I'm getting the following message while trying to compile a CUDA program:
/usr/bin/ld: cannot find -lcudart
Never got this kind of error before. Is it any problem with my PATH?
Here's the compile command:
gfortran -g -O2 -ffree-line-length-none -I../shared/ -o ../../bin/xspecfem3D ../../obj/spec/program_specfem3D.o ../../obj/spec/specfem3D_par.o ../../obj/spec/PML_init.o ../../obj/spec/compute_boundary_kernel.o ../../obj/spec/compute_kernels.o ../../obj/spec/compute_forces_acoustic.o ../../obj/spec/compute_forces_acoustic_pot.o ../../obj/spec/compute_forces_acoustic_PML.o ../../obj/spec/compute_forces_elastic.o ../../obj/spec/compute_forces_elastic_Dev.o ../../obj/spec/compute_forces_elastic_noDev.o ../../obj/spec/compute_forces_elastic_Dev_openmp.o ../../obj/spec/compute_add_sources_acoustic.o ../../obj/spec/compute_add_sources_elastic.o ../../obj/spec/compute_coupling_acoustic_el.o ../../obj/spec/compute_coupling_elastic_ac.o ../../obj/spec/compute_stacey_acoustic.o ../../obj/spec/compute_stacey_elastic.o ../../obj/spec/compute_gradient.o ../../obj/spec/compute_interpolated_dva.o ../../obj/spec/initialize_simulation.o ../../obj/spec/read_mesh_databases.o ../../obj/spec/setup_GLL_points.o ../../obj/spec/write_movie_output.o ../../obj/spec/create_color_image.o ../../obj/spec/write_seismograms.o ../../obj/spec/write_output_ASCII.o ../../obj/spec/detect_mesh_surfaces.o ../../obj/spec/setup_movie_meshes.o ../../obj/spec/read_topography_bathymetry.o ../../obj/spec/setup_sources_receivers.o ../../obj/spec/prepare_timerun.o ../../obj/spec/iterate_time.o ../../obj/spec/finalize_simulation.o ../../obj/spec/save_adjoint_kernels.o ../../obj/spec/specfem3D.o ../../obj/spec/assemble_MPI_vector.o ../../obj/spec/make_gravity.o ../../obj/spec/noise_tomography.o ../../lib/libspecfem.a ../../obj/spec/check_fields_cuda.cuda.o ../../obj/spec/compute_add_sources_acoustic_cuda.cuda.o ../../obj/spec/compute_add_sources_elastic_cuda.cuda.o ../../obj/spec/compute_coupling_cuda.cuda.o ../../obj/spec/compute_forces_acoustic_cuda.cuda.o ../../obj/spec/compute_forces_elastic_cuda.cuda.o ../../obj/spec/compute_kernels_cuda.cuda.o ../../obj/spec/compute_stacey_acoustic_cuda.cuda.o ../../obj/spec/compute_stacey_elastic_cuda.cuda.o ../../obj/spec/it_update_displacement_cuda.cuda.o ../../obj/spec/noise_tomography_cuda.cuda.o ../../obj/spec/prepare_mesh_constants_cuda.cuda.o ../../obj/spec/transfer_fields_cuda.cuda.o ../../obj/spec/write_seismograms_cuda.cuda.o ../../obj/spec/save_and_compare_cpu_vs_gpu.cudacc.o ../../obj/spec/serial.o -lcuda -lcudart -lcublas
/usr/bin/ld: cannot find -lcudart
collect2: ld returned 1 exit status
Edit:
I've changed the linkage command to
gfortran -g -O2 -ffree-line-length-none -I../shared/ -L$CUDA_HOME/lib [rest follows]
And all my PATHs seems to be OK, but I'm still having the same error.
Edit2:
The error was that the code I was trying to compile was in 64bit, so I needed to give the path a diferent set of .so's , which are located on
/usr/local/cuda/lib64
Now the compilation ran fine ans without any problems.
Answer
You are not specifying the location of the CUDA libraries to the linker, that is why the linkage is failing. nvcc is "automagically" configured to find the CUDA runtime library components, but if you are linking with a host compiler directly, you will need to explicitly specify their location to the compiler. Try changing the linkage command to this:
gfortran -g -O2 -ffree-line-length-none -I../shared/ -L$CUDA_HOME/lib [rest follows]
where CUDA_HOME denotes the path where the CUDA toolkit is installed. Typically this is /usr/local/cuda on linux and OS X systems. You will then need to make sure LD_LIBRARY_PATH contains an entry for $CUDA_HOME/lib so that the executable will dynamically load the required CUDA libraries at runtime.
EDIT:
Also see this