Simplify CUDAKernels.

Author: maleadt

src/CUDAKernels.jl

@@ -2,6 +2,7 @@ module CUDAKernels
 
 import KernelAbstractions
 import CUDA
+import CUDA: @device_override
 import UnsafeAtomicsLLVM
 import GPUCompiler
 
@@ -147,34 +148,6 @@ function (obj::Kernel{CUDABackend})(args...; ndrange=nothing, workgroupsize=noth
     return nothing
 end
 
-# list of overrides (only for Julia 1.6)
-const overrides = Expr[]
-
-macro device_override(ex)
-    ex = macroexpand(__module__, ex)
-    if Meta.isexpr(ex, :call)
-        @show ex = eval(ex)
-        error()
-    end
-    code = quote
-        Base.Experimental.@overlay($CUDA.method_table, $ex)
-    end
-    if isdefined(Base.Experimental, Symbol("@overlay"))
-        return esc(code)
-    else
-        push!(overrides, code)
-        return
-    end
-end
-
-function __init__()
-    precompiling = ccall(:jl_generating_output, Cint, ()) != 0
-    precompiling && return
-    # register device overrides
-    eval(Expr(:block, overrides...))
-    empty!(overrides)
-end
-
 import KernelAbstractions: CompilerMetadata, DynamicCheck, LinearIndices
 import KernelAbstractions: __index_Local_Linear, __index_Group_Linear, __index_Global_Linear, __index_Local_Cartesian, __index_Group_Cartesian, __index_Global_Cartesian, __validindex, __print
 import KernelAbstractions: mkcontext, expand, __iterspace, __ndrange, __dynamic_checkbounds