3 dimensional arrays don't do dot product of sub-matrices correctly

[mihbor]

3 dimensional arrays don't do dot product correctly (unless they have only one top-level element).
Reproducer:

import org.jetbrains.kotlinx.multik.api.linalg.dot
import org.jetbrains.kotlinx.multik.api.mk
import org.jetbrains.kotlinx.multik.api.ndarray
import org.jetbrains.kotlinx.multik.ndarray.data.get

fun main() {

  val a = mk.ndarray(
    mk[
      mk[
        mk[1.0],
        mk[0.0],
      ],
    ],
  )
  val b = mk.ndarray(
    mk[
      mk[
        mk[1.0],
        mk[0.0],
      ],
      mk[
        mk[1.0],
        mk[0.0],
      ],
    ],
  )
  println("a[0] dot a[0].T: (CORRECT)")
  println(a[0] dot a[0].transpose())

  println("\nb[0] dot b[0].T: (WRONG)")
  println(b[0] dot b[0].transpose())

  println("\nb[1] dot b[1].T: (WRONG)")
  println(b[1] dot b[1].transpose())
}

prints

a[0] dot a[0].T: (CORRECT)
[[1.0, 0.0],
[0.0, 0.0]]

b[0] dot b[0].T: (WRONG)
[[0.0, 0.0],
[0.0, 0.0]]

b[1] dot b[1].T: (WRONG)
[[0.0, 0.0],
[0.0, 0.0]]

[mihbor]

A couple of observations:
The problem goes away if I set the engine type to:
mk.setEngine(KEEngineType)
With the default engine the dot product doesn't work correctly for D2 x D1 shapes as well sometimes.

[devcrocod]

Thank you very much for the additional information.

I have reproduced the error. It indeed occurs with the native implementation, which is selected by default for tensor multiplication.

As far as I can see, this error is related to the incorrect passing of flags to the native implementation along with copying for sequential memory representation.

Unfortunately, it won’t be possible to quickly release an artifact with a fix for the bug, as there is a blocker due to linker errors on macOS.

[Tagakov]

There is another repro example with some additional workaround:

val left = mk.ndarray(
    arrayOf(
        doubleArrayOf(11.86,  13.85,  17.50, 0.0,    0.0,    0.0,    0.0),
        doubleArrayOf(0.0,    0.0,    17.50, 0.0,    0.0,    0.0,    0.0),
        doubleArrayOf(0.0,    0.0,    0.0,    0.0,   22.76, 0.0,    0.0),
        doubleArrayOf(0.0,    38.67, 0.0,    0.0,    28.73, 44.87, 0.0),
        doubleArrayOf(16.97, 0.0,    0.0,    0.0,    0.0,    0.0,    18.29),
        doubleArrayOf(0.0,    0.0,    0.0,    9.86, 0.0,    0.0,    0.0),
        doubleArrayOf(0.0,    0.0,    0.0,    13.01, 0.0,    18.40, 0.0),
        doubleArrayOf(0.0,    0.0,    0.0,    0.0,    0.0,    0.0,    32.37),
    )
)

val right = mk.ndarray(doubleArrayOf(200.0, 20.0, 40.0, 180.0, 200.0, 50.0, 69.0, 10.0))

val transposed = left.transpose()

val allZeroes = transposed dot right //Returns all zeroes
val allZeroes2 = transposed.copy() dot right //Returns all zeroes
val worksCorrectly = transposed.deepCopy() dot right