Variable access in kernel Is the access at variables in kernels mutually exclusive?

[chezze86]

Hi everybody, i’m new on cuda developing so maybe my question in very easy for you…

I’m trying to reimplement part of a software using cuda to improve the performance, but at the moment it goes slower…!!
What I was wondering is if in kernel the access at variables passed as parameters in mutually exclusive or is independent?

This is the code:

global void copyPartA(float *esm_jac_pinv, int esm_jac_rows, int esm_jac_pinv_cols, int jac2_inv_cols)
{

int index_thread=threadIdx.x;
int index_block=blockIdx.x;

esm_jac_pinv[index_threadesm_jac_pinv_cols+index_block]=tex1Dfetch(esm_jac_tx,2index_block)tex1Dfetch(jac2_inv_tx,index_threadjac2_inv_cols);
}

what I mean is: is it possible that just only i kernel execution at the time can simultaneously access at esm_jac_pinv? or cuda doesn’t check the access?

Thanks for the help

Andrea from Venice

[tera]

No it isn’t. Parameters are in shared memory on compute 1.x devices and in constant memory in 2.x. Both have a broadcast mechanism to send a word to all threads of a warp in one cycle.