decided to go with 128 for clustered (since it's less confined in range)
#endif
#define MAX_NR_COARSE_ENTRIES 64
#define MAX_NR_COARSE_ENTRIES 128
groupshared unsigned int coarseList[MAX_NR_COARSE_ENTRIES];
groupshared unsigned int clusterIdxs[MAX_NR_COARSE_ENTRIES/2];