Since commit v3.12.0-rc1~332^2 (FindCUDA/select_compute_arch: Add
support for CUDA as a language, 2018-03-15) this module sets
`CUDA_VERSION` based on `CMAKE_CUDA_COMPILER_VERSION` when the language
is enabled. Limit it to two components for consistency with the
normal `FindCUDA` documentation and behavior.
Fixes: #18231
Fix version comparisons to handle patch components. List and check
known archs for each version of CUDA so mismatching versions are not
suggested.
Fixes: #17921
Some are user facing.
Found using
codespell -q 3 --skip="./Utilities" -I .cmake-whitelist.txt`
whereby the whitelist contained:
ans
dum
helpfull
emmited
emmitted
buil
iff
isnt
nto
ot
pathes
substract
te
todays
upto
whitespaces
Set `CMP0007` to `NEW` for the entire script. The script comes with
CMake and is aware of the policy's behavior, so set the policy to ensure
that its warnings do not show up during the build phase.
Fixes: #16579
Fixes issue where include directories specified on the target are
not passed on to NVCC. This includes both target_include_directories()
as well as include directories added by dependency chaining.
Closes: #14201
Previously if a non-existent dependency file is found we set the file to "" and
then do if(NOT IS_DIRECTORY "${file}"). Later we call get_filename_component
on the empty file which returns basically the current build directory.
Having a dependency on the current build directory is really annoying, because
anything that compiles into that directory will change the file stamp and
cause your files to rebuild every time you call make. :(
I was previously appending to nvcc_flags inside the file loop. This
caused the flag to be appended multiple times which freaks out nvcc.
Now the flag is specifically handled per file.
When setting default CUDA_HOST_COMPILER we must dereference CMAKE_C_COMPILER,
i.e. /usr/bin/clang should be used instead /usr/bin/cc which is symlink.
Otherwise CUDA thinks it is GCC and issues -dumpspecs which is unknown option
to Clang.
Also in case neither CMAKE_C_COMPILER is defined (project does not use C
language) nor CUDA_HOST_COMPILER is specified manually we should skip -ccbin
and let nvcc use its own default C compiler.
The matches have already been calculated and can simply be taken from
CMAKE_MATCH_n variables. This avoids multiple compilations of the same or very
similar regular expressions.
Added a new CUDA variable for specifying the CUDA_HOST_COMPILER. This will allow users to
be able to specify which host compiler to use for invoking NVCC with. By default it will
use the compiler used for host compilation. This is convenient for when you want to
specify a different compiler than the default compiler. You end up using the same
compiler for both the NVCC compilation and the host compilation instead of using the
default compiler in the path for NVCC.
Ancient versions of CMake required else(), endif(), and similar block
termination commands to have arguments matching the command starting the
block. This is no longer the preferred style.
Run the following shell code:
for c in else endif endforeach endfunction endmacro endwhile; do
echo 's/\b'"$c"'\(\s*\)(.\+)/'"$c"'\1()/'
done >convert.sed &&
git ls-files -z -- bootstrap '*.cmake' '*.cmake.in' '*CMakeLists.txt' |
egrep -z -v '^(Utilities/cm|Source/kwsys/)' |
egrep -z -v 'Tests/CMakeTests/While-Endwhile-' |
xargs -0 sed -i -f convert.sed &&
rm convert.sed
Nvcc can emit '/path' instead of '//path' which can cause a lot of grief later. We test
to see if the file exists, if it doesn't then we see if the file exists with '/'
prepended. Files that don't exist won't be added to the list.
