Merge pull request #1122 from JuliaGPU/backports-release-3.4

Backports for 3.4.2

Project.toml

@@ -1,6 +1,6 @@
 name = "CUDA"
 uuid = "052768ef-5323-5732-b1bb-66c8b64840ba"
-version = "3.4.1"
+version = "3.4.2"
 
 [deps]
 AbstractFFTs = "621f4979-c628-5d54-868e-fcf4e3e8185c"

lib/utils/cache.jl

@@ -18,7 +18,7 @@ end
 function Base.pop!(f::Function, cache::HandleCache{K,V}, key) where {K,V}
     function check_cache(f::Function=()->nothing)
         try
-            GC.disable_finalizers()
+            GC.enable_finalizers(false)
             lock(cache.lock) do
                 handle = if !haskey(cache.idle_handles, key) || isempty(cache.idle_handles[key])
                     f()
@@ -33,7 +33,7 @@ function Base.pop!(f::Function, cache::HandleCache{K,V}, key) where {K,V}
                 return handle
             end
         finally
-            GC.enable_finalizers()
+            GC.enable_finalizers(true)
         end
     end
 

src/device/utils.jl

@@ -12,7 +12,7 @@ const overrides = Expr[]
 
 macro device_override(ex)
     code = quote
-        $GPUCompiler.@override($method_table, $ex)
+        $GPUCompiler.@override(CUDA.method_table, $ex)
     end
     if isdefined(Base.Experimental, Symbol("@overlay"))
         return esc(code)

src/pool.jl

@@ -527,7 +527,8 @@ Returns the cached amount of memory (in bytes) being held on by the CUDA allocat
 """
 function cached_memory()
   state = active_state()
-  if stream_ordered(state.device)
+  if version() >= v"11.3" && stream_ordered(state.device)
+    # we can only query the memory pool's reserved memory on CUDA 11.3 and later
     pool = memory_pool(state.device)
     Int(attribute(UInt64, pool, MEMPOOL_ATTR_RESERVED_MEM_CURRENT))
   else