Can't Initialize cuSolver (unknown error 53)

Accelerated Computing CUDA CUDA Programming and Performance
1

I am simply trying to run the cuSolverSp_LinearSolver example in CUDA Toolkit 7.5. I am on Debian Linux Jessie 64-bit with kernel 3.16.7 (3.16.0-4-amd64 #1 SMP Debian 3.16.7-ckt11-1+deb8u5 (2015-10-09) x86_64 GNU/Linux).

I don’t know how to solve this error because 53 is not even a valid cuSolver error.

The error message I get is:

GPU Device 0: “GeForce GTX TITAN Black” with compute capability 3.5

Using default input file [./lap2D_5pt_n100.mtx]
step 1: read matrix market format
sparse matrix A is 10000 x 10000 with 49600 nonzeros, base=1

You are running using the stub version of cusolver
CUDA error at cuSolverSp_LinearSolver.cpp:258 code=53() “cusolverSpCreate(&handle)”

In fact, I get similar errors using other CUDA libraries (but not core CUDA routines). In other words, cuFFT fails with error 50 but things like vectorAdd work fine.

2

You have somehow built the sample incorrectly.

For a standard install, the cusolver libraries should be in:

/usr/local/cuda/lib64

You have somehow linked the application against:

/usr/local/cuda/lib64/stubs

In that directory, the “stub” version of libcusolver.so contains the message:

“You are running using the stub version of cusolver”

I can only assume you built this sample code in some non-standard fashion. Did you build the samples using the provided Makefile at:

/usr/local/cuda/samples

?

3

Yes, you are correct. It was linked against the stub libraries. Simply deleting the “stub” directory are running sudo ldconfig and rebuilding the example fixed the problem.

Note, I was not specifically linking against the stubs. ld / gcc just decided to do it on its own.

Thanks!!