cuconvert

Author: wsmoses

ext/ReactantCUDAExt.jl

@@ -239,9 +239,12 @@ function Adapt.adapt_structure(
     )
 end
 
-Reactant.@reactant_overlay @noinline function CUDA.cudaconvert(arg)
+function recudaconvert(arg)
     return adapt(ReactantKernelAdaptor(), arg)
 end
+Reactant.@reactant_overlay @noinline function CUDA.cudaconvert(arg)
+    return recudaconvert(arg)
+end
 
 function Adapt.adapt_storage(::ReactantKernelAdaptor, xs::TracedRArray{T,N}) where {T,N}
     res = CuTracedArray{T,N,CUDA.AS.Global,size(xs)}(xs)
@@ -456,7 +459,7 @@ end
 
 Reactant.@reactant_overlay @noinline function (func::LLVMFunc{F,tt})(
     args...;
-    convert=Val(false),
+    convert=Val(true),
     blocks::CuDim=1,
     threads::CuDim=1,
     cooperative::Bool=false,
@@ -466,6 +469,10 @@ Reactant.@reactant_overlay @noinline function (func::LLVMFunc{F,tt})(
     blockdim = CUDA.CuDim3(blocks)
     threaddim = CUDA.CuDim3(threads)
 
+    if convert == Val(true)
+        args = recudaconvert.(args) 
+    end
+
     mlir_args = MLIR.IR.Value[]
     restys = MLIR.IR.Type[]
     aliases = MLIR.IR.Attribute[]