[PATCH] D110107: AMDGPU: Propagate amdgpu-waves-per-eu

Johannes Doerfert via Phabricator via llvm-commits llvm-commits at lists.llvm.org
Mon Sep 20 16:36:39 PDT 2021


jdoerfert added inline comments.


================
Comment at: llvm/lib/Target/AMDGPU/AMDGPUAttributor.cpp:651
+    intersectKnown(Range);
+  }
+
----------------
arsenm wrote:
> jdoerfert wrote:
> > Is it on purpose that you use the assumed/initial value of the flat-work-group-size as known range here?
> Yes, but I wasn't completely sure if this was the correct way to handle this. The amdgpu-flat-workgroup-size implies restrictions on amdgpu-waves-per-eu, but amdgpu-flat-workgroup-size should win if there's a conflict. There should always be an initially known range
So, the way it is now you will never go below the initial flat-workgroup-size.

>From what I read it seems like you want an upper bound? 
If so, you need to restrict the assumed value.
That way you'll go at most as big as the flat-workgroup-size initially is.

You might also need to consider doing this in update if you need to be bound by the final flat-workgroup-size and not only the initial one.


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D110107/new/

https://reviews.llvm.org/D110107



More information about the llvm-commits mailing list