Split in device and host folders.

Author: maleadt

src/GPUArrays.jl

@@ -12,24 +12,27 @@ using AbstractFFTs
 
 using Adapt
 
-# GPU interface
-## core definition
-include("abstractarray.jl")
-include("devices.jl")
-include("execution.jl")
-include("ondevice.jl")
-include("construction.jl")
+# device array
+include("device/abstractarray.jl")
+include("device/indexing.jl")
+include("device/gpu.jl")
+
+# host array
+include("host/abstractarray.jl")
+include("host/devices.jl")
+include("host/execution.jl")
+include("host/construction.jl")
 ## integrations and specialized functionality
-include("base.jl")
-include("indexing.jl")
-include("broadcast.jl")
-include("mapreduce.jl")
-include("linalg.jl")
-include("random.jl")
-
-# CPU implementation
+include("host/base.jl")
+include("host/indexing.jl")
+include("host/broadcast.jl")
+include("host/mapreduce.jl")
+include("host/linalg.jl")
+include("host/random.jl")
+include("host/quirks.jl")
+
+# CPU reference implementation
 include("array.jl")
 
-include("quirks.jl")
 
 end # module

src/array.jl

@@ -1,4 +1,4 @@
-# CPU implementation of the GPUArray interface
+# reference implementation of the GPUArray interfaces
 
 export JLArray
 
@@ -88,8 +88,6 @@ to_device(state, x::JLArray) = x.data
 to_device(state, x::Tuple) = to_device.(Ref(state), x)
 to_device(state, x::Base.RefValue{<: JLArray}) = Base.RefValue(to_device(state, x[]))
 to_device(state, x) = x
-# creates a `local` vector for each thread group
-to_device(state, x::LocalMemory{T}) where T = LocalMem(ntuple(i-> Vector{T}(x.size), blockdim_x(state)))
 
 to_blocks(state, x) = x
 # unpacks local memory for each block

src/device/abstractarray.jl

@@ -0,0 +1,26 @@
+# on-device functionality
+
+export AbstractDeviceArray
+
+
+## device array
+
+abstract type AbstractDeviceArray{T, N} <: AbstractArray{T, N} end
+
+Base.IndexStyle(::AbstractDeviceArray) = IndexLinear()
+
+@inline function Base.iterate(A::AbstractDeviceArray, i=1)
+    if (i % UInt) - 1 < length(A)
+        (@inbounds A[i], i + 1)
+    else
+        nothing
+    end
+end
+
+function Base.sum(A::AbstractDeviceArray{T}) where T
+    acc = zero(T)
+    for elem in A
+        acc += elem
+    end
+    acc
+end

src/abstractarray.jl renamed to src/host/abstractarray.jl

@@ -11,13 +11,6 @@ const GPUVector{T} = GPUArray{T, 1}
 const GPUMatrix{T} = GPUArray{T, 2}
 const GPUVecOrMat{T} = Union{GPUArray{T, 1}, GPUArray{T, 2}}
 
-# GPU Local Memory
-struct LocalMemory{T} <: GPUArray{T, 1}
-    size::Int
-    LocalMemory{T}(x::Integer) where T = new{T}(x)
-end
-
-
 # input/output
 
 ## serialization