event (firstDotCall) leak at xdot.c

Question

event (firstDotCall) leak at xdot.c

Closed this issue 8 years ago · 7 comments

Cause "out of memory" errors with intel opencl drivers.
Here's the forum thread:
https://software.intel.com/en-us/forums/opencl/topic/701907
And bugfix:
ache7@959372f

ache7 commented 8 years ago

Here #297

Answer 1 · 2017-01-20T16:13:30.000Z

Hi @ache7
Can you make a PR targeting our /develop branch? I'll merge this.

Answer 2 · 2017-01-21T08:03:04.000Z

Maybe it's better to initialize firstDotCall to NULL and release firstDotCall event before freeSolutionSeq(&seq). Differs if first makeSolutionSeq or executeSolutionSeq failed.

Answer 3 · 2017-01-21T14:40:33.000Z

Also, when I search for "eventWaitList, &first" in sources - there is more functions with unreleased events, all these events come to "step->event = events + j;" in makeSolutionSeq, so can be freed automatically somewhere in freeSolutionSeq -> freeSolutionStep -> releaseStepImgs, but I don't see clReleaseEvent() calls. Someone must check this, I don't know code well enough.

Answer 4 · 2017-01-24T20:08:41.000Z

Hi @ache7
I have made a new branch in my fork that fixes the many events you mention in this thread. It passes my unit tests.

If you have time, would you mind testing in your environment? If not, I'll merge my branch as it's already passing my tests.

Answer 5 · 2017-01-25T16:33:06.000Z

Sorry, I don't use library directly, only through other libraries (from cltorch), so can't test it thoroughly. Seems that neural nets use only xdot() from functions with event leaks.

Answer 6 · 2017-01-25T17:24:05.000Z

Thanks for replying; I'll close your PR and merge in my fixes. I think it should square you away.