[PATCH] D149393: [AMDGPU][IGLP] Parameterize the SchedGroup processing / linking in Solver

[Jeffrey Byrnes via Phabricator via llvm-commits]

jrbyrnes added inline comments.


================
Comment at: llvm/lib/Target/AMDGPU/AMDGPUIGroupLP.cpp:804
 
+class DemoOpt final : public IGLPStrategy {
+private:
----------------
kerbowa wrote:
> Is this strategy just a demo of this new feature? What are your thoughts on upstreaming this?
Hey -- thanks for bringing that up. 

It is to demo the new feature -- put another way, it is to facilitate lit testing of the new feature.

The assumption is that clients of iglp_opt have an intimate understanding of the i32s passed to the builtin. However, I see that mixing testing and features is less than ideal.

Maybe it makes the most sense to have incremental tests (under iglp_opt(1)) that is superseded by the singlewave strategy which incorporates all the new features?

Otherwise, we will need a mechanism to test new features without changing the existing behavior of the current builtins -- whether it be under iglp_opt or a separate intrinsic.  


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D149393/new/

https://reviews.llvm.org/D149393