topic I am submitting a bug. Even in OpenCL* for CPU

Memory leaks in clBuildProgram

Takaaki_H_ — Wed, 31 Jul 2013 08:48:53 GMT

Hello, Memory usage of attached program continue to increase. It doesn't happen if a line that contains clBuildProgram call is removed. I'm using Ubuntu 12.04.1, Intel SDK for OpenCL Applications 2013 (3.0.67279) and g++ 4.6.3.

Is there a reason why you are

Raghupathi_M_Intel — Tue, 13 Aug 2013 19:45:47 GMT

Is there a reason why you are creating and releasing the program objects so many times? Though you are releasing the program everytime there may be some internal state/objects that the runtime maintains that doesnt get released right away. Maybe that's causing the "leak". I'll investigate more but would like to hear why you are doing what you are doing.

Thanks,
Raghu

I am submitting a bug. Even

Raghupathi_M_Intel — Wed, 14 Aug 2013 18:09:45 GMT

I am submitting a bug. Even though what you are doing is not a best practice, nevertheless, you are releasing the program object and this shouldn't result in any memory leaks. Thanks for attaching the test case.

Raghu

I know that building kernel

Takaaki_H_ — Wed, 14 Aug 2013 20:20:35 GMT

I know that building kernel many times is not a best practice. I found this problem when I was testing my application. This test needs building CL kernel with various parameter combinations.
I hope this bug to be fixed.

Thanks,
Takaaki

We're also having problems

Andy_R — Wed, 05 Mar 2014 10:50:28 GMT

We're also having problems with the Intel OpenCL memory leak in clBuildProgram.

We're using version Intel OpenCL CPU Runtime version 3.0.0.63463 on Windows 7 and finding that every cycle of clCreateProgram->clBuildProgram->clReleaseProgram leaks significant memory (exactly in the same way that the intel-cl-leak.cpp file attached at the start of this thread demonstrates). This does not occur when the call to clBuildProgram is removed.

Has anyone found a solution to this or is a fix likely to be released?

This is highly problematic to us as we generate our OpenCL program source on the fly before each kernel execution and this can happen to the order of 1,000s of times within a single process space, which means we hit 'out of memory' errors as a result of this bug.

Thanks,

Andy

Hi Andy,

Raghupathi_M_Intel — Fri, 07 Mar 2014 20:19:47 GMT

Hi Andy,

I don't have a date yet on the release with the fix. But will keep you posted. I will also escalate the issue.

Thanks,
Raghu