* added CUDA_TOOLKIT_TARGET_DIR CMake variable that used
to locate headers and libraries for target platform
* added CUDA_TARGET_CPU_ARCH which is set to ARM for
arm cross-compilation and is used to add
--target-cpu-architecture NVCC flag
CMake's intended interface for linking to explicit object files (marked
with EXTERNAL_OBJECT) is that only those listed as target sources should
be linked. Drop FindCUDA's attempt to hide the .obj files from VS IDE
project files, which depends on VS-version-specific behavior of linking
custom command outputs that happen to be named "*.obj". CMake puts
external object files in a dedicated source group anyway.
Ubuntu install the CUDA libraries into a location that is different
than the default location provided by the NVidia installer. So we
teach the FindCUDA package to also find the Ubuntu install location.
As of CMake 2.8.8, the INCLUDE_DIRECTORIES() command does
not de-duplicate entries. Failure to do so can lead to an extremely
long and repetitive list of -I entries on the command line.
make CUDA_COMPUTE_BUILD_PATH use a relative path to the current
binary directory instead of the current source directory if the source
file considered is in the current binary directory. This is done to
shorten the paths given to the compiler.
Use normal custom command dependencies by default. Use the PRE_LINK
approach only to work around what seems to be a bug in msbuild handling
of custom build rules that chain together.
In addition to adding the cupti library, find_local_library_first has
been renamed to cuda_find_local_library_first with a backward
compatibility macro to find_local_library_first. Also added
cuda_find_local_library_first_with_path_ext to handle different paths.
This adds a new variable, CUDA_SEPARABLE_COMPILATION, and two new
functions, CUDA_COMPUTE_SEPARABLE_COMPILATION_OBJECT_FILE_NAME and
CUDA_LINK_SEPARABLE_COMPILATION_OBJECTS.
When CUDA_SEPARABLE_COMPILATION is specified then CUDA runtime objects
will be compiled with the separable compilation flag. These object
files are collected in a target named variable that can be used in
CUDA_COMPUTE_SEPARABLE_COMPILATION_OBJECT_FILE_NAME and
CUDA_LINK_SEPARABLE_COMPILATION_OBJECTS.
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.
Instead of directly passing $ENV{SOMEVAR} to a find_* call pass in ENV SOMEVAR.
This will make sure the paths will get correctly handled through different
platforms, especially on Windows.
Also fixes one place where paths with windows delimiters (\) were hardcoded to
use forward slashes.
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
This addresses Bug 11882 which provided a sample implementation for adding
support for cusparse. I went ahead and added all the libraries I thought
appropriate.
This allows you to have more than source file with the same name but different
directories. The intermediate and configuration files are now in this same directory.
Fix a long outstanding bug when a file in the dependency list wasn't found. This bug
wouldn't reset the dependencies, so the makefile would still want the missing file when
building. The work around was to configure twice, but this is no longer necessary.
The FindPackageHandleStandardArgs module was originally created outside
of CMake. It was added for CMake 2.6.0 by commit e118a627 (add a macro
FIND_PACKAGE_HANDLE_STANDARD_ARGS..., 2007-07-18). However, it also
proliferated into a number of other projects that at the time required
only CMake 2.4 and thus could not depend on CMake to provide the module.
CMake's own find modules started using the module in commit b5f656e0
(use the new FIND_PACKAGE_HANDLE_STANDARD_ARGS in some of the FindXXX
modules..., 2007-07-18).
Then commit d358cf5c (add 2nd, more powerful mode to
find_package_handle_standard_args, 2010-07-29) added a new feature to
the interface of the module that was fully optional and backward
compatible with all existing users of the module. Later commit 5f183caa
(FindZLIB: use the FPHSA version mode, 2010-08-04) and others shortly
thereafter started using the new interface in CMake's own find modules.
This change was also backward compatible because it was only an
implementation detail within each module.
Unforutnately these changes introduced a problem for projects that still
have an old copy of FindPackageHandleStandardArgs in CMAKE_MODULE_PATH.
When any such project uses one of CMake's builtin find modules the line
include(FindPackageHandleStandardArgs)
loads the copy from the project which does not have the new interface!
Then the including find module tries to use the new interface with the
old module and fails.
Whether this breakage can be considered a backward incompatible change
in CMake is debatable. The situation is analagous to copying a standard
library header from one version of a compiler into a project and then
observing problems when the next version of the compiler reports errors
in its other headers that depend on its new version of the original
header. Nevertheless it is a change to CMake that causes problems for
projects that worked with previous versions.
This problem was discovered during the 2.8.3 release candidate cycle.
It is an instance of a more general problem with projects that provide
their own versions of CMake modules when other CMake modules depend on
them. At the time we resolved this instance of the problem with commit
b0118402 (Use absolute path to FindPackageHandleStandardArgs.cmake
everywhere, 2010-09-28) for the 2.8.3 release.
In order to address the more general problem we introduced policy
CMP0017 in commit db44848f (Prefer files from CMAKE_ROOT when including
from CMAKE_ROOT, 2010-11-17). That change was followed by commit
ce28737c (Remove usage of CMAKE_CURRENT_LIST_DIR now that we have
CMP0017, 2010-12-20) which reverted the original workaround in favor of
using the policy. However, existing project releases do not set the
policy behavior to NEW and therefore still exhibit the problem.
We introduced in commit a364daf1 (Allow users to specify defaults for
unset policies, 2011-01-03) an option for users to build existing
projects by adding -DCMAKE_POLICY_DEFAULT_CMP0017=NEW to the command
line. Unfortunately this solution still does not allow such projects to
build out of the box, and there is no good way to suggest the use of the
new option.
The only remaining solution to keep existing projects that exhibit this
problem building is to restore the change originally made in commit
b0118402 (Use absolute path to FindPackageHandleStandardArgs.cmake
everywhere, 2010-09-28). This also avoids policy CMP0017 warnings for
this particular instance of the problem the policy addresses.
This puts the new search behaviour for included files in action, i.e.
now when a file from Modules/ include()s another file, it also gets the
one from Modules/ included, i.e. the one it expects.
Alex
CUDA 3.2 on Windows systems changed the layout of the library paths. This adds the extra
directories needed to locate the new files.
I also explicitly disable emulation mode for CUDA 3.1+. This feature was deprecated in
3.0 and ultimately removed in 3.1. The script errors out if CUDA_BUILD_EMULATION is
turned on. I didn't want to ignore emulation mode (even with a warning - which most
people may not even see) and have users confused as to why it wasn't working.
In version 3.0 of the CUDA toolkit when building code for emulation, you need to link
against a new version of the cuda run time library called cudartemu. This CL adds a check
for the new library and uses it when present and in emulation mode. Note that this
library is not present in previous or subsequent versions of the CUDA toolkit.
CUDA_VERSION_MAJOR and CUDA_VERSION_MINOR were only computed when CUDA_VERSION was first
computed. Subsequent runs of FindCUDA would not have CUDA_VERSION_MAJOR/MINOR set. We
now extract the major and minor versions from the CUDA_VERSION cache variable every run.
BUILD_SHARED_LIBS is now only recognized when calling CUDA_ADD_LIBRARY. If you want the CMAKE_SHARED_LIBRARY_C/CXX_FLAGS to be used, pass SHARED as an argument. This prevents -fPIC from being used on objects destined for executables by default.
Cutil was never intented to be used outside of the SDK. The removal of this
code is in support of this. The CUDA_SDK_ROOT_DIR will continue to be
supported, in case users wish to use this to find files in the SDK. There are
also two examples of how to use CUDA_SDK_ROOT_DIR to find header files and
libraries if users so wish.
The CUDA_PROPAGATE_HOST_FLAGS was incorrect in that it prevented the CUDA_NVCC_FLAGS_CONFIG variable from getting filled.
Also, added a search path for the CUDA SDK install on Macs.
- Finished updating and formatting documentation.
- Added CUDA_PROPAGATE_HOST_FLAGS (Default ON) that can disable the C flag
propagation to the host compiler.
_ Changed the output directory for support files from
${CMAKE_CURRENT_BINARY_DIR} to ${CMAKE_CURRENT_BINARY_DIR}/CMakeFiles. This
will hopefully reduce the clutter in the binary directory.
Vladislav Vinogradov
Kitware Robot
Brad King
Robert Maynard
Marcel Loose
Mathias Gaunard
James Bigler
David Cole
Rolf Eike Beer
Alex Neundorf
Kai Wasserbäch
James Bigler