Access violation : sounds a bug in the last SDK ?

[Polar01]

Hi,

I have the following code that works well :

#pragma unroll

for(uint i = 0; i < MLT_SAMPLING_TASK_PER_SAMPLER; i++, taskId++)

{

clTask task = tasks[taskId];

clMLTSample sample = task.mltSample;

if (sample.requestNewMLTSample && sample.score > bestScore)

{

bestScore = sample.score;

bestScoreTaskId = taskId;

}

}

#pragma unroll for(uint i = 0; i < MLT_SAMPLING_TASK_PER_SAMPLER; i++, taskId++) { clTask task = tasks[taskId]; clMLTSample sample = task.mltSample; if (sample.requestNewMLTSample && sample.score > bestScore) { bestScore = sample.score; bestScoreTaskId = taskId; } }

But when I replace 2 lines with (to minimize the memory access) I got an access violation !

clMLTSample sample = tasks[taskId].mltSample;

[sharpneli]

Is this OpenCL or Host code? From the looks of it it looks like a host code so it has nothing to do with Intel OpenCL SDK.

[Polar01]

No, it is OpenCL kernel code !!!!

[Polar01]

No reaction from the Intel Team ?

[Doron_S_Intel]

Hello,

I'm not sure it's legal to use #pragma unroll in OpenCL C code. Does the problem persist when you remove those instructions?

Doron Singer

[Polar01]

It is something that I have found on the Intel pages... so at least for the Intel SDK it should work.

Anyway, I have test without... and I still have the crash :-(

[Doron_S_Intel]

Good. Next question: What is a "clTask"? Is it a typedef to an OpenCL native type, or is it some user-defined struct?

[Polar01]

clTask is a typedef to a struct.

[Doron_S_Intel]

I'm not sure the assignment operator is defined for structs in C. You can use the printf extension to make sure the struct doesn't contain garbage data after the assignment.

[sharpneli]

Assignment operation for structs is very much defined in C (http://en.wikipedia.org/wiki/Struct_%28C_programming_language%29#Assignment).

These 2 pieces of code are identical in their functionality for a good compiler. Lazy compiler might allocate space from stack for clTask task but those days are long gone.

clTask task = tasks[taskId];

clMLTSample sample = task.mltSample;

clMLTSample sample = tasks[taskId].mltSample;

If the only difference between working and non working version are those pieces of code then this indeed is a bug in Intel's implementation.

[Doron_S_Intel]

In that case, we'll look into this.
Thanks for reporting (in case it's not obvious from my posts, I'm not on the compiler side of things :))

Doron