Merge pull request #523 from JuliaGPU/tb/threadfence

Add a system-wide thread fence to the exception flag setter.

Author: maleadt

src/device/runtime.jl

@@ -111,6 +111,7 @@ function signal_exception()
     ptr = exception_flag()
     if ptr !== C_NULL
         unsafe_store!(convert(Ptr{Int}, ptr), 1)
+        threadfence_system()
     else
         @cuprintf("""
             WARNING: could not signal exception status to the host, execution will continue.

src/exceptions.jl

@@ -25,7 +25,7 @@ function create_exceptions!(mod::CuModule)
     try
         flag_ptr = CuGlobal{Ptr{Cvoid}}(mod, "exception_flag")
         exception_flag = get!(exception_flags, ctx, Mem.alloc(Mem.Host, sizeof(Int),
-                            Mem.HOSTALLOC_DEVICEMAP))
+                              Mem.HOSTALLOC_DEVICEMAP))
         flag_ptr[] = reinterpret(Ptr{Cvoid}, convert(CuPtr{Cvoid}, exception_flag))
     catch err
         # modules that do not throw exceptions will not contain the indicator flag