topic Simple OpenCL program fails eventually due to memory leak in i.MX Processors

Simple OpenCL program fails eventually due to memory leak

andyp — Thu, 30 Apr 2015 17:05:33 GMT

I am using imx6 quad, running Windows Embedded Compact 7 (aka WEC7, Windows CE 7, CE7). The BSP came from Adeneo.

I have found what I believe is a memory leak in the OpenCL software stack. A very simple Vector_Add program runs in a loop ~8000 times and fails with CL error -6.

This file is attached. The .cl file is also attached.

If anyone is an expert in OpenCL and can spot a fault with the Vector_Add program, let me know.

Otherwise, if the problem looks to be in the OpenCL software stack, what is the support model?
How do I get this fixed?

andyp

Original Attachment has been moved to: vector_add_kernel.cl.txt.zip

Original Attachment has been moved to: Vector_Add1.cpp.txt.zip

Re: Simple OpenCL program fails eventually due to memory leak

CarlosCasillas — Mon, 04 May 2015 14:26:46 GMT

Deactivated user, could you please comment?

Re: Simple OpenCL program fails eventually due to memory leak

andyp — Mon, 04 May 2015 16:13:11 GMT

I added some more checking to my simple OpenCL program.

The checking looked at the reference count for all the OpenCL data structures created with the clCreate* API's. This includes the context, kernel, program, memory buffers, etc.

I found a surprising reference count = 2 for both the kernel object and the program object.

Could there be an implicit increment of the ref count for these 2 objects?

I modified the program to do 2 clReleaseProgram(...) calls and do 2 clReleaseKernel(...) calls each time through the loop.

Now I am retesting.

Re: Simple OpenCL program fails eventually due to memory leak

andyp — Mon, 04 May 2015 16:25:39 GMT

Program still eventually fails. Still feels like something wrong with the OpenCL software stack.

I will try to add an attachment for the modified OpenCL program.

Re: Simple OpenCL program fails eventually due to memory leak

andyp — Wed, 17 Jun 2015 17:21:33 GMT

More investigation and characterization of the OpenCL memory leak bug reveals that if clCreateContext() is called once, and clReleaseContext() is never called, then the program runs correctly without a memory leak.

My conclusion is that the problem is with clReleaseContext().