Skip to content

ALTREP ALTINTEGER_SUM/MIN/MAX Return Value and Behavior

10 messages · iuke-tier@ey m@iii@g oii uiow@@edu, Bill Dunlap, Sebastian Martin Krantz +1 more

Original
Sebastian Martin Krantz
# 
Hello together, I'm working on some custom (grouped, weighted) sum, min and
max functions and I want them to support the special case of plain integer
sequences using ALTREP. I thereby encountered some behavior I cannot
explain to myself. The head of my fsum C function looks like this (g is
optional grouping vector, w is optional weights vector):

SEXP fsumC(SEXP x, SEXP Rng, SEXP g, SEXP w, SEXP Rnarm) {
  int l = length(x), tx = TYPEOF(x), ng = asInteger(Rng),
    narm = asLogical(Rnarm), nprotect = 1, nwl = isNull(w);
  if(ALTREP(x) && ng == 0 && nwl) {
    switch(tx) {
    case INTSXP: return ALTINTEGER_SUM(x, (Rboolean)narm);
    case LGLSXP: return ALTLOGICAL_SUM(x, (Rboolean)narm);
    case REALSXP: return ALTLOGICAL_SUM(x, (Rboolean)narm);
    default: error("ALTREP object must be integer or real typed");
    }
  }
// ...
}

when I let x <- 1:1e8, fsum(x) works fine and returns the correct value. If
I now make this a matrix dim(x) <- c(1e2, 1e6) and subsequently turn this
into a vector again, dim(x) <- NULL, fsum(x) gives  NULL and a warning
message 'converting NULL pointer to R NULL'. For functions fmin and fmax
(similarly defined using ALTINTEGER_MIN/MAX), I get this error right away
e.g. fmin(1:1e8) gives NULL and warning 'converting NULL pointer to R
NULL'. So what is going on here? What do these functions return? And how do
I make this a robust implementation?

Best regards,

Sebastian Krantz
iuke-tier@ey m@iii@g oii uiow@@edu
# 
ALTINTEGER_SUM and friends are _not_ intended for use in package code.
Once we get some time to clean up headers they will no longer be
visible to packages.

Best,

luke
On Tue, 29 Jun 2021, Sebastian Martin Krantz wrote:

            

      
      
      
    

        

      
      
    

        

  
    

      
      
    

  


  


      

    

    

      
      

        


  

        

      Bill Dunlap
    

    

      
      #
    

  


  

      
Adding the dimensions attribute takes away the altrep-ness.  Removing
dimensions
does not make it altrep.  E.g.,

            

      
      
      
    

        
[1] TRUE

            

      
      
      
    

        
[1] FALSE

            

      
      
      
    

        
[1] FALSE

where is_altrep() is defined by the following C code:

#include <R.h>
#include <Rinternals.h>

SEXP is_altrep(SEXP x)
{
    return Rf_ScalarLogical(ALTREP(x));
}

-Bill

On Tue, Jun 29, 2021 at 8:03 AM Sebastian Martin Krantz <

        
sebastian.krantz at graduateinstitute.ch> wrote:

        

            

      
      
      
    

        

      
      
    

        

  
  


  


      

    

    

      
      

        


  

        

      Sebastian Martin Krantz
    

    

      
      #
    

  


  

      
Thanks both. Is there a suggested way I can get this speedup in a package?
Or just leave it for now?

Thanks also for the clarification Bill. The issue I have with that is that
in my C code ALTREP(x) evaluates to true even after adding and removing
dimensions (otherwise it would be handled by the normal sum method and I?d
be fine). Also .Internal(inspect(x)) still shows the compact
representation.

-Sebastian

        
On Tue 29. Jun 2021 at 19:43, Bill Dunlap <williamwdunlap at gmail.com> wrote:

        

            

      
      
      
    

        

      
      
    

            

      
      
      
    

        

  
  


  


      

    

    

      
      

        


  

        

      iuke-tier@ey m@iii@g oii uiow@@edu
    

    

      
      #
    

  


  

      
It depends on the size. For a larger vector adding dim will create a
wrapper ALTREP.

Currently the wrapper does not try to use the payload's sum method;
this could be added.

Best,

luke

        
On Tue, 29 Jun 2021, Bill Dunlap wrote:

        

            

      
      
      
    

        

      
      
    

            

      
      
      
    

        

      
      
    

        

  
    

      
      
    

  


  


      

    

    

      
      

        


  

        

      iuke-tier@ey m@iii@g oii uiow@@edu
    

    

      
      #
    

  


  

      
Call the R sum() function, either before going to C code or by calling
back into R. You may only want to do this if the vector is long enough
for e possible savings to be worth while.

        
On Tue, 29 Jun 2021, Sebastian Martin Krantz wrote:

        

            

      
      
      
    

        
When you use a longer vector

            

      
      
      
    

        
A different representation (wrapper around a compact sequence).

Best,

luke

            

      
      
      
    

        

      
      
    

            

      
      
      
    

        

      
      
    

        

  
    

      
      
    

  


  


      

    

    

      
      

        


  

        

      Gabriel Becker
    

    

      
      #
    

  


  

      
Hi Sebastian,

So the way that it is currently factored, there isn't a good way of getting
what you want under the constraints of what Luke said (ALTINTEGER_SUM is
not part of the API).

I don't know what his reason are for saying that per say and would not want
to speak for him, but of the top of my head, I suspect it is because ALTREP
sum methods are allowed to return NULL (the C version) to say "I don't have
a sum method that is applicable here, please continue with the normal
code". So, just as an example, your exact code is likely to segfault, I
think, if you hit an ALTREP that chooses not to implement a sum method
because you'll be running around with a SEXP that has the value NULL (the C
one, not the R one).

One thing you could do, is check for altrepness and then construct and
evaluate a call to the R sum function in that case, but that probably isn't
quite what you want either, as this will hit the code you're trying to
bypass/speedup  in the case where the ALTREP class doesn't implement a sum
methods. I see that Luke just mentioned this as well but I'll leave it in
since I had already typed it.

I hope that helps clarify some things.

Best,
~G


On Tue, Jun 29, 2021 at 10:13 AM Sebastian Martin Krantz <

        
sebastian.krantz at graduateinstitute.ch> wrote:

        

            

      
      
      
    

        

      
      
    

            

      
      
      
    

        

      
      
    

        

  
  


  


      

    

    

      
      

        


  

        

      Gabriel Becker
    

    

      
      #
    

  


  

      
Also, @Luke Tierney <luke-tierney at uiowa.edu>  I can prepare a patch that
has wrappers delegate to payload's ALTREP class methods for things like
sum, min, max, etc once conference season calms down a bit.

Best,
~G

On Tue, Jun 29, 2021 at 11:07 AM Gabriel Becker <gabembecker at gmail.com>
wrote:

            

      
      
      
    

        

      
      
    

            

      
      
      
    

        

      
      
    

            

      
      
      
    

        

  
  


  


      

    

    

      
      

        


  

        

      Sebastian Martin Krantz
    

    

      
      #
    

  


  

      
Thanks Gabriel and Luke,

I understand now the functions return NULL if no method is applicable. I
wonder though why do ALTINTEGER_MIN and MAX return NULL on a plain integer
sequence? I also see that min() and max() are not optimized i.e. min(1:1e8)
appears to materialize the vector.

In general I expect my functions to mostly be applied to real data so this
is not a huge issue for me (I?d rather get rid of it again than calling
sum() or risking that the macros are removed from the API), but it could be
nice to have this speedup available to packages. If these macros have
matured and it can be made explicit that they return NULL if no method is
applicable, or, better, they internally dispatch to a normal sum method if
this is the case, they could become very manageable and useful.

Best,

Sebastian

        
On Tue 29. Jun 2021 at 21:09, Gabriel Becker <gabembecker at gmail.com> wrote:

        

            

      
      
      
    

        

      
      
    

            

      
      
      
    

        

      
      
    

            

      
      
      
    

        

  
  


  


      

    

    

      
      

        


  

        

      Gabriel Becker
    

    

      
      #
    

  


  

      
Hi Sebastian,

min/max do not materialize the vector, you will see it as compact after
same as before. It *does* however do a pass over the data chunked by
region, which is much more expensive than it need be for compact sequences,
that is true.

I think in some version of code that never made it out of the branch, I had
default min/max methods which took sortedness into account if it was known.
One thing that significantly complicated that cod ewas that you have to
find the edge of the NAs(/NaNs for the real case) if narm is TRUE, which
involves a binary search using ELT (or a linear one using
ITERATE_BY_REGION, I suppose).

That said a newer version of the count nas code did get in from a later
patch  to update, so it is available in r-devel and could be used to
revisit that approach.

That aside, it is true that compact sequences in particular never have NAs
so the min and max altrep methods for those classes would be trivial. I
kind of doubt people are creating compact sequences and then asking for the
min/max/mean of them very often in practice.

Best,
~G

On Tue, Jun 29, 2021 at 11:26 AM Sebastian Martin Krantz <

        
sebastian.krantz at graduateinstitute.ch> wrote: