Update softmax.cpp

Author: deepeshfujitsu

tests/benchdnn/softmax/softmax.cpp

@@ -236,6 +236,8 @@ void setup_cmp(compare::compare_t &cmp, const prb_t *prb, data_kind_t kind,
     // MIOpen and ACL softmax accumulate in F16, but oneDNN now expects accumulation in
     // F32, this partially reverts 6727bbe8. For more information on ACL softmax, see
     // https://github.com/oneapi-src/oneDNN/issues/1819
+    // Similarly, for bf16 on AArch64, the relaxed threshold is necessary due to 
+    // minor accuracy drops observed compared to f32
     const float trh = trh_f32;
 #else
     const bool is_strict_acc