Performance

[Quafadas]

As of June 2026, it seems t me, that it is not possible, to have a generic and performant implementation. Some reading;

scala/scala3#15532

scala/scala3#18044

scala/scala3#12161

In particular, the @specialised annotation does not appear to be effecitve. It appears to still emit Generic parameters into the emitted bytecode.

package mat

import scala.reflect.ClassTag

// LLM thinks this is performant
final class MatrixDoubleInternal(val raw: Array[Double], val rows: Int, val cols: Int)


// LLM thinks everything below here sucks. 
final class MatrixGeneric2[A](
  val raw: Array[A], 
  val rows: Int, 
  val cols: Int
)


// Generic final class  
trait MatrixGeneric[A](
  val raw: Array[A], 
  val rows: Int, 
  val cols: Int
)


type DoubleMatrix = MatrixGeneric[Double] 

object DoubleMatrix2{
  def apply(rows: Int, cols: Int): DoubleMatrix =  
    new MatrixGeneric[Double](Array.ofDim[Double](rows*cols), rows, cols){}
    

  def apply[A](raw: Array[Double], rows: Int, cols: Int): DoubleMatrix =
    new MatrixGeneric[Double](raw, rows, cols){}

}

It also seems to hate the current opaque class implementation for performance reasons.

In fact, there seems to be no compelling reason to not simply have "class" DoubleMatrix.

[Quafadas]

The ultimate solution here is "stride aware" which allows both cache locality and zero copy semantics for operations like transpose and submatricies.

inlining should deal with the performance problems.