[llvm] [AMDGPU] Enable scheduler mfma rewrite stage by default (PR #205901)

[via llvm-commits]

llvmorg-github-actions[bot] wrote:


<!--LLVM PR SUMMARY COMMENT-->

@llvm/pr-subscribers-backend-amdgpu

Author: Vigneshwar Jayakumar (VigneshwarJ)

<details>
<summary>Changes</summary>

Re-enable after #<!-- -->198555, #<!-- -->199708, and #<!-- -->200972 fix the correctness issues that led to disabling in #<!-- -->185604. Perf test with CK shows reduced spilling on MFMA kernels under VGPR pressure.

---
Full diff: https://github.com/llvm/llvm-project/pull/205901.diff


1 Files Affected:

- (modified) llvm/lib/Target/AMDGPU/GCNSchedStrategy.cpp (+1-1) 


``````````diff
diff --git a/llvm/lib/Target/AMDGPU/GCNSchedStrategy.cpp b/llvm/lib/Target/AMDGPU/GCNSchedStrategy.cpp
index a4f854beaeebe..c471b7026e851 100644
--- a/llvm/lib/Target/AMDGPU/GCNSchedStrategy.cpp
+++ b/llvm/lib/Target/AMDGPU/GCNSchedStrategy.cpp
@@ -101,7 +101,7 @@ static cl::opt<bool> PrintMaxRPRegUsageAfterScheduler(
 
 static cl::opt<bool> DisableRewriteMFMAFormSchedStage(
     "amdgpu-disable-rewrite-mfma-form-sched-stage", cl::Hidden,
-    cl::desc("Disable rewrite mfma rewrite scheduling stage"), cl::init(true));
+    cl::desc("Disable rewrite mfma rewrite scheduling stage"), cl::init(false));
 
 namespace {
 

``````````

</details>


https://github.com/llvm/llvm-project/pull/205901