Archives Discussions

[Fuxianjun]

Is there a max number of work-group quanity for given device ? If yes ,how to query the max number?

I can't find anything about querying max number of work-group quanity in the table of param_name of clGetDeviceInfo() function.

Thanks !

[malcolm3141]

From my reading of the OpenCL spec, the max global index in any dimension is 2^32-1.

The spec states that an error will be returned when a global work size is specified that is bigger than the size_t type on the device (see page 133 of the OpenCL 1.1 spec).

Malcolm

[himanshu_gautam]

hi all,

As per spec there is no limit on the number of workgroups each device need to support.Obviously we cannot exceed 2^32-1 as the global thread id would not be representable in 32 bits.

I tried to create a test program & it ran successfully for 2^25 workgroups.After which it failed due to the lack of system resources.

So I think Gpus can execute as many workgroups in round robin method as system resources allow.The only upper limit being 2^32-1.

I hope there are no more issues regarding this topic anymore.