fix deconvolution x86 unaligned bias load (#5112)

Tencent · Oct 27, 2023 · 6c6c40e · 6c6c40e
1 parent 16524ca
commit 6c6c40e
Showing 1 changed file with 3 additions and 3 deletions.
diff --git a/src/layer/x86/deconvolution_x86.cpp b/src/layer/x86/deconvolution_x86.cpp
@@ -300,7 +300,7 @@ int Deconvolution_x86::forward(const Mat& bottom_blob, Mat& top_blob, const Opti
                     }
                     else
                     {
-                        outm.fill(_mm512_load_ps((const float*)bias_data + p * 16));
+                        outm.fill(_mm512_loadu_ps((const float*)bias_data + p * 16));
                     }
 
                     for (int u = 0; u < kernel_h; u++)
@@ -344,7 +344,7 @@ int Deconvolution_x86::forward(const Mat& bottom_blob, Mat& top_blob, const Opti
                     }
                     else
                     {
-                        outm.fill(_mm256_load_ps((const float*)bias_data + p * 8));
+                        outm.fill(_mm256_loadu_ps((const float*)bias_data + p * 8));
                     }
 
                     for (int u = 0; u < kernel_h; u++)
@@ -388,7 +388,7 @@ int Deconvolution_x86::forward(const Mat& bottom_blob, Mat& top_blob, const Opti
                     }
                     else
                     {
-                        outm.fill(_mm_load_ps((const float*)bias_data + p * 4));
+                        outm.fill(_mm_loadu_ps((const float*)bias_data + p * 4));
                     }
 
                     for (int u = 0; u < kernel_h; u++)