R interface could use inplace prediction

[david-cortes]

The Python interface where possible uses inplace_predict, which is fater than creating a DMatrix object for making predictions. The R interface OTOH always uses DMatrix, which requires 2 redundant copies of the data. It should be possible to use functions such as XGBoosterPredictFromDense in the R interface too, avoiding an extra data copy, and wrapping them should be even simpler since R does not support as many array configurations as numpy (e.g. no strided array views).

[trivialfis]

Do you think we should add a parameter for the predict function to use inplace prediction optionally? The prediction using DMatrix can be cached as we can use its memory address as the cache key so it's not always slower than inplace prediction.

[david-cortes]

Do you think we should add a parameter for the predict function to use inplace prediction optionally? The prediction using DMatrix can be cached as we can use its memory address as the cache key so it's not always slower than inplace prediction.

Bu whether it's cached or not, it first needs to produce a data copy when it casts the double R array to a float vector, and then when it copies that float vector to the DMatrix that's being created.

[trivialfis]

Would you like to work on the interface? :-). The R package can use a lot of help.

[david-cortes]

I'll pass on this specific feature.