buffer: add SIMD Neon optimization for `byteLength`simd-bytelength

Co-authored-by: Keyhan Vakil <kvakil@sylph.kvakil.me>
Co-authored-by: Daniel Lemire <daniel@lemire.me>

1 files changed, 60 insertions, 0 deletions

diff --git a/src/node_simd.cc b/src/node_simd.cc
new file mode 100644
index 0000000000..a5265a95c0
--- /dev/null
+++ b/src/node_simd.cc
@@ -0,0 +1,60 @@
+#include "node_simd.h"
+
+#include <string_view>
+
+#if NODE_HAS_SIMD_NEON
+#include <arm_neon.h>
+#endif
+
+namespace node {
+namespace simd {
+
+#if NODE_HAS_SIMD_NEON
+uint32_t utf8_byte_length(const uint8_t* data, size_t length) {
+  uint64_t result{0};
+
+  const int lanes = sizeof(uint8x16_t);
+  const int max_sra_count = 256 / lanes;  // Avoid overflowing vaddvq_u8.
+  const int unrolls = max_sra_count;
+  const int unrolled_lanes = lanes * unrolls;
+
+  const uint8_t* unroll_end = data + (length / unrolled_lanes) * unrolled_lanes;
+  uint32_t length_after_unroll = length % unrolled_lanes;
+  for (; data < unroll_end;) {
+    uint8x16_t acc = {};
+    for (int i = 0; i < unrolls; ++i, data += lanes) {
+      uint8x16_t chunk = vld1q_u8(data);
+      acc = vsraq_n_u8(acc, chunk, 7);
+    }
+    result += vaddvq_u8(acc);
+  }
+
+  const uint8_t* simd_end = data + (length_after_unroll / lanes) * lanes;
+  uint32_t length_after_simd = length % lanes;
+  uint8x16_t acc = {};
+  for (; data < simd_end; data += lanes) {
+    uint8x16_t chunk = vld1q_u8(data);
+    acc = vsraq_n_u8(acc, chunk, 7);
+  }
+  result += vaddvq_u8(acc);
+
+  const uint8_t* scalar_end = data + length_after_simd;
+  for (; data < scalar_end; data += 1) {
+    result += *data >> 7;
+  }
+
+  return result + length;
+}
+#else
+uint32_t utf8_byte_length(const uint8_t* data, size_t length) {
+  uint32_t result = 0;
+  for (uint32_t i = 0; i < length; ++i) {
+    result += (data[i] >> 7);
+  }
+  result += length;
+  return result;
+}
+#endif
+
+}  // namespace simd
+}  // namespace node