Split by HWY_NATIVE_SCATTER/HWY_NATIVE_GATHER

Mousius · Mousius · commit f5258670685e · 2024-04-12T17:49:07.000Z
diff --git a/hwy/ops/generic_ops-inl.h b/hwy/ops/generic_ops-inl.h
@@ -2679,6 +2679,25 @@ HWY_API void MaskedScatterIndex(VFromD<D> v, MFromD<D> m, D d,
   }
 }
 
+template <class D, typename T = TFromD<D>>
+HWY_API void ScatterIndexN(VFromD<D> v, D d, T* HWY_RESTRICT base,
+                           VFromD<RebindToSigned<D>> index,
+                           const size_t max_lanes_to_store) {
+  const RebindToSigned<decltype(d)> di;
+  using TI = TFromD<decltype(di)>;
+  static_assert(sizeof(T) == sizeof(TI), "Index/lane size must match");
+
+  for (size_t i = 0; i < MaxLanes(d); ++i) {
+    if (i < max_lanes_to_store) base[ExtractLane(index, i)] = ExtractLane(v, i);
+  }
+}
+#else
+template <class D, typename T = TFromD<D>>
+HWY_API void ScatterIndexN(VFromD<D> v, D d, T* HWY_RESTRICT base,
+                           VFromD<RebindToSigned<D>> index,
+                           const size_t max_lanes_to_store) {
+  MaskedScatterIndex(v, FirstN(d, max_lanes_to_store), d, base, index);
+}
 #endif  // (defined(HWY_NATIVE_SCATTER) == defined(HWY_TARGET_TOGGLE))
 
 // ------------------------------ Gather
@@ -2774,23 +2793,6 @@ HWY_API VFromD<D> MaskedGatherIndexOr(VFromD<D> no, MFromD<D> m, D d,
   return Load(d, lanes);
 }
 
-#endif  // (defined(HWY_NATIVE_GATHER) == defined(HWY_TARGET_TOGGLE))
-
-// ------------------------------ ScatterN/GatherN
-
-template <class D, typename T = TFromD<D>>
-HWY_API void ScatterIndexN(VFromD<D> v, D d, T* HWY_RESTRICT base,
-                           VFromD<RebindToSigned<D>> index,
-                           const size_t max_lanes_to_store) {
-  const RebindToSigned<decltype(d)> di;
-  using TI = TFromD<decltype(di)>;
-  static_assert(sizeof(T) == sizeof(TI), "Index/lane size must match");
-
-  for (size_t i = 0; i < MaxLanes(d); ++i) {
-    if (i < max_lanes_to_store) base[ExtractLane(index, i)] = ExtractLane(v, i);
-  }
-}
-
 template <class D, typename T = TFromD<D>>
 HWY_API VFromD<D> GatherIndexN(D d, const T* HWY_RESTRICT base,
                                VFromD<RebindToSigned<D>> index,
@@ -2806,6 +2808,14 @@ HWY_API VFromD<D> GatherIndexN(D d, const T* HWY_RESTRICT base,
   }
   return v;
 }
+#else
+template <class D, typename T = TFromD<D>>
+HWY_API VFromD<D> GatherIndexN(D d, const T* HWY_RESTRICT base,
+                               VFromD<RebindToSigned<D>> index,
+                               const size_t max_lanes_to_load) {
+  return MaskedGatherIndex(FirstN(d, max_lanes_to_load), d, base, index);
+}
+#endif  // (defined(HWY_NATIVE_GATHER) == defined(HWY_TARGET_TOGGLE))
 
 // ------------------------------ Integer AbsDiff and SumsOf8AbsDiff
 
