Use __CUDA__ macro with __NVCC__.

[trivialfis]

• __CUDA__ is defined in clang. Making the change won't make clang
  compile xgboost, but syntax checking for *.cu from clang is at least partially
  working.

[codecov-io]

Codecov Report

Merging #3539 into master will increase coverage by 0.09%.
The diff coverage is n/a.

@@             Coverage Diff              @@
##             master    #3539      +/-   ##
============================================
+ Coverage     45.64%   45.74%   +0.09%     
  Complexity      188      188              
============================================
  Files           166      166              
  Lines         13234    13233       -1     
  Branches        445      445              
============================================
+ Hits           6041     6053      +12     
+ Misses         6989     6976      -13     
  Partials        204      204

Impacted Files	Coverage Δ	Complexity Δ
include/xgboost/base.h	100% <ø> (ø)	0 <0> (ø)	⬇️
...ala/org/apache/spark/SparkParallelismTracker.scala	89.18% <0%> (-0.29%)	9% <0%> (ø)
src/data/sparse_page_dmatrix.cc	92.89% <0%> (+0.59%)	0% <0%> (ø)	⬇️
src/data/sparse_page_raw_format.cc	64.15% <0%> (+22.64%)	0% <0%> (ø)	⬇️

---

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 6cf97b4...e1a4292. Read the comment docs.

[hcho3]

Can you elaborate your reasoning for replacing __NVCC__ with __CUDA__? Does this improve compatibility?

[trivialfis]

@hcho3 To some degree. __CUDA__ is a macro defined by clang to indicate the current scope, __NVCC__ is a similar macro defined by NVCC. As you can see from the patch, it's not really replacing, just checking for the right one to use. With this patch clang can compile a larger portion of the code, but still not the whole project.

But the reason for myself is I can use better syntax checking and code static analyse tools from clang. Now I can see error/warning and do code navigation for CUDA interactively in my editor just like normal c++ code. NVCC provides very primitive error handling facility and confusing error message.

So the patch is mainly for easy development. :)

[RAMitchell]

I'm not sure I like substituting macros with more macros. It can be a little confusing. What about doing this:

#if defined(__NVCC__) || defined(__CUDA__)

[trivialfis]

@RAMitchell Done.