[R] Fix nthread in DMatrix constructor.

[trivialfis]

Close #7098 .

Restricting the construction of DMatrix to be single threaded.

[hcho3]

Is it possible to make XGDMatrixCreateFromMat_R aware of nthread setting?

[trivialfis]

That's either a new API or a breaking change.

[hcho3]

@trivialfis I don't think changing the signature for XGDMatrixCreateFromMat_R would break any user-facing functionality. We'd just need to modify the content of xgb.DMatrix function in xgb.DMatrix.R. What do you think?

[trivialfis]

I was told that the C function in R is also considered as part of the interface when working on #6819 . That's why the old predict C function for R is still here.

[codecov-commenter]

Codecov Report

Merging #7127 (e5b2e5b) into master (caa9e52) will decrease coverage by 0.51%.
The diff coverage is n/a.

@@            Coverage Diff             @@
##           master    #7127      +/-   ##
==========================================
- Coverage   82.71%   82.19%   -0.52%     
==========================================
  Files          13       13              
  Lines        3963     3999      +36     
==========================================
+ Hits         3278     3287       +9     
- Misses        685      712      +27     

Impacted Files	Coverage Δ
python-package/xgboost/tracker.py	86.33% <0.00%> (-8.77%)	⬇️
python-package/xgboost/core.py	83.93% <0.00%> (ø)
python-package/xgboost/dask.py	82.33% <0.00%> (+0.02%)	⬆️
python-package/xgboost/sklearn.py	89.77% <0.00%> (+0.20%)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update caa9e52...e5b2e5b. Read the comment docs.

[hcho3]

The for loop in lines 96-102 is still using multiple threads. We must revise the loop also.

[trivialfis]

It's actually using the default.

XGDMatrixCreateFromMat_omp_R, that takes in the extra parameter nthread

Maybe a new JSON based API.

[hcho3]

"The default" is set by OMP_NUM_THREADS, not nthread.

[hcho3]

Also, let's create a new API function, XGDMatrixCreateFromMat_omp_R, that takes in the extra parameter nthread, so that R users do not lose out on the ability to use multiple threads when loading from a matrix. The old function XGDMatrixCreateFromMat_R should just call XGDMatrixCreateFromMat_omp_R with nthread=1.

[hcho3]

Let's just change the signature for XGDMatrixCreateFromMat_R, since this function is only used by the R package.