CUDA_ADD_LIBRARY is deprecated after 3.10 (reference), and it limits the CUDA parallel compilation. Because many architectures are compiled serially, warp-ctc requires almost 2 minutes to be compiled.
In this PR,
Use add_library + set project language with CUDA instead of cuda_add_library. Then user can pass -t0 through -DCMAKE_CUDA_FLAGS to enable CUDA parallel compilation. The compile time can be decreased from 2min to tens of second (~10x).
Use CMAKE_CXX_STANDARD and CMAKE_CUDA_STANDARD to replace manually adding -std=c++11
CUDA_ADD_LIBRARY
is deprecated after 3.10 (reference), and it limits the CUDA parallel compilation. Because many architectures are compiled serially,warp-ctc
requires almost 2 minutes to be compiled.In this PR,
add_library
+ set project language with CUDA instead ofcuda_add_library
. Then user can pass-t0
through-DCMAKE_CUDA_FLAGS
to enable CUDA parallel compilation. The compile time can be decreased from 2min to tens of second (~10x).CMAKE_CXX_STANDARD
andCMAKE_CUDA_STANDARD
to replace manually adding-std=c++11