react-native-quick-crypto 1.0.19 → 1.1.1

package/deps/simdutf/src/simdutf/lsx/simd.h

@@ -0,0 +1,421 @@
+#ifndef SIMDUTF_LSX_SIMD_H
+#define SIMDUTF_LSX_SIMD_H
+
+#include "simdutf.h"
+#include "simdutf/lsx/bitmanipulation.h"
+
+namespace simdutf {
+namespace SIMDUTF_IMPLEMENTATION {
+namespace {
+namespace simd {
+
+template <typename T> struct simd8;
+
+//
+// Base class of simd8<uint8_t> and simd8<bool>, both of which use __m128i
+// internally.
+//
+template <typename T, typename Mask = simd8<bool>> struct base_u8 {
+  __m128i value;
+  static const int SIZE = sizeof(value);
+
+  // Conversion from/to SIMD register
+  simdutf_really_inline base_u8(const __m128i _value) : value(_value) {}
+  simdutf_really_inline operator const __m128i &() const { return this->value; }
+  simdutf_really_inline operator __m128i &() { return this->value; }
+
+  // Bit operations
+  simdutf_really_inline simd8<T> operator|(const simd8<T> other) const {
+    return __lsx_vor_v(this->value, other);
+  }
+  simdutf_really_inline simd8<T> operator&(const simd8<T> other) const {
+    return __lsx_vand_v(this->value, other);
+  }
+  simdutf_really_inline simd8<T> operator^(const simd8<T> other) const {
+    return __lsx_vxor_v(this->value, other);
+  }
+  simdutf_really_inline simd8<T> operator~() const { return *this ^ 0xFFu; }
+  simdutf_really_inline simd8<T> &operator|=(const simd8<T> other) {
+    auto this_cast = static_cast<simd8<T> *>(this);
+    *this_cast = *this_cast | other;
+    return *this_cast;
+  }
+
+  friend simdutf_really_inline Mask operator==(const simd8<T> lhs,
+                                               const simd8<T> rhs) {
+    return __lsx_vseq_b(lhs, rhs);
+  }
+
+  template <int N = 1>
+  simdutf_really_inline simd8<T> prev(const simd8<T> prev_chunk) const {
+    return __lsx_vor_v(__lsx_vbsll_v(this->value, N),
+                       __lsx_vbsrl_v(prev_chunk.value, 16 - N));
+  }
+};
+
+// SIMD byte mask type (returned by things like eq and gt)
+template <> struct simd8<bool> : base_u8<bool> {
+  typedef uint16_t bitmask_t;
+  typedef uint32_t bitmask2_t;
+
+  static simdutf_really_inline simd8<bool> splat(bool _value) {
+    return __lsx_vreplgr2vr_b(uint8_t(-(!!_value)));
+  }
+
+  simdutf_really_inline simd8(const __m128i _value) : base_u8<bool>(_value) {}
+  // False constructor
+  simdutf_really_inline simd8() : simd8(__lsx_vldi(0)) {}
+  // Splat constructor
+  simdutf_really_inline simd8(bool _value) : simd8(splat(_value)) {}
+  simdutf_really_inline void store(uint8_t dst[16]) const {
+    return __lsx_vst(this->value, dst, 0);
+  }
+
+  simdutf_really_inline uint32_t to_bitmask() const {
+    return __lsx_vpickve2gr_wu(__lsx_vmsknz_b(*this), 0);
+  }
+};
+
+// Unsigned bytes
+template <> struct simd8<uint8_t> : base_u8<uint8_t> {
+  static simdutf_really_inline simd8<uint8_t> splat(uint8_t _value) {
+    return __lsx_vreplgr2vr_b(_value);
+  }
+  static simdutf_really_inline simd8<uint8_t> zero() { return __lsx_vldi(0); }
+  static simdutf_really_inline simd8<uint8_t> load(const uint8_t *values) {
+    return __lsx_vld(values, 0);
+  }
+  simdutf_really_inline simd8(const __m128i _value)
+      : base_u8<uint8_t>(_value) {}
+  // Zero constructor
+  simdutf_really_inline simd8() : simd8(zero()) {}
+  // Array constructor
+  simdutf_really_inline simd8(const uint8_t values[16]) : simd8(load(values)) {}
+  // Splat constructor
+  simdutf_really_inline simd8(uint8_t _value) : simd8(splat(_value)) {}
+  // Member-by-member initialization
+  simdutf_really_inline
+  simd8(uint8_t v0, uint8_t v1, uint8_t v2, uint8_t v3, uint8_t v4, uint8_t v5,
+        uint8_t v6, uint8_t v7, uint8_t v8, uint8_t v9, uint8_t v10,
+        uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14, uint8_t v15)
+      : simd8((__m128i)v16u8{v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11,
+                             v12, v13, v14, v15}) {}
+
+  // Repeat 16 values as many times as necessary (usually for lookup tables)
+  simdutf_really_inline static simd8<uint8_t>
+  repeat_16(uint8_t v0, uint8_t v1, uint8_t v2, uint8_t v3, uint8_t v4,
+            uint8_t v5, uint8_t v6, uint8_t v7, uint8_t v8, uint8_t v9,
+            uint8_t v10, uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14,
+            uint8_t v15) {
+    return simd8<uint8_t>(v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12,
+                          v13, v14, v15);
+  }
+
+  // Store to array
+  simdutf_really_inline void store(uint8_t dst[16]) const {
+    return __lsx_vst(this->value, dst, 0);
+  }
+
+  // Order-specific operations
+  simdutf_really_inline simd8<bool>
+  operator>=(const simd8<uint8_t> other) const {
+    return __lsx_vsle_bu(other, *this);
+  }
+  simdutf_really_inline simd8<bool>
+  operator>(const simd8<uint8_t> other) const {
+    return __lsx_vslt_bu(other, *this);
+  }
+  simdutf_really_inline simd8 &operator-=(const simd8<uint8_t> other) {
+    value = __lsx_vsub_b(value, other.value);
+    return *this;
+  }
+  // Same as >, but instead of guaranteeing all 1's == true, false = 0 and true
+  // = nonzero. For ARM, returns all 1's.
+  simdutf_really_inline simd8<uint8_t>
+  gt_bits(const simd8<uint8_t> other) const {
+    return simd8<uint8_t>(*this > other);
+  }
+
+  // Bit-specific operations
+  simdutf_really_inline simd8<bool> any_bits_set(simd8<uint8_t> bits) const {
+    return __lsx_vslt_bu(__lsx_vldi(0), __lsx_vand_v(this->value, bits));
+  }
+  simdutf_really_inline bool is_ascii() const {
+    return __lsx_vpickve2gr_hu(__lsx_vmskgez_b(this->value), 0) == 0xFFFF;
+  }
+
+  simdutf_really_inline bool any_bits_set_anywhere() const {
+    return __lsx_vpickve2gr_hu(__lsx_vmsknz_b(this->value), 0) > 0;
+  }
+  template <int N> simdutf_really_inline simd8<uint8_t> shr() const {
+    return __lsx_vsrli_b(this->value, N);
+  }
+  template <int N> simdutf_really_inline simd8<uint8_t> shl() const {
+    return __lsx_vslli_b(this->value, N);
+  }
+
+  // Perform a lookup assuming the value is between 0 and 16 (undefined behavior
+  // for out of range values)
+  template <typename L>
+  simdutf_really_inline simd8<L> lookup_16(simd8<L> lookup_table) const {
+    return lookup_table.apply_lookup_16_to(*this);
+  }
+
+  template <typename L>
+  simdutf_really_inline simd8<L>
+  lookup_16(L replace0, L replace1, L replace2, L replace3, L replace4,
+            L replace5, L replace6, L replace7, L replace8, L replace9,
+            L replace10, L replace11, L replace12, L replace13, L replace14,
+            L replace15) const {
+    return lookup_16(simd8<L>::repeat_16(
+        replace0, replace1, replace2, replace3, replace4, replace5, replace6,
+        replace7, replace8, replace9, replace10, replace11, replace12,
+        replace13, replace14, replace15));
+  }
+
+  template <typename T>
+  simdutf_really_inline simd8<uint8_t>
+  apply_lookup_16_to(const simd8<T> original) const {
+    __m128i original_tmp = __lsx_vand_v(original, __lsx_vldi(0x1f));
+    return __lsx_vshuf_b(__lsx_vldi(0), *this, simd8<uint8_t>(original_tmp));
+  }
+
+  simdutf_really_inline uint64_t sum_bytes() const {
+    const auto sum_u16 = __lsx_vhaddw_hu_bu(value, value);
+    const auto sum_u32 = __lsx_vhaddw_wu_hu(sum_u16, sum_u16);
+    const auto sum_u64 = __lsx_vhaddw_du_wu(sum_u32, sum_u32);
+
+    return uint64_t(__lsx_vpickve2gr_du(sum_u64, 0)) +
+           uint64_t(__lsx_vpickve2gr_du(sum_u64, 1));
+  }
+};
+
+// Signed bytes
+template <> struct simd8<int8_t> {
+  __m128i value;
+
+  static const int SIZE = sizeof(value);
+
+  static simdutf_really_inline simd8<int8_t> splat(int8_t _value) {
+    return __lsx_vreplgr2vr_b(_value);
+  }
+  static simdutf_really_inline simd8<int8_t> zero() { return __lsx_vldi(0); }
+  static simdutf_really_inline simd8<int8_t> load(const int8_t values[16]) {
+    return __lsx_vld(values, 0);
+  }
+
+  template <endianness big_endian>
+  simdutf_really_inline void store_ascii_as_utf16(char16_t *p) const {
+    __m128i zero = __lsx_vldi(0);
+    if simdutf_constexpr (match_system(big_endian)) {
+      __lsx_vst(__lsx_vilvl_b(zero, (__m128i)this->value),
+                reinterpret_cast<uint16_t *>(p), 0);
+      __lsx_vst(__lsx_vilvh_b(zero, (__m128i)this->value),
+                reinterpret_cast<uint16_t *>(p + 8), 0);
+    } else {
+      __lsx_vst(__lsx_vilvl_b((__m128i)this->value, zero),
+                reinterpret_cast<uint16_t *>(p), 0);
+      __lsx_vst(__lsx_vilvh_b((__m128i)this->value, zero),
+                reinterpret_cast<uint16_t *>(p + 8), 0);
+    }
+  }
+
+  simdutf_really_inline void store_ascii_as_utf32(char32_t *p) const {
+    __m128i zero = __lsx_vldi(0);
+    __m128i in16low = __lsx_vilvl_b(zero, (__m128i)this->value);
+    __m128i in16high = __lsx_vilvh_b(zero, (__m128i)this->value);
+    __m128i in32_0 = __lsx_vilvl_h(zero, in16low);
+    __m128i in32_1 = __lsx_vilvh_h(zero, in16low);
+    __m128i in32_2 = __lsx_vilvl_h(zero, in16high);
+    __m128i in32_3 = __lsx_vilvh_h(zero, in16high);
+    __lsx_vst(in32_0, reinterpret_cast<uint32_t *>(p), 0);
+    __lsx_vst(in32_1, reinterpret_cast<uint32_t *>(p + 4), 0);
+    __lsx_vst(in32_2, reinterpret_cast<uint32_t *>(p + 8), 0);
+    __lsx_vst(in32_3, reinterpret_cast<uint32_t *>(p + 12), 0);
+  }
+
+  // In places where the table can be reused, which is most uses in simdutf, it
+  // is worth it to do 4 table lookups, as there is no direct zero extension
+  // from u8 to u32.
+  simdutf_really_inline void store_ascii_as_utf32_tbl(char32_t *p) const {
+    const simd8<uint8_t> tb1{0, 255, 255, 255, 1, 255, 255, 255,
+                             2, 255, 255, 255, 3, 255, 255, 255};
+    const simd8<uint8_t> tb2{4, 255, 255, 255, 5, 255, 255, 255,
+                             6, 255, 255, 255, 7, 255, 255, 255};
+    const simd8<uint8_t> tb3{8,  255, 255, 255, 9,  255, 255, 255,
+                             10, 255, 255, 255, 11, 255, 255, 255};
+    const simd8<uint8_t> tb4{12, 255, 255, 255, 13, 255, 255, 255,
+                             14, 255, 255, 255, 15, 255, 255, 255};
+
+    // encourage store pairing and interleaving
+    const auto shuf1 = this->apply_lookup_16_to(tb1);
+    const auto shuf2 = this->apply_lookup_16_to(tb2);
+    shuf1.store(reinterpret_cast<int8_t *>(p));
+    shuf2.store(reinterpret_cast<int8_t *>(p + 4));
+
+    const auto shuf3 = this->apply_lookup_16_to(tb3);
+    const auto shuf4 = this->apply_lookup_16_to(tb4);
+    shuf3.store(reinterpret_cast<int8_t *>(p + 8));
+    shuf4.store(reinterpret_cast<int8_t *>(p + 12));
+  }
+  // Conversion from/to SIMD register
+  simdutf_really_inline simd8(const __m128i _value) : value(_value) {}
+
+  // Zero constructor
+  simdutf_really_inline simd8() : simd8(zero()) {}
+  // Splat constructor
+  simdutf_really_inline simd8(int8_t _value) : simd8(splat(_value)) {}
+  // Array constructor
+  simdutf_really_inline simd8(const int8_t *values) : simd8(load(values)) {}
+
+  // Store to array
+  simdutf_really_inline void store(int8_t dst[16]) const {
+    return __lsx_vst(value, dst, 0);
+  }
+
+  simdutf_really_inline operator simd8<uint8_t>() const {
+    return ((__m128i)this->value);
+  }
+
+  simdutf_really_inline simd8<int8_t>
+  operator|(const simd8<int8_t> other) const {
+    return __lsx_vor_v((__m128i)value, (__m128i)other.value);
+  }
+
+  simdutf_really_inline bool is_ascii() const {
+    return (__lsx_vpickve2gr_hu(__lsx_vmskgez_b((__m128i)this->value), 0) ==
+            0xffff);
+  }
+
+  // Order-sensitive comparisons
+  simdutf_really_inline simd8<bool> operator>(const simd8<int8_t> other) const {
+    return __lsx_vslt_b((__m128i)other.value, (__m128i)value);
+  }
+  simdutf_really_inline simd8<bool> operator<(const simd8<int8_t> other) const {
+    return __lsx_vslt_b((__m128i)value, (__m128i)other.value);
+  }
+
+  template <int N = 1>
+  simdutf_really_inline simd8<int8_t>
+  prev(const simd8<int8_t> prev_chunk) const {
+    return __lsx_vor_v(__lsx_vbsll_v(this->value, N),
+                       __lsx_vbsrl_v(prev_chunk.value, 16 - N));
+  }
+
+  template <typename T>
+  simdutf_really_inline simd8<int8_t>
+  apply_lookup_16_to(const simd8<T> original) const {
+    __m128i original_tmp = __lsx_vand_v(original, __lsx_vldi(0x1f));
+    return __lsx_vshuf_b(__lsx_vldi(0), (__m128i)this->value,
+                         simd8<uint8_t>(original_tmp));
+  }
+};
+
+template <typename T> struct simd8x64 {
+  static constexpr int NUM_CHUNKS = 64 / sizeof(simd8<T>);
+  static_assert(
+      NUM_CHUNKS == 4,
+      "LoongArch kernel should use four registers per 64-byte block.");
+  simd8<T> chunks[NUM_CHUNKS];
+
+  simd8x64(const simd8x64<T> &o) = delete; // no copy allowed
+  simd8x64<T> &
+  operator=(const simd8<T> other) = delete; // no assignment allowed
+  simd8x64() = delete;                      // no default constructor allowed
+
+  simdutf_really_inline simd8x64(const simd8<T> chunk0, const simd8<T> chunk1,
+                                 const simd8<T> chunk2, const simd8<T> chunk3)
+      : chunks{chunk0, chunk1, chunk2, chunk3} {}
+  simdutf_really_inline simd8x64(const T *ptr)
+      : chunks{simd8<T>::load(ptr),
+               simd8<T>::load(ptr + sizeof(simd8<T>) / sizeof(T)),
+               simd8<T>::load(ptr + 2 * sizeof(simd8<T>) / sizeof(T)),
+               simd8<T>::load(ptr + 3 * sizeof(simd8<T>) / sizeof(T))} {}
+
+  simdutf_really_inline void store(T *ptr) const {
+    this->chunks[0].store(ptr + sizeof(simd8<T>) * 0 / sizeof(T));
+    this->chunks[1].store(ptr + sizeof(simd8<T>) * 1 / sizeof(T));
+    this->chunks[2].store(ptr + sizeof(simd8<T>) * 2 / sizeof(T));
+    this->chunks[3].store(ptr + sizeof(simd8<T>) * 3 / sizeof(T));
+  }
+
+  simdutf_really_inline simd8x64<T> &operator|=(const simd8x64<T> &other) {
+    this->chunks[0] |= other.chunks[0];
+    this->chunks[1] |= other.chunks[1];
+    this->chunks[2] |= other.chunks[2];
+    this->chunks[3] |= other.chunks[3];
+    return *this;
+  }
+
+  simdutf_really_inline simd8<T> reduce_or() const {
+    return (this->chunks[0] | this->chunks[1]) |
+           (this->chunks[2] | this->chunks[3]);
+  }
+
+  simdutf_really_inline bool is_ascii() const { return reduce_or().is_ascii(); }
+
+  template <endianness endian>
+  simdutf_really_inline void store_ascii_as_utf16(char16_t *ptr) const {
+    this->chunks[0].template store_ascii_as_utf16<endian>(ptr +
+                                                          sizeof(simd8<T>) * 0);
+    this->chunks[1].template store_ascii_as_utf16<endian>(ptr +
+                                                          sizeof(simd8<T>) * 1);
+    this->chunks[2].template store_ascii_as_utf16<endian>(ptr +
+                                                          sizeof(simd8<T>) * 2);
+    this->chunks[3].template store_ascii_as_utf16<endian>(ptr +
+                                                          sizeof(simd8<T>) * 3);
+  }
+
+  simdutf_really_inline void store_ascii_as_utf32(char32_t *ptr) const {
+    this->chunks[0].store_ascii_as_utf32_tbl(ptr + sizeof(simd8<T>) * 0);
+    this->chunks[1].store_ascii_as_utf32_tbl(ptr + sizeof(simd8<T>) * 1);
+    this->chunks[2].store_ascii_as_utf32_tbl(ptr + sizeof(simd8<T>) * 2);
+    this->chunks[3].store_ascii_as_utf32_tbl(ptr + sizeof(simd8<T>) * 3);
+  }
+
+  simdutf_really_inline uint64_t to_bitmask() const {
+    __m128i mask = __lsx_vbsll_v(__lsx_vmsknz_b(this->chunks[3]), 6);
+    mask = __lsx_vor_v(mask, __lsx_vbsll_v(__lsx_vmsknz_b(this->chunks[2]), 4));
+    mask = __lsx_vor_v(mask, __lsx_vbsll_v(__lsx_vmsknz_b(this->chunks[1]), 2));
+    mask = __lsx_vor_v(mask, __lsx_vmsknz_b(this->chunks[0]));
+    return __lsx_vpickve2gr_du(mask, 0);
+  }
+
+  simdutf_really_inline uint64_t lt(const T m) const {
+    const simd8<T> mask = simd8<T>::splat(m);
+    return simd8x64<bool>(this->chunks[0] < mask, this->chunks[1] < mask,
+                          this->chunks[2] < mask, this->chunks[3] < mask)
+        .to_bitmask();
+  }
+  simdutf_really_inline uint64_t gt(const T m) const {
+    const simd8<T> mask = simd8<T>::splat(m);
+    return simd8x64<bool>(this->chunks[0] > mask, this->chunks[1] > mask,
+                          this->chunks[2] > mask, this->chunks[3] > mask)
+        .to_bitmask();
+  }
+  simdutf_really_inline uint64_t gteq(const T m) const {
+    const simd8<T> mask = simd8<T>::splat(m);
+    return simd8x64<bool>(this->chunks[0] >= mask, this->chunks[1] >= mask,
+                          this->chunks[2] >= mask, this->chunks[3] >= mask)
+        .to_bitmask();
+  }
+  simdutf_really_inline uint64_t gteq_unsigned(const uint8_t m) const {
+    const simd8<uint8_t> mask = simd8<uint8_t>::splat(m);
+    return simd8x64<bool>(simd8<uint8_t>(this->chunks[0].value) >= mask,
+                          simd8<uint8_t>(this->chunks[1].value) >= mask,
+                          simd8<uint8_t>(this->chunks[2].value) >= mask,
+                          simd8<uint8_t>(this->chunks[3].value) >= mask)
+        .to_bitmask();
+  }
+}; // struct simd8x64<T>
+
+#include "simdutf/lsx/simd16-inl.h"
+#include "simdutf/lsx/simd32-inl.h"
+#include "simdutf/lsx/simd64-inl.h"
+
+} // namespace simd
+} // unnamed namespace
+} // namespace SIMDUTF_IMPLEMENTATION
+} // namespace simdutf
+
+#endif // SIMDUTF_LSX_SIMD_H

package/deps/simdutf/src/simdutf/lsx/simd16-inl.h

@@ -0,0 +1,242 @@
+template <typename T> struct simd16;
+
+template <typename T, typename Mask = simd16<bool>> struct base_u16 {
+  __m128i value;
+  static const size_t SIZE = sizeof(value);
+  static const size_t ELEMENTS = sizeof(value) / sizeof(T);
+
+  // Conversion from/to SIMD register
+  simdutf_really_inline base_u16() = default;
+  simdutf_really_inline base_u16(const __m128i _value) : value(_value) {}
+  // Bit operations
+  simdutf_really_inline simd16<T> operator|(const simd16<T> other) const {
+    return __lsx_vor_v(this->value, other.value);
+  }
+  simdutf_really_inline simd16<T> operator&(const simd16<T> other) const {
+    return __lsx_vand_v(this->value, other.value);
+  }
+  simdutf_really_inline simd16<T> operator~() const {
+    return __lsx_vxori_b(this->value, 0xFF);
+  }
+
+  friend simdutf_really_inline Mask operator==(const simd16<T> lhs,
+                                               const simd16<T> rhs) {
+    return __lsx_vseq_h(lhs.value, rhs.value);
+  }
+
+  template <unsigned N>
+  simdutf_really_inline simd16<T> byte_right_shift() const {
+    return __lsx_vbsrl_v(this->value, N);
+  }
+
+  simdutf_really_inline uint16_t first() const {
+    return uint16_t(__lsx_vpickve2gr_w(value, 0));
+  }
+};
+
+template <typename T, typename Mask = simd16<bool>>
+struct base16 : base_u16<T> {
+  using bitmask_type = uint16_t;
+
+  simdutf_really_inline base16() : base_u16<T>() {}
+  simdutf_really_inline base16(const __m128i _value) : base_u16<T>(_value) {}
+  template <typename Pointer>
+  simdutf_really_inline base16(const Pointer *ptr)
+      : base16(__lsx_vld(ptr, 0)) {}
+
+  static const int SIZE = sizeof(base_u16<T>::value);
+
+  template <unsigned N = 1>
+  simdutf_really_inline simd16<T> prev(const simd16<T> prev_chunk) const {
+    return __lsx_vor_v(__lsx_vbsll_v(*this, N * 2),
+                       __lsx_vbsrl_v(prev_chunk, 16 - N * 2));
+  }
+};
+
+// SIMD byte mask type (returned by things like eq and gt)
+template <> struct simd16<bool> : base16<bool> {
+  static simdutf_really_inline simd16<bool> splat(bool _value) {
+    return __lsx_vreplgr2vr_h(uint16_t(-(!!_value)));
+  }
+
+  simdutf_really_inline simd16() : base16() {}
+  simdutf_really_inline simd16(const __m128i _value) : base16<bool>(_value) {}
+
+  simdutf_really_inline bitmask_type to_bitmask() const {
+    __m128i mask = __lsx_vmsknz_b(this->value);
+    bitmask_type mask0 = bitmask_type(__lsx_vpickve2gr_wu(mask, 0));
+    return mask0;
+  }
+
+  simdutf_really_inline bool is_zero() const { return __lsx_bz_v(this->value); }
+};
+
+template <typename T> struct base16_numeric : base16<T> {
+  static simdutf_really_inline simd16<T> splat(T _value) {
+    return __lsx_vreplgr2vr_h(_value);
+  }
+  static simdutf_really_inline simd16<T> zero() { return __lsx_vldi(0); }
+
+  template <typename Pointer>
+  static simdutf_really_inline simd16<T> load(const Pointer values) {
+    return __lsx_vld(values, 0);
+  }
+
+  simdutf_really_inline base16_numeric(const __m128i _value)
+      : base16<T>(_value) {}
+
+  // Store to array
+  simdutf_really_inline void store(T dst[8]) const {
+    return __lsx_vst(this->value, dst, 0);
+  }
+
+  // Override to distinguish from bool version
+  simdutf_really_inline simd16<T> operator~() const {
+    return __lsx_vxori_b(this->value, 0xFF);
+  }
+};
+
+// Unsigned code unitstemplate<>
+template <> struct simd16<uint16_t> : base16_numeric<uint16_t> {
+  simdutf_really_inline simd16(const __m128i _value)
+      : base16_numeric<uint16_t>((__m128i)_value) {}
+
+  // Splat constructor
+  simdutf_really_inline simd16(uint16_t _value) : simd16(splat(_value)) {}
+
+  // Array constructor
+  simdutf_really_inline simd16(const uint16_t *values) : simd16(load(values)) {}
+  simdutf_really_inline simd16(const char16_t *values)
+      : simd16(load(reinterpret_cast<const uint16_t *>(values))) {}
+
+  // Copy constructor
+  simdutf_really_inline simd16(const simd16<bool> mask) : simd16(mask.value) {}
+
+  // Order-specific operations
+  simdutf_really_inline simd16 &operator+=(const simd16 other) {
+    value = __lsx_vadd_h(value, other.value);
+    return *this;
+  }
+
+  template <unsigned N>
+  static simdutf_really_inline simd8<uint8_t>
+  pack_shifted_right(const simd16<uint16_t> &v0, const simd16<uint16_t> &v1) {
+    return __lsx_vssrlni_bu_h(v1.value, v0.value, N);
+  }
+
+  // Pack with the unsigned saturation of two uint16_t code units into single
+  // uint8_t vector
+  static simdutf_really_inline simd8<uint8_t> pack(const simd16<uint16_t> &v0,
+                                                   const simd16<uint16_t> &v1) {
+    return pack_shifted_right<0>(v0, v1);
+  }
+
+  // Change the endianness
+  simdutf_really_inline simd16<uint16_t> swap_bytes() const {
+    return __lsx_vshuf4i_b(this->value, 0b10110001);
+  }
+
+  simdutf_really_inline uint64_t sum() const {
+    const auto sum_u32 = __lsx_vhaddw_wu_hu(value, value);
+    const auto sum_u64 = __lsx_vhaddw_du_wu(sum_u32, sum_u32);
+
+    return uint64_t(__lsx_vpickve2gr_du(sum_u64, 0)) +
+           uint64_t(__lsx_vpickve2gr_du(sum_u64, 1));
+  }
+};
+
+simdutf_really_inline simd16<bool> operator<(const simd16<uint16_t> a,
+                                             const simd16<uint16_t> b) {
+  return __lsx_vslt_hu(a.value, b.value);
+}
+
+simdutf_really_inline simd16<bool> operator>(const simd16<uint16_t> a,
+                                             const simd16<uint16_t> b) {
+  return __lsx_vslt_hu(b.value, a.value);
+}
+
+simdutf_really_inline simd16<bool> operator<=(const simd16<uint16_t> a,
+                                              const simd16<uint16_t> b) {
+  return __lsx_vsle_hu(a.value, b.value);
+}
+
+simdutf_really_inline simd16<bool> operator>=(const simd16<uint16_t> a,
+                                              const simd16<uint16_t> b) {
+  return __lsx_vsle_hu(b.value, a.value);
+}
+
+template <typename T> struct simd16x32 {
+  static constexpr int NUM_CHUNKS = 64 / sizeof(simd16<T>);
+  static_assert(
+      NUM_CHUNKS == 4,
+      "LOONGARCH kernel should use four registers per 64-byte block.");
+  simd16<T> chunks[NUM_CHUNKS];
+
+  simd16x32(const simd16x32<T> &o) = delete; // no copy allowed
+  simd16x32<T> &
+  operator=(const simd16<T> other) = delete; // no assignment allowed
+  simd16x32() = delete;                      // no default constructor allowed
+
+  simdutf_really_inline
+  simd16x32(const simd16<T> chunk0, const simd16<T> chunk1,
+            const simd16<T> chunk2, const simd16<T> chunk3)
+      : chunks{chunk0, chunk1, chunk2, chunk3} {}
+  simdutf_really_inline simd16x32(const T *ptr)
+      : chunks{simd16<T>::load(ptr),
+               simd16<T>::load(ptr + sizeof(simd16<T>) / sizeof(T)),
+               simd16<T>::load(ptr + 2 * sizeof(simd16<T>) / sizeof(T)),
+               simd16<T>::load(ptr + 3 * sizeof(simd16<T>) / sizeof(T))} {}
+
+  simdutf_really_inline void store(T *ptr) const {
+    this->chunks[0].store(ptr + sizeof(simd16<T>) * 0 / sizeof(T));
+    this->chunks[1].store(ptr + sizeof(simd16<T>) * 1 / sizeof(T));
+    this->chunks[2].store(ptr + sizeof(simd16<T>) * 2 / sizeof(T));
+    this->chunks[3].store(ptr + sizeof(simd16<T>) * 3 / sizeof(T));
+  }
+
+  simdutf_really_inline void swap_bytes() {
+    this->chunks[0] = this->chunks[0].swap_bytes();
+    this->chunks[1] = this->chunks[1].swap_bytes();
+    this->chunks[2] = this->chunks[2].swap_bytes();
+    this->chunks[3] = this->chunks[3].swap_bytes();
+  }
+  simdutf_really_inline uint64_t to_bitmask() const {
+    uint64_t r0 = uint32_t(this->chunks[0].to_bitmask());
+    uint64_t r1 = this->chunks[1].to_bitmask();
+    uint64_t r2 = this->chunks[2].to_bitmask();
+    uint64_t r3 = this->chunks[3].to_bitmask();
+    return r0 | (r1 << 16) | (r2 << 32) | (r3 << 48);
+  }
+  simdutf_really_inline uint64_t gteq(const T m) const {
+    const simd16<T> mask = simd16<T>::splat(m);
+    return simd16x32<bool>(this->chunks[0] >= mask, this->chunks[1] >= mask,
+                           this->chunks[2] >= mask, this->chunks[3] >= mask)
+        .to_bitmask();
+  }
+  simdutf_really_inline uint64_t lteq(const T m) const {
+    const simd16<T> mask = simd16<T>::splat(m);
+    return simd16x32<bool>(this->chunks[0] <= mask, this->chunks[1] <= mask,
+                           this->chunks[2] <= mask, this->chunks[3] <= mask)
+        .to_bitmask();
+  }
+}; // struct simd16x32<T>
+
+simdutf_really_inline simd16<uint16_t> operator^(const simd16<uint16_t> a,
+                                                 uint16_t b) {
+  const auto bv = __lsx_vreplgr2vr_h(b);
+  return __lsx_vxor_v(a.value, bv);
+}
+
+simdutf_really_inline simd16<bool> operator^(const simd16<bool> a,
+                                             const simd16<bool> b) {
+  return __lsx_vxor_v(a.value, b.value);
+}
+
+simdutf_really_inline simd16<uint16_t> min(const simd16<uint16_t> a,
+                                           const simd16<uint16_t> b) {
+  return __lsx_vmin_hu(a.value, b.value);
+}
+
+simdutf_really_inline simd16<uint16_t> as_vector_u16(const simd16<bool> x) {
+  return x.value;
+}