RADV's ACO Back-End Can Be A Massive Win For Vulkan Compute - Not Just Gaming

[phoronix]

Phoronix: RADV's ACO Back-End Can Be A Massive Win For Vulkan Compute - Not Just Gaming

While the Mesa "RADV" Radeon Vulkan driver's "ACO" back-end was developed and funded by Valve with gaming in mind to optimize game load times and help with delivering optimal performance, it turns out ACO works damn well for some Vulkan compute workloads too...

http://www.phoronix.com/scan.php?page=news_item&px=RADV-ACO-NN-Win

[Volta]

They should burn llvm back-end with fire, so no one will have dilemma what to focus on.

[Alexmitter]

They should burn llvm back-end with fire, so no one will have dilemma what to focus on.

LLVM does everything, but nothing well and that is the issue. Who even came up with the idea of using it to compile shader?

[Volta]

LLVM does everything, but nothing well and that is the issue. Who even came up with the idea of using it to compile shader?

I remember someone said it's because llvm is modular. However, it sucks a lot.

[ms178]

I wonder if these results do spur some discussions if ACO is also a better foundation for other compute workloads than LLVM. If that is deemed to be viable ACO could evolve into a full featured compiler for graphics and compute in a few years time, though the ACO developers have focused on other tasks and general compute support is not even on their to-do-list.

[geearf]

Who even came up with the idea of using it to compile shader?

Wasn't it Tom?

[xxmitsu]

Can't wait for clover (Mesa-OpenCL Frontend) to be more mature now that there's a raise in interest lately.. And then to wire up ACO as a backend!

[xxmitsu]

Wasn't it Tom?

Actually, due to the modular nature of the compiler, has been used by multiple parties: Microsoft, nvidia, AMD.

[user1]

They should burn llvm back-end with fire, so no one will have dilemma what to focus on.

AMD might not agree with you here.

But seriously, I do wonder if it is possible to convince them to ditch LLVM and switch to ACO, because I think the main reason they invested so heavily in LLVM is that this way they already have an existing compiler infrastructure, which means lower maintenance. I mean, ACO has recently reached feature parity with LLVM and it performs better in every way.

[microcode]

LLVM does everything, but nothing well and that is the issue. Who even came up with the idea of using it to compile shader?

Funny enough, using it compile shaders was an extremely early interest, back when it was Apple's baby.