Huawei ExaGear: x86_64 apps on Arm 64-bit Linux systems(threedots.ovh)
threedots.ovh
Huawei ExaGear: x86_64 apps on Arm 64-bit Linux systems
https://threedots.ovh/blog/2021/01/huawei-exagear-x86_64-apps-on-arm64/
58 comments
Some more information:
A chinese-speaking internet user put in some research and according to them it appears to be Eltech's software. They also appear to be somewhat angry that Huawei did not properly credit Eltech.
Note that apparently there was a lot of controversy around this on Chinese boards, with people arguing vehemently for either side (based on Eltech's software vs. not based on it).
Chinese: https://zhuanlan.zhihu.com/p/263519125
Translated: https://translate.google.com/translate?sl=auto&tl=en&u=https...
HN user 'oefrha' managed to find the official project page and documentation:
https://news.ycombinator.com/item?id=25750553
A chinese-speaking internet user put in some research and according to them it appears to be Eltech's software. They also appear to be somewhat angry that Huawei did not properly credit Eltech.
Note that apparently there was a lot of controversy around this on Chinese boards, with people arguing vehemently for either side (based on Eltech's software vs. not based on it).
Chinese: https://zhuanlan.zhihu.com/p/263519125
Translated: https://translate.google.com/translate?sl=auto&tl=en&u=https...
HN user 'oefrha' managed to find the official project page and documentation:
https://news.ycombinator.com/item?id=25750553
Eltechs' ExaGear at the time only provided x86_32 compatibility on top of 32-bit Arm.
This is a major departure from that, my best guess would be that Huawei purchased the company and then poured lots of resources to it.
This is a major departure from that, my best guess would be that Huawei purchased the company and then poured lots of resources to it.
Given the timeline Eltech had likely been shut down for at least a year. Then Apple came around with their x86 to ARM translation layer, prompting Huawei to go shopping for a competitor, seeing the defunct Eltech, and buying their IP.
Sounds like a late happy ending for a failed company that was just a few years too early with their product. I hope the founders got a good deal.
Sounds like a late happy ending for a failed company that was just a few years too early with their product. I hope the founders got a good deal.
...or Huawei strategically decided to take advantage of their vertical integration (with HiSilicon ARM chips), by moving its processors up the value chain into laptops, thereby capturing more profit instead of paying Intel/AMD? Such strategic decisions are usually years in the making - and everyone could see the trajectory of ARM vs x86 for years
ARM processors in laptops predates the M1 by many years - you don't need inspiration from Apple to figure out that efficiently running x86 code on ARM hardware is useful.
ARM processors in laptops predates the M1 by many years - you don't need inspiration from Apple to figure out that efficiently running x86 code on ARM hardware is useful.
Also check out Box86: https://github.com/ptitSeb/box86
Impressive to see people playing Half-Life 2, Unreal Tournament 99 and 2004 on a Raspberry Pi 4.
Curious if this could be extended to Emscripten (JIT for WebAssembly/Emscripten?) which would be perfect for archiving and portability. Closest project I have seen is DOSBox-X which seems to already have an Emscripten port: https://github.com/joncampbell123/dosbox-x
HL2: https://www.youtube.com/watch?v=Lq5dJ0oc_6k
UT2004: https://www.youtube.com/watch?v=4VOAeOQcrdo
Impressive to see people playing Half-Life 2, Unreal Tournament 99 and 2004 on a Raspberry Pi 4.
Curious if this could be extended to Emscripten (JIT for WebAssembly/Emscripten?) which would be perfect for archiving and portability. Closest project I have seen is DOSBox-X which seems to already have an Emscripten port: https://github.com/joncampbell123/dosbox-x
HL2: https://www.youtube.com/watch?v=Lq5dJ0oc_6k
UT2004: https://www.youtube.com/watch?v=4VOAeOQcrdo
Another fun thing about box86 is it isn't specifically limited to ARM. Its JIT translator is ARM-specific, but it includes a pure emulator which will work on any CPU. There's additional work to be done for each port, but I got it working on 32 bit PowerPC LE, allowing one to run x86 apps on IBM's POWER processors. For all the few dozens of people this may be relevant to :-)
Sadly, box86 doesn't support aarch64.
Does Raspbian yet?
Somewhat, they have a beta aarch64 image: https://downloads.raspberrypi.org/raspios_arm64/images/
More info here about box86 https://www.giantpockets.com/box86-run-x86-code-and-games-on...
An interesting excerpt from the article:
"ExaGear does not follow TSO completely by default, using heuristics instead." TSO emulation can be enabled fully, using heuristics, or turned off using a configuration option. It would be interesting to see benchmarks that analyze the performance (and correctness) impact of full TSO emulation vs. heuristics.
Another interesting bit that is mentioned is that like Apple's M1, NVidia's Tegra K1 64-bit and Tegra Xavier support sequential consistency as the memory model.
"ExaGear does not follow TSO completely by default, using heuristics instead." TSO emulation can be enabled fully, using heuristics, or turned off using a configuration option. It would be interesting to see benchmarks that analyze the performance (and correctness) impact of full TSO emulation vs. heuristics.
Another interesting bit that is mentioned is that like Apple's M1, NVidia's Tegra K1 64-bit and Tegra Xavier support sequential consistency as the memory model.
Apple M1 has togglable TSO with ACTLR_EL1, not sequential consistency.
You also don't run as TSO by default, has to be toggled on by the kernel when context switching to a given task.
You also don't run as TSO by default, has to be toggled on by the kernel when context switching to a given task.
This would make sense, as Eltech's engine was used extensively across multiple ports done by Nvidia for their Shield platform (RE5, MGRR)
What qemu-user makes so special for me is that is can be pre-loaded by the kernel, allowing any foreign arch chroot to directly work, nice for containers¹. I've not read if this here does support it, too.
¹ Once `qemu-user-static` is installed on your system, try `docker run --rm -ti docker.io/arm64v8/alpine` on amd64 or vice versa
¹ Once `qemu-user-static` is installed on your system, try `docker run --rm -ti docker.io/arm64v8/alpine` on amd64 or vice versa
That's just a general kernel feature named binfmt_misc [1], just register the interpreter with the O flag. You can write your own binary translator and have the kernel run it for you, too.
[1] https://www.kernel.org/doc/html/latest/admin-guide/binfmt-mi...
[1] https://www.kernel.org/doc/html/latest/admin-guide/binfmt-mi...
This supports it too. See the ExaGear4Docker folder in the install archive.
I wonder if this is related to Huawei's effort to bring forth desktop PCs and laptops based on their Kirin/Kunpeng ARM CPUs? Running a desktop version of their HarmonyOS with x86 apps in a compability layer.
Any solutions for doing the reverse, namely running arm64 binaries on x86_64? Use case: building ARM software packages (like Docker containers) on x86_64 CI servers.
qemu?
QEMU's sysemu is a little unwieldy for CI, but the linux userspace support is perfect for this use case.
In case you haven't encountered it -- it's terribly clever. Userspace instructions are translated just like they are in full system mode, but when a system call is encountered, the arguments are translated between the host and guest system. So the guest architecture linux program 'just works'. There's also hooks you can enable in binfmt-misc to make the loader invoke the qemu binary automagically.
In case you haven't encountered it -- it's terribly clever. Userspace instructions are translated just like they are in full system mode, but when a system call is encountered, the arguments are translated between the host and guest system. So the guest architecture linux program 'just works'. There's also hooks you can enable in binfmt-misc to make the loader invoke the qemu binary automagically.
Sounds interesting. Do you have a link with more info about these things?
[1] [2] are handy links.
[1] https://www.qemu.org/docs/master/user/main.html
[2] https://wiki.debian.org/QemuUserEmulation
[1] https://www.qemu.org/docs/master/user/main.html
[2] https://wiki.debian.org/QemuUserEmulation
BTW, how does ExaGear compare to qemu for running x86_64 on arm64 hosts ?
For something faster than Qemu, there's this: https://threedots.ovh/blog/2020/12/houdini-run-arm-32-bit-an...
Every time I travel abroad, I'm amazed by how much Huawei has penetrated every major market except U.S.
The reason is obvious: the US is enacting barriers to the internal market. But Huawei is developing all kinds of technology around the world.
The US is responding to asymmetrical access. Facebook, Google, Twitter, Instagram et al have been blocked for years in China. As of this week I'm ok with that but I digress. Apple's the only foreign cell phone maker that's really done alright in absolute dollar terms, but their market share is rather small. Access to the tantalizing "rising chinese middle class" is tantalizing and highly conditional.
https://www.counterpointresearch.com/china-smartphone-share/
> Facebook, Google, Twitter, Instagram et al have been blocked for years in China.
The Chinese government did not decide this on a whim, these companies are not willing to comply with Chinese rules within China. Whatever we might think of these rules the bottom line is that these companies are not in China because they have decided not to be. And when they try to move in, like Google did, their very employees argue against it.
China is Apple's largest market, iirc, they are doing very well there. For smartphones their market share is 3 times their market share in India. Still, a lot of Chinese still cannot afford iPhones and China has many domestic manufacturers: Actual competition is fierce, this has nothing to do with politics.
The argument that the US are retaliating against China blocking the FAANG is largely for the show but not reality.
The Chinese government did not decide this on a whim, these companies are not willing to comply with Chinese rules within China. Whatever we might think of these rules the bottom line is that these companies are not in China because they have decided not to be. And when they try to move in, like Google did, their very employees argue against it.
China is Apple's largest market, iirc, they are doing very well there. For smartphones their market share is 3 times their market share in India. Still, a lot of Chinese still cannot afford iPhones and China has many domestic manufacturers: Actual competition is fierce, this has nothing to do with politics.
The argument that the US are retaliating against China blocking the FAANG is largely for the show but not reality.
That’s a bit rich. The same could be argued about Huawei. The rules in the US are don’t be agent for a foreign government, but Huawei wouldn’t agree to that. So it’s also their own choice.
That's (bad) whataboutism, and nothing to do with my comment.
> The rules in the US are don’t be agent for a foreign government, but Huawei wouldn’t agree to that.
I'll bite: If you have proof, or even actual evidence, of any part of that statement, I am sure the public at large would be very interested.
Unfortunately this topic always leads to this sort of low effort reply and a sensible discussion is never possible.
> The rules in the US are don’t be agent for a foreign government, but Huawei wouldn’t agree to that.
I'll bite: If you have proof, or even actual evidence, of any part of that statement, I am sure the public at large would be very interested.
Unfortunately this topic always leads to this sort of low effort reply and a sensible discussion is never possible.
> https://www.telegraph.co.uk/news/2019/07/05/huawei-staff-cvs...
> https://www.forbes.com/sites/zakdoffman/2019/07/06/huawei-em...
> https://www.wsj.com/articles/u-s-officials-say-huawei-can-co...
This is not news.
Aside from the spying, Huawei has zero regard for Intellectual Property law, as demonstrated by numerous cases of stolen code, designs, and even wholesale products.
> https://www.forbes.com/sites/zakdoffman/2019/07/06/huawei-em...
> https://www.wsj.com/articles/u-s-officials-say-huawei-can-co...
This is not news.
Aside from the spying, Huawei has zero regard for Intellectual Property law, as demonstrated by numerous cases of stolen code, designs, and even wholesale products.
"US officials say ..." - That counts for very little. Where's the evidence?
The other stories are about some of Huawei's employees having on their CV that they've worked with the PLA and other government agencies. By that metric, what major US company isn't connected to the US government?
> numerous cases of stolen code, designs, and even wholesale products.
This is a huge exaggeration. Huawei has had IP disputes with other companies, and there have been some allegations like T-Mobile's claims about its "Tappy" robot being copied, but this is all relatively minor stuff for a massive company. Compared to the IP disputes between Apple and Samsung, for example, this is peanuts.
On the other hand, Huawei is one of the largest R&D spenders in the world now.
The other stories are about some of Huawei's employees having on their CV that they've worked with the PLA and other government agencies. By that metric, what major US company isn't connected to the US government?
> numerous cases of stolen code, designs, and even wholesale products.
This is a huge exaggeration. Huawei has had IP disputes with other companies, and there have been some allegations like T-Mobile's claims about its "Tappy" robot being copied, but this is all relatively minor stuff for a massive company. Compared to the IP disputes between Apple and Samsung, for example, this is peanuts.
On the other hand, Huawei is one of the largest R&D spenders in the world now.
What about GPL violations:
https://twitter.com/huaweimobile/status/256672638154457088
https://www.geekzone.co.nz/forums.asp?forumid=66&topicid=242...
https://twitter.com/huaweimobile/status/256672638154457088
https://www.geekzone.co.nz/forums.asp?forumid=66&topicid=242...
It looks like you're throwing spaghetti at the wall to see what sticks. Pick any large company and you'll find numerous tiny incidents like this.
It's surprising how facts like this are downvoted simply because people don't like hearing the truth.
I can understand it happening on Reddit, but really thought HN would have more knowledgeable users.
The fact is Microsoft and Apple follow Chinese laws and they operate there with no issues, while Facebook and Twitter do not.
Google was even going to reenter the Chinese market but was shamed into abandoning that project.
I can understand it happening on Reddit, but really thought HN would have more knowledgeable users.
The fact is Microsoft and Apple follow Chinese laws and they operate there with no issues, while Facebook and Twitter do not.
Google was even going to reenter the Chinese market but was shamed into abandoning that project.
I wonder if it’s partly due to the US being able to afford alternatives, wether they are cost effective or not.
For many installments, the Huawei option is just cheaper.
Keep in mind China does the same for US technology companies.
For many installments, the Huawei option is just cheaper.
Keep in mind China does the same for US technology companies.
Companies in the US largely just dont trust them. Huawei used to be on my line card when I was an SE like 7 years ago, and I'd always give people a quote for the huawei stuff, and then they'd always went with the HP stuff at like 1.5X the price.
And Europe.
Sure, there are some Huawei phones and other devices around, but they are not very common.
(mobile phone market share in Europe is ~ 15%, and lower in the western Europe)
Sure, there are some Huawei phones and other devices around, but they are not very common.
(mobile phone market share in Europe is ~ 15%, and lower in the western Europe)
Huawei had nearly 25% of the smartphone market in Europe before US sanctions hit. It was essentially even with Samsung and Apple.
Anybody know why this is showing up with Huawei branding after Eltechs shut down? I presume this must be because ExaGear was acquired, but I cannot find any information online about this, at least in English.
This is really interesting.
"ExaGear is a dynamic binary translation layer to run x86_32 and x86_64 applications on Arm 64-bit Linux systems."
Couldn't this eventually be used in combination with Asahi Linux (project aiming to run Linux on Apple M1)? That would be incredible. Too bad it's from Huawei and doesn't seem to be open source though? I'm unable to find any additional information about this.
"ExaGear is a dynamic binary translation layer to run x86_32 and x86_64 applications on Arm 64-bit Linux systems."
Couldn't this eventually be used in combination with Asahi Linux (project aiming to run Linux on Apple M1)? That would be incredible. Too bad it's from Huawei and doesn't seem to be open source though? I'm unable to find any additional information about this.
That is possible today with qemu/TCG. The assumption with exagear and some of these other JITs is that they are faster.
And they are, Qemu isn't even remotely in the same performance category.
(it was intended to be easily retargetable to other platforms instead of being tailored to them, not even using the hardware FPU at all for example)
(it was intended to be easily retargetable to other platforms instead of being tailored to them, not even using the hardware FPU at all for example)
We do use the host FPU for a subset of floating point operations now. However it only really works for clean 32/64 IEEE FP so anything that goes through the x87 still needs software emulation.
Wow, since which release did that finally happen?
(I might also look at the other issues that I have with TCG code quality sometime)
Might be worth it to update the wiki too.
(I might also look at the other issues that I have with TCG code quality sometime)
Might be worth it to update the wiki too.
Since v4.0.0 (see https://git.qemu.org/?p=qemu.git;a=commitdiff;h=a94b783952cc...). We are always up for improving the code generation quality but of course there is a trade off with JITs given we are not compilers. I suspect there are still big wins if we can come up with a reasonable solution for re-generating a hot-path of basic blocks with much better optimisation.
Thank you. A tiered JIT will indeed cause much more optimisations to be made possible.
I made a Qemu backend targeting LLVM before, but that turned out to be way too heavy to be usable, I wonder if it's worth revisiting that idea nowadays...
I made a Qemu backend targeting LLVM before, but that turned out to be way too heavy to be usable, I wonder if it's worth revisiting that idea nowadays...
I wonder if Huawei plan to open source this? That they make it publicly available makes it seem like they don't plan to monetise it, so the next step would be to open source it, unless patents or third-party code prevent that.
how to get this new EXAHEAR VERSION located here?
https://www.huaweicloud.com/kunpeng/software/exagear.html
i TRYED to contact huawei france support but they asked me to call China support. Tryed to create a china account but it require a China phone line number. Then how to get it?
i TRYED to contact huawei france support but they asked me to call China support. Tryed to create a china account but it require a China phone line number. Then how to get it?
This is the first I'm hearing of a dynamic binary translation tool from Huawei. This link only has a binary download and search results are scant; does anybody know any more about ExaGear?
Project page is here: https://www.huaweicloud.com/kunpeng/software/exagear.html Google (en-US) stupidly ranks it very low when searching for “Huawei exagear”, below some sketchy looking results, possibly due to language?
Documentation is here https://support.huaweicloud.com/ug-exagear-kunpengdevps/kunp... and appear to be Chinese-only. Google Translate does a passable translation though.
Documentation is here https://support.huaweicloud.com/ug-exagear-kunpengdevps/kunp... and appear to be Chinese-only. Google Translate does a passable translation though.
ExaGear was originally developed by a company called Eltechs. I presume this must have been somehow acquired by Huawei.
See my comment here: https://news.ycombinator.com/item?id=25750402
Does it really matter for Linux world where you jest compile software from open sources anyway?
There's plenty of software where the source isn't available, and plenty of open source software where it won't compile for a different architecture without significant work because of runtimes or other dependencies
I don't know whether they've been bought by Huawei, who is now reviving the product under their own name.
Original announcement of discontinuation:
Source: https://web.archive.org/web/20190227095445/http://forum.elte...