Provide co-resident gcc 8 toolchain for CUDA #867
Conversation
madisongh
force-pushed
the
gcc-8-for-cuda-only
branch
2 times, most recently
from
bcf9a9b
to
18896ee
Compare
|
After further testing and examining how Debian handles this, it looks like we can use the main toolchain for everything in most cases, and just pass the old compiler to I've also updated the gcc-8 recipes to version 8.5.0 to pick up all of the available fixes, including some CVE patches that had been back-ported to 8.3.0 in the original recipes. |
madisongh
changed the title
madisongh
force-pushed
the
gcc-8-for-cuda-only
branch
from
ca70458
to
bb3f8f6
Compare
Thanks @madisongh for doing this. |
This is a trimmed-down version of the gcc recipes, modified
to be co-resident with the normal gcc toolchain, for use with
recipes that need to compile CUDA code.
Main differences:
* Recipes are based on the 8.3.0 OE-Core recipes, but the
toolchain is updated to 8.5.0 to pick up the latest
available fixes.
* Binaries are suffixed with the version number, so they
can be installed alongside the normal gcc toolchain.
A version-suffixed directory under ${bindir} is also
created with non-version-suffixed symlinks for the
tools, in case that's needed somewhere.
* No shared libraries for the target are packaged, as we
use the ones from the primary toolchain instead. The
dev package does include static libraries and header files.
Signed-off-by: Matt Madison <[email protected]>
for redirecting the gcc toolchain used by nvcc (and, if necessary, directly compiled CC/C++ code) for compatibility with CUDA. This class can be inherited by recipes that need to compile CUDA-compatible code but do not need the CUDA toolkit dependencies. The class sets variables CC_FOR_CUDA and CXX_FOR_CUDA. Recipes needing to use this version of the compiler must set CC and/or CXX to point to these variables. Signed-off-by: Matt Madison <[email protected]>
* Use CXX_FOR_CUDA to determine the compiler to pass to nvcc * Add support for setting CMAKE_CUDA_ARCHITECTURES based on the CUDA_ARCHITECTURES variable Signed-off-by: Matt Madison <[email protected]>
Signed-off-by: Matt Madison <[email protected]>
Signed-off-by: Matt Madison <[email protected]>
for tegra platforms. Signed-off-by: Matt Madison <[email protected]>
Recent versions of CMake require us to pass just the architecture numbers through a variable setting, so break that out into a separate bitbake variable. Signed-off-by: Matt Madison <[email protected]>
The samples need to be compiled with the older compiler. Signed-off-by: Matt Madison <[email protected]>
* The CXX_FOR_CUDA variable is used to set CUDAHOSTCXX * CMAKE_CUDA_ARCHITECTURES support is added Signed-off-by: Matt Madison <[email protected]>
to set OPENCV_CUDA_DETECTION_NVCC_FLAGS, which tells the OpenCV CUDA module which host compiler to use without generating a warning during the configure step. Signed-off-by: Matt Madison <[email protected]>
Signed-off-by: Matt Madison <[email protected]>
to remove the reference to the gcc-8 toolchain recipes, which are no longer provided. Signed-off-by: Matt Madison <[email protected]>
Now that we can build with the latest toolchain from OE-Core, we no longer need this recipe. Signed-off-by: Matt Madison <[email protected]>
as some recipes in OE-Core now expect a provider for this. Signed-off-by: Matt Madison <[email protected]>
to point to libglvnd. Signed-off-by: Matt Madison <[email protected]>
to add the gcc-8 cross-compiler dependency on linux-libc-headers, mirroring what's done in OE-Core for the main toolchain. Signed-off-by: Matt Madison <[email protected]>
madisongh
force-pushed
the
gcc-8-for-cuda-only
branch
from
49a59cc
to
bec6cb9
Compare
|
Everything looks good with build, SDK, and runtime testing I've done. |
OE-Core now includes Rust support, and that toolchain requires gcc 11 in order to build its standard library. Because of that change, we need to change how we support building CUDA libraries and programs, since CUDA 10.2 is stuck at gcc 8.
This PR proposes that we bite the bullet and provide the gcc 8 toolchain directly in the layer. The recipes are modified to allow this toolchain to co-reside with the default toolchain from the core, and are intended to be used only when CUDA compatibility is required. A new
cuda-gcc.bbclassis provided; recipes that need to be built with the older toolchain can inherit that class (andcuda.bbclassis updated to do that).This allows us to build everything else with the latest toolchain available. We no longer need to force users to switch to the older toolchain at the distro level.