reduction thing

Author: jhuber6

clang/include/clang/Basic/AttrDocs.td

@@ -1128,15 +1128,18 @@ general-purpose code.
 
   template <typename T, uint32_t N>
   constexpr T simd_reduce(T [[clang::ext_vector_type(N)]] v) {
-    T sum{};
-    for (uint32_t i = 0; i < N; ++i) {
-      sum += v[i];
+    static_assert((N & (N - 1)) == 0, "N must be a power of two");
+    if constexpr (N == 1) {
+      return v[0];
+    } else {
+      T [[clang::ext_vector_type(N / 2)]] reduced = v.hi + v.lo;
+      return simd_reduce<T, N / 2>(reduced);
     }
-    return sum;
   }
 
 The vector type also supports swizzling up to sixteen elements. This can be done
-using the object accessors.
+using the object accessors. The OpenCL documentation lists the full list of
+accepted values.
 .. code-block:: c++
 
   using f16_x16 = _Float16 __attribute__((ext_vector_type(16)));