I agree with using Smalltalk when we can, and avoid FFI, except maybe
one thing...
In Smalltalk, we could completely eliminate the Float primitives and
emulate them with Integer arithmetic.
But we don't.
IMO we should adopt same pragmatic approach with basic operations on
vectors/matrices, and maybe also for more involved algebra.
I think so too.
A bit in the direction of WordArray or FloatArray?
The guys of LAPACK decided to use a "primitive" layer BLAS because the
FORTRAN compiler was too naive to compete with specially crafted
optimizations like that of MKL.
What is MKL?
Marcus is telling to me since years that we should use a good math
libraries like the one developed by some inria groups.
It does not make sense to compete in this area, our language is not
taylored for number crunching whatever the JIT efforts we can make,
Agreed.
Now what would be interesting to know is: what kind of operations do we
need to attack vectorised CPU
I do not remember their name but I was talking with Philippe Mougin the
father of F-Script and he wanted
to map his F-Script messages to vector on CPU to get all the power for
crunching numbers.
So let's use some kind of numeric primitives.
Which ones?
Do you have some cycles to produce a prototype so that we can learn?
That's how Matlab, NumPy, R or evry other interpreter would do it,
don't they?
I would love to have
- knowledge
- and time
Sadly I just have the energy for simple things.
It may be interesting to have some fallback code written in Smalltalk,
but it ain't gonna scale.
Indeed.
So don't forget that Smallapack could leverage number crunching,
that's what it's made for...
Should we put under the umbrella once we clean SciPharo?
Because it should be reorganized first.
stef
