pika
pika copied to clipboard
pika-org
pika builds on C++ std::execution with fiber, CUDA, HIP, and MPI support.
The runtime configuration is currently in a pretty messy state with many ad-hoc workarounds. It parses configurations multiple times, environment variables provide defaults instead of overriding configuration options, and many...
Allow setting `CXX` and `CUDA` standards to different values. This may help work around problems with `nvcc` and C++20 (#303).
Same as #302, but for nvcc/nvc++. - [x] nvc++ (#606) - [ ] nvcc (requires CUDA 12)
This is not ok: ``` > ls -lh test/unit/eigensolver/test_tridiag_solver_full -rwxr-xr-x 1 simbergm csstaff 1.5G Sep 6 11:49 test/unit/eigensolver/test_tridiag_solver_full ```
Using the default scheduler blank gaps appear in the trace when one of the threads is scheduling the tasks. The gaps disappear immediately after the scheduling task finished. 
Such as these: https://cdash.cscs.ch/viewTest.php?onlyfailed&buildid=88467 (segfaults in `split_continuation_clear`, `thread_pool_scheduler`, `condition_variable2`).
It fails at least on the GCC 12 configuration on Piz Daint with the following: ``` {os-thread}: pika/worker-thread/pool:worker/global:3/local:2 {what}: mmap() failed to allocate thread stack: pika(unhandled_exception) ``` The test is...
https://docs.nvidia.com/hpc-sdk/compilers/hpc-compilers-user-guide/index.html#acc-mem-unified
``` ld: error: undefined symbol: environ >>> referenced by unity_0_cxx.cxx >>> libs/pika/init_runtime/CMakeFiles/pika_init_runtime.dir/Unity/unity_0_cxx.cxx.o:(pika::init(std::__1::function, int, char const* const*, pika::init_params const&)) >>> referenced by unity_0_cxx.cxx >>> libs/pika/init_runtime/CMakeFiles/pika_init_runtime.dir/Unity/unity_0_cxx.cxx.o:(pika::init(std::__1::function, int, char const* const*, pika::init_params const&))...
