ansaurus

Question

How can I inprove this function under CUDA?

Answer 1

+1 A:

remove if statements by converting them to Boolean expressions.

flag += (DN[ELM1] == ans[0])

make sure your arrays are in registers or shared memory rather than global

also, on such simple algorithm you got way too complicated logic. change layout of D arrays to be D[N][6] as it will simplify many things

by the way, you may want to crop your post a little bit, way too much to read

3 x3 example

     A
  |0 0 0|           |x x 0
D |0 0 0| -> ... -> |x x 0 -> reduce down -> |x x o| -> reduce across -> 2x
  |0 0 0|           |x x x

basically you are setting matrix cells to true if match A is in array D. on each iteration you reduce columns to true if entire column is true. then you count number of true.

aaa 2010-06-10 23:32:46

well, the D1...D6 arrays are constant, which seems to be faster than shared. With D[N][6] would be more complicated, I tried it and it is slower.

Mark 2010-06-11 00:25:34

with flag += (DN[ELM1] == ans[0]) takes 700 ms more

Mark 2010-06-11 00:46:02

@Mark how many threads are you using?

aaa 2010-06-11 01:26:38

more than 13000000

Mark 2010-06-11 01:30:55

@Mark I assume each thread does its own work? how many threads per block?if I were you I would change algorithm to produce 6x6 matrix of true/false which you check at the end of each iteration to see if any three rows is all true. if does not make sense, I will post picture

aaa 2010-06-11 01:42:18

yes, please post the picture

Mark 2010-06-11 01:43:32

@Mark I need to know how many threads per block you using

aaa 2010-06-11 01:47:36

256 threads per block

Mark 2010-06-11 02:13:49

I tried again with D[N][6], but the kernel gives me the out of memory error, also I did not understand how to implement the 3x3 example

Mark 2010-06-11 15:42:44

ansaurus

tags:

views:

answers:

How can I inprove this function under CUDA?

related questions