Archives Discussions

[ctwang_thomas]

I wrote a opencl kernel for testing:

.....

if   (A==B)   {   copy private memory data to global memory }

......}

I am sure that  A definitely not equal B for all work-item, but the If statement still decreases the performance seriously.

Can anyone give me a hint?

Thank you!

[maximmoroz]

Is A not equal B for all work-items?

[ctwang_thomas]

I am sure that A definitely not equal B for all work-items.

[maximmoroz]

Is this "If" inside a tight cycle? Could you show a little more of your kernel?

[ctwang_thomas]

//global_size=1000

//__global uint B;

//__global uint b;

//__private uint A;

//__private uint a; 

...__kernel ......

for (i=0;i<=1000000;i++)

{

  //generating __private a;

  //generating __private A;

if (A==B)     b=a;

//but A!=B for all i=0~1000000 and global_id=0~999;

}

-----------------------------------------

thank you so much!

[maximmoroz]

Well my guess is that when you remove "if (A==B) b=a;" line the compiler throws away the code which generates a and A as usless code because it doesn't affect anything global.

You will be able to check this by looking at ISA code in the profiler.

[MicahVillmow]

If A != B for all work-items, and A and B are not compile time constants, the compiler cannot optimize the code away, so although the code does not get executed, you still are using resources which can limit the number of work-groups that you can execute in parallel.