Fix multithreading offsets. Fixes #14.

Project.toml

@@ -1,7 +1,7 @@
 name = "VectorizedRNG"
 uuid = "33b4df10-0173-11e9-2a0c-851a7edac40e"
 authors = ["Chris Elrod <elrodc@gmail.com>"]
-version = "0.2.9"
+version = "0.2.10"
 
 [deps]
 Distributed = "8ba89e20-285c-5b6f-9357-94700520ee1b"

src/VectorizedRNG.jl

@@ -33,46 +33,27 @@ include("xoshiro.jl")
 
 const GLOBAL_vRNGs = Ref{Ptr{UInt64}}()
 
-local_rng(i) = Xoshift{XREGISTERS}(i*4simd_integer_register_size()*XREGISTERS + GLOBAL_vRNGs[])
+
+const RNG_MEM_SIZE = (5(simd_integer_register_size()*XREGISTERS + 2048*3))
+local_rng(i) = Xoshift{XREGISTERS}(i*(RNG_MEM_SIZE) + GLOBAL_vRNGs[])
 local_rng() = local_rng(Base.Threads.threadid() - 1)
 
 # include("precompile.jl")
 # _precompile_()
 
-# const RANDBUFFER32 = Float32[]
-# const RANDNBUFFER32 = Float32[]
-# const RANDBUFFER64 = Float64[]
-# const RANDNBUFFER64 = Float64[]
-# const RANDBUFFERCOUNTER = UInt8[]
-# const RANDNBUFFER32COUNTER = UInt8[]
-# const RANDBUFFER64COUNTER = UInt8[]
-# const RANDNBUFFER64COUNTER = UInt8[]
-
-
-
 function __init__()
     # ccall(:jl_generating_output, Cint, ()) == 1 && return
-    nthreads = Base.Threads.nthreads()
-    nstreams = XREGISTERS * nthreads * simd_integer_register_size()
-    GLOBAL_vRNGs[] = ptr = VectorizationBase.valloc(5nstreams + 256 * 3nthreads, UInt64)
-    initXoshift!(ptr, nstreams)
-
-    for tid ∈ 0:nthreads-1
-        rng = local_rng(tid)
-        setrandu64counter!(rng, 0x00)
-        # setrandn32counter!(rng, 0x00)
-        setrand64counter!(rng, 0x00)
-        setrandn64counter!(rng, 0x00)
-    end
-    # resize!(RANDBUFFER32, 256nthreads)
-    # resize!(RANDNBUFFER32, 256nthreads)
-    # resize!(RANDBUFFER64, 256nthreads)
-    # resize!(RANDNBUFFER64, 256nthreads)
-
-    # resize!(RANDBUFFERCOUNTER, VectorizationBase.CACHELINE_SIZE*nthreads); fill!(RANDBUFFERCOUNTER, 0)
-    # resize!(RANDNBUFFER32COUNTER, VectorizationBase.CACHELINE_SIZE*nthreads)
-    # resize!(RANDBUFFER64COUNTER, VectorizationBase.CACHELINE_SIZE*nthreads)
-    # resize!(RANDNBUFFER64COUNTER, VectorizationBase.CACHELINE_SIZE*nthreads)
+  nthreads = Base.Threads.nthreads()
+  GLOBAL_vRNGs[] = ptr = VectorizationBase.valloc((RNG_MEM_SIZE ÷ 8)*nthreads, UInt64)
+  nstreams = XREGISTERS * nthreads * simd_integer_register_size()
+  initXoshift!(ptr, nstreams)
+  for tid ∈ 0:nthreads-1
+    rng = local_rng(tid)
+    setrandu64counter!(rng, 0x00)
+    # setrandn32counter!(rng, 0x00)
+    setrand64counter!(rng, 0x00)
+    setrandn64counter!(rng, 0x00)
+  end
 end