Code rollback

[vosen]

I understand that the topic might generate strong opinions, but please refrain from:

• using profanities
• posting the links to forks or other means for getting the code

[ethanc8]

Are you able to share the email that AMD originally sent you giving you permission to release ZLUDA 3?

[erkinalp]

@vosen If you manage to formally restore your permission to release this code, is it just a rollback away, or do you have further changes you privately made?

[sempervictus]

Sorry, but the sender is not allowing me to share the email publicly

Could you please share the spam filter logs or MTA logs from the third parties which do have copies since those are not the original correspondence but intermediate provider data about the emails sent and received between the TO/FROM fields in the headers? When i dealt with the prior incarnation of the "graphics legal team" nearly 30y ago, they had not figured out email encryption to avoid that loss of confidentiality and something tells me that they still have not given how they're handling this.

Any party with access to your mailbox is also not beholden to the terms of the work-for-hire contract under which the compensation for this was presumably structured - quite likely that one or two tech-pub journalists can be found who can be trusted not to abuse access to the mailbox(es) containing this thread and incidentally may find a very wide and animated audience for the contents of that specific thread. Same goes for backup providers, IT people, etc - unless the contract stipulated TS-grade security measures for work product, communications, and systems; the other party cannot have any expectation of confidentiality given their own well-publicized abysmal security practices (and products) despite the immense budget and talent pool available at their level.

[RealVarHar]

The worst thing is, nvidia will not legally shut this down, because of a risk of a really mean antitrust legislation.
But since its amd shooting its own foot this project will likely not survive this setback.

[Elise96nl]

you guys dont get it. No AMD is not interested in zluda all of a sudden. Zluda makes us use cuda. even if @vosen made a perfect piece of software. it's using nvidia trash. They want their own cuda. They thought zluda was okish, but not ready for real world use for guys all around the world making nsfw ai content. So it was harmless.

Now they see it works pretty great, with some tweaks. and even me as a non programmer can use it for stable diffusion. So they want it gone. They want to release their own alternative. And for you and me, to early adopt their thing. So they could start competing with nvidia and their golden boy CUDA.

So zluda release was ok if it was shit, now it's pretty decent they dont want us to use it.

[TinfoilyTim]

Why wouldn't amd want the code to be out tho? I know they are working on their own thing but this closes the gap just a little bit hardware wise. Nvidia have already been in trouble because of Cuda so they most likely won't go after amd, so I don't see a reason they need to take this down.

[powerpuffboysz]

Is gamework support completely shot or will it just take a long time to rebuild?

[vosen]

Well, if I could do this, then someone sufficiently motivated could do this too. The problem is that any org that is interested in funding ZLUDA is not interested in funding it for GameWorks

[Koth-87]

Well, if I could do this, then someone sufficiently motivated could do this too. The problem is that any org that is interested in funding ZLUDA is not interested in funding it for GameWorks

Is it possible to crowdfund you to rebuild it? I know that might be asking a lot...

[vosen]

It's not the matter of money, but rather time. I need to prioritise time spent on things for which there's a business case. Doesn't mean I can't take a quick detour to enable an easy workload, but GameWorks is neither quick nor easy

[qwertychouskie]

Would be nice to see CUDA implemented on top of Vulkan Compute so it's truly cross-platform and cross-vendor, though I'm not sure what the logistics are there.

[vosen]

Vulkan Compute is missing features. More detailed answer here: #115 (comment)

[mirh]

SYCL is already aspirationally meant to be "truly cross-platform and cross-vendor".

Besides, if you mentioned Vulkan because your purpose was gaming-related (physx) that would still be pretty hard considering everything is win32 based. But that's no biggie, really, considering wine-staging can pass through just about any call that is available to the native system (ironically including CUDA itself btw). So there's no problem here.

[vosen]

I wrote tens of thousand lines of SYCL code and my opinion is that it's not going to challenge CUDA at all.
Wine can pass calls, but for gaming specificially it might be more complex (depending if it's original PhysX, GameWorks or DLSS). In the case of gaming, most of the time you rely on CUDA <-> DX interop. To implement this on AMD HW you need to have HIP <-> DX interop too. Interop boils down to HIP having a knowledge of internal layout of various DX data structures and understanding its internal operations to inject its own data/operations.
HIP <-> DX works (just barely) on Windows, but on Linux HIP <-> VK is not going to be very useful, because AFAIK everyone uses Mesa and not AMDVLK. And the fact that you are going throught adds another layer (because you don't want to interop with DX object but underlying Vulkan one).

[Koth-87]

I am absolutely heartbroken :(

I know this probably trivial compared to what other people were gonna use it for, but Arkham Knight is my favorite game and I was dying to try this out. I even commented on it a few weeks ago: #31 (comment)

Truly a sad day :(

[niansa]

How can we donate to show our support as simple users?

[teamolhuang]

Really needs a mean of donation for us to support ZLUDA. I will not choose AMD for my future devices/replacements again since this is such an unacceptable move, pulling the rug under the creator worked for them.
But before my current card dies, I still need to generate my anime wife images, and your work has always helped me a lot. Please let us support your magnificent work.

[vosen]

Thanks, I've been in contact with other, longer-running open source projects of similar size and my conclusion is that the best way forward is funding by an organization. Which I'm in the process of finalizing, so you can buy yourself a daki or a fig

[BaalNetbek]

That was awesome run. I hope this hit doesn't set you down for long.

[sempervictus]

If anyone from AMD is reading this thread, i am RageLtMan of the ancient Rage3D ATI driver project.
I told ATI's people at the turn of the century that if they did not implement unified drivers (thankfully that went though but not for the codebase we were manhandling) and a language abstraction for GPGPU good enough to beat the competition, then Nvidia would eat ATI's lunch. They did, then AMD ate the starved to death carcass.
Please stop making ATIs mistakes again from beyond its grave and embrace the risks of battling for market share especially when you wage that battle by FOSS proxy; because your hardware quality has no bearing on adoption if CSPs and AI/HPC scalers won't buy it (as their users expect CUDA, various NV interfaces, etc). Its more cost effective to play the legal game if it comes to that and spend 0.1% of the profit to be gained in the HPC/AI market dragging out the process while gaining market share than to keep "playing it safe" (from a 2nd place position so far behind 1st that its barely relevant). The legal department is rarely the one to effect market share acquisition... engineers who take calculated risks are better decision makers in a dynamic market than people prescriptively fearful of the consequences whether they fail or succeed. Source: these days, run a consultancy building/securing/orchestrating these fabrics within the scalers and their HPC peers where AMD kit is almost nowhere to be seen. This market needs legitimate competition to ensure ongoing innovation of capability vs mere increases in complexity and throughput of packaging.

[erkinalp]

@saadrahim

[sempervictus]

@saadrahim

I'm happy to take my case all the way to Dr Su if she can find the time and bandwidth to speak.

[roastedpeacock]

Thank you for developing ZLUDA!

I think it is unfortunate this project had to be rolled back. Understood if you cannot comment but I would push back harder on censorship of this project. I hope a correct decision will be made eventually.

In any case continuing research into vendor-agnostic GPU computation is great. However I fear any further effort on your part would be on an unequal footing considering the existing code is already out there and a potential fork by third-parties is theoretically possible given enough skill and time. Not going to go into specifics however previous open-source projects that have ran into questionable legal territory have been forked and continued by unaffiliated developers.

[wiki-me]

Did you talk to the organisation that wants to sponsor the project? writing in a contract you can open the code and saying its not legally binding sounds like a slam dunk case to me. They might have a lawyer on retainer so the cost to them is basically zero and they might be able to do a better risk benefit analysis because they are lawyers.

[vosen]

Yes, they are aware of all the details. I talked to a lawyer. It's not a mega-corporation, the costs would really hurt

[vosen]

And one more thing: it was not written explictly into the contract, but as a separate email. I know that the previous README mentioned it being part of the contract, but I was trying to write a good explanation without being pedantic

[wiki-me]

oh well, at least we learnt a lesson i guess, one of the conditions for having an email become legally binding is apparently "A clear intention to be legally bound by the contract". I just hope with the new organisation the agreement to open source is in the contract.

[erkinalp]

next time have your employment contract notarised

[dbermond]

The latest version is still being shown as v3, which contains the AMD code published some months ago.

Shouldn't the latest version also be reverted back to v2? Will this AMD code, which is still present in the online repository history, be somehow rebased/removed?

[roastedpeacock]

The newest commit in my checkout is below

9e56862 (Update README with CGBN information (#237), 2024-05-23)

[vosen]

Hmmm, I'll leave the binary for now. I've got the request to remove the source code. Also please don't "post the links to forks or other means for getting the code"

[erkinalp]

in github's perspective, forks are branches of the same repository, not separate remotes like how forks work on other SCMs
so you should request a detachment or a takedown of other forks to truly break the connection

[vosen]

Yes, but I've been asked to rollback the code in the repo, not take down every trace of the project and its code from the internet. Let's leave it be

[erkinalp]

then, a detachment (email github support to request all forks of ZLUDA to be detached from you), so you won't be seen as the perpertrator in github's eyes

[phly95]

I guess I won't be buying AMD for my next PC. To not be able to games from 2008 (eg. Mirror's Edge) at max settings on a 2024 card is getting ridiculous. People do play old games, and will still choose Nvidia if it means they can play those old games at max settings.

Congrats AMD, you just cost yourselves billions of dollars in lost sales. (BTW, I have never bought AMD because of that slowdown in Mirror's Edge that happened when I was a kid and will not until it is fixed)

In fact, this is probably what almost drove AMD bankrupt until they were saved by the game consoles and then Ryzen. If AMD supported PhysX and Gameworks features, people wouldn't think of AMD as the scraps you buy because Nvidia is too expensive.

[oleid]

To not be able to games from 2008 (eg. Mirror's Edge) at max settings on a 2024 card is getting ridiculous

You mean because of PhysX? Other than that, I don't see why the game shouldn't run fast on today's GPUs.
It worked on mobile GPU 11 years ago: https://www.youtube.com/watch?v=scg9HQHwgpI

[phly95]

Yes, because of PhysX. It runs fine at first, but if you have AMD, once the glass is smashed, your frame rate is tanking. And ofc, you aren't running at max settings unless you've enabled PhysX. I have no idea if any other games have this kind of issue. Yes, Nvidia is to blame for the issue (limiting non-GPU PhysX to a single CPU core is clearly anti-competitive, given a modern CPU can easily run PhysX if that limitation did not exist), but AMD is actively preventing a fix now. That game is certainly not getting a PhysX update now unless there's a remaster.

[vosen]

I'm sad about GameWorks being gone too.
I can't comment on the Mirror's Edge CPU GameWorks, but my experience from enabling CUDA GameWorks is that it does some nonsensical things that are buggy, but just happen to work

[marcussacana]

So the project keep alive, you will just rewrite it...?
how much time you think will be needed to be at the same stability level of the current legacy version?

[AssassinK786]

This is Better option for everyone
Just ask for some extension @vosen !

[vosen]

Project is alive. you can see progress of the first big PR here: #267
I'm aiming for one year for the same approximate level of functionality. Meaning applications that work right now will work in as well (as long as thay are in the scope)

[TinfoilyTim]

Will we always be able to download the pre-takedown release?

[erkinalp]

there are forks we're not supposed to name here, feel free to look at them

[Elise96nl]

This is pretty standard practice, right? Even if it works great, CUDA is NVIDIA's golden ticket, and they don’t want that to change. The fact that you and I can make Stable Diffusion and some other AI programs work almost as well on AMD GPUs, which cost half as much as NVIDIA's, is great for us, but not for AMD. We are still using something that hooks into the competitor's golden ticket. Eventually, AMD will want to have their own CUDA equivalent. If they have something like that in the works and it's almost ready, they need to steer us away from ZLUDA and get early adopters to jump on their new solution for it to take off. They might try to compete with CUDA and NVIDIA.

So, even though @vosen did an amazing job with ZLUDA, I think it was always Plan B to have a way of running CUDA on AMD. Even if @vosen made the impossible work and ZLUDA would be complete and run perfectly, AMD would stil choose Plan A over Plan B. Plan A involves having their own CUDA alternative that they can push, own, and require their hardware for, which means more money. Companies always choose Plan A, even if it’s worse, flawed, or completely stupid.

[phly95]

Would it be possible to open source the Gameworks support code, even if it doesn't do anything useful or is completely broken? It might be useful for other developers interested in studying how Gameworks works, since it doesn't seem working Gameworks support will be a priority for a long time coming.
It shouldn't be a problem if it doesn't include the AMD code needed for it to work, given the Gameworks code was made after leaving AMD. Perhaps it could be used to make a non-CUDA Gameworks framework in Vulkan or something.

[dobi70]

Hi. I might have a stupid question but maybe vosen or someone else can help/answer it.
First things first. I do not game at all. I use my PC for video editing, maybe colorizations with pixbim in the future....
I just returned a Rx 7600 xt after finding out that zluda was removed from here. I am an Intel CPU guy but always an ATI/AMD GPU person.
Now here is my question. I don't have a golden goose.
Should I get a Rx 7700 xt or a Rtx 4600 ti 16 gb.
I don't know when zluda or anything like it will come back that can run cuda on AMD.
Is there something else that can efficiently run cuda on AMD? When will zluda return?
Days? Months? Years? Or never?
What is AMD up to, to request zluda be taken down? I hate to blow money out on something I'll regret later, but I also don't want to support Nvidia's monopoly game.
Any input would be appreciated

[marcussacana]

Him said that his objective is to get zluda in the same level of the current version with one year.
If you have money to spent should just get the overpriced RTXs, otherwise AMD.

[nitefood]

I may definitely be way off in my understanding of what's at stake here, and in that case, I apologize in advance. I am not trying to downplay the importance of what's going on here.

But have you considered taking the repo private and switching to a binary only release model? I know this means effectively turning to the dark side of closed source, but then again perhaps there could be a solution to grant access to the repository to contributors that are willing to sign an NDA, or something?

I mean if what they asked of you was to make the source code repository (and nothing else) not publicly available anymore, one may speculate that once you make it private, and cherry pick who is allowed access to it, all the while keeping the binaries up, it could be enough to make them happy...

[mirh]

Have you considered checking the sources and noting that 15K lines of code were merged just last week?
No looking back now.

[vosen]

Keeping the source closed would not help. AMD position is that it's their intellectual property, taking it closed solurce would not resolve that claim. It's better to roll back than to fight it

[nitefood]

Keeping the source closed would not help. AMD position is that it's their intellectual property, taking it closed solurce would not resolve that claim. It's better to roll back than to fight it

I see your point. It's always sad when open and brilliant software has to get censored because a mega corporation can't take the right stance and support the community, let alone listen to it. Best of luck for your future work.

[erkinalp]

meanwhile the largest fork quietly adding ROCm 6 support (not going to link to them, but you can easily find them using github search)

[allzhukov]

How to run zluda on intel?

[Salleyi75]

Hello, this is not related to the "code rollback" topic, i just wanted to ask this question since you are an expert and happened to create Zluda. For the past 2 months i have been wondering what exactly is making people say Nvidia is better than Amd for ai (i am contemplating on buying some hardware to run local ai, plus am a programmer so i don't mind error messages here and there), my initial assumption was that Nvidia has some special hardware that works for ai that amd lacks, so i delved a bit into cuda programming and hip programming, i mean they seem to achieve the same goals in different ways, but they seem equal in my eyes, what exactly is AMD missing, looking at the fact that Zluda exists, tells me that amd hardware is good enough because if the hardware was not good enough i dont think there will be a point in creating zluda, so my final conclusion is, people are used to programming cuda than hip/rocm. Kind of link the ps3 cell processor situation repeating itself where people were not familiar with the style of programming on the cell processor hence they did not really want to make games for the ps3 and if they did, it was badly optimized, is that what is happening to amd ?

[vosen]

Consider this: all major desktop GPU vendors share lineage: sure they are used for compute now, but they evolved to serve gaming APIs. As such, most of their capabilities overlap. We see more and more compute-only hardware features and removal of certain graphics HW from compute cards, but the baseline is very very similar. They program similarly. Even NV and AMD ISAs are similar (Intel is a bit of an outlier).
I think the NVIDIA is the first to add new HW features. They range from genuinly useful (various ML low precision instruction) to harmful (independent thread scheduling). New HW features sometimes appear in Intel/AMD GPUs bit later, sometimes not at all, sometimes earlier, but if you are just starting GPU programming they will be quite esoteric.

The big divide is software. For NV the whole stack (compiler, runtime, profiler, libraries, debugger, etc). mostly works and works as you'd expect it to. You've asked specifically about AMD, so I'll talk specifically about AMD. They had the right idea: provide a GPU programming language and libraries that are sufficiently similar to CUDA that you can just rename the functions and have yur program "just work". The problem is the execution of that idea. The list of issues with their SW is very long. Let's talk just creme de la creme. When you are compiling a HIP program for an AMD GPU you have no forward-compatiblity. Meaning, if you compile a HIP program today, it will not work with the next generation of GPUs. HIP program written today will not work on AMD Radeon RX 8XXX. This is the sort of problem that does not happen anywhere else. Not on NV GPUs (compiles to PTX, which is a forward-compatible IR), not on Intel GPUs (compiles to SPIR-V, which is a forward-compatible IR), not on CPUs (new CPU generations are backward-compatible on the ISA level).

That's not all. When you compile for AMD GPU, you have to compile for each card model separately. For example RDNA 2 consists of gfx1030, gfx1031, gfx1032, gfx1033, gfx1034, gfx1035, gfx1036 (see here: https://llvm.org/docs/AMDGPUUsage.html). If you compiled just after RX 6900 XT (gfx1030) released, remember to recompile when RX 6700 XT (gfx1031) releases! And again when all those other gfx103* drop!

There is some progress with that particular issue (https://llvm.org/docs/AMDGPUUsage.html see "Generic Processors", https://www.phoronix.com/news/LLVM-AMDGCN-Flavored-SPIR-V), but I don't think it filtered already to the release stack. And it's just one of many, many issues.

[Salleyi75]

Oh ok, i see, makes sense, so i guess one of the issues is interoperability with their own hardware. Ok, thanks for the explanation, i wanted to decide between the two vendors before investing anytime. As a programmer I don't mind dealing with issues on the C and C++ level, like pointers, memory leaks, and allocation issues, as much as I am brave enough to tackle those issues, I am currently not brave enough to deal with LLVM, tried learning it in other to help with the open source zig programming language, but I gave up, maybe next year. Thanks for the insight at least I have learnt about the hardware enough to understand why ISA changes can be an issue. I think i will be able to further delve into this when I understand the LLVM IR.

[Quackdoc]

Is there a "current state" of the project? I am a bit confused on where exactly the project is compared to the past.

[marcussacana]

Him said it may take at least one year to the project be at the same level of the legacy version, it was just just 2 months since then, you will need to be patient

[vosen]

I plan to have an announcement about current state and project goals next week. Stay tuned.

[vosen]

Announcement here: https://vosen.github.io/ZLUDA/blog/zludas-third-life/