loop-unrolling and memory access performance

To improve memory performance in task kernels, you must unroll your memory access loop on the dimension in which the accesses are consecutive. Take the following example:

__global float input;
__local float data;
for (i = 0; i < N; i++)
{
 # pragma unroll 4
  for (j = 0; j < M; j++)
  {
    data = input;
  }
}

In this case, you will get one coalesced port to off-chip memory, but with a width of 128 bits. However, if the accesses are as follows:

__global float  input;
__local float data;
for (i = 0; i < N; i++)
{
 # pragma unroll 4
  for (j = 0; j < M; j++)
  {
    data = input;
  }
}

You will get four non-coalesced 32-bit ports to external memory. In this case, your external memory performance will hardly improve since you now have 4 accesses competing with each other to acquire the memory bus.

To maximize memory performance, you should minimize the number of accesses, but maximize the size of the accesses.