Synchronization

Shuffle functions

• can be used to bypass shared memory

synch of mem ops

GA[i] = ...;
__threadfence();
G = 10;

• __threadfence(); = To ensure GA array is written before G to memory

Memory access optimization

Caching

• overfetching G[id * 1000] = cache fetches for example 256B for each element (4B) instead of the gpus lower access 32B

Partition camping

Transfer between host and GPU memory

• need to minimize counts of transfers