Add test

test/Project.toml

@@ -1,2 +1,4 @@
 [deps]
 Test = "8dfed614-e22c-5e08-85e1-65c5234f0b40"
+CUDA = "052768ef-5323-5732-b1bb-66c8b64840ba"
+

test/test-basic-test.jl

@@ -1,3 +1,28 @@
+using InclusiveScans
+using CUDA
+
 @testset "InclusiveScans.jl" begin
-    @test InclusiveScans.hello_world() == "Hello, World!"
+    function _cpu_inclusive_cumsum(x::Vector{Float32})
+        s = 0.0f0
+        out = similar(x)
+        for i = 1:length(x)
+            s += x[i]
+            out[i] = s
+        end
+        return out
+    end
+    eps = 0.1
+    N = 25000
+    h_in = rand(Float32, N)
+    d_in = CuArray(h_in)
+    d_out = CUDA.zeros(Float32, N)
+
+    InclusiveScans.largeArrayScanInclusive!(d_out, d_in, N)
+    h_out = Array(d_out)
+
+    # CPU cumsum check
+    h_check = _cpu_inclusive_cumsum(h_in)
+
+    maxdiff = maximum(abs.(h_out .- h_check))
+    @test maxdiff < eps
 end