Archives Discussions

[Raistmer]

what values is too high?

"
FCStacks - The number of flow control stack size used by the kernel (only for GPU device). This number may affect the number of wavefronts in-flight. To reduce the stack size, reduce the amount of flow control nesting in the kernel.
"
Could someone give some numbers, please. How this number will affect the number of simultaneously executed wavefronts on SIMD ?
For example one of my kernels has FCStacks ==2, another one FCStacks ==6.
In what case this number will be limiting factor?

EDIT: the kernel with FCStacks ==6 has pretty high GPR usage too, 55 GPRs used.
What will limit number of wavefronts in this case? GPR usage or FCStacks value?
And how many wavefronts per SIMD I could expect ?

[himanshu_gautam]

raistmer,

I think these limiting numbers would depend on the model of your gpu.Anyways we should always try to reduce the value of FCstack and GPR.

[ryta1203]

Originally posted by: himanshu.gautam raistmer,

I think these limiting numbers would depend on the model of your gpu.Anyways we should always try to reduce the value of FCstack and GPR.

I think "always" is a strong word here since there are cases where a reduction in registers would worsen performance.

[himanshu_gautam]

Hi ryta1203,
Thanks for correcting. Indeed in some cases less GPR usage and FCStack value might degrade the performance. Therefore the best way is:
1. Go through the OpenCL Programming Guide before trying optimization.
2. Do performance checks of your code using ATI Stream Profiler and SKA to identify the best optimization strategy in your case depending on the bottlenecks in performance.

[Raistmer]

My GPU is HD4870.
IMO these values are interdependent. I could do some loop unrols but this will lead to GPR usage increase.
Thats's why I ask what parameter is limiting one in my case.

[himanshu_gautam]

Refer to openCL programming guide.

The various factors and there limiting values are given with respect to hd5870.

They might clear your doubts.

thanks