Stable Diffusion on AMD RDNA3
nod.ai
nod.ai
Google has been doing a good job advancing the IREE ML compiler project, which I think is what will bring other hw platforms like AMD and Intel to the ML game. Industry only has to benefit from increased hardware portability.
Versus a polyglot compiler infrastructure, IDE tooling that includes shader debugging, and a rich ecosytem of GPU based libraries.
Even with SYSCL and SPIR-V, that has hardly improved, and while Intel bases oneAPI on top of SYSCL, that naturally also goes beyond the standard.
Behind, it can use AMD and NVIDIA hardware alike. Thus, the idea is that through typically negligible effort porting to HiP, your code becomes vendor-independent.
In practice, I do not know how true this is.
Here, the big AMD mistake was to rename those function prefixes in the first place. It's a mistake that they could have avoided...
What a lot of SW codebases did to support AMD (see PyTorch code notably): codebase is still CUDA, have the conversion pass to HIP done at build time.
See https://github.com/ROCm-Developer-Tools/HIPIFY/blob/amd-stag... for the Perl script to do it.
Then comes the problem of AMD not supporting ROCm HIP on most of their hardware or user base.
On Windows, the ROCm HIP SDK is private and only available under NDA. This means that while you can use Blender w/ HIP on Windows, the Blender builds that you compile yourself will not be able to use ROCm HIP.
On Linux, the supported GPUs are few and far between, Vega20 onwards are supported today. APUs, RDNA1, and lower end RDNA2 w/o unsupported hacks (6700 XT and below) are excluded.
This is sort of echoed in AMD's stance on FSR2/upscaling, where they have explicitly stated they will not support any API that allows plugging, regardless of whether the API is open-source or not, or who owns it, because a pluggable API might allow plugging proprietary implementations. Their opinion is their solution is what's best for everyone, in every situation, so you don't really need DLLs or pluggability because why would you want to plug something worse? FSR2 is the best for everyone and you should really just be compiling it directly into your application.
https://youtu.be/8ve5dDQ6TQE?t=974
(this of course also makes it impossible to update versions of FSR2 if better ones come out subsequently - you can't do the DLSS thing where you swap in newer DLLs (at your own risk of course) and benefit from later improvements to a modular grouping of code. You know, sort of the whole concept of libraries in the first place...)
The HIP stuff is the same thing... AMD really wants you to convert once to HIP and be locked in forever, because Theirs Is The Best, Why Would You Need Anything Else? But of course HIP doesn't have a PTX-like concept so you really need to distribute as source and compile everything at runtime... because who would want library code or dynamic linking?
Anyway, like, I know it's not really a shocker but the "we love open-source!" thing is a bit of an act. They love it when it's an angle for them as the underdog to leverage their way into marketshare... and as the underdog when it's not favorable for them (like FSR) they'll abandon their pro-freeness stance. And they too have their closed, proprietary technologies (like their CXL alternative that only works with their CPU+peripherals and nobody else can use) that they don't open up either. Nor is AMD racing to open up chipsets (like the NForce or Abit days) either, that's all locked down and proprietary too.
I know that's not really a shocker when you put it like that, but, AMD really gets a ton of the benefit-of-the-doubt all the time. They have on multiple occasions shipped defective/marginal silicon at launch for example, and it all just gets brushed over and people forget all about it. Both Zen2 (low-quality silicon in the launch batch meant chips were missing advertised boost clocks by 10%+) and RDNA1 had massive incidents of the community downplaying very real problems because AMD Is Good Now, many of the affected users never had their problems resolved and they just kinda sighed and lived with it or sold the hardware and bought something better, and the fans swept it all under the rug and never talked of it again. Same for pandemic profiteering (while Intel cut prices), etc. There's just a ton of shit that people bend over backwards to find justifications for with AMD that just wouldn't fly with more reputable vendors.
SYCL solves the first issue, and SPIR-V solves the second one. (OpenCL mostly avoids the issue in general though by making you ship source which is then compiled by the driver, but SPIR-V allows you to ship a 'binary' instead).
No clue as for debugging and IDE tooling, but I did find a rocgdb binary on my Linux ROCm installation (which is for HIP, not SYCL). No clue what oneAPI offers for debugging.
Furthermore, Clang (and hence clangd) speaks HIP and I think SYCL too. So the non-runtime IDE tooling should work.
Finally, a lot of GPU libraries are I think available for ROCm/HIP too. It's unfortunate that the HIP stack sucks enormously in other ways.
I hope that now that generative AI is becoming mainstream AMD steps up their game both on their consumer and professional lineups. If I were to buy a video card right now ( mostly for gaming+ML hobbies projects + running stable diffusion) I wouldn't pick AMD because I could do just 1/3 of my use cases properly without headaches (gaming).
For OpenCL: cl_qcom_ml_ops (Qualcomm) notably, for Vulkan: VK_NV_cooperative_matrix (NVIDIA)
AIUI, what's in current git master is very different.
AMD claims to have HIP-RT working internally, but not yet suitable for posting publically. Intel is planning it, I think. Both should land around Blender 3.6, if I'm not mistaken.
If you take the raw FLOPS, CUDA (not OptiX) and HIP are actually nearly equivalent in performance last I remember. I think RDNA2 just does "more with less", at least in terms of gaming performance per FLOP (e.g. due to the huge cache).
Openness is totally secondary to functional. You know, the same kind of reasons as of why Linux on the desktop is not a mass market thing compared to Windows for a very long time.
The AI crowd cared more about that then the impact of tieing the entire ecosystem to a single company.
Who knows what openCL might have been if it would be the premier implementation language. I'd wager it would have gotten a LOT more love.
When you basically have doomed humanity to rely on a single (malicious) company for a technology that is as important as AI. Then maybe, just maybe, the trade off that it is a harder to implement is worth it.
And NV was far from malicious here, they are who made building this ecosystem possible.
Without NV what would have plausibly happened was not having AI training on GPUs at all, but on bespoke accelerators (which _did_ exist back then) at a totally inaccessible cost to customers. It's hard to understate their role in building this ecosystem.
With CUDA having unmatched performance compared to alternatives too.
I don't disagree, but how did this argument fare against Microsoft? Is there a reason you expect it to fare better against Nvidia? That sweaty guy jumping around yelling "Developers! Developers! Developers!" had a point.
I would be fine with CUDA if Nvidia would allow anyone(AMD/Intel) to make implementations for their GPUs as well.
AMD fucked up by not having a stable IR between GPU generations and not having a public Windows SDK. But that's their own problem, not NVIDIA's.
The lack of a stable IR is probably deliberate. Much like the "we won't support DLLs or pluggable APIs, only statically compiling it into your application" with FSR2, once you port to HIP you're locked in. AMD wants you working in HIP, compiling from HIP, not treating them as an IR - they don't want to be an alternate runtime for NVIDIA's ecosystem.
And again, much like FSR2, they are in fact willing to compromise end-user experience (no updates) or developer convenience (continual patching) in order to do it. No libraries, only distribute as source, ever.
It's not about library pluggability or runtime compatibility (after all GPU Ocelot already existed), what they want is you building the ROCm Ecosystem and not the CUDA Ecosystem or OneAPI Ecosystem.
That's understandable from a corporate strategy perspective, as a corporation you don't want to be building a product on someone else's platform, because that gives a lot of freedom for the platform owner to fuck with you. But like, the whole "we won't even do libraries/IR" is a little crass from a customer experience/developer experience perspective, and it kinda goes against the whole good-guy-AMD mythos they've built up.
The problem is that using machine code makes it machine-specific. So if I compile a HIP program for my gfx803 (RX 580) card, I won't be able to run the same binary on someone else's 6800 XT (gfx103x) system. (I think technically you can put both in a single binary, but that's still a terrible solution).
CUDA instead ships NVPTX, which is an IR that can be compiled by the driver to machine code as long as the GPU has the right compute capability, similar to how it works in the graphics world (you submit your GLSL/HLSL source code or SPIR bytecode) to the driver which compiles it for the right GPU.
Intel's oneAPI/Level Zero API ships SPIR-V, afaik (or maybe regular SPIR?). oneAPI can also work on top of OpenCL instead of L0. SPIR-V is neat because it's an open standard, so in theory you can get L0 working on non-Intel GPUs (and iirc Intel also uses it for e.g. FPGA's). But both SPIR-V and NVPTX solve the "machine-specific" problem AMD has.
> (I think they even have https://gpuopen.com/orochi/ which explicitly allows compiling a single binary for both ROCm and CUDA).
Orochi sidesteps this problem... by only supporting NVRTC-style runtime compilation with C++ as input.
And even then, the HIP C++ compiler library is bundled as part of Orochi instead of being part of the app. This means that your app using Orochi will not run on a future GPU gen unless it's updated against a newer Orochi runtime.
Ugh. Leave it to AMD to make something that technically works but is an absolute nightmare.
IIRC this machine code nonsense is also the reason that GPU support is such an issue for AMD: to 'support' a chip, they need to bake binaries for that chip in all libraries. So to enable RDNA1, they'd need to ship RDNA1 code in all their libraries, which would make the install size balloon to crazy levels. At least Intel got it right.
I do believe that running oneAPI on AMD is possible, but it still needs HIP/ROCm? Wonder if it would be possible to bake a L0 backend for AMD that just uses SPIR-V like the Intel stuff does, side-stepping this issue entirely.
Frankly I wish AMD and Intel just started working together more on this stuff. Both of them stand to gain from a cross-vendor standard that works well.
RDNA1? more like 3 binary slices. Navi10 (5700 XT), Navi12 (AWS G4ad) and Navi14 (5500 XT) require separate binaries!
> I do believe that running oneAPI on AMD is possible, but it still needs HIP/ROCm?
Yes, HIP runtimes for AMD GPUs rely on an underlying HIP implementation.
> Wonder if it would be possible to bake a L0 backend for AMD
Yes. But why would anybody not named AMD do that? It's AMD's hardware so AMD has to support it. OSS/hobbyists can only do so much.
> Frankly I wish AMD and Intel just started working together more on this stuff
Why? AMD truly does not care about GPGPU APIs for the masses. For their management it's a useless additional expense so they haven't been doing it.
A chunk of the community has wanted to consider AMD as an NV alternative for this, but AMD are not selling the same product. They think that their gaming GPU line is gaming centred w/ often bare minimum support for other markets if any, while NV cares about a much wider audience.
That's how the market ended up with: Q3 2022 Discrete GPU Market Share Report: NVIDIA Gains 88% Market Share Hold, AMD Now at 8% Followed By Intel at 4%.
AMD only officially supports GPGPU headless. That discounts 90% of the market. Old graphics cards lose support randomly. That discounts much of the rest. The whole thing is a horrible, bug-ridden mess.
I'd pick AMD over NVidia if it was e.g. 50% slower at the same price point -- open source is worth waiting for -- but I can't take nonworking.
AMD also has no support. I'm now building tooling reliant on NVidia, so if AMD ever gets their stuff working, we're many backports away from a working ecosystem. The longer AMD takes, the deeper the hole.
The fact that it’s not functional on non-NVIDIA platforms is the whole reason Blender dropped OpenCL support. If you’re going to write a bunch of implementation-specific code to handle AMD’s bugs/non-compliant runtime, why not just target CUDA directly?
“it’s open-source, if you want it fixed then pay someone to do it or just spend a month making a patch instead of doing your work???”
bugs-bunny-no.gif
like, same with ROCm, is AMD just wants to externalize all their costs onto customers yet still wants them to adopt it instead of the turnkey solution that everyone else already uses. Why would anyone do that? It’s great for NVIDIA but terrible for users.
(and as for the “AMD drivers have been good for like ten years now!!!” crowd… counterpoint: the entire 5700XT thing, drivers broken for the first 18 months, just like Vega before it. And oh look 7900XTX is turning into a trainwreck too. There’s just constant showstopping bugs with AMD drivers. Just like with ROCm too… patchwork support and endless bugs that don’t exist in the industry-standard solution. Nobody wants to spend their time doing AMD’s job for them.)
To their credit this is one thing Intel got right… they probably spent more dev time on oneAPI in the last year than AMD spent on ROCm and all their previous attempts/projects/resume-driven-development fodder combined.
Vulkan is a primarily for graphics but does have options for GPGPU too. Vulkan is however not like OpenGL in that it's fairly close to the hardware in terms of abstraction.
The chance of it winning over CUDA is at zero. And that's _before_ considering its API gaps compared to modern OpenCL.
(Yes, even OpenCL is a much better compute API choice than Vulkan. Vulkan does not even have SVM)
It's not a niche feature at all, but one that is essential to lower the barrier for developer adoption.
Vulkan is well supported by most GPUs because it's so low level. Performance tends to be good everywhere.
What would make vulkan succesful is having APIs that "compile" to this IR. Stuff like vulkan Kompute are good ideas in this direction.
CUDA is only for general purpose compute.
As an aside, I’ve been kinda surprised that this has existed for as long as it has, but I am probably biased and think Ml acceleration is more important than most large business do today.
There's a lot of established ecosystem for CUDA, thanks to Nvidia's investment.
This is incorrect. Those optimizations do identical computations, but leverage memory bandwidth on the gpu more effectively. So there is no accuracy tradeoff there.
That said we (Nod.ai team) will add support for xformers soon so you can opt in for xformers anyway.
[1] https://github.com/xuhuisheng/rocm-build [2] https://github.com/RadeonOpenCompute/ROCm/issues/1587
For them, GPGPU is a pro level feature not worth supporting on most customer GPUs. They are doing much more feature segmentation than NVIDIA ever did.
https://www.travelneil.com/stable-diffusion-windows-amd.html
For 50 iterations:
* ONNX on Windows was 4-5 minutes
* ROCm on Arch Linux was ~2.5 minutes
* SHARK on Windows is ~30 seconds
I wasn't aware that Flash Attention trades accuracy for performance. Either I have a wrong understanding of what FA is, or this statement is not fully accurate.
Either way - looks like great work
We also extend FlashAttention to block-sparse attention, yielding an approximate attention algorithm that is faster than any existing approximate attention method.
So I assume they are using the approximate version as they also have an exact version.