[KB] bf16 element-wise is done at F32 on CPU

rengolin · rengolin · commit a1e6d30068a8 · 2026-06-16T15:28:44.000+01:00
diff --git a/examples/KernelBench/schedules/x86_64/element_wise/bf16.yaml b/examples/KernelBench/schedules/x86_64/element_wise/bf16.yaml
@@ -1,5 +1,6 @@
 Pipeline:
+  # Element-wise is usually done at F32 precision anyway
   - include: common.yaml{
-                          register_tile=32
+                          register_tile=16
                           unroll_factor=16
                         }

Original file line number	Diff line number	Diff line change
`@@ -1,5 +1,6 @@`
`1`	`1`	`Pipeline:`
	`2`	`+ # Element-wise is usually done at F32 precision anyway`
`2`	`3`	`- include: common.yaml{`
`3`		`- register_tile=32`
	`4`	`+ register_tile=16`
`4`	`5`	`unroll_factor=16`
`5`	`6`	`}`