From a321287e41400f3fdbb71d9f540dd026cd5d6dcc Mon Sep 17 00:00:00 2001
From: nihui <nihuini@tencent.com>
Date: Thu, 14 Nov 2024 17:05:07 +0800
Subject: [PATCH] fix avx

---
 src/layer/x86/gemm_int8.h | 15 ++++++++++++---
 1 file changed, 12 insertions(+), 3 deletions(-)

diff --git a/src/layer/x86/gemm_int8.h b/src/layer/x86/gemm_int8.h
index bcf5a74db62..6a3e42715bc 100644
--- a/src/layer/x86/gemm_int8.h
+++ b/src/layer/x86/gemm_int8.h
@@ -3837,7 +3837,10 @@ static void pack_A_tile_fp32_to_int8(const Mat& A, Mat& AT, int i, int max_ii, i
         }
     }
 #if !__AVX2__
-    pp = pp1;
+    if (max_ii >= 8)
+    {
+        pp = pp1;
+    }
 #endif
 #endif // __AVX__
     for (; ii + 3 < max_ii; ii += 4)
@@ -7138,7 +7141,10 @@ static void transpose_pack_A_tile_fp32_to_int8(const Mat& A, Mat& AT, int i, int
         }
     }
 #if !__AVX2__
-    pp = pp1;
+    if (max_ii >= 8)
+    {
+        pp = pp1;
+    }
 #endif
 #endif // __AVX__
     for (; ii + 3 < max_ii; ii += 4)
@@ -15357,7 +15363,10 @@ static void unpack_output_tile_int32_to_fp32(const Mat& topT, const Mat& C, Mat&
         }
     }
 #if !__AVX2__
-    pp = pp1;
+    if (max_ii >= 8)
+    {
+        pp = pp1;
+    }
 #endif
 #endif // __AVX__
     for (; ii + 3 < max_ii; ii += 4)