Re: Working patch for FFTW 3.0.x and Nd FFT's

[Paul Kienzle]

On Feb 18, 2004, at 11:30 PM, John W. Eaton wrote:

> So the basic idea is what you suggested, to save a location and set up
> for the signal handler to jump back to this location, then restore any
> previous location that may have been saved, so if we are inside some
> nested calls we can keep jumping back up the stack of saved
> locations (to get this right, you also need to use the macros
>
>   BEGIN_INTERRUPT_WITH_EXCEPTIONS
>
> and
>
>   END_INTERRUPT_WITH_EXCEPTIONS
>
> also defined in quit.h.  For examples of their usage, look at any of
> the functions in Octave that call user-supplied functions (in
> liboctave, LSODE.cc, DASSL.cc, etc.)
>
> So, should we also surround the FFTW code with these immediate
> interrupt enabling macros?

The question I have is whether this should happen implicitly
around every binary function, or around every DEFUN_DLD,
or if the user is required to put it in where it is needed.

The only advantage of putting it only where needed is that it
doesn't cost anything where it isn't needed.  The advantage
of doing it everywhere is that it simplifies the code everywhere.
Making it implicit in DEFUN_DLD is probably a good
compromise, at least for the user-defined extensions.

I timed the following:

  for (int i=0; i < n; i++) {
    BEGIN_INTERRUPT_WITH_EXCEPTIONS;
    op
    END_INTERRUPT_WITH_EXCEPTIONS;
  }

Using total=total+1 as op (total an int) I got the following:

octave:12> t=time; interrupt(100000000); time-t
ans = 1.5228
octave:3> t=time; nointerrupt(100000000); time-t
ans = 0.83887

Using total=total+sin(total):

octave:3> t=time; interrupt(100000000); time-t
ans = 17.290
octave:2> t=time; nointerrupt(100000000); time-t
ans = 15.204

Using total=total+1.

octave:5> t=time; interrupt(100000000); time-t
ans = 2.1333
octave:6> t=time; nointerrupt(100000000); time-t
ans = 2.0113

Using total=total+args(0).double_value()
octave:3> t=time; interrupt(100000000); time-t
ans = 13.244
octave:4> t=time; nointerrupt(100000000); time-t
ans = 10.695


The results are a bit weird, but I believe they show
the cost of the interrupt protection code is trivial
compared to the cost of an oct-file function call.

Paul Kienzle
address@hidden