Default/OpenBLAS matrix.transpose() dot vector yields NaNs and Zeroes compared to Kotlin implementation

[nekrozofbrionac]

Multik version

0.3.1

Module

multik-default or multik-openblas

Platform

JVM

Operating System

Linux

Description

When using the default or openblas module (on linux), matrix vecctor multiplication yields a corrupted vector when transposing the matrix first, especially if the matrix is not square.

Interestingly, this does not happen when importing the kotlin only module.

Reproducer

// in build.gradle.kts
dependencies {
    implementation("org.jetbrains.kotlinx:multik-default:0.3.1")
    // or
    // implementation("org.jetbrains.kotlinx:multik-openblas:0.3.1")

    // Compare this to 
    // implementation("org.jetbrains.kotlinx:multik-kotlin:0.3.1")
    // when testing
}


// snippet to test 
fun main() {
  val m1 = mk.d2array<Float>(20, 1000) { Random.nextFloat() }
  val v1 = mk.d1array<Float>(20) { Random.nextFloat() }

  val bug = m1.transpose() dot v1

  println(bug.shape.contentToString())
  println("amount of NaNs: " + bug.count { isNaN(it)  })
  println("amount of zeroes: " + bug.count { it == 0f })
}

// This yields 20 NaNs, and 980 zeroes using default, but none of both on kotlin

Additional information

This also does not seem to happen when you reshape the vector as a Nx1 matrix first, but nonetheless it still deviates from expected behaviour.

Let me know if you need more infos to reproduce or similar, i was just certain that this should definitely not be expected