A kernel should access at least 32-bits of data at a time, from addresses that are aligned to 32-bit boundaries. A

char4

,

short2

,

int

, or

float

counts as

32

-bits of data. If you can, load two, three, or four

32

-bit quantities at a time, which may improve performance. Loading more than four

32

-bit quantities at a time may reduce performance.

Optimize

__global

memory and

__constant

memory accesses to minimize the number of cache lines read from the L3 cache. This typically involves carefully choosing your work-group dimensions, and how your array indices are computed from the work-item local or global id.

If you cannot access

__global

memory or

__constant

memory in an optimal manner, consider moving part of your data to

__local

memory, where more access patterns can execute with full performance.

Local memory is most beneficial when the access pattern favors the banked nature of the SLM hardware.

Optimize

__local

memory accesses to minimize the number of bank conflicts. Reading the same address from the same bank is OK, but reading different addresses from the same bank results in a bank conflict. Writes to the same bank always result in a bank conflict, even if the writes are going to the same address. Consider adding a column to two-dimensional local memory arrays if it avoids bank conflicts when accessing columns of data.

Avoid dynamically-indexed

__private

arrays if possible.