[Rd] ALTREP ALTINTEGER_SUM/MIN/MAX Return Value and Behavior
Sebastian Martin Krantz
@eb@@t|@n@kr@ntz @end|ng |rom gr@du@te|n@t|tute@ch
Tue Jun 29 20:26:18 CEST 2021
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 using gmail.com> wrote:
> Also, @Luke Tierney <luke-tierney using 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 using gmail.com>
> wrote:
>
>> 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 using graduateinstitute.ch> wrote:
>>
>>> 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 using gmail.com>
>>> wrote:
>>>
>>> > Adding the dimensions attribute takes away the altrep-ness. Removing
>>> > dimensions
>>> > does not make it altrep. E.g.,
>>> >
>>> > > a <- 1:10
>>> > > am <- a ; dim(am) <- c(2L,5L)
>>> > > amn <- am ; dim(amn) <- NULL
>>> > > .Call("is_altrep", a)
>>> > [1] TRUE
>>> > > .Call("is_altrep", am)
>>> > [1] FALSE
>>> > > .Call("is_altrep", amn)
>>> > [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 using graduateinstitute.ch> wrote:
>>> >
>>> >> 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
>>> >>
>>> >> [[alternative HTML version deleted]]
>>> >>
>>> >> ______________________________________________
>>> >> R-devel using r-project.org mailing list
>>> >> https://stat.ethz.ch/mailman/listinfo/r-devel
>>> >>
>>> >
>>>
>>> [[alternative HTML version deleted]]
>>>
>>> ______________________________________________
>>> R-devel using r-project.org mailing list
>>> https://stat.ethz.ch/mailman/listinfo/r-devel
>>>
>>
[[alternative HTML version deleted]]
More information about the R-devel
mailing list