Merge pull request #33 from JuliaSIMD/intializefor32bit

Tests pass locally on 32 bit

Author: chriselrod

src/ThreadingUtilities.jl

@@ -37,17 +37,30 @@ function initialize_task(tid)
   # set to tid, we have tasks 2...nthread, from 1-based ind perspective
   ccall(:jl_set_task_tid, Cvoid, (Any, Cint), t, tid % Cint)
   TASKS[tid] = t
+  return nothing
 end
 
+if Sys.WORD_SIZE == 32
+  retnothing(::Ptr{UInt}) = nothing
+end
 function __init__()
   _print_exclusivity_warning()
   nt = min(Threads.nthreads(), (Sys.CPU_THREADS)::Int) - 1
   resize!(THREADPOOL, (THREADBUFFERSIZE ÷ sizeof(UInt)) * nt + (LINESPACING ÷ sizeof(UInt)) - 1)
   copyto!(THREADPOOL, zero(UInt))
   # align to LINESPACING boundary, and then subtract THREADBUFFERSIZE to make the pointer 1-indexed
-  THREADPOOLPTR[] = reinterpret(Ptr{UInt}, (reinterpret(UInt, (pointer(THREADPOOL)))+LINESPACING-1) & (-LINESPACING)) - THREADBUFFERSIZE
+  THREADPOOLPTR[] = reinterpret(Ptr{UInt}, (reinterpret(UInt, pointer(THREADPOOL))+LINESPACING-1) & (-LINESPACING)) - THREADBUFFERSIZE
   resize!(TASKS, nt)
   foreach(initialize_task, 1:nt)
+  @static if Sys.WORD_SIZE == 32
+    if nt > 0
+      fptr = @cfunction(retnothing, Cvoid, (Ptr{UInt},))
+      store!(taskpointer(1), fptr, sizeof(UInt))
+      _atomic_xchg!(taskpointer(1), TASK)
+      wake_thread!(1)
+      wait(1)
+    end
+  end
 end
 
 end # module

src/threadtasks.jl

@@ -56,6 +56,7 @@ end
   if istaskfailed(t)
     display(t)
     dump(t)
+    println()
     initialize_task(tid)
     return true
   end

test/aqua.jl

@@ -1,3 +1,4 @@
+import Aqua
 @testset "Aqua.jl" begin
     @time Aqua.test_all(ThreadingUtilities)
 end

test/testsetup.jl

@@ -2,6 +2,5 @@ using ThreadingUtilities
 using Test
 using VectorizationBase
 
-import Aqua
 import InteractiveUtils
 

test/threadingutilities.jl

@@ -1,54 +1,50 @@
 struct Copy{P} end
 function (::Copy{P})(p::Ptr{UInt}) where {P}
-    _, (ptry,ptrx,N) = ThreadingUtilities.load(p, P, 2*sizeof(UInt))
-    N > 0 || throw("This function throws if N == 0 for testing purposes.")
-    @simd ivdep for n ∈ 1:N
-        vstore!(ptry, vload(ptrx, (n,)), (n,))
-    end
+  _, (ptry,ptrx,N) = ThreadingUtilities.load(p, P, 2*sizeof(UInt))
+  N > 0 || throw("This function throws if N == 0 for testing purposes.")
+  @simd ivdep for n ∈ 1:N
+    vstore!(ptry, vload(ptrx, (n,)), (n,))
+  end
 end
 @generated function copy_ptr(::A, ::B) where {A,B}
-    c = Copy{Tuple{A,B,Int}}()
-    quote
-        @cfunction($c, Cvoid, (Ptr{UInt},))
-    end
+  c = Copy{Tuple{A,B,Int}}()
+  quote
+    @cfunction($c, Cvoid, (Ptr{UInt},))
+  end
 end
 function setup_copy!(p, y, x)
-    N = length(y)
-    @assert length(x) == N
-    py = stridedpointer(y)
-    px = stridedpointer(x)
-    fptr = copy_ptr(py, px)
-    offset = ThreadingUtilities.store!(p, fptr, sizeof(UInt))
-    ThreadingUtilities.store!(p, (py,px,N), offset)
+  N = length(y)
+  @assert length(x) == N
+  py = stridedpointer(y)
+  px = stridedpointer(x)
+  fptr = copy_ptr(py, px)
+  offset = ThreadingUtilities.store!(p, fptr, sizeof(UInt))
+  ThreadingUtilities.store!(p, (py,px,N), offset)
 end
 
-@inline function launch_thread_copy!(tid, y, x)
-    ThreadingUtilities.launch(tid, y, x) do p, y, x
-        setup_copy!(p, y, x)
-    end
-end
+@inline launch_thread_copy!(tid, y, x) = ThreadingUtilities.launch(setup_copy!, tid, y, x)
 
 function test_copy(tid, N = 100_000)
-    a = rand(N);
-    b = rand(N);
-    c = rand(N);
-    x = similar(a) .= NaN;
-    y = similar(b) .= NaN;
-    z = similar(c) .= NaN;
-    GC.@preserve a b c x y z begin
-        launch_thread_copy!(tid, x, a)
-        yield()
-        @assert !ThreadingUtilities.wait(tid)
-        launch_thread_copy!(tid, y, b)
-        yield()
-        @assert !ThreadingUtilities.wait(tid)
-        launch_thread_copy!(tid, z, c)
-        yield()
-        @assert !ThreadingUtilities.wait(ThreadingUtilities.taskpointer(tid))
-    end
-    @test a == x
-    @test b == y
-    @test c == z
+  a = rand(N);
+  b = rand(N);
+  c = rand(N);
+  x = similar(a) .= NaN;
+  y = similar(b) .= NaN;
+  z = similar(c) .= NaN;
+  GC.@preserve a b c x y z begin
+    launch_thread_copy!(tid, x, a)
+    yield()
+    @assert !ThreadingUtilities.wait(tid)
+    launch_thread_copy!(tid, y, b)
+    yield()
+    @assert !ThreadingUtilities.wait(tid)
+    launch_thread_copy!(tid, z, c)
+    yield()
+    @assert !ThreadingUtilities.wait(ThreadingUtilities.taskpointer(tid))
+  end
+  @test a == x
+  @test b == y
+  @test c == z
 end
 
 @testset "ThreadingUtilities.jl" begin