A faster version of an application type (i.e. a multivariate version)

One of the major performance bottlenecks jsonlite

turns out to be applicable, so I was wondering if there are more efficient alternatives for applying functions on rows or columns of a matrix.

For one-dimensional structs, it is vapply

often significantly faster than lapply

or sapply

, since it can allocate output space at the beginning, rather than trying to simplify the output lists at the very end:

> x <- runif(5e6)
> system.time(y1 <- sapply(x, sqrt))
   user  system elapsed 
  7.231   0.142   7.370 
> system.time(y2 <- vapply(x, sqrt, numeric(1)))
   user  system elapsed 
  1.314   0.095   1.410

      

        
        
        
      

    

However vapply

does not work for matrix structures. Is there a similar feature for the family apply

? Currently, the type of output apply

depends on FUN

, for example:

#returns vector
apply(volcano, 1, mean)

#returns matrix
apply(volcano, 1, fivenum)

#returns list
apply(volcano, 1, function(x){list(mean=mean(x), sd=sd(x))})

      

        
        
        
      

    

This script has a simple example apply

in action.