Archives Discussions

[ri239]

We've got a very simple kernel here, which does a copy from one cl_float4 array into another, and yet it crashes when running on the CPU. The compiled x86 kernel consists basically of some address computation and a movaps call, and the movaps call gets an unaligned address (8-byte aligned instead of 16 byte.) It works fine on the GPU.

Is this a known bug? Anything we can do to work around this?

[nou]

are you sure it is right kernel? becaue on GPU you do not get memory protection so writing into wrong location did not cause segmentation fault as on CPU.

[ri239]

Just found the reason, the problem seems to be that when I use CL_MEM_USE_HOST_PTR, the host memory should be already aligned -- otherwise, the copy fails. Allocating it with correct alignment resolves this issue.

[omkaranathan]

ri239,

Failure is expected, but you should not encounter a crash. Could you post the test case which produce the crash?