AnsweredAssumed Answered

opencl: SIGSEGV with native kernels.

Question asked by ribalda on Mar 16, 2016
Latest reply on May 10, 2016 by dipak

Hi

 

We have an OpenCL application that makes use of Native Kernels. This apps was SIGSEGV from time to time. After a lot of debugging we have narrowed down the issue to the Native Kernel scheduling on the AMD Inplementation of OpenCL.

 

We are using fgrlx version 15.12 (linux 64 bits) and the AMD appsdk v 3.0. Clinfo output is attached to this discussion.

 

We have created an small testcase that triggers the error.  On that testcase we launch a native kernel that simply waits for a flag to return.  A correct exection will have the following output

(S, start, L launch E end F finish). It is easier to trigger the error if the app is launched like ./native_kernel_test 2>/tmp/out_log

 

 

L 0

S 0

E 0

F 0

L 1

S 1

E 1

F 1

L 2

S 2

E 2

F 2

L 0

S 0

E 0

F 0

 

and so on.

 

But after  a minute or so I get the folling output

L 1

S 1

E 1

F 1

L 2

S 2

E 2

S 8935824

F 2

L 0

Segmentation fault

or

 

L 2

S 2

E 2

F 2

L 0

S 9855280

F 0

L 1

Segmentation faul

Attachments

Outcomes