AcceleratedKernels accumulate and sort

Project.toml

@@ -3,6 +3,7 @@ uuid = "08131aa3-fb12-5dee-8b74-c09406e224a2"
 version = "0.10.9"
 
 [deps]
+AcceleratedKernels = "6a4ca0a5-0e36-4168-a932-d9be78d558f1"
 Adapt = "79e6a3ab-5dfb-504d-930d-738a2a938a0e"
 GPUArrays = "0c68f7d7-f131-5f86-a1c3-88cf8149b2d7"
 GPUCompiler = "61eb1bfa-7361-4325-ad38-22787b887f55"
@@ -25,6 +26,7 @@ StaticArrays = "90137ffa-7385-5640-81b9-e52037218182"
 SPIRVIntrinsics = {path = "lib/intrinsics"}
 
 [compat]
+AcceleratedKernels = "0.3.1, 0.4"
 Adapt = "4"
 GPUArrays = "11.2.1"
 GPUCompiler = "1.7.1"

src/OpenCL.jl

@@ -10,6 +10,7 @@ using Random
 using Preferences
 
 import KernelAbstractions: KernelAbstractions
+import AcceleratedKernels as AK
 
 using Core: LLVMPtr
 
@@ -47,4 +48,7 @@ include("random.jl")
 include("OpenCLKernels.jl")
 import .OpenCLKernels: OpenCLBackend
 export OpenCLBackend
+
+include("sorting.jl")
+include("accumulate.jl")
 end

src/accumulate.jl

@@ -0,0 +1,21 @@
+# Use a smaller block size to work around a scan correctness issue
+# with the Blelloch parallel prefix sum at larger block sizes (>=128).
+const _ACCUMULATE_BLOCK_SIZE = 64
+
+Base.accumulate!(
+    op, B::CLArray, A::CLArray; init = zero(eltype(A)),
+    block_size = _ACCUMULATE_BLOCK_SIZE, kwargs...
+) =
+    AK.accumulate!(op, B, A, OpenCLBackend(); init, block_size, kwargs...)
+
+Base.accumulate(
+    op, A::CLArray; init = zero(eltype(A)),
+    block_size = _ACCUMULATE_BLOCK_SIZE, kwargs...
+) =
+    AK.accumulate(op, A, OpenCLBackend(); init, block_size, kwargs...)
+
+Base.cumsum(src::CLArray; block_size = _ACCUMULATE_BLOCK_SIZE, kwargs...) =
+    AK.cumsum(src, OpenCLBackend(); block_size, kwargs...)
+
+Base.cumprod(src::CLArray; block_size = _ACCUMULATE_BLOCK_SIZE, kwargs...) =
+    AK.cumprod(src, OpenCLBackend(); block_size, kwargs...)

src/sorting.jl

@@ -0,0 +1,3 @@
+Base.sort!(x::CLArray; kwargs...) = (AK.sort!(x; kwargs...); return x)
+Base.sortperm!(ix::CLArray, x::CLArray; kwargs...) = (AK.sortperm!(ix, x; kwargs...); return ix)
+Base.sortperm(x::CLArray; kwargs...) = sortperm!(CLArray(1:length(x)), x; kwargs...)

test/setup.jl

@@ -90,7 +90,7 @@ function runtests(f, name, platform_filter)
         end
 
         # some tests require native execution capabilities
-        requires_il = name in ["atomics", "execution", "intrinsics", "kernelabstractions"] ||
+        requires_il = name in ["acceleratedkernels", "atomics", "execution", "intrinsics", "kernelabstractions"] ||
                       startswith(name, "gpuarrays/") || startswith(name, "device/")
 
         ex = quote