PGI accelerator: cuMemFree error: launch failed

running my program over several iterations, I get the error “call to cuMemFree returned error 700: Launch failed” after quite a while. As I am using PGI Accelerator Fortran, I am wondering what could be the reason.
I use 10.9 compiler and toolkit 3.1, CUDA driver version 3.2 on linux.

My accelerator region includes a nested loop and several private 3D-data (e.g. in1, in2 etc.):

!$acc region copyout(dis(0:numberPoints-1), counter(0:numberPoints-1),&
!$acc indices(0:numberPoints*3-1)), &
!$acc copyin(triangles(0:numberTriangles-1),points(0:numberPoints*4-1),&
!$acc wPoints(0:numberWPoints*4-1), normals(0:numberPoints*4-1))
!$acc do parallel, vector(64) private(o,d,v0,v1,v2,e1,e2,s,p,q,in1,in2,in3,dotProd,detInvers,u,v,tempDis)

The variables numberPoints at the like have different values for each accelerator execution.

Any idea?

Hi Sandra,

A “call to cuMemFree returned error 700: Launch failed” typically means the that the kernel abnormally aborted for some reason. I’ve seen it mostly if there was an out-of-bounds or other memory access violation.

Does the error occur at the same spot each time you run the program? One thing to try, is compile for the host using “-g -Mbounds” to check for out-of-bounds error. You can also try using Valgrind ( to check for memory errors.

Hope this helps and Go Ducks!