--- a/include/everett/select_groups.h
+++ b/include/everett/select_groups.h
@@ -24 +24 @@
-#if defined(__BMI2__)
+#if defined(__BMI2__) || defined(__AVX2__)
@@ -25,0 +26,3 @@
+#endif
+#if defined(__aarch64__) && defined(__ARM_NEON)
+#include <arm_neon.h>
@@ -41,0 +45,21 @@
+    }
+
+    inline std::array<unsigned, 4> populations4(std::uint64_t const * words) noexcept {
+#if defined(__aarch64__) && defined(__ARM_NEON)
+      auto a = vpaddlq_u32(vpaddlq_u16(vpaddlq_u8(vcntq_u8(vld1q_u8(reinterpret_cast<std::uint8_t const *>(words))))));
+      auto b = vpaddlq_u32(vpaddlq_u16(vpaddlq_u8(vcntq_u8(vld1q_u8(reinterpret_cast<std::uint8_t const *>(words + 2))))));
+      return {unsigned(vgetq_lane_u64(a, 0)), unsigned(vgetq_lane_u64(a, 1)),
+              unsigned(vgetq_lane_u64(b, 0)), unsigned(vgetq_lane_u64(b, 1))};
+#elif defined(__AVX2__)
+      auto table = _mm256_setr_epi8(0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4);
+      auto bits = _mm256_loadu_si256(reinterpret_cast<__m256i const *>(words));
+      auto mask = _mm256_set1_epi8(15);
+      auto bytes = _mm256_add_epi8(_mm256_shuffle_epi8(table, _mm256_and_si256(bits, mask)),
+        _mm256_shuffle_epi8(table, _mm256_and_si256(_mm256_srli_epi16(bits, 4), mask)));
+      auto counts = _mm256_sad_epu8(bytes, _mm256_setzero_si256());
+      return {unsigned(_mm256_extract_epi64(counts, 0)), unsigned(_mm256_extract_epi64(counts, 1)),
+              unsigned(_mm256_extract_epi64(counts, 2)), unsigned(_mm256_extract_epi64(counts, 3))};
+#else
+      return {unsigned(std::popcount(words[0])), unsigned(std::popcount(words[1])),
+              unsigned(std::popcount(words[2])), unsigned(std::popcount(words[3]))};
+#endif
@@ -107,3 +131,20 @@
-        for (; source.size() - at >= inputs; at += inputs)
-          low_tile<W>(source.data() + at, out.data() + (at / inputs) * outputs,
-                      std::make_index_sequence<outputs>{});
+        for (; source.size() - at >= inputs; at += inputs) {
+          auto destination = out.data() + (at / inputs) * outputs;
+#if defined(__aarch64__) && defined(__ARM_NEON) && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
+          if constexpr (W == 32) {
+            vst1_u32(reinterpret_cast<std::uint32_t *>(destination), vmovn_u64(vld1q_u64(source.data() + at)));
+          } else if constexpr (W == 16) {
+            auto a = vmovn_u64(vld1q_u64(source.data() + at));
+            auto b = vmovn_u64(vld1q_u64(source.data() + at + 2));
+            vst1_u16(reinterpret_cast<std::uint16_t *>(destination), vmovn_u32(vcombine_u32(a, b)));
+          } else if constexpr (W == 8) {
+            auto a = vmovn_u64(vld1q_u64(source.data() + at));
+            auto b = vmovn_u64(vld1q_u64(source.data() + at + 2));
+            auto c = vmovn_u64(vld1q_u64(source.data() + at + 4));
+            auto d = vmovn_u64(vld1q_u64(source.data() + at + 6));
+            vst1_u8(reinterpret_cast<std::uint8_t *>(destination),
+              vmovn_u16(vcombine_u16(vmovn_u32(vcombine_u32(a, b)), vmovn_u32(vcombine_u32(c, d)))));
+          } else
+#endif
+            low_tile<W>(source.data() + at, destination, std::make_index_sequence<outputs>{});
+        }
@@ -254,0 +296,10 @@
+        if (scanned && high_.size() - word >= 4 && scanned <= 61) {
+          auto populations = select_groups_detail::populations4(high_.data() + word);
+          auto total = populations[0] + populations[1] + populations[2] + populations[3];
+          if (remaining >= total) { remaining -= total; scanned += 3; word += 3; continue; }
+          unsigned lane = 0;
+          while (remaining >= populations[lane]) { remaining -= populations[lane]; ++lane; }
+          word += lane;
+          value = high_[word];
+          population = populations[lane];
+        }
@@ -286,2 +337,9 @@
-      for (std::size_t i = 1; i < residuals.size(); ++i)
-        if (residuals[i] < residuals[i - 1])
+      std::size_t checked = 1;
+#if defined(__aarch64__) && defined(__ARM_NEON)
+      for (; residuals.size() - checked >= 2; checked += 2)
+        if (vmaxvq_u32(vreinterpretq_u32_u64(vcltq_u64(vld1q_u64(residuals.data() + checked),
+                                                      vld1q_u64(residuals.data() + checked - 1)))))
+          throw std::invalid_argument("select_groups nonmonotone offsets");
+#endif
+      for (; checked < residuals.size(); ++checked)
+        if (residuals[checked] < residuals[checked - 1])