Solved: Strange crash... access to __local memory - Page 2

Polar01 · ‎05-31-2011

Hi,

I have implement a 'scan' algorithm in OpenCL. Because it is an open source library I test in on several machines and OpenCL-SDK.

But, it crash with the Intel SDK ! (Not with the other ones).

What I have discover is that the problem is maybe related to the "__local" memory !

I have the following kernel :

__kernelvoid kernel__ExclusivePrefixScan(...,__local T* localBuffer,...)

And I set up my buffer with the following command :

clStatus = clSetKernelArg(_kernel_Scan, 2, _workgroupSize * 2 * sizeof(int), 0);

checkCLStatus(clStatus); <= CL_SUCCESS !!!

Where _workgroupSize= 128; So, I reserve 1024 bytes only !!!

You can find the code at :http://code.google.com/p/clpp/

Krys

Evgeny_F_Intel · ‎06-01-2011

Hi,

I investigatedon rev41.

Well, the issue here is not the __local memory but the memory overrun during write in line:
line 175, clppScan.cl : blockSums[bid] = localBuffer[localBufferFullSize-1];

Looking on the host code I saw that you allocate memory buffer that is not sufficient for the operation of the algorithm.

One of the issues is buffer size calculation (line 219, clppScan.cpp). You are use workgroup size of 128, while providing local size of 64 (line 72) to NDRange. Thus, causes number of workgroups to be greater than size of the allocated buffer and as a result you have memory overrun.

After the change first NDRange passed, I added clFinish() after it, but then the next NDRange failed. This is because the same reason. The intermidiate buffer size doesn't match the number of work groups, probably you should decrease the global size in the next pass.

Regards,
Evgeny

View solution in original post

Polar01 · ‎06-07-2011

Great,

Thanks for your support