Error changing cuda version and PGI version

Accelerated Computing HPC Compilers Legacy PGI Compilers
1

Hi I’m updating my system from CUDA 6.5 PGI 15.10 GTX 970
to CUDA 7.5 PGI 16.7 GTX 1080 with the old configuration
I don’t have any problem with the new when i run the codes
I get this error:

0: copyin Symbol Memcpy (dev=0x0x735840, host=0x0x615258, size=8, offset=1179648) FAILED: 30(unknown error)


or

0: ALLOCATE: 1600000000 bytes requested; status = 30(unknown error)

I’m compiling my code in this way:

pgf90 -Mcuda=7.5 -o code.f90 exec_code -lcurand

How i can solve this problem ?

2

Hi dpera,

A GTX 1080 uses the Pascal architecture (compute capability 6.0) which requires CUDA 8.0. Please update your CUDA Device driver (if not done already) to at least CUDA 8.0 and compile with:

pgf90 -Mcuda=8.0,cc60 -o code.f90 exec_code -lcurand

Hope this helps,
Mat

3

Hi Mat,

I used these configurations

CUDA 8.0 (44) PGI 16.7 GTX 1080 driver 387.26
CUDA 8.0 (61) PGI 16.7 GTX 1080 driver 387.26

compile command

pgf90 -Mcuda=8.0,cc60 -o code.f90 exec_code -lcurand


In both cases i have this error

0: ALLOCATE: 1600000000 bytes requested; status = 30(unknown error)


With C code I don’t have any problem

Any idea? Regards D.

4

We need a little more information to help dig into this failure.
Does this occur on the very first allocation? If so, he error might come from the device not getting initialized correctly.

Does the size reported match the size you think you should be allocating?

Can you create a small reproducer and send it to us?

  • Brent