Use per-BestSplitAllocs rand source to prevent contention

* When learning on many threads, there is a _lot_ of contention
  on the mutex in global rand source for no real reason.
* In my use case, results in speed up from 4m30s to 33s
  on 64 core system (7.7x faster)

densecatfeature.go

@@ -657,7 +657,7 @@ func (f *DenseCatFeature) BestCatSplit(target Target,
 		bits = i
 		if !useExhaustive {
 			//generate random partition
-			bits = rand.Int()
+			bits = allocs.Rnd.Int()
 		}
 
 		// //check the value of the j'th bit of i and
@@ -859,7 +859,6 @@ func (f *DenseCatFeature) BestCatSplitBig(target Target, cases *[]int, parentImp
 
 	bits := big.NewInt(1)
 
-	var randgn *rand.Rand
 	var maxPart *big.Int
 	useExhaustive := nCats <= maxEx
 	nPartitions := big.NewInt(2)
@@ -871,7 +870,6 @@ func (f *DenseCatFeature) BestCatSplitBig(target Target, cases *[]int, parentImp
 		nPartitions.Lsh(nPartitions, uint(maxEx-2))
 		maxPart = big.NewInt(2)
 		maxPart.Lsh(maxPart, uint(nCats-2))
-		randgn = rand.New(rand.NewSource(0))
 	}
 
 	//iteratively build a combination of categories until they
@@ -881,7 +879,7 @@ func (f *DenseCatFeature) BestCatSplitBig(target Target, cases *[]int, parentImp
 		bits.Set(i)
 		if !useExhaustive {
 			//generate random partition
-			bits.Rand(randgn, maxPart)
+			bits.Rand(allocs.Rnd, maxPart)
 		}
 
 		//check the value of the j'th bit of i and

densenumfeature.go

@@ -313,7 +313,7 @@ func (f *DenseNumFeature) BestNumSplit(target Target,
 		lasti := leafSize - 1
 
 		if randomSplit {
-			leafSize = leafSize + rand.Intn(stop-leafSize)
+			leafSize = leafSize + allocs.Rnd.Intn(stop-leafSize)
 			lasti = leafSize - 1
 			stop = leafSize + 1
 

featurematrix.go

@@ -198,7 +198,7 @@ func (fm *FeatureMatrix) BestSplitter(target Target,
 		if lcans > nDrawnConstants+lastSample {
 
 			randi = lastSample
-			randi += rand.Intn(lcans - nDrawnConstants - lastSample)
+			randi += allocs.Rnd.Intn(lcans - nDrawnConstants - lastSample)
 			//randi = lastSample + rand.Intn(nnonconstant-lastSample)
 			if randi >= lcans-nConstants {
 				nDrawnConstants++

splitallocations.go

@@ -1,6 +1,8 @@
 package CloudForest
 
-import ()
+import (
+	"math/rand"
+)
 
 //BestSplitAllocs contains reusable allocations for split searching and evaluation.
 //Seprate instances should be used in each go routing doing learning.
@@ -27,6 +29,7 @@ type BestSplitAllocs struct {
 	SortVals       []float64
 	Sorter         *SortableFeature //for learning from numerical features
 	ContrastTarget Target
+	Rnd            *rand.Rand //prevent contention on global rand source
 }
 
 //NewBestSplitAllocs initializes all of the reusable allocations for split
@@ -63,6 +66,8 @@ func NewBestSplitAllocs(nTotalCases int, target Target) (bsa *BestSplitAllocs) {
 		make([]float64, nTotalCases, nTotalCases),
 		&SortableFeature{make([]float64, nTotalCases, nTotalCases),
 			nil},
-		target.(Feature).Copy().(Target)}
+		target.(Feature).Copy().(Target),
+		rand.New(rand.NewSource(rand.Int63())),
+	}
 	return
 }