Archives Discussions

[dschaa]

Currently we're using the environment variable GPU_MAX_WORKGROUP_SIZE to get around the 256 work item limit per work group.  I understand that there are performance implications for requiring too many work items to be active on a compute unit at a time, however, is there any chance of execution failure or incorrect computation?  We'd like the work group size to be up to 768. 

Thanks!

[d_a_a_]

Why don't you just increase the granularity of each work-item, so each one would process three elements (768/256) instead of only one?

[dschaa]

Originally posted by: d.a.a. Why don't you just increase the granularity of each work-item, so each one would process three elements (768/256) instead of only one?

Thanks for the reply.  I'm interested in this question in general, regardless if it's possible for a work item to process multiple elements.  Also, the code would become much more complex.  We have a bunch of different kernels that require greater than 256 work items, and multiple conditionals would be required for some of the kernels to ensure correct execution.