hi all,

I'm working on project using OpenXLA <https://github.com/openxla/xla> (c++ 
library for fast execution of computation graphs) in Go, and while managing 
their implementation of "tensors" (high dimensional data) in Go, I bumped 
into an odd behavior of slices pointing to data that is managed by the C++.

The full example is unfortunately too large (it would include C/C++ code), 
hopefully the relevant part I copy below is sufficient for someone to spot 
what could be happening -- I'm hoping I'm simply misunderstanding something:


func (l *Literal) Data() any {
  if l.IsNil() {
    return nil
  }
  rawData := unsafe.Pointer(l.cLiteralPtr.data)
  len := int(l.cLiteralPtr.size)
  return unsafe.Slice((int)(rawData), len)
}

  ...
    indices = indicesT.Literal().Data().([]int)
    fmt.Printf("\tindices=%v, ptr=%0X\n", indices, &indices[0])
    runtime.GC()   // Apparently changes content of indices
    fmt.Printf("\tindices=%v, ptr=%0X\n", indices, &indices[0])
    ...

The `indicesT` is generated (in C++ land)  as a small vector of random 
indices, and it gets converted to a Go slice in `indices`, using 
`unsafe.Slice`.

Executing the above I get something like:

indices=[421 5309 10924 8068 2193 4380 3475 8713], ptr=2F66300 indices=[0 
5309 10924 8068 2193 4380 3475 8713], ptr=2F66300

Notice the value of `indices[0]` changed (!) to zero. There is no other 
goroutine that knows or interacts in any way with `indices` (or the 
underlying C++ object) -- at least that I'm aware.

Curiously if I remove the call to `runtime.GC()` everything works (!). Took 
me a while to discover and isolate the issue related to garbage collection 
(it could be just correlation)  since it would otherwise happen only very 
occasionally, I usually use these slices into CGO just after I create them.

Now also very curious is that if I add a second conversion for `indices`, 
after the `fmt.Printf` (!), then the issue goes away. As in:

   ...
    indices = indicesT.Literal().Data().([]int)
    fmt.Printf("\tindices=%v, ptr=%0X\n", indices, &indices[0])
    runtime.GC()
    fmt.Printf("\tindices=%v, ptr=%0X\n", indices, &indices[0])
    indices = indicesT.Literal().Data().([]int)
    ...

A typical output now is (notice the two lines are identical, as expected):

indices=[563 9434 222 8007 9770 11613 9906 11786], ptr=179C300 indices=[563 
9434 222 8007 9770 11613 9906 11786], ptr=179C300

I observed the same issue with other larger tensors I was working on (some 
garbled images stored as tensors). 

Any thoughts ?

many thanks!
Jan

-- 
You received this message because you are subscribed to the Google Groups 
"golang-nuts" group.
To unsubscribe from this group and stop receiving emails from it, send an email 
to golang-nuts+unsubscr...@googlegroups.com.
To view this discussion on the web visit 
https://groups.google.com/d/msgid/golang-nuts/b85e5ce7-cadc-4afd-b161-b74f923c7589n%40googlegroups.com.

Reply via email to