Merge pull request #3112 from JuliaGPU/tb/multi

Temporarily disable cross-device memory access from kernels

Author: maleadt

CUDACore/lib/cudadrv/state.jl

@@ -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

CUDACore/src/memory.jl

@@ -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