max_work_group_size vs reqd_work_group_size

Altera_Forum · ‎03-24-2018

Hello,

Can anybody tell what is real difference between these two attributes: max_work_group_size and reqd_work_group_size?

for example between max_work_group_size(256) and reqd_work_group_size(256)? Is there any benefit in using reqd_work_group_size, while max_work_group_size is a super-set of it?

I guessed the reqd_work_group_size leads to more optimized hardware, but after some tests I saw even the hardware size did not change that much.

thanks

Altera_Forum · ‎03-24-2018

"max_work_group_size" is for cases where you want to change your work-group size at runtime, but you know the maximum limit for the work-group size and hence, supply the compiler with this info to prevent excessive local memory allocation. "reqd_work_group_size" is for cases where you are absolutely sure your work-group size is fixed at runtime, and supply the compiler with this info to minimize area utilization. In this case, if you try to invoke the kernel with a work-group size that is different from what is set in the kernel, you will get a runtime error. These attributes mainly affect local memory utilization and if you are not using local memory in your kernel, the area utilization difference will be minimal.