What's the most trivial function that would benfit from being computed on a GPU?

Question

I'm just starting out learning OpenCL. I'm trying to get a feel for what performance gains to expect when moving functions/algorithms to the GPU.

The most basic kernel given in most tutorials is a kernel that takes two arrays of numbers and sums the value at the corresponding indexes and adds them to a third array, like so:

__kernel void 
add(__global float *a,
    __global float *b,
    __global float *answer)
{
    int gid = get_global_id(0);
    answer[gid] = a[gid] + b[gid];
}

__kernel void
sub(__global float* n,
    __global float* answer)
{
    int gid = get_global_id(0);
    answer[gid] = n[gid] - 2;
}

__kernel void
ranksort(__global const float *a,
         __global float *answer)
{
  int gid = get_global_id(0);
  int gSize = get_global_size(0);
  int x = 0;
  for(int i = 0; i < gSize; i++){
    if(a[gid] > a[i]) x++;
  }
  answer[x] = a[gid];
}

I am assuming that you could never justify computing this on the GPU, the memory transfer would out weight the time it would take computing this on the CPU by magnitudes (I might be wrong about this, hence this question).

What I am wondering is what would be the most trivial example where you would expect significant speedup when using a OpenCL kernel instead of the CPU?

Answer 1

if you have a sufficiently large set of matrices you intend to perform linear algebra operations on, or that you're essentially performing the same operation on each element, i would regard that as a trivial example. matrix multiplication, addition, fft's, convolution, etc. you'll see a bit of a speedup without doing much work. now if you want to see the 100x speedups then you need to delve into the memory management and know a fair bit about what's going on behind the scenes.

for getting started, i would recommend starting with pycuda since it is fairly simple to get started since it provides a very high level of abstraction and will allow you to jump in very quickly. check out this course on parallel computing using cuda from the university of illinois http://courses.ece.illinois.edu/ece498/al/ when you are ready to dive in further.