API changes

Author: pxl-th

src/ROCKernels.jl

@@ -168,7 +168,7 @@ end
 
 # Reduction.
 
-KA.supports_warp_reduction(::ROCBackend) = true
+@device_override @inline KA.__supports_warp_reduction() = true
 
 @device_override @inline function KA.__shfl_down(val, offset)
     AMDGPU.Device.shfl_down(val, offset)

t.jl

@@ -0,0 +1,18 @@
+using AMDGPU
+using KernelAbstractions
+
+@kernel cpu=false function groupreduce_1!(y, x, op, neutral)
+    i = @index(Global)
+    val = i > length(x) ? neutral : x[i]
+    res = @groupreduce(op, val, neutral)
+    i == 1 && (y[1] = res)
+end
+
+function main()
+    x = ROCArray(ones(Float32, 256))
+    y = ROCArray(zeros(Float32, 1))
+    groupreduce_1!(ROCBackend(), 256)(y, x, +, 0f0; ndrange=256)
+    @show y
+    return
+end
+main()