@ohos-ports/react-native-quick-crypto 0.0.0-stage → 1.1.8-beta.0

package/deps/blake3/c/blake3_impl.h

@@ -0,0 +1,333 @@
+#ifndef BLAKE3_IMPL_H
+#define BLAKE3_IMPL_H
+
+#include <assert.h>
+#include <stdbool.h>
+#include <stddef.h>
+#include <stdint.h>
+#include <string.h>
+
+#include "blake3.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// internal flags
+enum blake3_flags {
+  CHUNK_START         = 1 << 0,
+  CHUNK_END           = 1 << 1,
+  PARENT              = 1 << 2,
+  ROOT                = 1 << 3,
+  KEYED_HASH          = 1 << 4,
+  DERIVE_KEY_CONTEXT  = 1 << 5,
+  DERIVE_KEY_MATERIAL = 1 << 6,
+};
+
+// This C implementation tries to support recent versions of GCC, Clang, and
+// MSVC.
+#if defined(_MSC_VER)
+#define INLINE static __forceinline
+#else
+#define INLINE static inline __attribute__((always_inline))
+#endif
+
+#ifdef __cplusplus
+#define NOEXCEPT noexcept
+#else
+#define NOEXCEPT
+#endif
+
+#if (defined(__x86_64__) || defined(_M_X64)) && !defined(_M_ARM64EC)
+#define IS_X86
+#define IS_X86_64
+#endif
+
+#if defined(__i386__) || defined(_M_IX86)
+#define IS_X86
+#define IS_X86_32
+#endif
+
+#if defined(__aarch64__) || defined(_M_ARM64) || defined(_M_ARM64EC)
+#define IS_AARCH64
+#endif
+
+#if defined(IS_X86)
+#if defined(_MSC_VER)
+#include <intrin.h>
+#endif
+#endif
+
+#if !defined(BLAKE3_USE_NEON) 
+  // If BLAKE3_USE_NEON not manually set, autodetect based on AArch64ness
+  #if defined(IS_AARCH64)
+    #if defined(__ARM_BIG_ENDIAN)
+      #define BLAKE3_USE_NEON 0
+    #else
+      #define BLAKE3_USE_NEON 1
+    #endif
+  #else
+    #define BLAKE3_USE_NEON 0
+  #endif
+#endif
+
+#if defined(IS_X86)
+#define MAX_SIMD_DEGREE 16
+#elif BLAKE3_USE_NEON == 1
+#define MAX_SIMD_DEGREE 4
+#else
+#define MAX_SIMD_DEGREE 1
+#endif
+
+// There are some places where we want a static size that's equal to the
+// MAX_SIMD_DEGREE, but also at least 2.
+#define MAX_SIMD_DEGREE_OR_2 (MAX_SIMD_DEGREE > 2 ? MAX_SIMD_DEGREE : 2)
+
+static const uint32_t IV[8] = {0x6A09E667UL, 0xBB67AE85UL, 0x3C6EF372UL,
+                               0xA54FF53AUL, 0x510E527FUL, 0x9B05688CUL,
+                               0x1F83D9ABUL, 0x5BE0CD19UL};
+
+static const uint8_t MSG_SCHEDULE[7][16] = {
+    {0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15},
+    {2, 6, 3, 10, 7, 0, 4, 13, 1, 11, 12, 5, 9, 14, 15, 8},
+    {3, 4, 10, 12, 13, 2, 7, 14, 6, 5, 9, 0, 11, 15, 8, 1},
+    {10, 7, 12, 9, 14, 3, 13, 15, 4, 0, 11, 2, 5, 8, 1, 6},
+    {12, 13, 9, 11, 15, 10, 14, 8, 7, 2, 5, 3, 0, 1, 6, 4},
+    {9, 14, 11, 5, 8, 12, 15, 1, 13, 3, 0, 10, 2, 6, 4, 7},
+    {11, 15, 5, 0, 1, 9, 8, 6, 14, 10, 2, 12, 3, 4, 7, 13},
+};
+
+/* Find index of the highest set bit */
+/* x is assumed to be nonzero.       */
+static unsigned int highest_one(uint64_t x) {
+#if defined(__GNUC__) || defined(__clang__)
+  return 63 ^ (unsigned int)__builtin_clzll(x);
+#elif defined(_MSC_VER) && defined(IS_X86_64)
+  unsigned long index;
+  _BitScanReverse64(&index, x);
+  return index;
+#elif defined(_MSC_VER) && defined(IS_X86_32)
+  if(x >> 32) {
+    unsigned long index;
+    _BitScanReverse(&index, (unsigned long)(x >> 32));
+    return 32 + index;
+  } else {
+    unsigned long index;
+    _BitScanReverse(&index, (unsigned long)x);
+    return index;
+  }
+#else
+  unsigned int c = 0;
+  if(x & 0xffffffff00000000ULL) { x >>= 32; c += 32; }
+  if(x & 0x00000000ffff0000ULL) { x >>= 16; c += 16; }
+  if(x & 0x000000000000ff00ULL) { x >>=  8; c +=  8; }
+  if(x & 0x00000000000000f0ULL) { x >>=  4; c +=  4; }
+  if(x & 0x000000000000000cULL) { x >>=  2; c +=  2; }
+  if(x & 0x0000000000000002ULL) {           c +=  1; }
+  return c;
+#endif
+}
+
+// Count the number of 1 bits.
+INLINE unsigned int popcnt(uint64_t x) {
+#if defined(__GNUC__) || defined(__clang__)
+  return (unsigned int)__builtin_popcountll(x);
+#else
+  unsigned int count = 0;
+  while (x != 0) {
+    count += 1;
+    x &= x - 1;
+  }
+  return count;
+#endif
+}
+
+// Largest power of two less than or equal to x. As a special case, returns 1
+// when x is 0. 
+INLINE uint64_t round_down_to_power_of_2(uint64_t x) {
+  return 1ULL << highest_one(x | 1);
+}
+
+INLINE uint32_t counter_low(uint64_t counter) { return (uint32_t)counter; }
+
+INLINE uint32_t counter_high(uint64_t counter) {
+  return (uint32_t)(counter >> 32);
+}
+
+INLINE uint32_t load32(const void *src) {
+  const uint8_t *p = (const uint8_t *)src;
+  return ((uint32_t)(p[0]) << 0) | ((uint32_t)(p[1]) << 8) |
+         ((uint32_t)(p[2]) << 16) | ((uint32_t)(p[3]) << 24);
+}
+
+INLINE void load_key_words(const uint8_t key[BLAKE3_KEY_LEN],
+                           uint32_t key_words[8]) {
+  key_words[0] = load32(&key[0 * 4]);
+  key_words[1] = load32(&key[1 * 4]);
+  key_words[2] = load32(&key[2 * 4]);
+  key_words[3] = load32(&key[3 * 4]);
+  key_words[4] = load32(&key[4 * 4]);
+  key_words[5] = load32(&key[5 * 4]);
+  key_words[6] = load32(&key[6 * 4]);
+  key_words[7] = load32(&key[7 * 4]);
+}
+
+INLINE void load_block_words(const uint8_t block[BLAKE3_BLOCK_LEN],
+                             uint32_t block_words[16]) {
+  for (size_t i = 0; i < 16; i++) {
+      block_words[i] = load32(&block[i * 4]);
+  }
+}
+
+INLINE void store32(void *dst, uint32_t w) {
+  uint8_t *p = (uint8_t *)dst;
+  p[0] = (uint8_t)(w >> 0);
+  p[1] = (uint8_t)(w >> 8);
+  p[2] = (uint8_t)(w >> 16);
+  p[3] = (uint8_t)(w >> 24);
+}
+
+INLINE void store_cv_words(uint8_t bytes_out[32], uint32_t cv_words[8]) {
+  store32(&bytes_out[0 * 4], cv_words[0]);
+  store32(&bytes_out[1 * 4], cv_words[1]);
+  store32(&bytes_out[2 * 4], cv_words[2]);
+  store32(&bytes_out[3 * 4], cv_words[3]);
+  store32(&bytes_out[4 * 4], cv_words[4]);
+  store32(&bytes_out[5 * 4], cv_words[5]);
+  store32(&bytes_out[6 * 4], cv_words[6]);
+  store32(&bytes_out[7 * 4], cv_words[7]);
+}
+
+void blake3_compress_in_place(uint32_t cv[8],
+                              const uint8_t block[BLAKE3_BLOCK_LEN],
+                              uint8_t block_len, uint64_t counter,
+                              uint8_t flags);
+
+void blake3_compress_xof(const uint32_t cv[8],
+                         const uint8_t block[BLAKE3_BLOCK_LEN],
+                         uint8_t block_len, uint64_t counter, uint8_t flags,
+                         uint8_t out[64]);
+
+void blake3_xof_many(const uint32_t cv[8],
+                     const uint8_t block[BLAKE3_BLOCK_LEN],
+                     uint8_t block_len, uint64_t counter, uint8_t flags,
+                     uint8_t out[64], size_t outblocks);
+
+void blake3_hash_many(const uint8_t *const *inputs, size_t num_inputs,
+                      size_t blocks, const uint32_t key[8], uint64_t counter,
+                      bool increment_counter, uint8_t flags,
+                      uint8_t flags_start, uint8_t flags_end, uint8_t *out);
+
+size_t blake3_simd_degree(void);
+
+BLAKE3_PRIVATE size_t blake3_compress_subtree_wide(const uint8_t *input, size_t input_len,
+                                                   const uint32_t key[8],
+                                                   uint64_t chunk_counter, uint8_t flags,
+                                                   uint8_t *out, bool use_tbb);
+
+#if defined(BLAKE3_USE_TBB)
+BLAKE3_PRIVATE void blake3_compress_subtree_wide_join_tbb(
+    // shared params
+    const uint32_t key[8], uint8_t flags, bool use_tbb,
+    // left-hand side params
+    const uint8_t *l_input, size_t l_input_len, uint64_t l_chunk_counter,
+    uint8_t *l_cvs, size_t *l_n,
+    // right-hand side params
+    const uint8_t *r_input, size_t r_input_len, uint64_t r_chunk_counter,
+    uint8_t *r_cvs, size_t *r_n) NOEXCEPT;
+#endif
+
+// Declarations for implementation-specific functions.
+void blake3_compress_in_place_portable(uint32_t cv[8],
+                                       const uint8_t block[BLAKE3_BLOCK_LEN],
+                                       uint8_t block_len, uint64_t counter,
+                                       uint8_t flags);
+
+void blake3_compress_xof_portable(const uint32_t cv[8],
+                                  const uint8_t block[BLAKE3_BLOCK_LEN],
+                                  uint8_t block_len, uint64_t counter,
+                                  uint8_t flags, uint8_t out[64]);
+
+void blake3_hash_many_portable(const uint8_t *const *inputs, size_t num_inputs,
+                               size_t blocks, const uint32_t key[8],
+                               uint64_t counter, bool increment_counter,
+                               uint8_t flags, uint8_t flags_start,
+                               uint8_t flags_end, uint8_t *out);
+
+#if defined(IS_X86)
+#if !defined(BLAKE3_NO_SSE2)
+void blake3_compress_in_place_sse2(uint32_t cv[8],
+                                   const uint8_t block[BLAKE3_BLOCK_LEN],
+                                   uint8_t block_len, uint64_t counter,
+                                   uint8_t flags);
+void blake3_compress_xof_sse2(const uint32_t cv[8],
+                              const uint8_t block[BLAKE3_BLOCK_LEN],
+                              uint8_t block_len, uint64_t counter,
+                              uint8_t flags, uint8_t out[64]);
+void blake3_hash_many_sse2(const uint8_t *const *inputs, size_t num_inputs,
+                           size_t blocks, const uint32_t key[8],
+                           uint64_t counter, bool increment_counter,
+                           uint8_t flags, uint8_t flags_start,
+                           uint8_t flags_end, uint8_t *out);
+#endif
+#if !defined(BLAKE3_NO_SSE41)
+void blake3_compress_in_place_sse41(uint32_t cv[8],
+                                    const uint8_t block[BLAKE3_BLOCK_LEN],
+                                    uint8_t block_len, uint64_t counter,
+                                    uint8_t flags);
+void blake3_compress_xof_sse41(const uint32_t cv[8],
+                               const uint8_t block[BLAKE3_BLOCK_LEN],
+                               uint8_t block_len, uint64_t counter,
+                               uint8_t flags, uint8_t out[64]);
+void blake3_hash_many_sse41(const uint8_t *const *inputs, size_t num_inputs,
+                            size_t blocks, const uint32_t key[8],
+                            uint64_t counter, bool increment_counter,
+                            uint8_t flags, uint8_t flags_start,
+                            uint8_t flags_end, uint8_t *out);
+#endif
+#if !defined(BLAKE3_NO_AVX2)
+void blake3_hash_many_avx2(const uint8_t *const *inputs, size_t num_inputs,
+                           size_t blocks, const uint32_t key[8],
+                           uint64_t counter, bool increment_counter,
+                           uint8_t flags, uint8_t flags_start,
+                           uint8_t flags_end, uint8_t *out);
+#endif
+#if !defined(BLAKE3_NO_AVX512)
+void blake3_compress_in_place_avx512(uint32_t cv[8],
+                                     const uint8_t block[BLAKE3_BLOCK_LEN],
+                                     uint8_t block_len, uint64_t counter,
+                                     uint8_t flags);
+
+void blake3_compress_xof_avx512(const uint32_t cv[8],
+                                const uint8_t block[BLAKE3_BLOCK_LEN],
+                                uint8_t block_len, uint64_t counter,
+                                uint8_t flags, uint8_t out[64]);
+
+void blake3_hash_many_avx512(const uint8_t *const *inputs, size_t num_inputs,
+                             size_t blocks, const uint32_t key[8],
+                             uint64_t counter, bool increment_counter,
+                             uint8_t flags, uint8_t flags_start,
+                             uint8_t flags_end, uint8_t *out);
+
+#if !defined(_WIN32) && !defined(__CYGWIN__)
+void blake3_xof_many_avx512(const uint32_t cv[8],
+                            const uint8_t block[BLAKE3_BLOCK_LEN],
+                            uint8_t block_len, uint64_t counter, uint8_t flags,
+                            uint8_t* out, size_t outblocks);
+#endif
+#endif
+#endif
+
+#if BLAKE3_USE_NEON == 1
+void blake3_hash_many_neon(const uint8_t *const *inputs, size_t num_inputs,
+                           size_t blocks, const uint32_t key[8],
+                           uint64_t counter, bool increment_counter,
+                           uint8_t flags, uint8_t flags_start,
+                           uint8_t flags_end, uint8_t *out);
+#endif
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif /* BLAKE3_IMPL_H */

package/deps/blake3/c/blake3_neon.c

@@ -0,0 +1,367 @@
+#include "blake3_impl.h"
+
+#include <arm_neon.h>
+
+#ifdef __ARM_BIG_ENDIAN
+#error "This implementation only supports little-endian ARM."
+// It might be that all we need for big-endian support here is to get the loads
+// and stores right, but step zero would be finding a way to test it in CI.
+#endif
+
+INLINE uint32x4_t loadu_128(const uint8_t src[16]) {
+  // vld1q_u32 has alignment requirements. Don't use it.
+  return vreinterpretq_u32_u8(vld1q_u8(src));
+}
+
+INLINE void storeu_128(uint32x4_t src, uint8_t dest[16]) {
+  // vst1q_u32 has alignment requirements. Don't use it.
+  vst1q_u8(dest, vreinterpretq_u8_u32(src));
+}
+
+INLINE uint32x4_t add_128(uint32x4_t a, uint32x4_t b) {
+  return vaddq_u32(a, b);
+}
+
+INLINE uint32x4_t xor_128(uint32x4_t a, uint32x4_t b) {
+  return veorq_u32(a, b);
+}
+
+INLINE uint32x4_t set1_128(uint32_t x) { return vld1q_dup_u32(&x); }
+
+INLINE uint32x4_t set4(uint32_t a, uint32_t b, uint32_t c, uint32_t d) {
+  uint32_t array[4] = {a, b, c, d};
+  return vld1q_u32(array);
+}
+
+INLINE uint32x4_t rot16_128(uint32x4_t x) {
+  // The straightforward implementation would be two shifts and an or, but that's
+  // slower on microarchitectures we've tested. See
+  // https://github.com/BLAKE3-team/BLAKE3/pull/319.
+  // return vorrq_u32(vshrq_n_u32(x, 16), vshlq_n_u32(x, 32 - 16));
+  return vreinterpretq_u32_u16(vrev32q_u16(vreinterpretq_u16_u32(x)));
+}
+
+INLINE uint32x4_t rot12_128(uint32x4_t x) {
+  // See comment in rot16_128.
+  // return vorrq_u32(vshrq_n_u32(x, 12), vshlq_n_u32(x, 32 - 12));
+  return vsriq_n_u32(vshlq_n_u32(x, 32-12), x, 12);
+}
+
+INLINE uint32x4_t rot8_128(uint32x4_t x) {
+  // See comment in rot16_128.
+  // return vorrq_u32(vshrq_n_u32(x, 8), vshlq_n_u32(x, 32 - 8));
+#if defined(__clang__)
+  return vreinterpretq_u32_u8(__builtin_shufflevector(vreinterpretq_u8_u32(x), vreinterpretq_u8_u32(x), 1,2,3,0,5,6,7,4,9,10,11,8,13,14,15,12));
+#elif __GNUC__ * 10000 + __GNUC_MINOR__ * 100 >=40700
+  static const uint8x16_t r8 = {1,2,3,0,5,6,7,4,9,10,11,8,13,14,15,12};
+  return vreinterpretq_u32_u8(__builtin_shuffle(vreinterpretq_u8_u32(x), vreinterpretq_u8_u32(x), r8));
+#else 
+  return vsriq_n_u32(vshlq_n_u32(x, 32-8), x, 8);
+#endif
+}
+
+INLINE uint32x4_t rot7_128(uint32x4_t x) {
+  // See comment in rot16_128.
+  // return vorrq_u32(vshrq_n_u32(x, 7), vshlq_n_u32(x, 32 - 7));
+  return vsriq_n_u32(vshlq_n_u32(x, 32-7), x, 7);
+}
+
+// TODO: compress_neon
+
+// TODO: hash2_neon
+
+/*
+ * ----------------------------------------------------------------------------
+ * hash4_neon
+ * ----------------------------------------------------------------------------
+ */
+
+INLINE void round_fn4(uint32x4_t v[16], uint32x4_t m[16], size_t r) {
+  v[0] = add_128(v[0], m[(size_t)MSG_SCHEDULE[r][0]]);
+  v[1] = add_128(v[1], m[(size_t)MSG_SCHEDULE[r][2]]);
+  v[2] = add_128(v[2], m[(size_t)MSG_SCHEDULE[r][4]]);
+  v[3] = add_128(v[3], m[(size_t)MSG_SCHEDULE[r][6]]);
+  v[0] = add_128(v[0], v[4]);
+  v[1] = add_128(v[1], v[5]);
+  v[2] = add_128(v[2], v[6]);
+  v[3] = add_128(v[3], v[7]);
+  v[12] = xor_128(v[12], v[0]);
+  v[13] = xor_128(v[13], v[1]);
+  v[14] = xor_128(v[14], v[2]);
+  v[15] = xor_128(v[15], v[3]);
+  v[12] = rot16_128(v[12]);
+  v[13] = rot16_128(v[13]);
+  v[14] = rot16_128(v[14]);
+  v[15] = rot16_128(v[15]);
+  v[8] = add_128(v[8], v[12]);
+  v[9] = add_128(v[9], v[13]);
+  v[10] = add_128(v[10], v[14]);
+  v[11] = add_128(v[11], v[15]);
+  v[4] = xor_128(v[4], v[8]);
+  v[5] = xor_128(v[5], v[9]);
+  v[6] = xor_128(v[6], v[10]);
+  v[7] = xor_128(v[7], v[11]);
+  v[4] = rot12_128(v[4]);
+  v[5] = rot12_128(v[5]);
+  v[6] = rot12_128(v[6]);
+  v[7] = rot12_128(v[7]);
+  v[0] = add_128(v[0], m[(size_t)MSG_SCHEDULE[r][1]]);
+  v[1] = add_128(v[1], m[(size_t)MSG_SCHEDULE[r][3]]);
+  v[2] = add_128(v[2], m[(size_t)MSG_SCHEDULE[r][5]]);
+  v[3] = add_128(v[3], m[(size_t)MSG_SCHEDULE[r][7]]);
+  v[0] = add_128(v[0], v[4]);
+  v[1] = add_128(v[1], v[5]);
+  v[2] = add_128(v[2], v[6]);
+  v[3] = add_128(v[3], v[7]);
+  v[12] = xor_128(v[12], v[0]);
+  v[13] = xor_128(v[13], v[1]);
+  v[14] = xor_128(v[14], v[2]);
+  v[15] = xor_128(v[15], v[3]);
+  v[12] = rot8_128(v[12]);
+  v[13] = rot8_128(v[13]);
+  v[14] = rot8_128(v[14]);
+  v[15] = rot8_128(v[15]);
+  v[8] = add_128(v[8], v[12]);
+  v[9] = add_128(v[9], v[13]);
+  v[10] = add_128(v[10], v[14]);
+  v[11] = add_128(v[11], v[15]);
+  v[4] = xor_128(v[4], v[8]);
+  v[5] = xor_128(v[5], v[9]);
+  v[6] = xor_128(v[6], v[10]);
+  v[7] = xor_128(v[7], v[11]);
+  v[4] = rot7_128(v[4]);
+  v[5] = rot7_128(v[5]);
+  v[6] = rot7_128(v[6]);
+  v[7] = rot7_128(v[7]);
+
+  v[0] = add_128(v[0], m[(size_t)MSG_SCHEDULE[r][8]]);
+  v[1] = add_128(v[1], m[(size_t)MSG_SCHEDULE[r][10]]);
+  v[2] = add_128(v[2], m[(size_t)MSG_SCHEDULE[r][12]]);
+  v[3] = add_128(v[3], m[(size_t)MSG_SCHEDULE[r][14]]);
+  v[0] = add_128(v[0], v[5]);
+  v[1] = add_128(v[1], v[6]);
+  v[2] = add_128(v[2], v[7]);
+  v[3] = add_128(v[3], v[4]);
+  v[15] = xor_128(v[15], v[0]);
+  v[12] = xor_128(v[12], v[1]);
+  v[13] = xor_128(v[13], v[2]);
+  v[14] = xor_128(v[14], v[3]);
+  v[15] = rot16_128(v[15]);
+  v[12] = rot16_128(v[12]);
+  v[13] = rot16_128(v[13]);
+  v[14] = rot16_128(v[14]);
+  v[10] = add_128(v[10], v[15]);
+  v[11] = add_128(v[11], v[12]);
+  v[8] = add_128(v[8], v[13]);
+  v[9] = add_128(v[9], v[14]);
+  v[5] = xor_128(v[5], v[10]);
+  v[6] = xor_128(v[6], v[11]);
+  v[7] = xor_128(v[7], v[8]);
+  v[4] = xor_128(v[4], v[9]);
+  v[5] = rot12_128(v[5]);
+  v[6] = rot12_128(v[6]);
+  v[7] = rot12_128(v[7]);
+  v[4] = rot12_128(v[4]);
+  v[0] = add_128(v[0], m[(size_t)MSG_SCHEDULE[r][9]]);
+  v[1] = add_128(v[1], m[(size_t)MSG_SCHEDULE[r][11]]);
+  v[2] = add_128(v[2], m[(size_t)MSG_SCHEDULE[r][13]]);
+  v[3] = add_128(v[3], m[(size_t)MSG_SCHEDULE[r][15]]);
+  v[0] = add_128(v[0], v[5]);
+  v[1] = add_128(v[1], v[6]);
+  v[2] = add_128(v[2], v[7]);
+  v[3] = add_128(v[3], v[4]);
+  v[15] = xor_128(v[15], v[0]);
+  v[12] = xor_128(v[12], v[1]);
+  v[13] = xor_128(v[13], v[2]);
+  v[14] = xor_128(v[14], v[3]);
+  v[15] = rot8_128(v[15]);
+  v[12] = rot8_128(v[12]);
+  v[13] = rot8_128(v[13]);
+  v[14] = rot8_128(v[14]);
+  v[10] = add_128(v[10], v[15]);
+  v[11] = add_128(v[11], v[12]);
+  v[8] = add_128(v[8], v[13]);
+  v[9] = add_128(v[9], v[14]);
+  v[5] = xor_128(v[5], v[10]);
+  v[6] = xor_128(v[6], v[11]);
+  v[7] = xor_128(v[7], v[8]);
+  v[4] = xor_128(v[4], v[9]);
+  v[5] = rot7_128(v[5]);
+  v[6] = rot7_128(v[6]);
+  v[7] = rot7_128(v[7]);
+  v[4] = rot7_128(v[4]);
+}
+
+INLINE void transpose_vecs_128(uint32x4_t vecs[4]) {
+  // Individually transpose the four 2x2 sub-matrices in each corner.
+  uint32x4x2_t rows01 = vtrnq_u32(vecs[0], vecs[1]);
+  uint32x4x2_t rows23 = vtrnq_u32(vecs[2], vecs[3]);
+
+  // Swap the top-right and bottom-left 2x2s (which just got transposed).
+  vecs[0] =
+      vcombine_u32(vget_low_u32(rows01.val[0]), vget_low_u32(rows23.val[0]));
+  vecs[1] =
+      vcombine_u32(vget_low_u32(rows01.val[1]), vget_low_u32(rows23.val[1]));
+  vecs[2] =
+      vcombine_u32(vget_high_u32(rows01.val[0]), vget_high_u32(rows23.val[0]));
+  vecs[3] =
+      vcombine_u32(vget_high_u32(rows01.val[1]), vget_high_u32(rows23.val[1]));
+}
+
+INLINE void transpose_msg_vecs4(const uint8_t *const *inputs,
+                                size_t block_offset, uint32x4_t out[16]) {
+  out[0] = loadu_128(&inputs[0][block_offset + 0 * sizeof(uint32x4_t)]);
+  out[1] = loadu_128(&inputs[1][block_offset + 0 * sizeof(uint32x4_t)]);
+  out[2] = loadu_128(&inputs[2][block_offset + 0 * sizeof(uint32x4_t)]);
+  out[3] = loadu_128(&inputs[3][block_offset + 0 * sizeof(uint32x4_t)]);
+  out[4] = loadu_128(&inputs[0][block_offset + 1 * sizeof(uint32x4_t)]);
+  out[5] = loadu_128(&inputs[1][block_offset + 1 * sizeof(uint32x4_t)]);
+  out[6] = loadu_128(&inputs[2][block_offset + 1 * sizeof(uint32x4_t)]);
+  out[7] = loadu_128(&inputs[3][block_offset + 1 * sizeof(uint32x4_t)]);
+  out[8] = loadu_128(&inputs[0][block_offset + 2 * sizeof(uint32x4_t)]);
+  out[9] = loadu_128(&inputs[1][block_offset + 2 * sizeof(uint32x4_t)]);
+  out[10] = loadu_128(&inputs[2][block_offset + 2 * sizeof(uint32x4_t)]);
+  out[11] = loadu_128(&inputs[3][block_offset + 2 * sizeof(uint32x4_t)]);
+  out[12] = loadu_128(&inputs[0][block_offset + 3 * sizeof(uint32x4_t)]);
+  out[13] = loadu_128(&inputs[1][block_offset + 3 * sizeof(uint32x4_t)]);
+  out[14] = loadu_128(&inputs[2][block_offset + 3 * sizeof(uint32x4_t)]);
+  out[15] = loadu_128(&inputs[3][block_offset + 3 * sizeof(uint32x4_t)]);
+  transpose_vecs_128(&out[0]);
+  transpose_vecs_128(&out[4]);
+  transpose_vecs_128(&out[8]);
+  transpose_vecs_128(&out[12]);
+}
+
+INLINE void load_counters4(uint64_t counter, bool increment_counter,
+                           uint32x4_t *out_low, uint32x4_t *out_high) {
+  uint64_t mask = (increment_counter ? ~0 : 0);
+  *out_low = set4(
+      counter_low(counter + (mask & 0)), counter_low(counter + (mask & 1)),
+      counter_low(counter + (mask & 2)), counter_low(counter + (mask & 3)));
+  *out_high = set4(
+      counter_high(counter + (mask & 0)), counter_high(counter + (mask & 1)),
+      counter_high(counter + (mask & 2)), counter_high(counter + (mask & 3)));
+}
+
+static void blake3_hash4_neon(const uint8_t *const *inputs, size_t blocks,
+                              const uint32_t key[8], uint64_t counter,
+                              bool increment_counter, uint8_t flags,
+                              uint8_t flags_start, uint8_t flags_end, 
+                              uint8_t *out) {
+  uint32x4_t h_vecs[8] = {
+      set1_128(key[0]), set1_128(key[1]), set1_128(key[2]), set1_128(key[3]),
+      set1_128(key[4]), set1_128(key[5]), set1_128(key[6]), set1_128(key[7]),
+  };
+  uint32x4_t counter_low_vec, counter_high_vec;
+  load_counters4(counter, increment_counter, &counter_low_vec,
+                 &counter_high_vec);
+  uint8_t block_flags = flags | flags_start;
+
+  for (size_t block = 0; block < blocks; block++) {
+    if (block + 1 == blocks) {
+      block_flags |= flags_end;
+    }
+    uint32x4_t block_len_vec = set1_128(BLAKE3_BLOCK_LEN);
+    uint32x4_t block_flags_vec = set1_128(block_flags);
+    uint32x4_t msg_vecs[16];
+    transpose_msg_vecs4(inputs, block * BLAKE3_BLOCK_LEN, msg_vecs);
+
+    uint32x4_t v[16] = {
+        h_vecs[0],       h_vecs[1],        h_vecs[2],       h_vecs[3],
+        h_vecs[4],       h_vecs[5],        h_vecs[6],       h_vecs[7],
+        set1_128(IV[0]), set1_128(IV[1]),  set1_128(IV[2]), set1_128(IV[3]),
+        counter_low_vec, counter_high_vec, block_len_vec,   block_flags_vec,
+    };
+    round_fn4(v, msg_vecs, 0);
+    round_fn4(v, msg_vecs, 1);
+    round_fn4(v, msg_vecs, 2);
+    round_fn4(v, msg_vecs, 3);
+    round_fn4(v, msg_vecs, 4);
+    round_fn4(v, msg_vecs, 5);
+    round_fn4(v, msg_vecs, 6);
+    h_vecs[0] = xor_128(v[0], v[8]);
+    h_vecs[1] = xor_128(v[1], v[9]);
+    h_vecs[2] = xor_128(v[2], v[10]);
+    h_vecs[3] = xor_128(v[3], v[11]);
+    h_vecs[4] = xor_128(v[4], v[12]);
+    h_vecs[5] = xor_128(v[5], v[13]);
+    h_vecs[6] = xor_128(v[6], v[14]);
+    h_vecs[7] = xor_128(v[7], v[15]);
+
+    block_flags = flags;
+  }
+
+  transpose_vecs_128(&h_vecs[0]);
+  transpose_vecs_128(&h_vecs[4]);
+  // The first four vecs now contain the first half of each output, and the
+  // second four vecs contain the second half of each output.
+  storeu_128(h_vecs[0], &out[0 * sizeof(uint32x4_t)]);
+  storeu_128(h_vecs[4], &out[1 * sizeof(uint32x4_t)]);
+  storeu_128(h_vecs[1], &out[2 * sizeof(uint32x4_t)]);
+  storeu_128(h_vecs[5], &out[3 * sizeof(uint32x4_t)]);
+  storeu_128(h_vecs[2], &out[4 * sizeof(uint32x4_t)]);
+  storeu_128(h_vecs[6], &out[5 * sizeof(uint32x4_t)]);
+  storeu_128(h_vecs[3], &out[6 * sizeof(uint32x4_t)]);
+  storeu_128(h_vecs[7], &out[7 * sizeof(uint32x4_t)]);
+}
+
+/*
+ * ----------------------------------------------------------------------------
+ * hash_many_neon
+ * ----------------------------------------------------------------------------
+ */
+
+void blake3_compress_in_place_portable(uint32_t cv[8],
+                                       const uint8_t block[BLAKE3_BLOCK_LEN],
+                                       uint8_t block_len, uint64_t counter,
+                                       uint8_t flags);
+
+INLINE void hash_one_neon(const uint8_t *input, size_t blocks,
+                          const uint32_t key[8], uint64_t counter,
+                          uint8_t flags, uint8_t flags_start, uint8_t flags_end,
+                          uint8_t out[BLAKE3_OUT_LEN]) {
+  uint32_t cv[8];
+  memcpy(cv, key, BLAKE3_KEY_LEN);
+  uint8_t block_flags = flags | flags_start;
+  while (blocks > 0) {
+    if (blocks == 1) {
+      block_flags |= flags_end;
+    }
+    // TODO: Implement compress_neon. However note that according to
+    // https://github.com/BLAKE2/BLAKE2/commit/7965d3e6e1b4193438b8d3a656787587d2579227,
+    // compress_neon might not be any faster than compress_portable.
+    blake3_compress_in_place_portable(cv, input, BLAKE3_BLOCK_LEN, counter,
+                                      block_flags);
+    input = &input[BLAKE3_BLOCK_LEN];
+    blocks -= 1;
+    block_flags = flags;
+  }
+  memcpy(out, cv, BLAKE3_OUT_LEN);
+}
+
+void blake3_hash_many_neon(const uint8_t *const *inputs, size_t num_inputs,
+                           size_t blocks, const uint32_t key[8],
+                           uint64_t counter, bool increment_counter,
+                           uint8_t flags, uint8_t flags_start,
+                           uint8_t flags_end, uint8_t *out) {
+  while (num_inputs >= 4) {
+    blake3_hash4_neon(inputs, blocks, key, counter, increment_counter, flags,
+                      flags_start, flags_end, out);
+    if (increment_counter) {
+      counter += 4;
+    }
+    inputs += 4;
+    num_inputs -= 4;
+    out = &out[4 * BLAKE3_OUT_LEN];
+  }
+  while (num_inputs > 0) {
+    hash_one_neon(inputs[0], blocks, key, counter, flags, flags_start,
+                  flags_end, out);
+    if (increment_counter) {
+      counter += 1;
+    }
+    inputs += 1;
+    num_inputs -= 1;
+    out = &out[BLAKE3_OUT_LEN];
+  }
+}