IV82940: CUDADEVICE FUNCTIONS INCORRECTLY RECORD THE PRESENCE OF AN INITIALIZED CUDA DEVICE CONTEXT

APAR status

• Closed as program error.

Error description

• Error Message: com.ibm.cuda.CudaException: invalid device
  context
  .
  Stack Trace: N/A
  .
  

Local fix

• Ensure that the first use of CUDA4J made by any particular
  thread is not via one of these functions.
  CudaDevice.canAccessPeer()
  CudaDevice.disablePeerAccess()
  CudaDevice.enablePeerAccess()
  CudaDevice.getAttribute()
  For example, using CudaDevice.getFreeMemory() will ensure a
  valid device context.
  

Problem summary

• Each thread requires a valid CUDA device context for certain
  interactions with a GPU. When a thread uses the CUDA4J API, the
  context will be initialized as necessary. A successful use of
  any of the following CudaDevice functions was incorrectly
  marking the context as valid causing to the initialization to be
  omitted in other places where it was required.
  CudaDevice.canAccessPeer()
  CudaDevice.disablePeerAccess()
  CudaDevice.enablePeerAccess()
  CudaDevice.getAttribute()
  

Problem conclusion

• The code which tracks the validity of device contexts has been
  corrected.
  .
  This APAR will be fixed in the following Java Releases:
     8    SR3       (8.0.3.0)
  .
  Contact your IBM Product's Service Team for these Service
  Refreshes and Fix Packs.
  For those running stand-alone, information about the available
  Service Refreshes and Fix Packs can be found at:
             https://www.ibm.com/developerworks/java/jdk/
  

Temporary fix

Comments

APAR Information

• APAR is sysrouted FROM one or more of the following:

• APAR is sysrouted TO one or more of the following:

Fix information

• Fixed component name

  J9 COMMON CODE

• Fixed component ID

  620700127

Applicable component levels

• R270 PSY

     UP