In this test i tried to adapt the classic Levenshtein_distance algorithm to Aparapi. What i have did was to basically calculate in parallel the values for each color.
Lest say:
A = the first string,
B = the second string
then my algorithm runs the same kernel N = ( len(A)+len(B)-1) times ( this number is equals with the number of different colors ).
Lest say:
A = the first string,
B = the second string
then my algorithm runs the same kernel N = ( len(A)+len(B)-1) times ( this number is equals with the number of different colors ).
The GPU time is 37X slower than the iterative version for the 1024*4 length strings and 14X slower for the 1024*14 length string. Why is this?
I did some testing and the 2 most probably caused are:
a) Kernel start time, if you run the same kernel multiple times even if it doesn't do nothing it takes around 0.6 ms when run in GPU mode and 0.125 ms when run in CPU mode ( the 2 times are without buffer copy and java to opencl code translation ).
b) Low number of operations executed by each kernel run. The maximum number of operations executed at on time is len(B).
What did i learned from this? When you start a kernel make sure it has a lot of data to process at least millions if not billions of operations. If you start a kernel to make 10 000, 100 000 even 1 000 000 if will be slower than running that code on the CPU.
Results
Java code
