pgaccelinfo: FAILED to create/destroy device context error?

Accelerated Computing HPC Compilers Legacy PGI Compilers
1

Please help:

  • cannot run f1.f90 example with Tesla C870 card. Output from f1.exe:

call to cuCtxCreate returned error 2: Out of memory

\

  • NVIDIA SDK examples work fine

  • output from make f1

make f1
pgfortran -ta=nvidia -Minfo=accel -fast f1.f90 -o f1
NOTE: your trial license will expire in 14 days, 9.59 hours.
NOTE: your trial license will expire in 14 days, 9.59 hours.
main:
38, Generating copyin(a(1:n))
Generating copyout(r(1:n))
39, Loop is parallelizable
Accelerator kernel generated
39, !$acc do parallel, vector(256)

  • output from pgaccelinfo shows FAILED to create/destroy device context errors

Device Number: 0
Device Name: Tesla C870
Device Revision Number: 1.0
Global Memory Size: 1610350592
Number of Multiprocessors: 16
Number of Cores: 128
Concurrent Copy and Execution: No
Total Constant Memory: 65536
Total Shared Memory per Block: 16384
Registers per Block: 8192
Warp Size: 32
Maximum Threads per Block: 512
Maximum Block Dimensions: 512 x 512 x 64
Maximum Grid Dimensions: 65535 x 65535 x 1
Maximum Memory Pitch: 262144B
Texture Alignment 256B
Clock Rate: 1350 MHz
Initialization time: 5153 microseconds
Current free memory 1564213248
Upload time (4MB) 2195 microseconds (1381 ms pinned)
Download time 3604 microseconds (3729 ms pinned)
Upload bandwidth 1910 MB/sec (3037 MB/sec pinned)
Download bandwidth 1163 MB/sec (1124 MB/sec pinned)

Device Number: 1
Device Name: Quadro FX 370
Device Revision Number: 1.1
Global Memory Size: 267714560
Number of Multiprocessors: 2
Number of Cores: 16
Concurrent Copy and Execution: Yes
Total Constant Memory: 65536
Total Shared Memory per Block: 16384
Registers per Block: 8192
Warp Size: 32
Maximum Threads per Block: 512
Maximum Block Dimensions: 512 x 512 x 64
Maximum Grid Dimensions: 65535 x 65535 x 1
Maximum Memory Pitch: 262144B
Texture Alignment 256B
Clock Rate: 720 MHz
Initialization time: 5153 microseconds
FAILED to create device context
FAILED to destroy device context

2

Hi Sorin Mitran,

Our default is to target cards that support compute capability 1.3. For older cards like these, you will need to compile using “-ta=nvidia,cc10” or “-ta=nvidia,cc11”.

Hope this helps,
Mat

3

Thank you for the suggestion. The behavior is the same though. Also,
set COMPUTECAP=10; was specified in sitenvrc file.

4

Updating NVidia driver from version 180 to 190 solved this problem.

5

Hi Sorin,

Thanks for the update.

FYI, I just got the following response from our engineers:

pgaccelinfo FAILED to create context on the Quadro, not on the Tesla.

The problem is that the runtime tries, by default, to run on the device with the highest “Revision” number. The Tesla C870 is Revision 1.0, but the Quadro FX 370 is Revision 1.1, so the runtime tries to run on device 1. But, the Quadro isn’t enabled for CUDA.

Workaround:
set the environment variable ACC_DEVICE to 0.

  • Mat