On Sun, Aug 15, 2010 at 10:34 AM, Douglas Bates<bates at stat.wisc.edu>
wrote:
On Sun, Aug 15, 2010 at 3:39 AM, Romain Francois
<romain at r-enthusiasts.com> wrote:
Hi,
Could you provide some minimal example using inline, etc ...
I enclose a comparison of several different versions that I wrote.
Interestingly the old code, called by fx3, does reasonably well,
although I'm not sure if this is a fair comparison as the levels of
optimization by the compiler may not be the same for the code
generated by inline and for the code that is part of R.
As we have seen in the past, extracting the pointers and calling
scalar functions on individual elements does well. The direct use of
the sugar functions is much easier to understand in the code (it is
essentially the same as writing R expressions) except that ifelse() is
expensive in R but not as a sugar expression.
Here are a few hints in the meantime. In the next version of Rcpp, I've
added a few things that help generating sugar functions. This is how for
example the sugar version of choose is currently implemented :
SUGAR_BLOCK_2(choose , ::Rf_choose )
The SUGAR_BLOCK_2 macro (perhaps I should prefix it with RCPP_) lives in
Rcpp/sugar/SugarBlock.h
You can promote y_log_y to a sugar function like this :
inc<- '
double y_log_y(double y, double mu){
return (y) ? (y * log(y/mu)) : 0;
}
SUGAR_BLOCK_2( y_log_y , ::y_log_y )
'
I added a couple of other tests based on SUGAR_BLOCK_2 and cleaned up
the logic a bit. I enclose the results.
fx<- cxxfunction( signature( y = "numeric", mu = "numeric" ), '
NumericVector res = y_log_y(
NumericVector(y),
NumericVector(mu)
) ;
return res ;
', plugin = "Rcpp", includes = inc )
fx( runif(11) , seq(0, 1, .1 ) )
This gives you 3 Rcpp::y_log_y functions :
- one for when y is a double and mu is a NumericVector (or any numeric
sugar
expression)
- one for when y is a NumericVector and mu is a double
- one for both are NumericVector.
However, at the moment it does not take care about recycling so it is up
to
the user to make sure that y and mu have the same length. I'm currently
thinking about how to deal with recycling.
rep is a possibility here, but consider this hint in the TODO file:
o not sure rep should be lazy, i.e. rep( x, 4 ) fetches x[i] 4
times,
maybe we should use LazyVector like in outer to somehow cache the
result when it is judged expensive to calculate
One other thing about sugar is that it has to do many checks for missing
values so if (with the one defined above) you did call this expression:
2 * wt * (y_log_y(y, mu) + y_log_y(1.-y, 1.-mu))
(and again, currently binary operators +,*,.. don't take care of
recycling)
you would have many checks for missing values. That is fine if you may
have
missing values, because it will propagate them correctly, but it can
slow
things down because they are tested for at every step.
If you are sure that y and mu don't contain missing values, then perhaps
one
thing we can do is embed that information in the data so that sugar does
not
have to check for missing values because it just assumes there are not
any.
Most of the code in sugar contains version that ignores missing values,
controlled by a template parameter. For example seq_len creates a sugar
expression where we know for sure that it does not contain missing
values.
One way perhaps to short circuit this is to first write the code with
double's and then promote it to sugar:
double resid( double y, double mu, double w){
return 2 * wt * (y_log_y(y, mu) + y_log_y(1.-y, 1.-mu) ;
}
But then you need someone to write SUGAR_BLOCK_3 or write it manually.
Another idea would be to have something like
NumericVector::import_transform
but that would take 3 vector parameters instead of 1.
Sorry if this email is a bit of a mess, I sort of wrote the ideas as
they
came.
Romain
Le 14/08/10 20:28, Douglas Bates a ?crit :
In profiling code for generalized linear mixed models for binary
responses I found that a substantial portion of the execution time was
being spend in evaluating the R functions for inverse links and for
the deviance residuals. As I result I wrote C code (in
$RSRC/src/library/stats/src/family.c) or some of those.
The way that some research is going I expect that I will soon be in
the position where I need to evaluate such expressions even more
frequently so it is worthwhile to me to tune it up.
In general there will be three NumericVector objects -- y (observed
responses), wt (weights) and eta (linear predictors) -- plus an
IntegerVector ind. y, wt and ind will all have length n. eta's
length can be a multiple of n.
The index vector, ind, is a factor with k< n levels so all of its
elements are between 1 and k.
The objective is to apply the inverse link function to eta, producing
the predicted mean response, mu, evaluate the deviance residuals for
y, mu and wt and sum the deviance residuals according to the indices
in ind.
The simplest inverse link is for the logit link
NumericVector mu = 1./(1. + exp(-eta));
For the deviance residuals I defined an helper function
static R_INLINE
double y_log_y(double y, double mu)
{
return (y) ? (y * log(y/mu)) : 0;
}
which has the appropriate limiting behavior when y is zero. Using
that, the deviance residuals can be evaluated as
2 * wt * (y_log_y(y, mu) + y_log_y(1.-y, 1.-mu))
That last expression could have different lengths of y and mu but I
could use rep to extend y, wt and ind to the desired length.
The conditional evaluation in y_log_y is not something that can be
ignored because, in many cases y consists only of 0's and 1's so
either y_log_y(y, mu) or y_log_y(1.-y, 1.-mu) will fail the condition
in the ? expression.
Are there suggestions on how best to structure the calculation to make
it blazingly fast?