There is a problem in operating floating-point data in K210's IOMEM.

.vscode/settings.json

@@ -0,0 +1,11 @@
+{
+    "files.associations": {
+        "util.h": "c",
+        "iomem.h": "c",
+        "utils.h": "c",
+        "sysctl.h": "c",
+        "printf.h": "c",
+        "nncase.h": "c",
+        "*.def": "cpp"
+    }
+}

lib/nncase/include/kernels/neutral/neutral_kernels.h

@@ -42,8 +42,12 @@ namespace kernels
                             const auto in_a_off = kernels::details::get_reduced_offset(in_off, in_a_shape);
                             const auto in_b_off = kernels::details::get_reduced_offset(in_off, in_b_shape);
                             const auto a = input_a[offset(in_a_shape, in_a_off)];
+#if FIX_CACHE
+                            float b;
+                            memcpy(&b, &input_b[offset(in_b_shape, in_b_off)], sizeof(float));
+#else
                             const auto b = input_b[offset(in_b_shape, in_b_off)];
-
+#endif
                             output[offset(out_shape, in_off)] = kernels::details::apply_activation(op(a, b), fused_activation);
                         }
                     }