The default prefix map settings have switched to using -ffile-prefix-map,
but the gcc 10 compiler we're building here needs the old -fmacro-prefix-map
and -fdebug-prefix-map settings, so add them back in for these recipes.
Signed-off-by: Matt Madison <matt@madison.systems>
nvcc needs the cicc binary provided in the cuda-nvvm package, so
make sure that gets installed with the host-side SDK tools.
Signed-off-by: Matt Madison <matt@madison.systems>
OE-Core master has udpated how the toolchain virtual providers
are handled, so update the layer accordingly.
This affects the gcc-for-nvcc extra toolchain build we have for
CUDA builds, and a few other recipes that have some explicit
references to the old toolchain virtual/<something> names.
Signed-off-by: Matt Madison <matt@madison.systems>
sinpi/cospi/sinpif/cospif are now defined in glibc as part of C23 extentions
but they conflict with CUDA declrations, therefore use the glibc definitions
when its provided by glibc
Fixes https://github.com/OE4T/meta-tegra/issues/1784
Signed-off-by: Khem Raj <raj.khem@gmail.com>
Looks like VPI is dynamically loading one of the libnpp libraries
at runtime, rather than linking against it, so make sure it gets
pulled in.
Signed-off-by: Matt Madison <matt@madison.systems>
* Fixes for the core and plugins-prebuilt recipes
* Licensing is now all Apache-2.0
* Rework the patch for the samples so they compile cleanly
Signed-off-by: Matt Madison <matt@madison.systems>
to take advantage of the automatic license whitelisting buried in
base.bbclass that checks for PN starting with 'gcc-source-', and
avoid the issue mentioned in #1711.
Signed-off-by: Matt Madison <matt@madison.systems>
CUDA 12 added 32-bit x86 tools and libraries to the cuda-sanitizer
package. They're not likely needed in our builds or SDKs, so just
remove them to avoid packaging issues.
Signed-off-by: Matt Madison <matt@madison.systems>
Add needed dependencies and linker flags as well as copy the data files
to target, needed to run c++11_cuda sample, in order to run this the
limitation is that one must be in /usr/bin/cuda-samples subdir when
executing this sample, so it can find warandpeace.txt under ./data dir
Signed-off-by: Khem Raj <raj.khem@gmail.com>
nvcc needs nvvm-dev and crt-dev when installed on the target.
It really does need those -dev packages, so relax the dev-deps
packaging QA check.
Signed-off-by: Matt Madison <matt@madison.systems>
Move all of the extras/CUPTI files into the -dev package, and
add in missing runtime dependencies for the -dev package, which
has makefiles and a perl script.
Signed-off-by: Matt Madison <matt@madison.systems>
There is now a check that ${S} exists after running do_unpack, so
append the extraction of the sources from the debian package to
the do_unpack task, instead of breaking it out as a separate
task.
Signed-off-by: Matt Madison <matt@madison.systems>
Like the upstream gcc recipes, we can use the global DEBUG_PREFIX_MAP
setting, but for this older compiler we have to remove '-fcanon-prefix-map'.
Signed-off-by: Matt Madison <matt@madison.systems>
* No more 'S = "${WORKDIR}"', which is not permitted
* All ${WORKDIR} references for local files updated to ${UNPACKDIR}
* Custom do_unpack for the recipes that are in the public_sources
archive in the BSP adjusted for the new unpacking approach
* Some SRC_URIs updated with 'destsuffix=' parameters where needed
Signed-off-by: Matt Madison <matt@madison.systems>
The use of ${SRCPV} in recipe version (PV) strings has
been discontinued in OE-Core; source revisions are now
inserted directly into task hashes and into PKGV, and
SRCPV is a null string.
Signed-off-by: Matt Madison <matt@madison.systems>
Update the CUDACXXARCHFLAGS environment variable to include
the same glibc 2.38 workaround to SDKs that was added to
cuda.bbclass with 52f994fbb3.
Signed-off-by: Matt Madison <matt@madison.systems>
