Missing warp reduce functions in HIP

[Epliz]

Hi,

Warp reduce functions are available in CUDA (cf. https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#warp-reduce-functions) but not in HIP.

There is equivalent functionality in HLSL shader model 6.0 supported by all AMD GPUs after the Vega architecture as far as I know: https://github.com/microsoft/DirectXShaderCompiler/wiki/Wave-Intrinsics .

So I think it can be implemented for HIP as well.

Could you please add them?

Best regards,
Epliz

[Epliz]

@cjatin , you were kind enough to have a look at #2 ; this one is quite similar, so I would appreciate if you could have a look at this one as well :) .

[cjatin]

Hello,
Will raise an internal issue to track this and #9

[Epliz]

@ppanchad-amd , any update on this ticket?

[lucbruni-amd]

Hi @Epliz,

These functions have been implemented on our side and the changes are actively being tested and reviewed. I cannot promise an exact release version target, but I will keep watch on its progress and provide you with updates along the way.

Please let me know if you have any additional questions or concerns. Thanks!

[lucbruni-amd]

The current target release for these features is ROCm 6.5 - please look out for that.

If anything changes about that, I will update this issue. Thanks!

[Epliz]

Thank you for the update @lucbruni-amd !

[g-h-c]

@Epliz The reduce operations will be present in 7.0. They will try to make use of DPP (Data-Parallel Primitives) instructions which will make them quite fast, although whether DPP will be used or not will depend on the mask used. Stay tuned ☺

[Epliz]

Great :)

[lucbruni-amd]

@Epliz The reduce operations will be present in 7.0. They will try to make use of DPP (Data-Parallel Primitives) instructions which will make them quite fast, although whether DPP will be used or not will depend on the mask used. Stay tuned ☺

As these reduce functions are set for release in ROCm 7.0 (see ROCm 7.0 alpha release notes and warp_reduce.cc for details), I am closing this issue as resolved.

If you have additional questions/concerns upon release of ROCm 7.0, feel free to reopen this issue or open a new one. Many thanks!