BUG: Cannot cross-compile Numpy to OpenHarmony OS (AArch64) platform. #28861
Comments
|
The cross-file approach is preferred I think. The cross build is correctly configured: The first failure: The file extension for Python extension modules is I think the fix is to add The more robust but slower way of doing it is running the build as a native build under QEMU rather than as a cross build. |
|
@rgommers Really, really, really thank you! The problem that has been bothering me for almost a week is finally solved! Following your solution, I changed the meson cross file (specifying python as cross-python in crossenv, and changing the linking options that were problematic before): [constants]
ohos_sdk = '/home/xhw/Desktop/OH/oh-cross-toolchain/ohos-sdk/14'
toolchain = ohos_sdk / 'native/llvm/bin'
sysroot = ohos_sdk / 'native/sysroot'
host_cpu = 'aarch64'
host_arch = 'aarch64'
dist_root = '/home/xhw/Desktop/OH/oh-edu-python/dist.aarch64'
common_flags = ['--sysroot=' + sysroot, '--target=' + host_cpu + '-linux-ohos']
common_c_flags = ['-fPIC', '-D__MUSL__=1', '-I' + dist_root/'include', '-I' + dist_root/'include/lzma', '-I' + dist_root/'include/ncursesw', '-I' + dist_root/'include/readline', '-I' + dist_root/'ssl/include']
common_ld_flags = ['-fuse-ld=lld', '-L' + dist_root/'lib', '-L' + dist_root/'ssl/lib64', '-lpython3.11']
[built-in options]
c_args = common_flags + common_c_flags
cpp_args = common_flags + common_c_flags
c_link_args = common_flags + common_ld_flags
cpp_link_args = common_flags + common_ld_flags
[properties]
cmake_toolchain_file = ohos_sdk / 'native/build/cmake/ohos.toolchain.cmake'
sizeof_long_double = 8
longdouble_format = 'IEEE_DOUBLE_LE'
[binaries]
c = toolchain / 'clang'
cpp = toolchain / 'clang++'
python = '/home/xhw/Desktop/OH/oh-python-pkgs/crossenv_aarch64/cross/bin/python3'
cython = ''
cython3 = cython
as = toolchain / 'llvm-as'
ld = toolchain / 'ld.lld'
c_ld = ld
cpp_ld = ld
lld = toolchain / 'ld.lld'
strip = toolchain / 'llvm-strip'
ranlib = toolchain / 'llvm-ranlib'
objdump = toolchain / 'llvm-objdump'
objcopy = toolchain / 'llvm-objcopy'
readelf = toolchain / 'llvm-readelf'
nm = toolchain / 'llvm-nm'
ar = toolchain / 'llvm-ar'
profdata = toolchain / 'llvm-profdata'
[host_machine]
system = 'ohos'
kernel = 'linux'
cpu_family = host_cpu
cpu = host_cpu
endian = 'little'and compiled with crossenv. I ran the script as follows: $BUILD_PIP install crossenv
$BUILD_PYTHON -m crossenv \
$HOST_PYTHON \
crossenv_aarch64
. crossenv_aarch64/bin/activate
cd numpy
VENDORED_MESON=${p_DIR}/numpy/vendored-meson/meson/meson.py
python ${VENDORED_MESON} setup --reconfigure --prefix=${p_DIR}/dist --cross-file ../ohos-build.meson.ini build-ohos
cd build-ohos
python ${VENDORED_MESON} compile
python ${VENDORED_MESON} installthe problem was solved!
|
|
Great, glad you solved it and thanks for sharing the solution! |
|
Worth sticking into a wiki somewhere? |
We have docs on cross-compilation, I'll make a quick update to that. |
Problem: Cannot cross-compile Numpy to OpenHarmony OS (AArch64) platform.
Background
Hello! I currently need to cross-compile the numpy library to the OpenHarmony AArch64 platform. It uses a modified linux kernel and musl libc, so I can't just download the officially compiled binary.
But I'm having a lot of problems compiling it.
Checkings
Before trying to cross-compile numpy, I checked the following:
Try 1: Use Crossenv + pip
I then proceeded to prepare the environment as follows:
The log file for the above command is here: Log 1: Paste Bin.
In the crossenv environment I executed the following command:
And the log file: Log 2: Paste Bin;
Looks like it worked, right? The binary suffix of the output indicates that the architecture is correct.
But this is problematic:
readelfoutputs thex86_64architecture instead:This method didn't work, so I tried another.
Try 2: Use meson + Cross Compile
I then quit the crossenv environment, downloaded the numpy source code and tried again:
I defined my own
ohos-build.meson.iniconfiguration file as follows:It still failed, the logs are as follows (it seems that x86_64 was incorrectly used as the target platform):
Log 3: Paste Bin;
It seems that vendored-meson can't have failed to deal with the cross-compilation case?
So how should I cross-compile numpy correctly? Thanks a lot!
The text was updated successfully, but these errors were encountered: