Other backends

[Zaneham]

I've purposefully made the architecture of this compiler 'architecture agnostic' as much as humanely possible. If you'd like to see something supported I am more than happy to help. Just reach out here (if anyone knows where I can test this stuff or if anyone willing to use their own hardware let me know :-) )

[octaveoclx]

What about apple silicon? I am happy to help by trying on my devices after my busy next month.

[octaveoclx]

Also Arm devices like Arm Mali or Qualcomm Adreno? These can be tested on Termux.

[Jakdaw]

Is RDNA 2 out of the question? - still a lot of devices out there....

[Zaneham]

Hello @Jakdaw no it isn't! I was planning on allowing for a support on a broad range of GPU's. I am just one guy atm so Ill get onto it.

[Zaneham]

Hello @Jakdaw RDNA 2 support I've just released. If theres any issues please let me know.

[XiongjieDai]

Hi @Zaneham — great work!

If you’re considering additional backends, we can help test CDNA2/CDNA3 (MI250 + MI300). We have access to the hardware and can run tests + share results/logs. Let me know what you’d like us to try.

[Zaneham]

@XiongjieDai

Thank you! I'm actually reading up on the CDNA instruction set. Heads a bit full of ISA docs bit this is something I am interested in doing.

At a glance there are some key differences between RDNA and CDNA so I'm still in the research and design phase whereas I looked at RDNA before I released BarraCUDA. I'm just wondering what sort of workloads are you running?

[XiongjieDai]

@Zaneham

Thanks for looking into CDNA!

On our side, the main workload is LLM inference (Transformer decode + prefill). Concretely, we care about being able to run (or port with minimal changes) the kinds of CUDA kernels commonly used in LLM inference on AMD CDNA GPUs (MI250/MI300), so people who already have CUDA kernel expertise can keep iterating without needing to become ROCm specialists first.

Typical hot spots for us are:

• GEMM / matmul (FP16/BF16, sometimes INT8/INT4 paths)
• Attention (incl. KV-cache / paged attention style kernels)
• Fused kernels like RMSNorm/LayerNorm + bias/activation, softmax, reductions
• Memory-bound ops around KV-cache reads/writes and layout transforms

If it helps your design, we can test some representative kernels on MI250 + MI300, then report back with results and logs.

[Zaneham]

@XiongjieDai

Hey this is actually super helpful for what to prioritise. As of writing BarraCUDA covers fp32 fully and fp16 via __half. The fused ops should be able to work today using RDNA using warp shuffles +hardware exp2.

For the rest I'll have to add these, but this is actually really good to know and its achievable. Once I'm chucking out CDNA binaries I'll reach out, if you have a simple GEMM + Softmax kernel that would be a great first test target.

[Zaneham]

What about apple silicon? I am happy to help by trying on my devices after my busy next month.

I just thought I'd comment on this. I really appreciate the offer but the problem with Apple's GPU's is it's completely proprietary. The Reason why I targeted AMD and Tenstorrent is because they keep their ISA docs open. If a company wants to collaborate on this compiler I am open to that but other than that I can't do much except maybe reverse engineer their GPU instructions which would be a mammoth undertaking and break on the next macbook release.

[octaveoclx]

Interesting point. I am just wondering how they did it on OpenCL and GL. https://liliputing.com/asahi-linux-brings-support-for-aaa-gaming-to-apple-silicon-macs-running-linux/ Get BlueMail<https://bluemail.me> for Desktop On February 26, 2026, at 20:42, Zane Hambly ***@***.******@***.***>> wrote: [https://avatars.githubusercontent.com/u/222171382?s=20&v=4]Zaneham left a comment (Zaneham/BarraCUDA#16)<#16 (comment)> What about apple silicon? I am happy to help by trying on my devices after my busy next month. I just thought I'd comment on this. I really appreciate the offer but the problem with Apple's GPU's is it's completely proprietary. The Reason why I targeted AMD and Tenstorrent is because they keep their ISA docs open. If a company wants to collaborate on this compiler I am open to that but other than that I can't do much except maybe reverse engineer their GPU instructions which would be a mammoth undertaking and break on the next macbook release. — Reply to this email directly, view it on GitHub<#16 (comment)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/BUUCY7LKZQAGYYFXFE2C4I34N3S4NAVCNFSM6AAAAACVO5E74GVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZTSNRWGM4DOMRTGA>. You are receiving this because you commented.Message ID: ***@***.***>