Issues after 17.9->18.4 upgrade

Accelerated Computing HPC Compilers Legacy PGI Compilers
1

Hi,

I am experiencing some issues when I moved to 18.4 with cuda code:
0: ALLOCATE: copyin Symbol Memcpy FAILED:13(invalid device symbol)

The offending line of code is a simple Allocate statement on the host for a device variable from a module.

Have others experienced this? I can currently compile and run the code with 17.9 and cuda 8.0 but it segfaults with 18.4.

Thanks, Jan

2

Hi Jan,

I’m assuming this is a CUDA Fortran code?

What type of device are you using and is this allocate the first time the GPU is used?

My first thought is that the default target devices changed from 17.9 to 18.4. In 18.4 we no longer target cc20 or cc30 devices by default and this type of error could occur if newer device code was used on an older device. So if you have one of these older devices, try compiling “-Mcuda=cc20” or “-Mcuda=cc30”.

If that’s not the issue, can you give more detail and if possible either post or send to PGI Customer Service (trs@pgroup.com), a reproducing example?

-Mat

3

Hi Mat,

the device is a Tesla K80. On Ubuntu, adding cc30 to the compiler solved the issue. However, on Rocks Cluster 7.0, the issue persists.

I sent an email with code attached.

Thanks, Jan