From 1351d9cd1b88c3aad1ad56987613421a22eeb58d Mon Sep 17 00:00:00 2001
From: Qianfeng Zhang <Qianfeng.Zhang@amd.com>
Date: Wed, 16 Apr 2025 06:54:06 +0000
Subject: [PATCH] Use exp2() to calculate exp() for better performance

---
 .../ck_tile/18_hstu_attention/hstu_attention_fwd_pipeline.hpp   | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/example/ck_tile/18_hstu_attention/hstu_attention_fwd_pipeline.hpp b/example/ck_tile/18_hstu_attention/hstu_attention_fwd_pipeline.hpp
index 4942b67606..cb4d79b7a2 100644
--- a/example/ck_tile/18_hstu_attention/hstu_attention_fwd_pipeline.hpp
+++ b/example/ck_tile/18_hstu_attention/hstu_attention_fwd_pipeline.hpp
@@ -251,7 +251,7 @@ struct HstuAttentionFwdPipelineQRKSVS
         const auto f_silu = [](CompDataType& x) {
             const auto one = ck_tile::type_convert<CompDataType>(1.0f);
 
-            return x = x / (one + exp(-x));
+            return x = x / (one + exp2(-log2e_v<CompDataType> * x));
         };
 
         using OaccBlockTileType = decltype(gemm_1.MakeCBlockTile());