You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Duplicate of #839. All back ends now accept or imply always_inline (AMDGPU always inlines, CUDABackend and oneAPIBackend take it as a field); exposing it per @kernel is what #839 proposes.
Activity
lcw commentedon Jun 29, 2022
That would be great. I tried to add it to CUDA.jl and it is not working yet. See JuliaGPU/CUDA.jl#1554 for details.
vchuravy commentedon Sep 23, 2022
Thinking about this today, we have three options:
matinraayai commentedon Sep 26, 2022
@vchuravy I think adding it as a parameter to kernel construction is a good idea.
vchuravy commentedon Sep 26, 2022
@matinraayai would you be interested in trying to add that?
matinraayai commentedon Dec 15, 2022
@lcw it seems that your PR for adding force-inlining support has been merged. If so can we close this issue?
lcw commentedon Dec 16, 2022
I think we still need to add forced inlining to the other GPU backends and then add a kernel construction parameter to kernel abstractions.
matinraayai commentedon Dec 17, 2022
Okay I'll investigate GCN.
maleadt commentedon Oct 19, 2026
Duplicate of #839. All back ends now accept or imply
always_inline(AMDGPU always inlines,CUDABackendandoneAPIBackendtake it as a field); exposing it per@kernelis what #839 proposes.