(j3.2006) Pre-Processed Programs Poorly Designed? (Was: Select Rank)
Wed Feb 18 17:36:36 EST 2015
> On Feb 18, 2015, at 12:38 PM, Craig Dedo CAPM <craig at ctdedo.com> wrote:
> I disagree. There are several situations that require some kind of pre-processing, much as I dislike it. One of the more important is that the Fortran standard does not require any specific kind of any data type. This is explicitly left as processor-dependent.
I also find include files and pre-processing indispensable for another processor-dependent issue: support for the more recent standards. Based on the compiler being used (a choice that I hardwire into Makefiles or detect automatically in CMake files), my build system copies pre-processor definitions into the source directory. For example, if I?m using gfortran, then my make or CMake files to do the following:
cp include_path/gfortran_capabilities.txt source_path/compiler_capabilities.txt
Any source files with features that aren?t universally supported has the line
where the included file contains lines like
The above three are the most common cases of processor-dependence in my current codes. (I don?t have a reasonable workaround for missing coarray support, which I use in most of my projects.) The last #define above so important from a performance standpoint that I recommend users write their own collectives matching the proposed Fortran 2015 syntax and then use something like
use my_collectives_module, only : co_sum
This minimizes code changes once more compilers provide collective subroutines. (Currently only gfortran and Cray do.) Until the compiler landscape is more even in its support for recent standards, I don?t know a good alternative to the above approach other than abandoning the relevant features altogether.
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the J3