SVD threading issue (lapack.dgesdd)

[zorba128]

Hi

I noticed SVD decomposition retuns wrong results when
invoking this method concurrently on the multithreaded environment:

• fails on: com.github.fommil.netlib.NativeRefLAPACK (netlib-native_ref-win-x86_64.dll)
• fails on: com.github.fommil.netlib.NativeSystemLAPACK (netlib-native_system-win-x86_64.dll)
• success on: com.github.fommil.netlib.NativeSystemLAPACK (netlib-native_system-linux-x86_64.so)
• success on: com.github.fommil.netlib.F2jLAPACK

Not sure if this is matter of glue code or LAPACK itself. Wrapping SVD method invocation
in the synchronization block helps (produces same results as when running single-threaded).

I noticedthat this issue affects only left/right vectors,
singular values seem to be fine.

Please take a look at sample code below:

object BreezeMultiThreading extends App {
  implicit val ec: ExecutionContextExecutor = scala.concurrent.ExecutionContext.global
  val x = Seq.tabulate(1000) { i =>
    Future /*.successful*/ { // uncomment to run single-threaded
      println(s"Iteration $i")
      val rnd = new scala.util.Random(12345 + i)
      val input = DenseMatrix.tabulate[Double](100, 100)((_,_) => rnd.nextDouble())
      val svdec = svd(input)
      // val sum = svdec.singularValues.valuesIterator.sum
      val sum = svdec.leftVectors.valuesIterator.sum + svdec.rightVectors.valuesIterator.sum
      sum
    }
  }

  val sum = Await.result(Future.sequence(x).map(_.sum), Duration.Inf)
  // val exp = 291367.83234463 // singular values
  val exp = -19997.446618756203 // left/right vectors
  if (abs(sum - exp) < 0.00001)
    Console.out.println(s"Success: $sum")
  else
    Console.err.println(s"Failed: $sum, expected: $exp")
}

marcin

[dlwh]

sorry for being slow. it's suspicious to me this only happens on windows so i doubt it's the glue code. I can add a lock if we're running on windows, i guess?