GPU Compute Software
Ask questions about Intel® Graphics Compute software technologies, such as OpenCL* GPU driver and oneAPI Level Zero
222 Discussions

Understanding the partitioning of work-groups in multi-stacks PVC-based GPU systems

L30nardo
Novice
462 Views

I am reading the documentation on multi-GPU and multi-stack programming, specifically  Work scheduling and distribution . According to this, the partitioning of work-groups to stacks (in the composite mode) follows a simple rule:

 

... the slowest moving dimension is divided in number-of-stacks chunks and distributed round-robin between stacks ...

For SYCL, the slowest moving dimension is the first element of global range. 

 

I cannot understand what is meant by "first element of global range", even tough numerical examples are given:

 

For example, for global range {38,512,512} with local range {2,1,8} the driver would partition the y-dimension, while for local range {1,1,16} the driver would partition the z-dimension.

 

Could anybody give me some hints to understand what the slowest dimension is in this context?

Labels (1)
0 Replies
Reply