cancel
Showing results for 
Search instead for 
Did you mean: 

Archives Discussions

vmiura
Adept II

OpenCL max work group size on Southern Islands

Hello,

I always see CL_DEVICE_MAX_WORK_GROUP_SIZE set to 256 on the Southern Islands (e.g. Tahiti, Pitcairn), but according to the AMD Southern Islands ISA doc section 4.3, "Up to 16 wavefronts (1024 work-items) can be combined into a work-group.

Is it possible to enable large work group sizes up to 1024 in the driver?

Larger work group size is good for image processing with overlapped title border, for example I have one filter with a 4 pixel border.  With 1024 work-items, I can do 32*32 tiles with a 24*24 payload (56% efficiency), whereas with 256 work-items it would be 16*16 tiles with 8*8 payload (25% efficiency).

Thanks,

Victor

0 Likes
2 Replies
nou
Exemplar

try set GPU_MAX_WORKGROUP_SIZE enviroment variable. it may override max workgroup size.

0 Likes
LeeHowes
Staff

Remember that you don't have to do just one pixel per work-item... the only real reason to use big workgroup sizes is if you are using enough local memory that without doing so you lose latency hiding.

0 Likes