples of it being broken in certain SoCs, I don't think there is any hope of ARM being less of a trainwreck in this area than Intel#asahi-dev · 2021-10-08 12:09
arndalyssa: yes, exactly#asahi-dev · 2021-10-08 07:00
arndalyssa: I'd try to avoid the ## concatenation. While that does involve passing both (almost identical) identifiers in the DCP_THUNK_*() call, it does become possible to grep afterwards, which helps readers understand where the calls actually go.#asahi-dev · 2021-07-16 14:10
arndit looks a lot like the AMX instructions you documented, but I wonder what differences there are.#asahi-dev · 2021-07-16 14:09
arnddougall: have you seen https://developer.arm.com/documentation/ddi0602/2021-06/SME-Instructions https://community.arm.com/developer/ip-products/processors/b/processors-ip-blog/posts/scalable-matrix-extension-armv9-a-architecture ?#asahi-dev · 2021-07-13 08:30
arndRight. The idea of what a "bus" is in DT doesn't really match the reality of modern SoCs any more, where you have lots of point-to-point links that can have different behavior. It used to be that dma-ranges and coherency were fundamentally tied to the physical bus implementation#asahi-dev · 2021-07-13 08:21
arnddma-ranges should definitely be in all the bus nodes, it seems we went the other way with dma-coherent for some reason#asahi-dev · 2021-07-13 08:20
arndok, I'm probably misremembering it then#asahi-dev · 2021-07-13 08:07
arndso if there are hierarchical buses, put the properties in both the parent and child buses#asahi-dev · 2021-07-13 08:06
arndkettenis: IIRC "dma-coherent" should be a property of each bus node that is a parent of a dma-capable device, same as "dma-ranges"#asahi-dev · 2021-05-28 07:44
arndsven: I agree that there should be something in the dart driver to allow booting with 4K pages, but I don't think it's necessary to have isolation in that case for the initial merge. Falling back to a fixed linear map (plus swiotlb if necessary), but having proper iommu support for 16KB pages should give us enough confidence in the binding that we can change the iommu core code for small pages later