[AccelerateMatmul] Sync from upstream

Signed-off-by: Whitney Tsang <[email protected]>
intel · Jan 11, 2025 · 5f5feb0 · 5f5feb0
1 parent 5bab216
commit 5f5feb0
Showing 1 changed file with 4 additions and 1 deletion.
diff --git a/third_party/intel/lib/TritonIntelGPUTransforms/AccelerateMatmul.cpp b/third_party/intel/lib/TritonIntelGPUTransforms/AccelerateMatmul.cpp
@@ -421,7 +421,10 @@ class DecomposeScaledBlocked : public OpRewritePattern<tt::DotScaledOp> {
     if (!scale)
       return v;
 
-    return rewriter.create<ttg::UpcastMXFPOp>(v.getLoc(), v, scale, elemType);
+    auto retTy = triton::gpu::UpcastMXFPOp::deduceOutputType(
+        v, elemType, Builder(v.getContext()).getBF16Type());
+    return rewriter.create<ttg::UpcastMXFPOp>(v.getLoc(), retTy, v, scale,
+                                              elemType);
   }
 };