Merge branch 'master' into x86-conv-gemm-0

Tencent · Jan 25, 2025 · 285bd90 · 285bd90
2 parents 564cd1d + 48e1260
commit 285bd90
Show file tree

Hide file tree

Showing 9 changed files with 704 additions and 5,378 deletions.
diff --git a/src/layer/arm/convolution_arm.cpp b/src/layer/arm/convolution_arm.cpp
@@ -1376,7 +1376,16 @@ int Convolution_arm::forward_int8_arm(const Mat& bottom_blob, Mat& top_blob, con
 #if __ARM_NEON
     if (opt.use_packing_layout)
     {
-        out_elempack_int32 = num_output % 8 == 0 ? 8 : num_output % 4 == 0 ? 4 : 1;
+#if NCNN_ARM82
+        if (ncnn::cpu_support_arm_asimdhp() && opt.use_fp16_arithmetic)
+        {
+            out_elempack_int32 = num_output % 8 == 0 ? 8 : num_output % 4 == 0 ? 4 : 1;
+        }
+        else
+#endif // NCNN_ARM82
+        {
+            out_elempack_int32 = num_output % 4 == 0 ? 4 : 1;
+        }
     }
 #endif // __ARM_NEON