fix: eliminate threadid()

Project.toml

@@ -1,7 +1,7 @@
 name = "Stuffing"
 uuid = "4175e07e-e5b7-423e-8796-3ea7f6d48281"
 authors = ["guoyongzhi <[email protected]>"]
-version = "0.10.1"
+version = "0.10.2"
 
 [deps]
 Random = "9a3f8284-a2c9-5f02-9a11-845980a1fd5c"

src/fit.jl

@@ -241,16 +241,21 @@ function filttrain!(qtrees, inpool, outpool, nearlevel2; optimiser,
     colist = Vector{QTrees.CoItem}()
     sl1 = Threads.SpinLock()
     sl2 = Threads.SpinLock()
-    Threads.@threads for (i1, i2) in inpool |> shuffle!
-        que = queue[Threads.threadid()]
-        cp = QTrees._collision_randbfs(qtrees[i1], qtrees[i2], empty!(que))
-        if cp[1] >= nearlevel2
-            if outpool !== nothing
-                @Base.lock sl1 push!(outpool, (i1, i2))
-            end
-            if cp[1] > 0
-                @Base.lock sl2 push!(colist, (i1, i2) => cp)
-                nc1 += 1
+    shuffle!(inpool)
+    nchunks = min(length(queue), max(1, length(inpool)÷4))
+    Threads.@threads for ichunk in 1:nchunks
+        que = @inbounds queue[ichunk]
+        for ind in ichunk : nchunks : length(inpool)
+            i1, i2 = inpool[ind]
+            cp = QTrees._collision_randbfs(qtrees[i1], qtrees[i2], empty!(que))
+            if cp[1] >= nearlevel2
+                if outpool !== nothing
+                    @Base.lock sl1 push!(outpool, (i1, i2))
+                end
+                if cp[1] > 0
+                    @Base.lock sl2 push!(colist, (i1, i2) => cp)
+                    nc1 += 1
+                end
             end
         end
     end

src/qtree_functions.jl

@@ -67,34 +67,42 @@ function collision(Q1::AbstractStackedQTree, Q2::AbstractStackedQTree)
 end
 const CoItem = Pair{Tuple{Int,Int}, Index}
 const AbstractThreadQueue = AbstractVector{<:AbstractVector{Index}}
-thread_queue() = [Vector{Tuple{Int,Int,Int}}() for i = 1:Threads.nthreads()]
+thread_queue() = [Vector{Tuple{Int,Int,Int}}() for i = 1:2Threads.nthreads()]
 # assume inkernelbounds(qtree, at) is true
 function _totalcollisions_native(qtrees::AbstractVector, copairs; 
         colist=Vector{CoItem}(),
         queue::AbstractThreadQueue=thread_queue(),
         at=(length(qtrees[1]), 1, 1))
     sl = Threads.SpinLock()
-    Threads.@threads for (i1, i2) in copairs
-        que = @inbounds queue[Threads.threadid()]
-        empty!(que)
-        push!(que, at)
-        cp = @inbounds _collision_randbfs(qtrees[i1], qtrees[i2], que)
-        if cp[1] >= 0
-            @Base.lock sl push!(colist, (i1, i2) => cp)
+    nchunks = min(length(queue), max(1, length(copairs)÷4))
+    Threads.@threads for ichunk in 1:nchunks
+        que = @inbounds queue[ichunk]
+        for ind in ichunk : nchunks : length(copairs)
+            i1, i2 = copairs[ind]
+            empty!(que)
+            push!(que, at)
+            cp = @inbounds _collision_randbfs(qtrees[i1], qtrees[i2], que)
+            if cp[1] >= 0
+                @Base.lock sl push!(colist, (i1, i2) => cp)
+            end
         end
     end
     colist
 end
 function _totalcollisions_native(qtrees::AbstractVector, coitems::Vector{CoItem}; 
     colist=Vector{CoItem}(), queue::AbstractThreadQueue=thread_queue())
     sl = Threads.SpinLock()
-    Threads.@threads for ((i1, i2), at) in coitems
-        que = @inbounds queue[Threads.threadid()]
-        empty!(que)
-        push!(que, at)
-        cp = @inbounds _collision_randbfs(qtrees[i1], qtrees[i2], que)
-        if cp[1] >= 0
-            @Base.lock sl push!(colist, (i1, i2) => cp)
+    nchunks = min(length(queue), max(1, length(coitems)÷4))
+    Threads.@threads for ichunk in 1:nchunks
+        que = @inbounds queue[ichunk]
+        for ind in ichunk : nchunks : length(coitems)
+            (i1, i2), at = coitems[ind]
+            empty!(que)
+            push!(que, at)
+            cp = @inbounds _collision_randbfs(qtrees[i1], qtrees[i2], que)
+            if cp[1] >= 0
+                @Base.lock sl push!(colist, (i1, i2) => cp)
+            end
         end
     end
     colist