From 35b1ef182c497439056b703b4607e29178234545 Mon Sep 17 00:00:00 2001
From: Xintong Han <hixintonghan@gmail.com>
Date: Mon, 28 Mar 2022 13:23:01 +0800
Subject: [PATCH] Fix bug in MNNSigmoidLowp

When dataSize is not a multiple of 4, the calculation is wrong as it does not move the dst address after the for loop.
---
 source/backend/cpu/compute/CommonOptFunction.cpp | 1 +
 1 file changed, 1 insertion(+)

diff --git a/source/backend/cpu/compute/CommonOptFunction.cpp b/source/backend/cpu/compute/CommonOptFunction.cpp
index 485b0c57..1fc87b81 100644
--- a/source/backend/cpu/compute/CommonOptFunction.cpp
+++ b/source/backend/cpu/compute/CommonOptFunction.cpp
@@ -2417,6 +2417,7 @@ void MNNSigmoidLowp(float* dst, const float* src, size_t dataSize) {
         }
         out = vrecpeq_f32(vaddq_f32(value,out));
         vst1q_f32(dst, out);
+        dst += 4;
         dataSize = dataSize - 4 * dataC4;
     }
 #endif