Use single precision math functions for CUDA variants

[MichaelSt98]

Use single precision math functions for CUDA variants when compiling for single precision, thus:

• pow -> powf
• fabs -> fabsf

[reuterbal]

Looks straightforward. How big is the expected performance gain?

[MichaelSt98]

Looks straightforward. How big is the expected performance gain?

bin/dwarf-cloudsc-c-cuda-k-caching 1 262144 128
     NUMOMP=1, NGPTOT=262144, NPROMA=128, NGPBLKS=2048
     NUMOMP    NGPTOT  #GP-cols     #BLKS    NPROMA tid# : Time(msec)  MFlops/s     col/s
          1    262144    262144         0       128    0 :         14   2235400  17908522 @ core#
          1    262144    262144      2048       128   -1 :       1723     18987    152116 TOTAL

vs. with the "correct" math functions:

bin/dwarf-cloudsc-c-cuda-k-caching 1 262144 128
     NUMOMP=1, NGPTOT=262144, NPROMA=128, NGPBLKS=2048
     NUMOMP    NGPTOT  #GP-cols     #BLKS    NPROMA tid# : Time(msec)  MFlops/s     col/s
          1    262144    262144         0       128    0 :         10   3146297  25206016 @ core#
          1    262144    262144      2048       128   -1 :       1729     18924    151607 TOTAL

[reuterbal]

Very nice!