Is padding needed for Malloc?

Accelerated Computing CUDA CUDA Programming and Performance
1

When malloc:ing say 10-member vectors a() ,b() and c(), and multiplying c(i)=a(i)*b(i) with 16 threads, what happens with the 10th to 16th thread? Do they read from outside of a() and b() and save that to past of c()? I can’t see any methods to control what happens when there are more threads than members in vectors. Is the only possibility to Malloc extra members for the remaining threads?

2

you must set boundary condition explicitly in your kernel code, for example

[codebox]global void add( float *C, float *A, float *B , unsigned int N )

{

unsigned int idx = blockIdx.x * BLOCK_DIM + threadIdx.x;



if ( idx < N ){ // N is size of A and B

	C[i] = A[i] * B[i] ;

}

}[/codebox]

3

Hi. Thanks for example. It is still a bit unclear to me, what happens to the “overflow” threads between N and blocsize. As far as I understand, all threads in a block are executing exactly the same commands. How is it then prevented that the overflow threads do nothing?

TimoS