unspecified Launch Failure Kernel fails often

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi all,

I am having kernel launch failure issue and i am discussing it with developers on the other thread.

BY the by I wanted to know the restrictions of accessing the global memory.

If my threads try to access data from here and there in the global memory, what worse it can happen?

Can it led to “unspecified launch failure”.

As i read the CUDA Docs, I come across this (Please find the Attachment).
I see there is restrictions on global memory access.
I am having a 1.0 CUDA device.
Please help me out.

Thanking you in Advance,
Nabarun Paul.
GPU.bmp (1.49 MB)

2

Here and there inside allocated global memory is fine…

but here and there accessing everywhere will cause launch failures

3

You meant segfault ?

I am trying to find, if there is segfault.

But coming back to my point,

in the documentations, there seems to be restrictions on global memory access patterns.

I am doing some kind of permutation and combination.

So with just hundreds of input parameters, i need to generate lakhs of threads.

And hence any thread in the grid have to access any data from the memory.

regards,

Nabarun

4

There is only a performance penalty if you don’t coalesce accesses to global memory, it will not crash your kernel. You will get segfaults only if you access memory that has not been allocated.

Might want to change that to hundreds of thousands. People on here might not get the concept of lakhs.

Global memory is shared among all of the threads in the kernel, as long as it is within an allocated range.

5

Arun,

Just in case u did not understand gregory’s comment:

Brits, Indians call 1,00,000 as One Lakh. However US people and some others call it “one hundred thousand”. So, Greg was asking you to re-phrase your sentence…

btw,

If you are spawning lakhs of threads - make sure that you are NOT crossing block or grid limits…Check out

6

Sarnath and Gregory,

Hmmm Okay ! UK and US English differences.

I will definitely take care of these when I post.

Now I am going through my entire code.

Will get back to you soon.