Skip to content

Temporarily disable cross-device memory access from kernels #3112

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
maleadt merged 2 commits into from
Apr 21, 2026
+11 −5
Merged

Temporarily disable cross-device memory access from kernels #3112

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
844c53f
PerDevice cache: store device's own context, not caller's.
maleadt Apr 21, 2026
08bb31d
Drop cuMemPoolSetAccess from cross-device pointer conversion.
maleadt Apr 21, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
7 changes: 6 additions & 1 deletion CUDACore/lib/cudadrv/state.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -459,7 +459,12 @@ function Base.get!(constructor::F, x::PerDevice{T}, dev::CuDevice) where {F <: B
if y[id] === nothing || (y[id]::Tuple)[1] !== ctx
Base.@lock x.lock begin
if y[id] === nothing || (y[id]::Tuple)[1] !== ctx
y[id] = (context(), constructor())
# store the device's own context (it may be created during `constructor()`),
# so subsequent lookups — which compare against `device_context(id)`, not
# the currently-active context — hit the cache regardless of which context
# was active when the value was constructed.
value = constructor()
y[id] = (context(dev), value)
end
end
end
Expand Down
9 changes: 5 additions & 4 deletions CUDACore/src/memory.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -604,10 +604,11 @@ function Base.convert(::Type{CuPtr{T}}, managed::Managed{M}) where {T,M}
end

# set pool visibility
if stream_ordered(source_device)
pool = pool_create(source_device)
access!(pool, state.device, ACCESS_FLAGS_PROT_READWRITE)
end
# XXX: disabled because of NVIDIA bug #6098762
#if stream_ordered(source_device)
# pool = pool_create(source_device)
# access!(pool, state.device, ACCESS_FLAGS_PROT_READWRITE)
#end
end

# accessing memory on another stream: ensure the data is ready and take ownership
Expand Down