Radeon Linux Driver Seeing "MALL" Feature For Big Navi

[mdedetrich]

NO! Go! Go is the perfect rewrite everything in it bubble hype language!

Jokes aside, as long as you cant compile Rust to proper bytecode aka use GCC to compile it, its a worthless language for systems. Keep that LLVM trash far far away from anything that needs to be reliable, like a kernel driver.

Jokes aside, seems like Linux devs disagree and are open to trying out Rust within the Kernel (initially for drivers actually)

[reavertm]

They should rewrite it in Rust.

Rust is a measureable waste of electricity compared to C:
https://thenewstack.io/which-program...FGfCSrh9yKrhBo

[Ironmask]

Rust is a measureable waste of electricity compared to C:
https://thenewstack.io/which-program...FGfCSrh9yKrhBo

Just rewrite the laws of physics in Rust.

[pal666]

A core part of Rust is that you dont' have to rewrite you can migrate.... and this is exactly what Firefox did, it has Rust and C / C++ code coexisting.

firefox had c++ and javascript code coexisting for decades. coexisting is very low bar, c++ can include c headers.

[cb88]

firefox had c++ and javascript code coexisting for decades. coexisting is very low bar, c++ can include c headers.

Not in the same way... not even remotely the same thing.

[intelfx]

Jokes aside, as long as you cant compile Rust to proper bytecode aka use GCC to compile it, its a worthless language for systems.

May I ask what exactly is "proper bytecode" and how is GCC even related to "bytecode"?

Keep that LLVM trash far far away from anything that needs to be reliable, like a kernel driver.

Where did you get this nonsensical idea?

[oiaohm]

Nice, now implement FLR so that AMD GPU's aren't such a joke when it comes to virtualization/crashes leaving the GPU in an unfixable state.

Not 100 percent unfixable. Problem is the way you dig AMD card out of a stuck state is different to a Nvidia one.

https://forum.level1techs.com/t/navi-reset-kernel-patch/147547/47

This is where not 100 percent unfix-able appears.

Yes AMD cards have something different a BACO reset path. BACO reset is basicly does the same as Nvidia FLR just not hooked up to the FLR interface.

https://xdc2020.x.org/event/9/contributions/633/attachments/731/1336/Dont_Bake_Your_Gfx_Cards_.pdf

Interesting read of page 8 of that pdf. The modern AMD GPUs have 5 different reset functions. Yes BACO reset is part of mainline AMDGPU driver now.. Yes you can bring a VM crashed GPU back from screwed by loading up a Linux kernel in a VM set to do AMDGPU reset with BACO on the card.

Of course I would love to be able todo this main kernel by say modload amdgpu with command line option to BACO reset the card then let go of the card so it can go back to VM.

Yes big Navi support BACO reset.

So there is 2 problems here.
1) mdedetrich you and others are saying AMD has to implement function level reset because that is what Nvidia does. AMD is not Nvidia so does not have do it the same way. There can be odd things in legals with patents and the like that may forbid implementing something the same way. So you need to look if something equal is already implemented. That what I did when I hit the same problem and found the BACO reset. Nvidia cards don't have a BACO reset that use FLR for that.

2) AMD BACO reset is not as user-friendly as it could be for VM users. See you are not asking for AMDGPU reset options to be made simpler to use for VM users instead you are being like a brat throwing a tantrum to get your own way. If in the process of making BACO reset more usable to VM users they implement Function Level Reset that will be nice. The good part is asking for AMD to make BACO reset more user friendly for VM users will not be asking anything that could be patent infringement or anything else as they have already implemented BACO reset.

Remember with BACO reset it would in theory be possible to extract more information out card for why it locked up than a Nvidia PCIe FLR allows.

[mdedetrich]

1) mdedetrich you and others are saying AMD has to implement function level reset because that is what Nvidia does. AMD is not Nvidia so does not have do it the same way.

You do realize that FLR (Function Level Reset) is an optional part of the PCIe standard, i.e. its nothing specific to NVidia.

AMD is the one that is doing something non standard here/specific to their architecture here, not NVidia.

Also even with all of the workarounds existing (Linux or otherwise), the story for AMD currently when it comes to virtualization is kinda pathetic, both on their CPU's and GPU's. This is yet another reason why so many people are going to NVidia (at least for the GPU side)

[microcode]

Cool. Weird name but I guess this is about accessing memory directly from an L3 or equivalent cache, so you don't, thrash your L1 (or equivalent). If the throughput is sufficient for the specific operation (essentially memory-to-memory or close), the latency will be hidden anyway.

[pal666]

Not in the same way... not even remotely the same thing.

that's exactly my point. rust can't include c++ headers