weifengliu-ssslab / benchmark_spgemm_using_csr Goto Github PK

Hi everyone, here below is how I use CUSP when compile bhSPARSE. One can use it as a reference when errors are encountered.
Test environment:
Ubuntu 18.04, CUDA 10.2, gcc 8.4.0, CUSP 0.5.0, Thrust 1.13.0 (i.e HPC SDK 21.7).

(1) modify the makefile in SpGEMM_CUDA and include the cusp and thrust folder path.

(2) when compiling spgemm, one might meet the following errors:
(a) cannot find csr.h.
search csr.h path in the cusp folder and replace the path with the newly found one.
(b) cannot find launch_calculator.h.
search launch_calculator.h path in the thrust folder and replace the path with the newly found one.
(c) the function __thrust_compiler_fence() in cusp/system/cuda/detail/runtime_introspection.inl is undefined.
search the function __thrust_compiler_fence() on the thrust mainpage on Github and one will find the __thrust_compiler_fence() is defined in thrust/detail/config/compiler_fence.h. Hence include this file in the runtime_introspection.inl.
(d) enable_if is ambiguous.
one can find the solution in cusplibrary/cusplibrary#101 (comment).

Good luck.

nvcc -O3 -m64 -gencode=arch=compute_70,code=sm_70 main.cu -o spgemm -I/usr/local/cuda/include -I/usr/local/cuda/samples/common/inc -L/usr/local/cuda/lib64 -lcudart
In file included from main.cu:19:0:
ref_spgemm.h:17:44: fatal error: cusp/detail/host/reference/csr.h: No such file or directory
compilation terminated.
Makefile:16: recipe for target 'all' failed
make: *** [all] Error 1

./spgemm -cuda -spgemm cage4.mtx

A: cage4.mtx
B: cage4.mtx
A: ( 9 by 9, nnz = 49 )
B: ( 9 by 9, nnz = 49 )
Device [0] GeForce GTX TITAN @ 875.5MHz. 14 SMXs.
err = invalid device function
warmup error = -1

Found an err, code = -1