as passing/escaping variables with special characters is far too painful.

is present. CUDA outputs can be directed into a separate directory, which can
be specified via CUDABUILDDIR

a build directory. Project Makefiles can set a BUILDDIR variable to their
build directory - this will be used in the generated target names

to -rpath flags

define here rather than in every project which uses ciftilib (which in turn
uses libxml++)

into a CUDA lib / exe - so that USRNVCCFLAGS, and CUDA include paths are added
to compiler flags for these targets.

support c++17

evaluated against $@="", so was matching all targets

specified in environment. This is to support local development, so that
exes can be executed from $FSLDIR/src/<project>/<exe> and that libs in
$FSLDIR/src/<project>/ will take precedence over those in $FSLDIR/lib/

default since ~2010. Explicitly set rpath to pwd, pwd/../lib, and
$FSLDIR/lib. Set install name for shared libs on macOS so that executables
will refer to them via rpath.

These last two changes are explicitly performed by conda build, but are needed
for stand-alone/development compilation

the $NVCC variable needs to be set - CUDA_VER / CUDA_HOME are set
automatically from nvcc

runtime/toolkit

newlines to make maintenance easier.

Always add -fPIC on all platforms.

For NVCC, use same variable naming conventions as used in conda-forge build
env.

Use -ccbin rather than --compiler-bindir (they are equivalent), as the
conda-forge nvcc wrapper only checks for the former.

Fix path to supportedGencodes.sh script.

flags to be set by the environment, but always set -std option to fix C/C++
standard. Compiler option precedence is 1. Project, 2. ARCH* flags (which
currently just contain -std), and 3. Environment flags.