Merge 4e81bb972b into 0816badf3a

benches/bench.rs

@@ -85,6 +85,14 @@ fn bench_single_compression_avx512(b: &mut Bencher) {
     }
 }
 
+#[bench]
+#[cfg(feature = "neon")]
+fn bench_single_compression_neon(b: &mut Bencher) {
+    if let Some(platform) = Platform::neon() {
+        bench_single_compression_fn(b, platform);
+    }
+}
+
 fn bench_many_chunks_fn(b: &mut Bencher, platform: Platform) {
     let degree = platform.simd_degree();
     let mut inputs = Vec::new();

c/blake3_c_rust_bindings/src/lib.rs

@@ -289,6 +289,21 @@ pub mod ffi {
     pub mod neon {
         extern "C" {
             // NEON low level functions
+            pub fn blake3_compress_xof_neon(
+                cv: *const u32,
+                block: *const u8,
+                block_len: u8,
+                counter: u64,
+                flags: u8,
+                out: *mut u8,
+            );
+            pub fn blake3_compress_in_place_neon(
+                cv: *mut u32,
+                block: *const u8,
+                block_len: u8,
+                counter: u64,
+                flags: u8,
+            );
             pub fn blake3_hash_many_neon(
                 inputs: *const *const u8,
                 num_inputs: usize,

c/blake3_dispatch.c

@@ -4,9 +4,12 @@
 
 #include "blake3_impl.h"
 
-#if defined(IS_X86)
 #if defined(_MSC_VER)
 #include <Windows.h>
+#endif
+
+#if defined(IS_X86)
+#if defined(_MSC_VER)
 #include <intrin.h>
 #elif defined(__GNUC__)
 #include <immintrin.h>
@@ -188,6 +191,12 @@ void blake3_compress_in_place(uint32_t cv[8],
   }
 #endif
 #endif
+
+#if BLAKE3_USE_NEON == 1
+  blake3_compress_in_place_neon(cv, block, block_len, counter, flags);
+  return;
+#endif
+
   blake3_compress_in_place_portable(cv, block, block_len, counter, flags);
 }
 
@@ -217,6 +226,12 @@ void blake3_compress_xof(const uint32_t cv[8],
   }
 #endif
 #endif
+
+#if BLAKE3_USE_NEON == 1
+  blake3_compress_xof_neon(cv, block, block_len, counter, flags, out);
+  return;
+#endif
+
   blake3_compress_xof_portable(cv, block, block_len, counter, flags, out);
 }
 

c/blake3_impl.h

@@ -28,7 +28,7 @@ enum blake3_flags {
 #define INLINE static inline __attribute__((always_inline))
 #endif
 
-#if defined(__x86_64__) || defined(_M_X64) 
+#if (defined(__x86_64__) || defined(_M_X64)) && !defined(_M_ARM64EC)
 #define IS_X86
 #define IS_X86_64
 #endif
@@ -38,7 +38,7 @@ enum blake3_flags {
 #define IS_X86_32
 #endif
 
-#if defined(__aarch64__) || defined(_M_ARM64)
+#if defined(__aarch64__) || defined(_M_ARM64) || defined(_M_ARM64EC)
 #define IS_AARCH64
 #endif
 
@@ -274,6 +274,16 @@ void blake3_hash_many_avx512(const uint8_t *const *inputs, size_t num_inputs,
 #endif
 
 #if BLAKE3_USE_NEON == 1
+void blake3_compress_in_place_neon(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_neon(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_neon(const uint8_t *const *inputs, size_t num_inputs,
                            size_t blocks, const uint32_t key[8],
                            uint64_t counter, bool increment_counter,

c/blake3_neon.c

@@ -66,10 +66,263 @@ INLINE uint32x4_t rot7_128(uint32x4_t x) {
   return vsriq_n_u32(vshlq_n_u32(x, 32-7), x, 7);
 }
 
-// TODO: compress_neon
-
 // TODO: hash2_neon
 
+INLINE void g1(uint32x4_t *row0, uint32x4_t *row1, uint32x4_t *row2,
+               uint32x4_t *row3, uint32x4_t m) {
+  *row0 = vaddq_u32(vaddq_u32(*row0, m), *row1);
+  *row3 = veorq_u32(*row3, *row0);
+  *row3 = rot16_128(*row3);
+  *row2 = vaddq_u32(*row2, *row3);
+  *row1 = veorq_u32(*row1, *row2);
+  *row1 = rot12_128(*row1);
+}
+
+INLINE void g2(uint32x4_t *row0, uint32x4_t *row1, uint32x4_t *row2,
+               uint32x4_t *row3, uint32x4_t m) {
+  *row0 = vaddq_u32(vaddq_u32(*row0, m), *row1);
+  *row3 = veorq_u32(*row3, *row0);
+  *row3 = rot8_128(*row3);
+  *row2 = vaddq_u32(*row2, *row3);
+  *row1 = veorq_u32(*row1, *row2);
+  *row1 = rot7_128(*row1);
+}
+
+INLINE void diagonalize(uint32x4_t *row0, uint32x4_t *row2, uint32x4_t *row3) {
+  *row0 = vextq_u32(*row0, *row0, 3);
+  *row3 = vextq_u32(*row3, *row3, 2);
+  *row2 = vextq_u32(*row2, *row2, 1);
+}
+
+INLINE void undiagonalize(uint32x4_t *row0, uint32x4_t *row2, uint32x4_t *row3) {
+  *row0 = vextq_u32(*row0, *row0, 1);
+  *row3 = vextq_u32(*row3, *row3, 2);
+  *row2 = vextq_u32(*row2, *row2, 3);
+}
+
+#define unpacklo_32(a, b) \
+  vzip1q_u32(a, b)
+
+#define unpackhi_32(a, b) \
+  vzip2q_u32(a, b)
+
+#define unpacklo_64(a, b) \
+  vreinterpretq_u64_u32(vzip1q_u64(vreinterpretq_u32_u64(a), vreinterpretq_u32_u64(b)))
+
+#define shuffle_128(a, m3, m2, m1, m0) \
+  (__builtin_shufflevector(a, a, m0, m1, m2, m3))
+
+#define shuffle_256(a, b, m3, m2, m1, m0) \
+  (__builtin_shufflevector(a, b, m0, m1, m2 + 4, m3 + 4))
+
+#define blend_16(a, b, mask)      \
+  (vreinterpretq_u32_u16(         \
+    __builtin_shufflevector(      \
+      vreinterpretq_u16_u32(a),   \
+      vreinterpretq_u16_u32(b),   \
+      0 + ((mask >> 0) & 1) * 8,  \
+      1 + ((mask >> 1) & 1) * 8,  \
+      2 + ((mask >> 2) & 1) * 8,  \
+      3 + ((mask >> 3) & 1) * 8,  \
+      4 + ((mask >> 4) & 1) * 8,  \
+      5 + ((mask >> 5) & 1) * 8,  \
+      6 + ((mask >> 6) & 1) * 8,  \
+      7 + ((mask >> 7) & 1) * 8   \
+      )))
+
+INLINE void compress_pre(uint32x4_t rows[4], const uint32_t cv[8],
+                         const uint8_t block[BLAKE3_BLOCK_LEN],
+                         uint8_t block_len, uint64_t counter, uint8_t flags) {
+  rows[0] = loadu_128((uint8_t *)&cv[0]);
+  rows[1] = loadu_128((uint8_t *)&cv[4]);
+  rows[2] = set4(IV[0], IV[1], IV[2], IV[3]);
+  rows[3] = set4(counter_low(counter), counter_high(counter),
+                 (uint32_t)block_len, (uint32_t)flags);
+
+  uint32x4_t m0 = loadu_128(&block[sizeof(uint32x4_t) * 0]);
+  uint32x4_t m1 = loadu_128(&block[sizeof(uint32x4_t) * 1]);
+  uint32x4_t m2 = loadu_128(&block[sizeof(uint32x4_t) * 2]);
+  uint32x4_t m3 = loadu_128(&block[sizeof(uint32x4_t) * 3]);
+
+  uint32x4_t t0, t1, t2, t3, tt;
+
+  // Round 1. The first round permutes the message words from the original
+  // input order, into the groups that get mixed in parallel.
+  t0 = shuffle_256(m0, m1, 2, 0, 2, 0); //  6  4  2  0
+  g1(&rows[0], &rows[1], &rows[2], &rows[3], t0);
+  t1 = shuffle_256(m0, m1, 3, 1, 3, 1); //  7  5  3  1
+  g2(&rows[0], &rows[1], &rows[2], &rows[3], t1);
+  diagonalize(&rows[0], &rows[2], &rows[3]);
+  t2 = shuffle_256(m2, m3, 2, 0, 2, 0); // 14 12 10  8
+  t2 = shuffle_128(t2, 2, 1, 0, 3);     // 12 10  8 14
+  g1(&rows[0], &rows[1], &rows[2], &rows[3], t2);
+  t3 = shuffle_256(m2, m3, 3, 1, 3, 1); // 15 13 11  9
+  t3 = vextq_u32(t3, t3, 3);
+  g2(&rows[0], &rows[1], &rows[2], &rows[3], t3);
+  undiagonalize(&rows[0], &rows[2], &rows[3]);
+  m0 = t0;
+  m1 = t1;
+  m2 = t2;
+  m3 = t3;
+
+  // Round 2. This round and all following rounds apply a fixed permutation
+  // to the message words from the round before.
+  t0 = shuffle_256(m0, m1, 3, 1, 1, 2);
+  t0 = vextq_u32(t0, t0, 1);
+  g1(&rows[0], &rows[1], &rows[2], &rows[3], t0);
+  t1 = shuffle_256(m2, m3, 3, 3, 2, 2);
+  tt = shuffle_128(m0, 0, 0, 3, 3);
+  t1 = blend_16(tt, t1, 0xCC);
+  g2(&rows[0], &rows[1], &rows[2], &rows[3], t1);
+  diagonalize(&rows[0], &rows[2], &rows[3]);
+  t2 = unpacklo_64(m3, m1);
+  tt = blend_16(t2, m2, 0xC0);
+  t2 = shuffle_128(tt, 1, 3, 2, 0);
+  g1(&rows[0], &rows[1], &rows[2], &rows[3], t2);
+  t3 = unpackhi_32(m1, m3);
+  tt = unpacklo_32(m2, t3);
+  t3 = shuffle_128(tt, 0, 1, 3, 2);
+  g2(&rows[0], &rows[1], &rows[2], &rows[3], t3);
+  undiagonalize(&rows[0], &rows[2], &rows[3]);
+  m0 = t0;
+  m1 = t1;
+  m2 = t2;
+  m3 = t3;
+
+  // Round 3
+  t0 = shuffle_256(m0, m1, 3, 1, 1, 2);
+  t0 = vextq_u32(t0, t0, 1);
+  g1(&rows[0], &rows[1], &rows[2], &rows[3], t0);
+  t1 = shuffle_256(m2, m3, 3, 3, 2, 2);
+  tt = shuffle_128(m0, 0, 0, 3, 3);
+  t1 = blend_16(tt, t1, 0xCC);
+  g2(&rows[0], &rows[1], &rows[2], &rows[3], t1);
+  diagonalize(&rows[0], &rows[2], &rows[3]);
+  t2 = unpacklo_64(m3, m1);
+  tt = blend_16(t2, m2, 0xC0);
+  t2 = shuffle_128(tt, 1, 3, 2, 0);
+  g1(&rows[0], &rows[1], &rows[2], &rows[3], t2);
+  t3 = unpackhi_32(m1, m3);
+  tt = unpacklo_32(m2, t3);
+  t3 = shuffle_128(tt, 0, 1, 3, 2);
+  g2(&rows[0], &rows[1], &rows[2], &rows[3], t3);
+  undiagonalize(&rows[0], &rows[2], &rows[3]);
+  m0 = t0;
+  m1 = t1;
+  m2 = t2;
+  m3 = t3;
+
+  // Round 4
+  t0 = shuffle_256(m0, m1, 3, 1, 1, 2);
+  t0 = vextq_u32(t0, t0, 1);
+  g1(&rows[0], &rows[1], &rows[2], &rows[3], t0);
+  t1 = shuffle_256(m2, m3, 3, 3, 2, 2);
+  tt = shuffle_128(m0, 0, 0, 3, 3);
+  t1 = blend_16(tt, t1, 0xCC);
+  g2(&rows[0], &rows[1], &rows[2], &rows[3], t1);
+  diagonalize(&rows[0], &rows[2], &rows[3]);
+  t2 = unpacklo_64(m3, m1);
+  tt = blend_16(t2, m2, 0xC0);
+  t2 = shuffle_128(tt, 1, 3, 2, 0);
+  g1(&rows[0], &rows[1], &rows[2], &rows[3], t2);
+  t3 = unpackhi_32(m1, m3);
+  tt = unpacklo_32(m2, t3);
+  t3 = shuffle_128(tt, 0, 1, 3, 2);
+  g2(&rows[0], &rows[1], &rows[2], &rows[3], t3);
+  undiagonalize(&rows[0], &rows[2], &rows[3]);
+  m0 = t0;
+  m1 = t1;
+  m2 = t2;
+  m3 = t3;
+
+  // Round 5
+  t0 = shuffle_256(m0, m1, 3, 1, 1, 2);
+  t0 = vextq_u32(t0, t0, 1);
+  g1(&rows[0], &rows[1], &rows[2], &rows[3], t0);
+  t1 = shuffle_256(m2, m3, 3, 3, 2, 2);
+  tt = shuffle_128(m0, 0, 0, 3, 3);
+  t1 = blend_16(tt, t1, 0xCC);
+  g2(&rows[0], &rows[1], &rows[2], &rows[3], t1);
+  diagonalize(&rows[0], &rows[2], &rows[3]);
+  t2 = unpacklo_64(m3, m1);
+  tt = blend_16(t2, m2, 0xC0);
+  t2 = shuffle_128(tt, 1, 3, 2, 0);
+  g1(&rows[0], &rows[1], &rows[2], &rows[3], t2);
+  t3 = unpackhi_32(m1, m3);
+  tt = unpacklo_32(m2, t3);
+  t3 = shuffle_128(tt, 0, 1, 3, 2);
+  g2(&rows[0], &rows[1], &rows[2], &rows[3], t3);
+  undiagonalize(&rows[0], &rows[2], &rows[3]);
+  m0 = t0;
+  m1 = t1;
+  m2 = t2;
+  m3 = t3;
+
+  // Round 6
+  t0 = shuffle_256(m0, m1, 3, 1, 1, 2);
+  t0 = vextq_u32(t0, t0, 1);
+  g1(&rows[0], &rows[1], &rows[2], &rows[3], t0);
+  t1 = shuffle_256(m2, m3, 3, 3, 2, 2);
+  tt = shuffle_128(m0, 0, 0, 3, 3);
+  t1 = blend_16(tt, t1, 0xCC);
+  g2(&rows[0], &rows[1], &rows[2], &rows[3], t1);
+  diagonalize(&rows[0], &rows[2], &rows[3]);
+  t2 = unpacklo_64(m3, m1);
+  tt = blend_16(t2, m2, 0xC0);
+  t2 = shuffle_128(tt, 1, 3, 2, 0);
+  g1(&rows[0], &rows[1], &rows[2], &rows[3], t2);
+  t3 = unpackhi_32(m1, m3);
+  tt = unpacklo_32(m2, t3);
+  t3 = shuffle_128(tt, 0, 1, 3, 2);
+  g2(&rows[0], &rows[1], &rows[2], &rows[3], t3);
+  undiagonalize(&rows[0], &rows[2], &rows[3]);
+  m0 = t0;
+  m1 = t1;
+  m2 = t2;
+  m3 = t3;
+
+  // Round 7
+  t0 = shuffle_256(m0, m1, 3, 1, 1, 2);
+  t0 = vextq_u32(t0, t0, 1);
+  g1(&rows[0], &rows[1], &rows[2], &rows[3], t0);
+  t1 = shuffle_256(m2, m3, 3, 3, 2, 2);
+  tt = shuffle_128(m0, 0, 0, 3, 3);
+  t1 = blend_16(tt, t1, 0xCC);
+  g2(&rows[0], &rows[1], &rows[2], &rows[3], t1);
+  diagonalize(&rows[0], &rows[2], &rows[3]);
+  t2 = unpacklo_64(m3, m1);
+  tt = blend_16(t2, m2, 0xC0);
+  t2 = shuffle_128(tt, 1, 3, 2, 0);
+  g1(&rows[0], &rows[1], &rows[2], &rows[3], t2);
+  t3 = unpackhi_32(m1, m3);
+  tt = unpacklo_32(m2, t3);
+  t3 = shuffle_128(tt, 0, 1, 3, 2);
+  g2(&rows[0], &rows[1], &rows[2], &rows[3], t3);
+  undiagonalize(&rows[0], &rows[2], &rows[3]);
+}
+
+void blake3_compress_in_place_neon(uint32_t cv[8],
+                                   const uint8_t block[BLAKE3_BLOCK_LEN],
+                                   uint8_t block_len, uint64_t counter,
+                                   uint8_t flags) {
+  uint32x4_t rows[4];
+  compress_pre(rows, cv, block, block_len, counter, flags);
+  storeu_128(veorq_u32(rows[0], rows[2]), (uint8_t *)&cv[0]);
+  storeu_128(veorq_u32(rows[1], rows[3]), (uint8_t *)&cv[4]);
+}
+
+void blake3_compress_xof_neon(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]) {
+  uint32x4_t rows[4];
+  compress_pre(rows, cv, block, block_len, counter, flags);
+  storeu_128(veorq_u32(rows[0], rows[2]), &out[0]);
+  storeu_128(veorq_u32(rows[1], rows[3]), &out[16]);
+  storeu_128(veorq_u32(rows[2], loadu_128((uint8_t *)&cv[0])), &out[32]);
+  storeu_128(veorq_u32(rows[3], loadu_128((uint8_t *)&cv[4])), &out[48]);
+}
+
 /*
  * ----------------------------------------------------------------------------
  * hash4_neon
@@ -232,6 +485,47 @@ INLINE void transpose_msg_vecs4(const uint8_t *const *inputs,
   transpose_vecs_128(&out[12]);
 }
 
+// NOTE: The version below avoids the explicit transposes by relying on the interleaving from
+// `vst4q_u32` but it seems to make no difference, or perhaps might be even a little slower.
+
+// INLINE void transpose_msg_vecs4(const uint8_t *const *inputs,
+//                                 size_t block_offset, uint32x4_t out[4]) {
+//   uint8x16x4_t l0 = vld1q_u8_x4(&inputs[0][block_offset]);
+//   uint8x16x4_t l1 = vld1q_u8_x4(&inputs[1][block_offset]);
+//   uint8x16x4_t l2 = vld1q_u8_x4(&inputs[2][block_offset]);
+//   uint8x16x4_t l3 = vld1q_u8_x4(&inputs[3][block_offset]);
+
+//   uint32x4x4_t s0 = {
+//     vreinterpretq_u32_u8(l0.val[0]),
+//     vreinterpretq_u32_u8(l1.val[0]),
+//     vreinterpretq_u32_u8(l2.val[0]),
+//     vreinterpretq_u32_u8(l3.val[0]),
+//   };
+//   uint32x4x4_t s1 = {
+//     vreinterpretq_u32_u8(l0.val[1]),
+//     vreinterpretq_u32_u8(l1.val[1]),
+//     vreinterpretq_u32_u8(l2.val[1]),
+//     vreinterpretq_u32_u8(l3.val[1]),
+//   };
+//   uint32x4x4_t s2 = {
+//     vreinterpretq_u32_u8(l0.val[2]),
+//     vreinterpretq_u32_u8(l1.val[2]),
+//     vreinterpretq_u32_u8(l2.val[2]),
+//     vreinterpretq_u32_u8(l3.val[2]),
+//   };
+//   uint32x4x4_t s3 = {
+//     vreinterpretq_u32_u8(l0.val[3]),
+//     vreinterpretq_u32_u8(l1.val[3]),
+//     vreinterpretq_u32_u8(l2.val[3]),
+//     vreinterpretq_u32_u8(l3.val[3]),
+//   };
+
+//   vst4q_u32((uint32_t *)&out[0], s0);
+//   vst4q_u32((uint32_t *)&out[4], s1);
+//   vst4q_u32((uint32_t *)&out[8], s2);
+//   vst4q_u32((uint32_t *)&out[12], s3);
+// }
+
 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);
@@ -310,11 +604,6 @@ void blake3_hash4_neon(const uint8_t *const *inputs, size_t blocks,
  * ----------------------------------------------------------------------------
  */
 
-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,
@@ -326,11 +615,8 @@ INLINE void hash_one_neon(const uint8_t *input, size_t blocks,
     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);
+    blake3_compress_in_place_neon(cv, input, BLAKE3_BLOCK_LEN, counter,
+                                  block_flags);
     input = &input[BLAKE3_BLOCK_LEN];
     blocks -= 1;
     block_flags = flags;

src/ffi_neon.rs

@@ -1,5 +1,34 @@
 use crate::{CVWords, IncrementCounter, BLOCK_LEN, OUT_LEN};
 
+pub unsafe fn compress_in_place(
+    cv: &mut CVWords,
+    block: &[u8; BLOCK_LEN],
+    block_len: u8,
+    counter: u64,
+    flags: u8,
+) {
+    ffi::blake3_compress_in_place_neon(cv.as_mut_ptr(), block.as_ptr(), block_len, counter, flags)
+}
+
+pub unsafe fn compress_xof(
+    cv: &CVWords,
+    block: &[u8; BLOCK_LEN],
+    block_len: u8,
+    counter: u64,
+    flags: u8,
+) -> [u8; 64] {
+    let mut out = [0u8; 64];
+    ffi::blake3_compress_xof_neon(
+        cv.as_ptr(),
+        block.as_ptr(),
+        block_len,
+        counter,
+        flags,
+        out.as_mut_ptr(),
+    );
+    out
+}
+
 // Unsafe because this may only be called on platforms supporting NEON.
 pub unsafe fn hash_many<const N: usize>(
     inputs: &[&[u8; N]],
@@ -29,31 +58,23 @@ pub unsafe fn hash_many<const N: usize>(
     )
 }
 
-// blake3_neon.c normally depends on blake3_portable.c, because the NEON
-// implementation only provides 4x compression, and it relies on the portable
-// implementation for 1x compression. However, we expose the portable Rust
-// implementation here instead, to avoid linking in unnecessary code.
-#[no_mangle]
-pub extern "C" fn blake3_compress_in_place_portable(
-    cv: *mut u32,
-    block: *const u8,
-    block_len: u8,
-    counter: u64,
-    flags: u8,
-) {
-    unsafe {
-        crate::portable::compress_in_place(
-            &mut *(cv as *mut [u32; 8]),
-            &*(block as *const [u8; 64]),
-            block_len,
-            counter,
-            flags,
-        )
-    }
-}
-
 pub mod ffi {
     extern "C" {
+        pub fn blake3_compress_in_place_neon(
+            cv: *mut u32,
+            block: *const u8,
+            block_len: u8,
+            counter: u64,
+            flags: u8,
+        );
+        pub fn blake3_compress_xof_neon(
+            cv: *const u32,
+            block: *const u8,
+            block_len: u8,
+            counter: u64,
+            flags: u8,
+            out: *mut u8,
+        );
         pub fn blake3_hash_many_neon(
             inputs: *const *const u8,
             num_inputs: usize,

src/platform.rs

@@ -133,9 +133,11 @@ impl Platform {
             Platform::AVX512 => unsafe {
                 crate::avx512::compress_in_place(cv, block, block_len, counter, flags)
             },
-            // No NEON compress_in_place() implementation yet.
+            // Safe because detect() checked for platform support.
             #[cfg(blake3_neon)]
-            Platform::NEON => portable::compress_in_place(cv, block, block_len, counter, flags),
+            Platform::NEON => unsafe {
+                crate::neon::compress_in_place(cv, block, block_len, counter, flags)
+            },
         }
     }
 
@@ -165,9 +167,11 @@ impl Platform {
             Platform::AVX512 => unsafe {
                 crate::avx512::compress_xof(cv, block, block_len, counter, flags)
             },
-            // No NEON compress_xof() implementation yet.
+            // Safe because detect() checked for platform support.
             #[cfg(blake3_neon)]
-            Platform::NEON => portable::compress_xof(cv, block, block_len, counter, flags),
+            Platform::NEON => unsafe {
+                crate::neon::compress_xof(cv, block, block_len, counter, flags)
+            },
         }
     }