extra

Author: serge-sans-paille

include/xsimd/arch/xsimd_rvv.hpp

@@ -1517,13 +1517,35 @@ namespace xsimd
         {
             XSIMD_IF_CONSTEXPR((8 * sizeof(T)) >= batch_bool<T, A>::size)
             {
+                // (A) Easy case: the number of slots fits in T.
                 const auto zero = detail::broadcast<as_unsigned_integer_t<T>, types::detail::rvv_width_m1>(T(0));
                 auto ones = detail::broadcast<as_unsigned_integer_t<T>, A::width>(1);
                 auto iota = detail::vindex<A, as_unsigned_integer_t<T>>();
                 auto upowers = detail::rvvsll(ones, iota);
                 auto r = __riscv_vredor(self.data.as_mask(), upowers, (typename decltype(zero)::register_type)zero, batch_bool<T, A>::size);
                 return detail::reduce_scalar<A, as_unsigned_integer_t<T>>(r);
             }
+            else XSIMD_IF_CONSTEXPR((2 * 8 * sizeof(T)) == batch_bool<T, A>::size)
+            {
+                // (B) We need two rounds, one for the low part, one for the high part.
+
+                // The low part is similar to the approach in (A).
+                const auto zero = detail::broadcast<as_unsigned_integer_t<T>, types::detail::rvv_width_m1>(T(0));
+                auto ones = detail::broadcast<as_unsigned_integer_t<T>, A::width>(1);
+                auto iota = ::vindex<A, as_unsigned_integer_t<T>>();
+                auto upowers_low = detail::rvvsll(ones, iota);
+                auto low_mask = self & (batch<T, A>(detail::vindex<A, T>()) < batch_bool<T, A>::size / 2);
+                auto r_low = __riscv_vredor(low_mask.as_mask(), upowers_low, (typename decltype(zero)::register_type)zero, batch_bool<T, A>::size);
+
+                // The high part requires a sub before the shift.
+                auto iota_high = __riscv_vsub(iota, 8 * sizeof(T), batch_bool<T, A>::size);
+                auto upowers_high = detail::rvvsll(ones, iota_high);
+                auto high_mask = self & (batch<T, A>(detail::vindex<A, T>()) >= batch_bool<T, A>::size / 2);
+                auto r_high = __riscv_vredor(high_mask.as_mask(), upowers_high, (typename decltype(zero)::register_type)zero, batch_bool<T, A>::size);
+
+                // Agglomerate the two parts.
+                return detail::reduce_scalar<A, as_unsigned_integer_t<T>>(r_low) | (detail::reduce_scalar<A, as_unsigned_integer_t<T>>(r_high) << (8 * sizeof(T)));
+            }
             else
             {
                 return mask(self, common {});