Khronos Public Bugzilla
Bug 498 - Desription of mapping work-groups to compute units
Summary: Desription of mapping work-groups to compute units
Status: NEW
Alias: None
Product: OpenCL
Classification: Unclassified
Component: Specification (show other bugs)
Version: 1.1
Hardware: All All
: P3 normal
Target Milestone: ---
Assignee: Aaftab Munshi
QA Contact: OpenCL Working Group
Depends on:
Reported: 2011-07-23 19:06 PDT by Sean Settle
Modified: 2011-07-23 19:06 PDT (History)
0 users

See Also:


Note You need to log in before you can comment on or make changes to this bug.
Description Sean Settle 2011-07-23 19:06:30 PDT
In a prior bug report someone requested some statement about the execution order of work-groups (order of starting, not necessarily finishing).  The discussion seemed to end without agreement on stating any requirements on this ordering.

However, given the existance of NUMA architectures and caches, I think at least there should be a statement on the order that work-groups are assigned to compute units.  Are they assigned randomly or in a round-robin fashion with possibly some chunksize of work-groups (e.g., assigning work-groups in contiguous pairs).  Without any statement I don‘t see how we can control the load balancing; and a poorly implemented ICD could potentially starve all compute units except one.

Please take a stance and specify how work-groups must be mapped to compute units, and provide a means to query the chunksize if such a concept exists in the mapping scheme.  Especially in the case where the number of work-groups is less than or equal to the number of compute units.