/* Writes an output datum to an indexed buffer.
* Operates on a block of model instances in parallel.
* quantities is an array of output data of length (outputsize * blocksize)
* participate is an array of length blocksize indicating which threads are active
*/
__DEVICE__ void buffer_indexed_output (unsigned int modelid, unsigned int outputid, unsigned int outputsize, CDATAFORMAT *quantities, indexed_output_buffer *pos, unsigned int threadid, unsigned int blocksize, int participate) {
unsigned int i, offset;
CDATAFORMAT *buffer;
indexed_sort_data *sort;
int *index = pos->scratch;
index[threadid] = !!participate; // ensures index is 1 or 0
parallel_scan(index,threadid,blocksize);
if (participate) {
offset = pos->size + index[threadid] - 1;
buffer = pos->buffer;
for (i=0; i<outputsize; i++) {
buffer[i+offset] = quantities[VEC_IDX(outputsize,i,blocksize,threadid)];
}
sort = pos->sort + offset;
sort->modelid = modelid;
sort->outputid = outputid;
sort->offset = offset;
}
if (0 == threadid) {
pos->size += index[blocksize-1] * outputsize;
}
}
TestScan( const WorkSpec & Start , const WorkSpec & N )
{
typedef Kokkos::RangePolicy<execution_space> exec_policy ;
parallel_scan( exec_policy( Start , N ) , *this );
}
TestScan( const WorkSpec & N )
{ parallel_scan( N , *this ); }