Cgo drnm2 panics with zero-length slice
btracey opened this issue · 8 comments
Dnrm2 allows a zero-length slice, but cgo does not check for zero length before taking &x[0].
This will be true of all the functions that take an n and a vector. I can add a check for n == 0 before the cgo call and return early then. If the ret type is int, return -1; this catches I*amax. The native implementations have some potential panics prior to the equivalent test, so presumably I should mimic that, but what is the basis for those?
I'm not sure what you mean by "the basis for those". In Idamax right now the code is:
if n < 2 {
if n == 1 {
return math.Abs(x[0])
}
if n == 0 {
return 0
}
if n < 1 {
panic(negativeN)
}
}
It returns 0 and not -1 because the refernce BLAS returns 1 when the size is zero.
Sorry, BLAS returns 1 for Idamax when the slice length is zero, and also returns 0 when the slice is zero for Dnrm2.
Easy to change if we'd rather it return -1. As far as I'm concerned, as long as we're doing something reasonable and cgo and native have the same behavior, I'm happy to deviate from the strict blas standard.
There are panics for incX < 0 when n == 0.
Yes.
Also, sorry, I misread the code. LAPACK does return an invalid index when n == 0.
OK, so what would you like. I have the hooks for int, float and none - in the float64 case we return 0, for int?
(What on earth is the basis for returning 1 when n == 0?).
I think Idamax should return -1. This is consistent with the current behavior and the reference standard. (Reference returns 0, but it's one-indexed)