[Rd] 1.6x speedup for requal() function (in R/src/main/unique.c)

[Duncan Murdoch]

On 11-12-01 8:40 PM, Hervé Pagès wrote:
> Hi,
>
> FWIW:
>
> /* Taken from R/src/main/unique.c */
> static int requal(SEXP x, int i, SEXP y, int j)
> {
>       if (i<  0 || j<  0) return 0;
>       if (!ISNAN(REAL(x)[i])&&  !ISNAN(REAL(y)[j]))
>           return (REAL(x)[i] == REAL(y)[j]);
>       else if (R_IsNA(REAL(x)[i])&&  R_IsNA(REAL(y)[j])) return 1;
>       else if (R_IsNaN(REAL(x)[i])&&  R_IsNaN(REAL(y)[j])) return 1;
>       else return 0;
> }
>
> /* Between 1.34x and 1.37x faster on my 64-bit Ubuntu laptop */
> static int requal2(SEXP x, int i, SEXP y, int j)
> {
>       double xi, yj;
>
>       if (i<  0 || j<  0) return 0;
>       xi = REAL(x)[i];
>       yj = REAL(y)[j];
>       if (!ISNAN(xi)&&  !ISNAN(yj)) return xi == yj;
>       if (R_IsNA(xi)&&  R_IsNA(yj)) return 1;
>       if (R_IsNaN(xi)&&  R_IsNaN(yj)) return 1;
>       return 0;
> }

That looks like a valid improvement.

>
> /* Another extra 1.18x speedup. So overall requal3() is about 1.6x
>      faster than requal() for me. requal3() uses a simpler logic than
>      requal() but this logic should be equivalent to the logic used
>      by requal(), based on the following facts:
>        (a) If *one* of xi or yi is a number (i.e. not NA or NaN),
>            then xi and yi can be compared with xi == yi. They don't
>            need to *both* be numbers for this comparison to be valid.
>        (b) Otherwise (i.e. if each of them is not a number) then each
>            of them is either NA or NaN (only 2 possible values for
>            each), so comparing them with R_IsNA(xi) == R_IsNA(yj)
>            should do the trick. */

I think this one is probably correct, but it's too tricky for my taste.

> static int requal3(SEXP x, int i, SEXP y, int j)
> {
>       double xi, yj;
>
>       if (i<  0 || j<  0) return 0;
>       xi = REAL(x)[i];
>       yj = REAL(y)[j];
>       if (!ISNAN(xi) || !ISNAN(yj)) return xi == yj;
>       return R_IsNA(xi) == R_IsNA(yj);
> }

Duncan Murdoch

>
> The logic of the cequal() function (in the same file) could also be
> cleaned up in a similar way, probably for an even greater speedup.
>
> This will benefit duplicated(), anyDuplicated() and unique() on numeric
> and complex vectors.
>
> Cheers,
> H.
>