Re: JIT compilation with callbacks into Scheme

[Jan Wedekind]

On Fri, 21 Oct 2016, Amirouche Boubekki wrote:

> Héllo Jan!
>
> On 2016-10-20 22:14, Jan Wedekind wrote:
> > Hi,
> >   I managed to implement JIT compilation of method calls into the
> > Guile interpreter (AIscm 0.8.2 [1]):
> >
> >     (use-modules (oop goops) (aiscm sequence) (aiscm obj) (aiscm jit))
> >     (<< 1 (* 10 (seq <obj> 1 2 4 8)))
> >     ;#<sequence<obj>>:
> >     ;(1024 1048576 1099511627776 1208925819614629174706176)
> >
> >   Types can also be composed using Scheme objects:
> >
> >     (use-modules (oop goops) (aiscm sequence) (aiscm obj) (aiscm
> > complex) (aiscm jit))
> >     (* (seq (complex <obj>) 1+2i 3+4i) 5)
> >     ;#<sequence<complex<obj>>>:
> >     ;(5.0+10.0i 15.0+20.0i)
> >
> >   In the same way native types can be composed:
> >
> >     (use-modules (oop goops) (aiscm sequence) (aiscm rgb) (aiscm int)
> > (aiscm jit))
> >     (make (multiarray (rgb (integer 8 unsigned)) 2) #:shape '(3 2))
> >     ;#<sequence<sequence<rgb<int<8,unsigned>>>>>:
> >     ;(((rgb 160 85 191) (rgb 10 0 0) (rgb 0 0 0))
> >     ; ((rgb 0 0 0) (rgb 0 0 0) (rgb 0 0 0)))
> >
> > Let me know if you have any suggestions or comments.
> >
> > Regards
> > Jan
> >
> > [1] http://wedesoft.github.io/aiscm/
>
> How does this compare with numpy?
>
> TIA!
>
> --
> Amirouche ~ amz3 ~ http://www.hyperdev.fr

  First of all numpy supports native floating point operations and this 
work does not yet do that ;)
  In terms of performance I previously posted some benchmarks comparing 
with C programs [1].
  Numpy is statically compiled and uses generic functions and function 
pointers to implement binary operations for combinations of types [1].

    np.array([1,2,3],dtype=np.uint8) + np.array([1,2,3],dtype=np.int16)
    # array([2, 4, 6], dtype=int16)

  When instead using a JIT (e.g. Python Theano) the resulting software is 
more generic and composable in terms of operations (tensors, composite 
operations) and in terms of datatypes (RGB, complex values, arrays, 
hypercomplex values, ...).
  Here's a technical report [3] with more details and motivation (in that 
case libJIT and Ruby was used).
  Please let me know if you have some interesting applications in mind :)

Regards
Jan

[1] https://lists.gnu.org/archive/html/guile-user/2016-06/msg00019.html
[2] 
https://github.com/numpy/numpy/blob/master/numpy/core/src/multiarray/calculation.c#L534
[3] http://www.wedesoft.de/downloads/tensors_libjit.pdf