cuda-memcheck what exactly does --continue do unspecified launch failure Address is out of bounds de

[wlangdon]

The nVidia user manual says --continue means it will “continue executing the rest of the kernel”

but it seems to do more than that. I have an as yet undiagnosed bug in a kernel which was working

kernel<<<grid_size0, block_siz0, shared_size>>>

    (nl,na,mlen,

     &d_A[data1start],d_B);

  cutilCheckMsg("kernel() execution failed.\n");

With out --continue I get

cutilCheckMsg cudaThreadSynchronize error: kernel() execution failed.

in file <kernel.cu>, line 1455 : unspecified launch failure.

========= ERROR SUMMARY: 0 errors

With I get

cutilCheckMsg cudaThreadSynchronize error: kernel() execution failed.

in file <kernel.cu>, line 1455 : unspecified launch failure.

========= Invalid read of size 4

========= at 0x00000120 in kernel

========= by thread (0,0,0) in block (81,0)

========= Address 0x349f8680 is out of bounds

=========

========= ERROR SUMMARY: 1 error

Is --continue expected to give better diagnostic error messages? Why without --continue does it say “0 errors”. Is this a real error message or just some side effect of trying to contine a kernel which has already died because of something else?

Any help you can give would be most welcome

Bill

[wlangdon]

I have an as yet undiagnosed bug in a kernel which was working

The bug turned out to be the kernel attempting to read outside the bounds of a global variable.