Expose {ml, mllib}-private f2jBLAS and use that

Author: Nick Pentreath

mllib-local/src/main/scala/org/apache/spark/ml/linalg/BLAS.scala

@@ -29,7 +29,7 @@ private[spark] object BLAS extends Serializable {
   @transient private var _nativeBLAS: NetlibBLAS = _
 
   // For level-1 routines, we use Java implementation.
-  private def f2jBLAS: NetlibBLAS = {
+  private[ml] def f2jBLAS: NetlibBLAS = {
     if (_f2jBLAS == null) {
       _f2jBLAS = new F2jBLAS
     }

mllib/src/main/scala/org/apache/spark/ml/recommendation/ALS.scala

@@ -27,7 +27,6 @@ import scala.util.{Sorting, Try}
 import scala.util.hashing.byteswap64
 
 import com.github.fommil.netlib.BLAS.{getInstance => blas}
-import com.github.fommil.netlib.F2jBLAS
 import org.apache.hadoop.fs.Path
 import org.json4s.DefaultFormats
 import org.json4s.JsonDSL._
@@ -36,6 +35,7 @@ import org.apache.spark.{Dependency, Partitioner, ShuffleDependency, SparkContex
 import org.apache.spark.annotation.{DeveloperApi, Since}
 import org.apache.spark.internal.Logging
 import org.apache.spark.ml.{Estimator, Model}
+import org.apache.spark.ml.linalg.BLAS
 import org.apache.spark.ml.param._
 import org.apache.spark.ml.param.shared._
 import org.apache.spark.ml.util._
@@ -399,7 +399,7 @@ class ALSModel private[ml] (
         srcIter.foreach { case (srcId, srcFactor) =>
           dstIter.foreach { case (dstId, dstFactor) =>
             // We use F2jBLAS which is faster than a call to native BLAS for vector dot product
-            val score = ALSModel._f2jBLAS.sdot(rank, srcFactor, 1, dstFactor, 1)
+            val score = BLAS.f2jBLAS.sdot(rank, srcFactor, 1, dstFactor, 1)
             pq += dstId -> score
           }
           pq.foreach { case (dstId, score) =>
@@ -439,8 +439,6 @@ class ALSModel private[ml] (
 @Since("1.6.0")
 object ALSModel extends MLReadable[ALSModel] {
 
-  @transient private[recommendation] val _f2jBLAS = new F2jBLAS
-
   private val NaN = "nan"
   private val Drop = "drop"
   private[recommendation] final val supportedColdStartStrategies = Array(NaN, Drop)

mllib/src/main/scala/org/apache/spark/mllib/linalg/BLAS.scala

@@ -31,7 +31,7 @@ private[spark] object BLAS extends Serializable with Logging {
   @transient private var _nativeBLAS: NetlibBLAS = _
 
   // For level-1 routines, we use Java implementation.
-  private def f2jBLAS: NetlibBLAS = {
+  private[mllib] def f2jBLAS: NetlibBLAS = {
     if (_f2jBLAS == null) {
       _f2jBLAS = new F2jBLAS
     }

mllib/src/main/scala/org/apache/spark/mllib/recommendation/MatrixFactorizationModel.scala

@@ -22,7 +22,6 @@ import java.lang.{Integer => JavaInteger}
 
 import com.clearspring.analytics.stream.cardinality.HyperLogLogPlus
 import com.github.fommil.netlib.BLAS.{getInstance => blas}
-import com.github.fommil.netlib.F2jBLAS
 import org.apache.hadoop.fs.Path
 import org.json4s._
 import org.json4s.JsonDSL._
@@ -32,6 +31,7 @@ import org.apache.spark.SparkContext
 import org.apache.spark.annotation.Since
 import org.apache.spark.api.java.{JavaPairRDD, JavaRDD}
 import org.apache.spark.internal.Logging
+import org.apache.spark.mllib.linalg.BLAS
 import org.apache.spark.mllib.rdd.MLPairRDDFunctions._
 import org.apache.spark.mllib.util.{Loader, Saveable}
 import org.apache.spark.rdd.RDD
@@ -246,8 +246,6 @@ object MatrixFactorizationModel extends Loader[MatrixFactorizationModel] {
 
   import org.apache.spark.mllib.util.Loader._
 
-  @transient private val _f2jBLAS = new F2jBLAS
-
   /**
    * Makes recommendations for a single user (or product).
    */
@@ -299,7 +297,7 @@ object MatrixFactorizationModel extends Loader[MatrixFactorizationModel] {
       srcIter.foreach { case (srcId, srcFactor) =>
         dstIter.foreach { case (dstId, dstFactor) =>
           // We use F2jBLAS which is faster than a call to native BLAS for vector dot product
-          val score = _f2jBLAS.ddot(rank, srcFactor, 1, dstFactor, 1)
+          val score = BLAS.f2jBLAS.ddot(rank, srcFactor, 1, dstFactor, 1)
           pq += dstId -> score
         }
         pq.foreach { case (dstId, score) =>