The target properties `CUDA_SEPARABLE_COMPILATION` and `CUDA_PTX_COMPILATION`
now aren't mutually exclusive and can now be used together on the same
target.
Some of the things are set in Platform/NVIDIA-CUDA and since we aren't setting
them to different values there's no point in overriding them.
Also fixed an unset() copy-paste error.
Since commit 0d0145138f (CUDA: Add abstraction for cuda runtime
selection, 2019-11-29, v3.17.0-rc1~83^2) we add CUDA runtime library
selection flags by default.
To maintain backwards compatibility the default CUDA runtime
library needs to be computed based on what libraries are found
on the initial compiler invocation. For example a toolchain
could establish initial flags that have all CUDA compilations
using the runtime version, and if we don't detect this we will
try to link to both the static and shared runtime.
Co-Author: Brad King <brad.king@kitware.com>
Fixes: #20708
When crosscompiling we pass the sysroot.
We need to try various architecture flags. Clang doesn't automatically
select one that works. First try the ones that are more likely to work
for modern installations:
* <=sm_50 is deprecated since CUDA 10.2, try sm_52 first for
future compatibility.
* <=sm_20 is removed since CUDA 9.0, try sm_30.
Otherwise fallback to Clang's current default. Currently that's `sm_20`,
the lowest it supports.
Separable compilation isn't supported yet.
Fixes: #16586
properties LINK_OPTIONS and INTERFACE_LINK_OPTIONS are propagated
to the device link step.
To control which options are selected for normal link and device link steps,
the $<DEVICE_LINK> and $<HOST_LINK> generator expressions can be used.
Fixes: #18265
`link.exe /lib` is an undocumented flag and it just calls `lib.exe`.
Also `link.exe` doesn't parse the `/lib` option correctly when in a
response file.
Fixes#17559
Replace our hard-coded default of cudart=static with a first-class abstraction to select the runtime library from an enumeration of logical names.
Starting with CUDA 10.2 the nvcc compiler has gained support
to automatically forward unknown flags to the host compiler.
This behavior is highly desired as projcts that mix CUDA, C, C++
run into situation where flags such as `-pthread` which aren't
supported by nvcc, are being applied to all source files and
therefore break CUDA compilation.
In commit fb3370b6a1 (MSVC: Add abstraction for runtime library
selection, 2019-04-10, v3.15.0-rc1~229^2) we overlooked updating flags
for CUDA on Windows, where nvcc uses MSVC as the host compiler. Add
them now and update the MSVCRuntimeLibrary test to cover CUDA.
Fixes: #19428
We do not add default warning flags on other compilers, and having
a warning flag in the default flags makes it hard for projects to
customize the warning level. They need to use string processing
to remove `/W3` from `CMAKE_{C,CXX}_FLAGS`. Therefore we should
drop it.
However, projects may be using string processing to replace `/W3`
with another flag, so we cannot simply drop it. Add a policy to
drop it in a compatible way.
Fixes: #18317
CMake commands vs_link_dll and vs_link_exe, performing linking on MSVC,
are responsible for calling resource compiler and manifest tool.
Before this commit, both of these tools were called directly, with the
expectation that they are available in the `PATH`. This has been fixed
by respecting CMake variables `CMAKE_RC_COMPILER` and `CMAKE_MT`
defining paths to these tools.
Fixes: #17804
Use `curand_static` to identify directories containing device libraries
because `cublas_device` is deprecated and will be removed in the future.
Issue: #18290
The `<CMAKE_CUDA_LINK_FLAGS>` placeholder in CUDA rule variables comes
from the `<CMAKE_CXX_LINK_FLAGS>` placeholder from which the CUDA rule
variables were originally derived. It is not a public interface for
adding link flags so no projects should be using it. It is needed for
platform information modules to specify platform-specific link flags for
the language. None of our platform modules set it, so it is unused.
Furthermore, it is broken as currently implemented. Some of the
contexts in which it is used need `-Xlinker` and some do not.
Therefore it is not possible to use the placeholder at all.
Simply remove it for now. If some need for platform-specific CUDA link
flags arises a new solution will be needed.
In some environments, libraries containing device code are installed in
directories searched by the host linker but not passed by nvcc to
nvlink. Make these libraries available by explicitly passing the
host linker search directories during device linking.
Issue: #16317
The CUDA compiler automatic deprecation warnings are pure noise when
doing device linking, and should be suppressed to reduce the amount of
confusion from users.
Fix the default values of `CMAKE_CUDA_FLAGS[_<CONFIG>]` on Windows to
make the host compiler flags match those produced for C++ by the
`Platform/Windows-MSVC` module. This makes the flags consistent with
those used for C++.
On Windows with MSVC-like host compilers we must honor the standard
libraries chosen by the `Platform/Windows-MSVC` module. Otherwise C
code linked into the CUDA binary that expects to have these libraries
available may not link.
Previously we had a two issues when building cuda executables
that required separable compilation. The first was that we didn't
propagate FLAGS causing any -arch / -gencode flags to be dropped, and
secondly generators such as ninja would use the CXX language flags
instead of CUDA when the executable was mixed language.
Port Windows-specific compilation and linking rules over from the
`Platform/Windows-MSVC` module and adapt it for NVIDIA CUDA. On Windows
nvcc and its host compiler (MSVC) do not understand or use options like
`-fPIC` or `-std=`, so condition those out.
