Link errors lc0 on Pi4

[p2dewit]

When following Pi4 compile description using latest lc0 clone (and full updated buster) results in link errors.[110/133] Linking target lc0 FAILED: lc0 clang++-6.0 -o lc0 lc0.p/src_main.cc.o lc0.p/src_utils_protomessage.cc.o lc0.p/src_engine.cc.o lc0.p/src_version.cc.o lc0.p/src_benchmark_backendbench.cc.o lc0.p/src_benchmark_benchmark.cc.o lc0.p/src_chess_bitboard.cc.o lc0.p/src_chess_board.cc.o lc0.p/src_chess_position.cc.o lc0.p/src_chess_uciloop.cc.o lc0.p/src_mcts_node.cc.o lc0.p/src_mcts_params.cc.o lc0.p/src_mcts_search.cc.o lc0.p/src_mcts_stoppers_common.cc.o lc0.p/src_mcts_stoppers_factory.cc.o lc0.p/src_mcts_stoppers_legacy.cc.o lc0.p/src_mcts_stoppers_alphazero.cc.o lc0.p/src_mcts_stoppers_smooth.cc.o lc0.p/src_mcts_stoppers_stoppers.cc.o lc0.p/src_mcts_stoppers_timemgr.cc.o lc0.p/src_neural_cache.cc.o lc0.p/src_neural_encoder.cc.o lc0.p/src_neural_factory.cc.o lc0.p/src_neural_loader.cc.o lc0.p/src_neural_network_check.cc.o lc0.p/src_neural_network_demux.cc.o lc0.p/src_neural_network_legacy.cc.o lc0.p/src_neural_network_mux.cc.o lc0.p/src_neural_network_random.cc.o lc0.p/src_neural_network_record.cc.o lc0.p/src_neural_network_rr.cc.o lc0.p/src_neural_writer.cc.o lc0.p/src_selfplay_game.cc.o lc0.p/src_selfplay_loop.cc.o lc0.p/src_selfplay_tournament.cc.o lc0.p/src_syzygy_syzygy.cc.o lc0.p/src_utils_commandline.cc.o lc0.p/src_utils_configfile.cc.o lc0.p/src_utils_esc_codes.cc.o lc0.p/src_utils_histogram.cc.o lc0.p/src_utils_logging.cc.o lc0.p/src_utils_optionsdict.cc.o lc0.p/src_utils_optionsparser.cc.o lc0.p/src_utils_random.cc.o lc0.p/src_utils_string.cc.o lc0.p/src_utils_weights_adapter.cc.o lc0.p/src_utils_filesystem.posix.cc.o lc0.p/src_neural_blas_convolution1.cc.o lc0.p/src_neural_blas_fully_connected_layer.cc.o lc0.p/src_neural_blas_se_unit.cc.o lc0.p/src_neural_blas_network_blas.cc.o lc0.p/src_neural_blas_winograd_convolution3.cc.o lc0.p/src_neural_opencl_network_opencl.cc.o lc0.p/src_neural_opencl_OpenCL.cc.o lc0.p/src_neural_opencl_OpenCLTuner.cc.o lc0.p/src_neural_opencl_OpenCLBuffers.cc.o lc0.p/src_neural_shared_activation.cc.o lc0.p/src_neural_shared_winograd_filter.cc.o -flto -Wl,--as-needed -Wl,--no-undefined -Wl,-O1 -pthread -Wl,--start-group /usr/lib/libopenblas.so /usr/bin/../lib/gcc/arm-linux-gnueabihf/8/../../../arm-linux-gnueabihf/libOpenCL.so /usr/lib/arm-linux-gnueabihf/libz.so -Wl,--end-group clang: error: unable to execute command: Segmentation fault clang: error: linker command failed due to signal (use -v to see invocation) [115/133] Compiling C++ object liblc0_lib.a.p/src_neural_opencl_OpenCL.cc.o ninja: build stopped: subcommand failed.

[Naphthalin]

Lc0 requires C++17 now. Minimum compiler versions:
Visual Studio: 2017
g++: v8.0
clang: v5.0
(you probably need to delete build/release after upgrading)

What version of g++ and ninja do you have?

[p2dewit]

g++ version 8.3.0 ninja version 1.10.0

[p2dewit]

clang v6.0.1 Deleting build/release does not help.

[ScallyBag]

Hi,

Lc0 compiles fine on my RPi 4 with gcc-8.3.0 and ninja 1.8.2

Al.

Sent from Scally's iPhone 11 Pro

On 26 Jul 2020, at 12:24, p2dewit notifications@github.com wrote:

 g++ version 8.3.0 ninja version 1.10.0

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub, or unsubscribe.

[p2dewit]

HI, Worked with me fine as well, until i completely redid all the prescribed steps as in the README.md on a fresh SD card with the latest git clone yesterday (25/7) (which gave me the latest versions of meson and ninja and clang v6.0 and the link errors). Pi steps not updated in line with latest changes ?

[borg323]

My guess is you run out of memory during compilation. You can try to do a ninja -j 1 in the build directory.

[p2dewit]

ninja -j 1 gives the same result (linking failed). Memory usage not going much over 10%, so does not seem the issue. Will try on a Pi with more memory though (just to be sure).

Update: Tried on a Pi4 4GB and same result.

[borg323]

Then it may be a clang issue - after all, it is a segmentation fault, not some build error. Can you try building with gcc? You need to clear the build directory, changing compiler after the initial configuration is not supported.

[p2dewit]

Switched to gcc. After resolving the libatomic dependency by adding it to meson.build all went fine and the build was successful.

[vishwanathan-iyer]

@p2dewit Can you please share how to resolve libatomic dependency in meson.build?

[borg323]

Can you try adding the following after line 512 of meson.build: deps += cc.find_library('libatomic', required: false)