implement synchronize correctly

Author: vchuravy

src/KernelAbstractions.jl

@@ -105,7 +105,9 @@ end
    @synchronize()
 """
 macro synchronize()
-    @error "@synchronize not captured or used outside @kernel"
+    quote
+        $__synchronize()
+    end
 end
 
 """
@@ -286,6 +288,10 @@ function SharedMemory(::Type{T}, ::Val{Dims}, ::Val{Id}) where {T, Dims, Id}
     throw(MethodError(ScratchArray, (T, Val(Dims), Val(Id))))
 end
 
+function __synchronize()
+    error("@synchronize used outside kernel or not captured")
+end
+
 ###
 # Backends/Implementation
 ###

src/backends/cuda.jl

@@ -214,3 +214,7 @@ end
 @inline function Cassette.overdub(ctx::CUDACtx, ::typeof(Scratchpad), ::Type{T}, ::Val{Dims}) where {T, Dims}
     MArray{__size(Dims), T}(undef)
 end
+
+@inline function Cassette.overdub(ctx::CUDACtx, ::typeof(__synchronize))
+    CUDAnative.sync_threads()
+end