Le 30/03/2011 15:14, Aidy a écrit :
>> Please, read.
>> __device__ is going to inline the function (or I'm just getting it
>> wrong, which is 90% likely).
> 
> Sorry I missed that bit. You're right, I've just looking through the nVidia
> documentation, and the link you sent me and it will inline it.
> 
> If it inlines it then, can I pass the structure itself and NOT have it create a
> copy within the function itself?? It shouldn't increase the overhead at all
> should it??
> 
I would be more concerned about the content of the structure: if it hold
any pointer (and use it), you will be once again in trouble.

(for the sack of my comprehension so far: the address space used by the
main CPU is not the address space used by the Cuda element, so any
pointer filled by the CPU is useless on the Cuda element. Which means
only linear/flat data(aka structure) can be exchanged at the interface
of a __device__ function.)

Now, I did not check the definition of TNORMAL and other.

For your question: as it is inline, no copy are created, it just use the
actual data. (unless a C expert wants to show up and put me in
embarassement right for that sentence)

You might get issue with patterns that use a cache, as 3.6 is not
thread-aware and Cuda code could run into collisions.

-- 
Software is like dirt - it costs time and money to change it and move it
around.

Just because you can't see it, it doesn't weigh anything,
and you can't drill a hole in it and stick a rivet into it doesn't mean
it's free.

Post a reply to this message