use 256 threads

SimonDanisch · SimonDanisch · commit f50a1ea64490 · 2017-08-22T15:14:38.000+02:00
diff --git a/src/backends/cudanative/cudanative.jl b/src/backends/cudanative/cudanative.jl
@@ -293,7 +293,7 @@ function acc_mapreduce{T, OT, N}(
     )
     dev = context(A).device
     @assert(CUDAdrv.capability(dev) >= v"3.0", "Current CUDA reduce implementation requires a newer GPU")
-    threads = 512
+    threads = 256
     blocks = min((length(A) + threads - 1) ÷ threads, 1024)
     out = similar(buffer(A), OT, (blocks,))
     args = map(unpack_cu_array, rest)

Original file line number	Diff line number	Diff line change
`@@ -293,7 +293,7 @@ function acc_mapreduce{T, OT, N}(`
`293`	`293`	`)`
`294`	`294`	`dev = context(A).device`
`295`	`295`	`@assert(CUDAdrv.capability(dev) >= v"3.0", "Current CUDA reduce implementation requires a newer GPU")`
`296`		`- threads = 512`
	`296`	`+ threads = 256`
`297`	`297`	`blocks = min((length(A) + threads - 1) ÷ threads, 1024)`
`298`	`298`	`out = similar(buffer(A), OT, (blocks,))`
`299`	`299`	`args = map(unpack_cu_array, rest)`