MasterJH5574 opened a new pull request, #14642:
URL: https://github.com/apache/tvm/pull/14642

   This PR adds a behavior to the MetaSchedule post-processor 
RewriteParallelVectorizeUnroll, so that it does not annotate spatial blocks 
with the unroll annotation.
   
   This is because the optimization for spatial blocks (standalone in a GPU 
kernel, for example) can be done by purely thread binding. As a result, 
annotating loop unrolling for spatial blocks does not help. In some case where 
the unroll factor is very large (e.g., 512 or 1024), unrolling the spatial 
blocks will consume much time during the kernel compilation and introduces 
unnecessary overhead.
   
   Therefore, we turn off the behavior of unrolling spatial blocks.


-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: [email protected]

For queries about this service, please contact Infrastructure at:
[email protected]

Reply via email to