Proposed speedup of ifelse
Thanks Radford. I concur with all your points. I've attempted to address the issues you raised through the github.io post. The new method appears to be slower for test lengths < 100 and possibly longer lengths (not just < 10). Of course length(test) < 100 is very quick, so I simply added this to the conditions that cause the old ifelse method to be invoked. I'll leave it to R-core to decide whether or not the benefits for longer vectors are worth it.
On Fri, 4 May 2018 at 01:01 Radford Neal <radford at cs.toronto.edu> wrote:
I propose a patch to ifelse that leverages anyNA(test) to achieve an improvement in performance. For a test vector of length 10, the change nearly halves the time taken and for a test of length 1 million, there is a tenfold increase in speed. Even for small vectors, the distributions of timings between the old and the proposed ifelse do not intersect.
For smaller vectors, your results are significantly affected by your invoking the old version via base::ifelse. You could try defining your new version as new_ifelse, and invoking the old version as just ifelse. There might still be some issues with the two versions having different context w.r.t environments, and hence looking up functions in different ways. You could copy the code of the old version and define it in the global environment just like new_ifelse. When using ifelse rather than base::ifelse, it seems the new version is slower for vectors of length 10, but faster for long vectors. Also, I'd use system.time rather than microbenchmark. The latter will mix invocations of the two functions in a way where it is unclear that garbage collection time will be fairly attributed. Also, it's a bit silly to plot the distributions of times, which will mostly reflect variations in when garbage collections at various levels occur - just the mean is what is relevant. Regards, Radford Neal