clsparseStatus dot(clsparse::array_base<T>& pR,
const clsparse::array_base<T>& pX,
const clsparse::array_base<T>& pY,
const clsparseControl control)
{
cl_int status;
//not necessary to have it, but remember to init the pR with the proper value
init_scalar(pR, (T)0, control);
// with REDUCE_BLOCKS_NUMBER = 256 final reduction can be performed
// within one block;
const cl_ulong REDUCE_BLOCKS_NUMBER = 256;
/* For future optimisation
//workgroups per compute units;
const cl_uint WG_PER_CU = 64;
const cl_ulong REDUCE_BLOCKS_NUMBER = control->max_compute_units * WG_PER_CU;
*/
const cl_ulong REDUCE_BLOCK_SIZE = 256;
cl_ulong xSize = pX.size();
cl_ulong ySize = pY.size();
assert (xSize == ySize);
cl_ulong size = xSize;
if (size > 0)
{
cl::Context context = control->getContext();
//partial result
clsparse::vector<T> partial(control, REDUCE_BLOCKS_NUMBER, 0,
CL_MEM_READ_WRITE, false);
status = inner_product<T>(partial, pX, pY, size, REDUCE_BLOCKS_NUMBER,
REDUCE_BLOCK_SIZE, control);
if (status != clsparseSuccess)
{
return clsparseInvalidKernelExecution;
}
status = atomic_reduce<T>(pR, partial, REDUCE_BLOCK_SIZE,
control);
if (status != CL_SUCCESS)
{
return clsparseInvalidKernelExecution;
}
}
return clsparseSuccess;
}
static C init_scalar_periodic(
int rnk_n, const INT *n, const INT *ind
)
{
INT *periodic_ind = PX(malloc_INT)(rnk_n);
/* assure periodicity in all directions */
for(int t=0; t<rnk_n; t++)
periodic_ind[t] = ind[t] % n[t];
C result = init_scalar(rnk_n, n, periodic_ind);
free(periodic_ind);
return result;
}