I'm compiling some OpenMP 4.5 code with the IBM XL C/C++ compiler with the intention of offloading some of its work to a GPU, like so:
xlc++ mycode.cpp -qsmp=omp -qreport -qoffload -std=c++11 -Wall
Compilation seems to be successful, giving me only the following messages:
mycode.cpp:
"mycode.cpp", line 284: 1586-358 (I) Loop was parallelized.
"mycode.cpp", line 293: 1586-358 (I) Loop was parallelized.
"mycode.cpp", line 309: 1586-358 (I) Loop was parallelized.
"mycode.cpp", line 324: 1586-358 (I) Loop was parallelized.
"mycode.cpp", line 126: 1586-674 (I) Remark: Simd or nested parallel directive requires OpenMP runtime
"" 1586-671 (I) GPU OpenMP Runtime is required for offloaded kernel '__xl__Z9MyCodeiii_l123_h44039046689_OL_1'
However, when I run the code, I get the following unpleasant message:
1587-169 No valid target devices available.
Using nvidia-smi, I have verified that target devices are, in fact, available:
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 384.59 Driver Version: 384.59 |
|-------------------------------+----------------------+----------------------+
| GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. |
|===============================+======================+======================|
| 0 Tesla P100-SXM2... Off | 00000002:01:00.0 Off | 0 |
| N/A 33C P0 29W / 300W | 10MiB / 16276MiB | 0% E. Process |
+-------------------------------+----------------------+----------------------+
| 1 Tesla P100-SXM2... Off | 00000003:01:00.0 Off | 0 |
| N/A 29C P0 30W / 300W | 10MiB / 16276MiB | 0% E. Process |
+-------------------------------+----------------------+----------------------+
| 2 Tesla P100-SXM2... Off | 00000006:01:00.0 Off | 0 |
| N/A 31C P0 28W / 300W | 10MiB / 16276MiB | 0% E. Process |
+-------------------------------+----------------------+----------------------+
| 3 Tesla P100-SXM2... Off | 00000007:01:00.0 Off | 0 |
| N/A 27C P0 29W / 300W | 10MiB / 16276MiB | 0% E. Process |
+-------------------------------+----------------------+----------------------+
My thought is that XL is somehow targeting the wrong accelerator, but I can't find an option to set this.
How can I get my code to recognize and utilize the available GPUs?
-qtgtarch specifies GPU architectures where the code may run. Please try -qtgtarch=auto if you would like the compiler to automatically detect the architecture of device 0 of the system on which the compiler is being executed. Alternatively, you might try setting it manually, for example -qtgtarch=sm_60.
More information at Knowledge Center.