util: bump blake3 from 1.3.3 to 1.5.1, improve armv7 and aarch64 performance

Steps for uprev:
 - copy files from BLAKE3/c src/util/blake3/
 - edit README
 - `for file in *.asm; do mv "$file" "${file%.asm}.masm"; done`
 - keep
  - blake3.h (no relevant changes), only change BLAKE3_VERSION_STRING
  - blake3_sse2_x86-64_unix.S (no changes)
  - blake3_avx512_x86-64_unix.S (no changes)
  - blake3_sse41_x86-64_unix.S (no changes)

Acked-by: Marek Olšák <marek.olsak@amd.com>
Signed-off-by: David Heidelberg <david@ixit.cz>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/29687>

14 files changed, 2762 insertions, 39 deletions

diff --git a/src/util/blake3/README b/src/util/blake3/README
index 02922dc0f49..712595cddf7 100644
--- a/src/util/blake3/README
+++ b/src/util/blake3/README
@@ -1,7 +1,7 @@
-This folder contains a local copy of BLAKE3 cryptographic hash library, version 1.3.3.
+This folder contains a local copy of BLAKE3 cryptographic hash library, version 1.5.1.
 
 Except for changes listed in the "Changes" section, this is a verbatim copy from
-https://github.com/BLAKE3-team/BLAKE3, tag 1.3.3.
+https://github.com/BLAKE3-team/BLAKE3, tag 1.5.1.
 
 Files will be periodically synchronized with the upstream, and any local changes should
 be clearly documented below.
@@ -13,4 +13,4 @@ Changes:
 
 - Add "static" to blake3_hash4_neon, to comply with -Werror=missing-prototypes.
 
-- Add mesa_blake3_visibility.h and set symbol visibility to hidden for assembly sources.
\ No newline at end of file
+- Add mesa_blake3_visibility.h and set symbol visibility to hidden for assembly sources.
diff --git a/src/util/blake3/blake3.c b/src/util/blake3/blake3.c
index dc343f91c51..1b44c719345 100644
--- a/src/util/blake3/blake3.c
+++ b/src/util/blake3/blake3.c
@@ -254,7 +254,7 @@ INLINE size_t compress_parents_parallel(const uint8_t *child_chaining_values,
 // As a special case when the SIMD degree is 1, this function will still return
 // at least 2 outputs. This guarantees that this function doesn't perform the
 // root compression. (If it did, it would use the wrong flags, and also we
-// wouldn't be able to implement exendable output.) Note that this function is
+// wouldn't be able to implement extendable output.) Note that this function is
 // not used when the whole input is only 1 chunk long; that's a different
 // codepath.
 //
@@ -341,21 +341,24 @@ INLINE void compress_subtree_to_parent_node(
   size_t num_cvs = blake3_compress_subtree_wide(input, input_len, key,
                                                 chunk_counter, flags, cv_array);
   assert(num_cvs <= MAX_SIMD_DEGREE_OR_2);
-
-  // If MAX_SIMD_DEGREE is greater than 2 and there's enough input,
+  // The following loop never executes when MAX_SIMD_DEGREE_OR_2 is 2, because
+  // as we just asserted, num_cvs will always be <=2 in that case. But GCC
+  // (particularly GCC 8.5) can't tell that it never executes, and if NDEBUG is
+  // set then it emits incorrect warnings here. We tried a few different
+  // hacks to silence these, but in the end our hacks just produced different
+  // warnings (see https://github.com/BLAKE3-team/BLAKE3/pull/380). Out of
+  // desperation, we ifdef out this entire loop when we know it's not needed.
+#if MAX_SIMD_DEGREE_OR_2 > 2
+  // If MAX_SIMD_DEGREE_OR_2 is greater than 2 and there's enough input,
   // compress_subtree_wide() returns more than 2 chaining values. Condense
   // them into 2 by forming parent nodes repeatedly.
   uint8_t out_array[MAX_SIMD_DEGREE_OR_2 * BLAKE3_OUT_LEN / 2];
-  // The second half of this loop condition is always true, and we just
-  // asserted it above. But GCC can't tell that it's always true, and if NDEBUG
-  // is set on platforms where MAX_SIMD_DEGREE_OR_2 == 2, GCC emits spurious
-  // warnings here. GCC 8.5 is particularly sensitive, so if you're changing
-  // this code, test it against that version.
-  while (num_cvs > 2 && num_cvs <= MAX_SIMD_DEGREE_OR_2) {
+  while (num_cvs > 2) {
     num_cvs =
         compress_parents_parallel(cv_array, num_cvs, key, flags, out_array);
     memcpy(cv_array, out_array, num_cvs * BLAKE3_OUT_LEN);
   }
+#endif
   memcpy(out, cv_array, 2 * BLAKE3_OUT_LEN);
 }
 
diff --git a/src/util/blake3/blake3.h b/src/util/blake3/blake3.h
index ff09151c512..4f1cca97166 100644
--- a/src/util/blake3/blake3.h
+++ b/src/util/blake3/blake3.h
@@ -8,7 +8,7 @@
 extern "C" {
 #endif
 
-#define BLAKE3_VERSION_STRING "1.3.3"
+#define BLAKE3_VERSION_STRING "1.5.1"
 #define BLAKE3_KEY_LEN 32
 #define BLAKE3_OUT_LEN 32
 #define BLAKE3_BLOCK_LEN 64
diff --git a/src/util/blake3/blake3_avx2.c b/src/util/blake3/blake3_avx2.c
new file mode 100644
index 00000000000..381e7c422f3
--- /dev/null
+++ b/src/util/blake3/blake3_avx2.c
@@ -0,0 +1,326 @@
+#include "blake3_impl.h"
+
+#include <immintrin.h>
+
+#define DEGREE 8
+
+INLINE __m256i loadu(const uint8_t src[32]) {
+  return _mm256_loadu_si256((const __m256i *)src);
+}
+
+INLINE void storeu(__m256i src, uint8_t dest[16]) {
+  _mm256_storeu_si256((__m256i *)dest, src);
+}
+
+INLINE __m256i addv(__m256i a, __m256i b) { return _mm256_add_epi32(a, b); }
+
+// Note that clang-format doesn't like the name "xor" for some reason.
+INLINE __m256i xorv(__m256i a, __m256i b) { return _mm256_xor_si256(a, b); }
+
+INLINE __m256i set1(uint32_t x) { return _mm256_set1_epi32((int32_t)x); }
+
+INLINE __m256i rot16(__m256i x) {
+  return _mm256_shuffle_epi8(
+      x, _mm256_set_epi8(13, 12, 15, 14, 9, 8, 11, 10, 5, 4, 7, 6, 1, 0, 3, 2,
+                         13, 12, 15, 14, 9, 8, 11, 10, 5, 4, 7, 6, 1, 0, 3, 2));
+}
+
+INLINE __m256i rot12(__m256i x) {
+  return _mm256_or_si256(_mm256_srli_epi32(x, 12), _mm256_slli_epi32(x, 32 - 12));
+}
+
+INLINE __m256i rot8(__m256i x) {
+  return _mm256_shuffle_epi8(
+      x, _mm256_set_epi8(12, 15, 14, 13, 8, 11, 10, 9, 4, 7, 6, 5, 0, 3, 2, 1,
+                         12, 15, 14, 13, 8, 11, 10, 9, 4, 7, 6, 5, 0, 3, 2, 1));
+}
+
+INLINE __m256i rot7(__m256i x) {
+  return _mm256_or_si256(_mm256_srli_epi32(x, 7), _mm256_slli_epi32(x, 32 - 7));
+}
+
+INLINE void round_fn(__m256i v[16], __m256i m[16], size_t r) {
+  v[0] = addv(v[0], m[(size_t)MSG_SCHEDULE[r][0]]);
+  v[1] = addv(v[1], m[(size_t)MSG_SCHEDULE[r][2]]);
+  v[2] = addv(v[2], m[(size_t)MSG_SCHEDULE[r][4]]);
+  v[3] = addv(v[3], m[(size_t)MSG_SCHEDULE[r][6]]);
+  v[0] = addv(v[0], v[4]);
+  v[1] = addv(v[1], v[5]);
+  v[2] = addv(v[2], v[6]);
+  v[3] = addv(v[3], v[7]);
+  v[12] = xorv(v[12], v[0]);
+  v[13] = xorv(v[13], v[1]);
+  v[14] = xorv(v[14], v[2]);
+  v[15] = xorv(v[15], v[3]);
+  v[12] = rot16(v[12]);
+  v[13] = rot16(v[13]);
+  v[14] = rot16(v[14]);
+  v[15] = rot16(v[15]);
+  v[8] = addv(v[8], v[12]);
+  v[9] = addv(v[9], v[13]);
+  v[10] = addv(v[10], v[14]);
+  v[11] = addv(v[11], v[15]);
+  v[4] = xorv(v[4], v[8]);
+  v[5] = xorv(v[5], v[9]);
+  v[6] = xorv(v[6], v[10]);
+  v[7] = xorv(v[7], v[11]);
+  v[4] = rot12(v[4]);
+  v[5] = rot12(v[5]);
+  v[6] = rot12(v[6]);
+  v[7] = rot12(v[7]);
+  v[0] = addv(v[0], m[(size_t)MSG_SCHEDULE[r][1]]);
+  v[1] = addv(v[1], m[(size_t)MSG_SCHEDULE[r][3]]);
+  v[2] = addv(v[2], m[(size_t)MSG_SCHEDULE[r][5]]);
+  v[3] = addv(v[3], m[(size_t)MSG_SCHEDULE[r][7]]);
+  v[0] = addv(v[0], v[4]);
+  v[1] = addv(v[1], v[5]);
+  v[2] = addv(v[2], v[6]);
+  v[3] = addv(v[3], v[7]);
+  v[12] = xorv(v[12], v[0]);
+  v[13] = xorv(v[13], v[1]);
+  v[14] = xorv(v[14], v[2]);
+  v[15] = xorv(v[15], v[3]);
+  v[12] = rot8(v[12]);
+  v[13] = rot8(v[13]);
+  v[14] = rot8(v[14]);
+  v[15] = rot8(v[15]);
+  v[8] = addv(v[8], v[12]);
+  v[9] = addv(v[9], v[13]);
+  v[10] = addv(v[10], v[14]);
+  v[11] = addv(v[11], v[15]);
+  v[4] = xorv(v[4], v[8]);
+  v[5] = xorv(v[5], v[9]);
+  v[6] = xorv(v[6], v[10]);
+  v[7] = xorv(v[7], v[11]);
+  v[4] = rot7(v[4]);
+  v[5] = rot7(v[5]);
+  v[6] = rot7(v[6]);
+  v[7] = rot7(v[7]);
+
+  v[0] = addv(v[0], m[(size_t)MSG_SCHEDULE[r][8]]);
+  v[1] = addv(v[1], m[(size_t)MSG_SCHEDULE[r][10]]);
+  v[2] = addv(v[2], m[(size_t)MSG_SCHEDULE[r][12]]);
+  v[3] = addv(v[3], m[(size_t)MSG_SCHEDULE[r][14]]);
+  v[0] = addv(v[0], v[5]);
+  v[1] = addv(v[1], v[6]);
+  v[2] = addv(v[2], v[7]);
+  v[3] = addv(v[3], v[4]);
+  v[15] = xorv(v[15], v[0]);
+  v[12] = xorv(v[12], v[1]);
+  v[13] = xorv(v[13], v[2]);
+  v[14] = xorv(v[14], v[3]);
+  v[15] = rot16(v[15]);
+  v[12] = rot16(v[12]);
+  v[13] = rot16(v[13]);
+  v[14] = rot16(v[14]);
+  v[10] = addv(v[10], v[15]);
+  v[11] = addv(v[11], v[12]);
+  v[8] = addv(v[8], v[13]);
+  v[9] = addv(v[9], v[14]);
+  v[5] = xorv(v[5], v[10]);
+  v[6] = xorv(v[6], v[11]);
+  v[7] = xorv(v[7], v[8]);
+  v[4] = xorv(v[4], v[9]);
+  v[5] = rot12(v[5]);
+  v[6] = rot12(v[6]);
+  v[7] = rot12(v[7]);
+  v[4] = rot12(v[4]);
+  v[0] = addv(v[0], m[(size_t)MSG_SCHEDULE[r][9]]);
+  v[1] = addv(v[1], m[(size_t)MSG_SCHEDULE[r][11]]);
+  v[2] = addv(v[2], m[(size_t)MSG_SCHEDULE[r][13]]);
+  v[3] = addv(v[3], m[(size_t)MSG_SCHEDULE[r][15]]);
+  v[0] = addv(v[0], v[5]);
+  v[1] = addv(v[1], v[6]);
+  v[2] = addv(v[2], v[7]);
+  v[3] = addv(v[3], v[4]);
+  v[15] = xorv(v[15], v[0]);
+  v[12] = xorv(v[12], v[1]);
+  v[13] = xorv(v[13], v[2]);
+  v[14] = xorv(v[14], v[3]);
+  v[15] = rot8(v[15]);
+  v[12] = rot8(v[12]);
+  v[13] = rot8(v[13]);
+  v[14] = rot8(v[14]);
+  v[10] = addv(v[10], v[15]);
+  v[11] = addv(v[11], v[12]);
+  v[8] = addv(v[8], v[13]);
+  v[9] = addv(v[9], v[14]);
+  v[5] = xorv(v[5], v[10]);
+  v[6] = xorv(v[6], v[11]);
+  v[7] = xorv(v[7], v[8]);
+  v[4] = xorv(v[4], v[9]);
+  v[5] = rot7(v[5]);
+  v[6] = rot7(v[6]);
+  v[7] = rot7(v[7]);
+  v[4] = rot7(v[4]);
+}
+
+INLINE void transpose_vecs(__m256i vecs[DEGREE]) {
+  // Interleave 32-bit lanes. The low unpack is lanes 00/11/44/55, and the high
+  // is 22/33/66/77.
+  __m256i ab_0145 = _mm256_unpacklo_epi32(vecs[0], vecs[1]);
+  __m256i ab_2367 = _mm256_unpackhi_epi32(vecs[0], vecs[1]);
+  __m256i cd_0145 = _mm256_unpacklo_epi32(vecs[2], vecs[3]);
+  __m256i cd_2367 = _mm256_unpackhi_epi32(vecs[2], vecs[3]);
+  __m256i ef_0145 = _mm256_unpacklo_epi32(vecs[4], vecs[5]);
+  __m256i ef_2367 = _mm256_unpackhi_epi32(vecs[4], vecs[5]);
+  __m256i gh_0145 = _mm256_unpacklo_epi32(vecs[6], vecs[7]);
+  __m256i gh_2367 = _mm256_unpackhi_epi32(vecs[6], vecs[7]);
+
+  // Interleave 64-bit lanes. The low unpack is lanes 00/22 and the high is
+  // 11/33.
+  __m256i abcd_04 = _mm256_unpacklo_epi64(ab_0145, cd_0145);
+  __m256i abcd_15 = _mm256_unpackhi_epi64(ab_0145, cd_0145);
+  __m256i abcd_26 = _mm256_unpacklo_epi64(ab_2367, cd_2367);
+  __m256i abcd_37 = _mm256_unpackhi_epi64(ab_2367, cd_2367);
+  __m256i efgh_04 = _mm256_unpacklo_epi64(ef_0145, gh_0145);
+  __m256i efgh_15 = _mm256_unpackhi_epi64(ef_0145, gh_0145);
+  __m256i efgh_26 = _mm256_unpacklo_epi64(ef_2367, gh_2367);
+  __m256i efgh_37 = _mm256_unpackhi_epi64(ef_2367, gh_2367);
+
+  // Interleave 128-bit lanes.
+  vecs[0] = _mm256_permute2x128_si256(abcd_04, efgh_04, 0x20);
+  vecs[1] = _mm256_permute2x128_si256(abcd_15, efgh_15, 0x20);
+  vecs[2] = _mm256_permute2x128_si256(abcd_26, efgh_26, 0x20);
+  vecs[3] = _mm256_permute2x128_si256(abcd_37, efgh_37, 0x20);
+  vecs[4] = _mm256_permute2x128_si256(abcd_04, efgh_04, 0x31);
+  vecs[5] = _mm256_permute2x128_si256(abcd_15, efgh_15, 0x31);
+  vecs[6] = _mm256_permute2x128_si256(abcd_26, efgh_26, 0x31);
+  vecs[7] = _mm256_permute2x128_si256(abcd_37, efgh_37, 0x31);
+}
+
+INLINE void transpose_msg_vecs(const uint8_t *const *inputs,
+                               size_t block_offset, __m256i out[16]) {
+  out[0] = loadu(&inputs[0][block_offset + 0 * sizeof(__m256i)]);
+  out[1] = loadu(&inputs[1][block_offset + 0 * sizeof(__m256i)]);
+  out[2] = loadu(&inputs[2][block_offset + 0 * sizeof(__m256i)]);
+  out[3] = loadu(&inputs[3][block_offset + 0 * sizeof(__m256i)]);
+  out[4] = loadu(&inputs[4][block_offset + 0 * sizeof(__m256i)]);
+  out[5] = loadu(&inputs[5][block_offset + 0 * sizeof(__m256i)]);
+  out[6] = loadu(&inputs[6][block_offset + 0 * sizeof(__m256i)]);
+  out[7] = loadu(&inputs[7][block_offset + 0 * sizeof(__m256i)]);
+  out[8] = loadu(&inputs[0][block_offset + 1 * sizeof(__m256i)]);
+  out[9] = loadu(&inputs[1][block_offset + 1 * sizeof(__m256i)]);
+  out[10] = loadu(&inputs[2][block_offset + 1 * sizeof(__m256i)]);
+  out[11] = loadu(&inputs[3][block_offset + 1 * sizeof(__m256i)]);
+  out[12] = loadu(&inputs[4][block_offset + 1 * sizeof(__m256i)]);
+  out[13] = loadu(&inputs[5][block_offset + 1 * sizeof(__m256i)]);
+  out[14] = loadu(&inputs[6][block_offset + 1 * sizeof(__m256i)]);
+  out[15] = loadu(&inputs[7][block_offset + 1 * sizeof(__m256i)]);
+  for (size_t i = 0; i < 8; ++i) {
+    _mm_prefetch((const void *)&inputs[i][block_offset + 256], _MM_HINT_T0);
+  }
+  transpose_vecs(&out[0]);
+  transpose_vecs(&out[8]);
+}
+
+INLINE void load_counters(uint64_t counter, bool increment_counter,
+                          __m256i *out_lo, __m256i *out_hi) {
+  const __m256i mask = _mm256_set1_epi32(-(int32_t)increment_counter);
+  const __m256i add0 = _mm256_set_epi32(7, 6, 5, 4, 3, 2, 1, 0);
+  const __m256i add1 = _mm256_and_si256(mask, add0);
+  __m256i l = _mm256_add_epi32(_mm256_set1_epi32((int32_t)counter), add1);
+  __m256i carry = _mm256_cmpgt_epi32(_mm256_xor_si256(add1, _mm256_set1_epi32(0x80000000)), 
+                                     _mm256_xor_si256(   l, _mm256_set1_epi32(0x80000000)));
+  __m256i h = _mm256_sub_epi32(_mm256_set1_epi32((int32_t)(counter >> 32)), carry);
+  *out_lo = l;
+  *out_hi = h;
+}
+
+static
+void blake3_hash8_avx2(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) {
+  __m256i h_vecs[8] = {
+      set1(key[0]), set1(key[1]), set1(key[2]), set1(key[3]),
+      set1(key[4]), set1(key[5]), set1(key[6]), set1(key[7]),
+  };
+  __m256i counter_low_vec, counter_high_vec;
+  load_counters(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;
+    }
+    __m256i block_len_vec = set1(BLAKE3_BLOCK_LEN);
+    __m256i block_flags_vec = set1(block_flags);
+    __m256i msg_vecs[16];
+    transpose_msg_vecs(inputs, block * BLAKE3_BLOCK_LEN, msg_vecs);
+
+    __m256i 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(IV[0]),     set1(IV[1]),      set1(IV[2]),   set1(IV[3]),
+        counter_low_vec, counter_high_vec, block_len_vec, block_flags_vec,
+    };
+    round_fn(v, msg_vecs, 0);
+    round_fn(v, msg_vecs, 1);
+    round_fn(v, msg_vecs, 2);
+    round_fn(v, msg_vecs, 3);
+    round_fn(v, msg_vecs, 4);
+    round_fn(v, msg_vecs, 5);
+    round_fn(v, msg_vecs, 6);
+    h_vecs[0] = xorv(v[0], v[8]);
+    h_vecs[1] = xorv(v[1], v[9]);
+    h_vecs[2] = xorv(v[2], v[10]);
+    h_vecs[3] = xorv(v[3], v[11]);
+    h_vecs[4] = xorv(v[4], v[12]);
+    h_vecs[5] = xorv(v[5], v[13]);
+    h_vecs[6] = xorv(v[6], v[14]);
+    h_vecs[7] = xorv(v[7], v[15]);
+
+    block_flags = flags;
+  }
+
+  transpose_vecs(h_vecs);
+  storeu(h_vecs[0], &out[0 * sizeof(__m256i)]);
+  storeu(h_vecs[1], &out[1 * sizeof(__m256i)]);
+  storeu(h_vecs[2], &out[2 * sizeof(__m256i)]);
+  storeu(h_vecs[3], &out[3 * sizeof(__m256i)]);
+  storeu(h_vecs[4], &out[4 * sizeof(__m256i)]);
+  storeu(h_vecs[5], &out[5 * sizeof(__m256i)]);
+  storeu(h_vecs[6], &out[6 * sizeof(__m256i)]);
+  storeu(h_vecs[7], &out[7 * sizeof(__m256i)]);
+}
+
+#if !defined(BLAKE3_NO_SSE41)
+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);
+#else
+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);
+#endif
+
+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) {
+  while (num_inputs >= DEGREE) {
+    blake3_hash8_avx2(inputs, blocks, key, counter, increment_counter, flags,
+                      flags_start, flags_end, out);
+    if (increment_counter) {
+      counter += DEGREE;
+    }
+    inputs += DEGREE;
+    num_inputs -= DEGREE;
+    out = &out[DEGREE * BLAKE3_OUT_LEN];
+  }
+#if !defined(BLAKE3_NO_SSE41)
+  blake3_hash_many_sse41(inputs, num_inputs, blocks, key, counter,
+                         increment_counter, flags, flags_start, flags_end, out);
+#else
+  blake3_hash_many_portable(inputs, num_inputs, blocks, key, counter,
+                            increment_counter, flags, flags_start, flags_end,
+                            out);
+#endif
+}
diff --git a/src/util/blake3/blake3_avx2_x86-64_windows_gnu.S b/src/util/blake3/blake3_avx2_x86-64_windows_gnu.S
index bb58d2ae64b..3d4be4a7d7c 100644
--- a/src/util/blake3/blake3_avx2_x86-64_windows_gnu.S
+++ b/src/util/blake3/blake3_avx2_x86-64_windows_gnu.S
@@ -1784,7 +1784,7 @@ blake3_hash_many_avx2:
         vmovdqu xmmword ptr [rbx+0x10], xmm1
         jmp     4b
 
-.section .rodata
+.section .rdata
 .p2align  6
 ADD0:
         .long  0, 1, 2, 3, 4, 5, 6, 7
diff --git a/src/util/blake3/blake3_avx512.c b/src/util/blake3/blake3_avx512.c
new file mode 100644
index 00000000000..d6b1ae9b183
--- /dev/null
+++ b/src/util/blake3/blake3_avx512.c
@@ -0,0 +1,1220 @@
+#include "blake3_impl.h"
+
+#include <immintrin.h>
+
+#define _mm_shuffle_ps2(a, b, c)                                               \
+  (_mm_castps_si128(                                                           \
+      _mm_shuffle_ps(_mm_castsi128_ps(a), _mm_castsi128_ps(b), (c))))
+
+INLINE __m128i loadu_128(const uint8_t src[16]) {
+  return _mm_loadu_si128((const __m128i *)src);
+}
+
+INLINE __m256i loadu_256(const uint8_t src[32]) {
+  return _mm256_loadu_si256((const __m256i *)src);
+}
+
+INLINE __m512i loadu_512(const uint8_t src[64]) {
+  return _mm512_loadu_si512((const __m512i *)src);
+}
+
+INLINE void storeu_128(__m128i src, uint8_t dest[16]) {
+  _mm_storeu_si128((__m128i *)dest, src);
+}
+
+INLINE void storeu_256(__m256i src, uint8_t dest[16]) {
+  _mm256_storeu_si256((__m256i *)dest, src);
+}
+
+INLINE __m128i add_128(__m128i a, __m128i b) { return _mm_add_epi32(a, b); }
+
+INLINE __m256i add_256(__m256i a, __m256i b) { return _mm256_add_epi32(a, b); }
+
+INLINE __m512i add_512(__m512i a, __m512i b) { return _mm512_add_epi32(a, b); }
+
+INLINE __m128i xor_128(__m128i a, __m128i b) { return _mm_xor_si128(a, b); }
+
+INLINE __m256i xor_256(__m256i a, __m256i b) { return _mm256_xor_si256(a, b); }
+
+INLINE __m512i xor_512(__m512i a, __m512i b) { return _mm512_xor_si512(a, b); }
+
+INLINE __m128i set1_128(uint32_t x) { return _mm_set1_epi32((int32_t)x); }
+
+INLINE __m256i set1_256(uint32_t x) { return _mm256_set1_epi32((int32_t)x); }
+
+INLINE __m512i set1_512(uint32_t x) { return _mm512_set1_epi32((int32_t)x); }
+
+INLINE __m128i set4(uint32_t a, uint32_t b, uint32_t c, uint32_t d) {
+  return _mm_setr_epi32((int32_t)a, (int32_t)b, (int32_t)c, (int32_t)d);
+}
+
+INLINE __m128i rot16_128(__m128i x) { return _mm_ror_epi32(x, 16); }
+
+INLINE __m256i rot16_256(__m256i x) { return _mm256_ror_epi32(x, 16); }
+
+INLINE __m512i rot16_512(__m512i x) { return _mm512_ror_epi32(x, 16); }
+
+INLINE __m128i rot12_128(__m128i x) { return _mm_ror_epi32(x, 12); }
+
+INLINE __m256i rot12_256(__m256i x) { return _mm256_ror_epi32(x, 12); }
+
+INLINE __m512i rot12_512(__m512i x) { return _mm512_ror_epi32(x, 12); }
+
+INLINE __m128i rot8_128(__m128i x) { return _mm_ror_epi32(x, 8); }
+
+INLINE __m256i rot8_256(__m256i x) { return _mm256_ror_epi32(x, 8); }
+
+INLINE __m512i rot8_512(__m512i x) { return _mm512_ror_epi32(x, 8); }
+
+INLINE __m128i rot7_128(__m128i x) { return _mm_ror_epi32(x, 7); }
+
+INLINE __m256i rot7_256(__m256i x) { return _mm256_ror_epi32(x, 7); }
+
+INLINE __m512i rot7_512(__m512i x) { return _mm512_ror_epi32(x, 7); }
+
+/*
+ * ----------------------------------------------------------------------------
+ * compress_avx512
+ * ----------------------------------------------------------------------------
+ */
+
+INLINE void g1(__m128i *row0, __m128i *row1, __m128i *row2, __m128i *row3,
+               __m128i m) {
+  *row0 = add_128(add_128(*row0, m), *row1);
+  *row3 = xor_128(*row3, *row0);
+  *row3 = rot16_128(*row3);
+  *row2 = add_128(*row2, *row3);
+  *row1 = xor_128(*row1, *row2);
+  *row1 = rot12_128(*row1);
+}
+
+INLINE void g2(__m128i *row0, __m128i *row1, __m128i *row2, __m128i *row3,
+               __m128i m) {
+  *row0 = add_128(add_128(*row0, m), *row1);
+  *row3 = xor_128(*row3, *row0);
+  *row3 = rot8_128(*row3);
+  *row2 = add_128(*row2, *row3);
+  *row1 = xor_128(*row1, *row2);
+  *row1 = rot7_128(*row1);
+}
+
+// Note the optimization here of leaving row1 as the unrotated row, rather than
+// row0. All the message loads below are adjusted to compensate for this. See
+// discussion at https://github.com/sneves/blake2-avx2/pull/4
+INLINE void diagonalize(__m128i *row0, __m128i *row2, __m128i *row3) {
+  *row0 = _mm_shuffle_epi32(*row0, _MM_SHUFFLE(2, 1, 0, 3));
+  *row3 = _mm_shuffle_epi32(*row3, _MM_SHUFFLE(1, 0, 3, 2));
+  *row2 = _mm_shuffle_epi32(*row2, _MM_SHUFFLE(0, 3, 2, 1));
+}
+
+INLINE void undiagonalize(__m128i *row0, __m128i *row2, __m128i *row3) {
+  *row0 = _mm_shuffle_epi32(*row0, _MM_SHUFFLE(0, 3, 2, 1));
+  *row3 = _mm_shuffle_epi32(*row3, _MM_SHUFFLE(1, 0, 3, 2));
+  *row2 = _mm_shuffle_epi32(*row2, _MM_SHUFFLE(2, 1, 0, 3));
+}
+
+INLINE void compress_pre(__m128i 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);
+
+  __m128i m0 = loadu_128(&block[sizeof(__m128i) * 0]);
+  __m128i m1 = loadu_128(&block[sizeof(__m128i) * 1]);
+  __m128i m2 = loadu_128(&block[sizeof(__m128i) * 2]);
+  __m128i m3 = loadu_128(&block[sizeof(__m128i) * 3]);
+
+  __m128i 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 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(2, 0, 2, 0)); //  6  4  2  0
+  g1(&rows[0], &rows[1], &rows[2], &rows[3], t0);
+  t1 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(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 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(2, 0, 2, 0)); // 14 12 10  8
+  t2 = _mm_shuffle_epi32(t2, _MM_SHUFFLE(2, 1, 0, 3));   // 12 10  8 14
+  g1(&rows[0], &rows[1], &rows[2], &rows[3], t2);
+  t3 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 1, 3, 1)); // 15 13 11  9
+  t3 = _mm_shuffle_epi32(t3, _MM_SHUFFLE(2, 1, 0, 3));   // 13 11  9 15
+  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 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2));
+  t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1));
+  g1(&rows[0], &rows[1], &rows[2], &rows[3], t0);
+  t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2));
+  tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3));
+  t1 = _mm_blend_epi16(tt, t1, 0xCC);
+  g2(&rows[0], &rows[1], &rows[2], &rows[3], t1);
+  diagonalize(&rows[0], &rows[2], &rows[3]);
+  t2 = _mm_unpacklo_epi64(m3, m1);
+  tt = _mm_blend_epi16(t2, m2, 0xC0);
+  t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0));
+  g1(&rows[0], &rows[1], &rows[2], &rows[3], t2);
+  t3 = _mm_unpackhi_epi32(m1, m3);
+  tt = _mm_unpacklo_epi32(m2, t3);
+  t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(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 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2));
+  t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1));
+  g1(&rows[0], &rows[1], &rows[2], &rows[3], t0);
+  t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2));
+  tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3));
+  t1 = _mm_blend_epi16(tt, t1, 0xCC);
+  g2(&rows[0], &rows[1], &rows[2], &rows[3], t1);
+  diagonalize(&rows[0], &rows[2], &rows[3]);
+  t2 = _mm_unpacklo_epi64(m3, m1);
+  tt = _mm_blend_epi16(t2, m2, 0xC0);
+  t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0));
+  g1(&rows[0], &rows[1], &rows[2], &rows[3], t2);
+  t3 = _mm_unpackhi_epi32(m1, m3);
+  tt = _mm_unpacklo_epi32(m2, t3);
+  t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(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 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2));
+  t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1));
+  g1(&rows[0], &rows[1], &rows[2], &rows[3], t0);
+  t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2));
+  tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3));
+  t1 = _mm_blend_epi16(tt, t1, 0xCC);
+  g2(&rows[0], &rows[1], &rows[2], &rows[3], t1);
+  diagonalize(&rows[0], &rows[2], &rows[3]);
+  t2 = _mm_unpacklo_epi64(m3, m1);
+  tt = _mm_blend_epi16(t2, m2, 0xC0);
+  t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0));
+  g1(&rows[0], &rows[1], &rows[2], &rows[3], t2);
+  t3 = _mm_unpackhi_epi32(m1, m3);
+  tt = _mm_unpacklo_epi32(m2, t3);
+  t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(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 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2));
+  t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1));
+  g1(&rows[0], &rows[1], &rows[2], &rows[3], t0);
+  t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2));
+  tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3));
+  t1 = _mm_blend_epi16(tt, t1, 0xCC);
+  g2(&rows[0], &rows[1], &rows[2], &rows[3], t1);
+  diagonalize(&rows[0], &rows[2], &rows[3]);
+  t2 = _mm_unpacklo_epi64(m3, m1);
+  tt = _mm_blend_epi16(t2, m2, 0xC0);
+  t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0));
+  g1(&rows[0], &rows[1], &rows[2], &rows[3], t2);
+  t3 = _mm_unpackhi_epi32(m1, m3);
+  tt = _mm_unpacklo_epi32(m2, t3);
+  t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(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 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2));
+  t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1));
+  g1(&rows[0], &rows[1], &rows[2], &rows[3], t0);
+  t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2));
+  tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3));
+  t1 = _mm_blend_epi16(tt, t1, 0xCC);
+  g2(&rows[0], &rows[1], &rows[2], &rows[3], t1);
+  diagonalize(&rows[0], &rows[2], &rows[3]);
+  t2 = _mm_unpacklo_epi64(m3, m1);
+  tt = _mm_blend_epi16(t2, m2, 0xC0);
+  t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0));
+  g1(&rows[0], &rows[1], &rows[2], &rows[3], t2);
+  t3 = _mm_unpackhi_epi32(m1, m3);
+  tt = _mm_unpacklo_epi32(m2, t3);
+  t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(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 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2));
+  t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1));
+  g1(&rows[0], &rows[1], &rows[2], &rows[3], t0);
+  t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2));
+  tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3));
+  t1 = _mm_blend_epi16(tt, t1, 0xCC);
+  g2(&rows[0], &rows[1], &rows[2], &rows[3], t1);
+  diagonalize(&rows[0], &rows[2], &rows[3]);
+  t2 = _mm_unpacklo_epi64(m3, m1);
+  tt = _mm_blend_epi16(t2, m2, 0xC0);
+  t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0));
+  g1(&rows[0], &rows[1], &rows[2], &rows[3], t2);
+  t3 = _mm_unpackhi_epi32(m1, m3);
+  tt = _mm_unpacklo_epi32(m2, t3);
+  t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(0, 1, 3, 2));
+  g2(&rows[0], &rows[1], &rows[2], &rows[3], t3);
+  undiagonalize(&rows[0], &rows[2], &rows[3]);
+}
+
+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]) {
+  __m128i rows[4];
+  compress_pre(rows, cv, block, block_len, counter, flags);
+  storeu_128(xor_128(rows[0], rows[2]), &out[0]);
+  storeu_128(xor_128(rows[1], rows[3]), &out[16]);
+  storeu_128(xor_128(rows[2], loadu_128((uint8_t *)&cv[0])), &out[32]);
+  storeu_128(xor_128(rows[3], loadu_128((uint8_t *)&cv[4])), &out[48]);
+}
+
+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) {
+  __m128i rows[4];
+  compress_pre(rows, cv, block, block_len, counter, flags);
+  storeu_128(xor_128(rows[0], rows[2]), (uint8_t *)&cv[0]);
+  storeu_128(xor_128(rows[1], rows[3]), (uint8_t *)&cv[4]);
+}
+
+/*
+ * ----------------------------------------------------------------------------
+ * hash4_avx512
+ * ----------------------------------------------------------------------------
+ */
+
+INLINE void round_fn4(__m128i v[16], __m128i 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(__m128i vecs[4]) {
+  // Interleave 32-bit lanes. The low unpack is lanes 00/11 and the high is
+  // 22/33. Note that this doesn't split the vector into two lanes, as the
+  // AVX2 counterparts do.
+  __m128i ab_01 = _mm_unpacklo_epi32(vecs[0], vecs[1]);
+  __m128i ab_23 = _mm_unpackhi_epi32(vecs[0], vecs[1]);
+  __m128i cd_01 = _mm_unpacklo_epi32(vecs[2], vecs[3]);
+  __m128i cd_23 = _mm_unpackhi_epi32(vecs[2], vecs[3]);
+
+  // Interleave 64-bit lanes.
+  __m128i abcd_0 = _mm_unpacklo_epi64(ab_01, cd_01);
+  __m128i abcd_1 = _mm_unpackhi_epi64(ab_01, cd_01);
+  __m128i abcd_2 = _mm_unpacklo_epi64(ab_23, cd_23);
+  __m128i abcd_3 = _mm_unpackhi_epi64(ab_23, cd_23);
+
+  vecs[0] = abcd_0;
+  vecs[1] = abcd_1;
+  vecs[2] = abcd_2;
+  vecs[3] = abcd_3;
+}
+
+INLINE void transpose_msg_vecs4(const uint8_t *const *inputs,
+                                size_t block_offset, __m128i out[16]) {
+  out[0] = loadu_128(&inputs[0][block_offset + 0 * sizeof(__m128i)]);
+  out[1] = loadu_128(&inputs[1][block_offset + 0 * sizeof(__m128i)]);
+  out[2] = loadu_128(&inputs[2][block_offset + 0 * sizeof(__m128i)]);
+  out[3] = loadu_128(&inputs[3][block_offset + 0 * sizeof(__m128i)]);
+  out[4] = loadu_128(&inputs[0][block_offset + 1 * sizeof(__m128i)]);
+  out[5] = loadu_128(&inputs[1][block_offset + 1 * sizeof(__m128i)]);
+  out[6] = loadu_128(&inputs[2][block_offset + 1 * sizeof(__m128i)]);
+  out[7] = loadu_128(&inputs[3][block_offset + 1 * sizeof(__m128i)]);
+  out[8] = loadu_128(&inputs[0][block_offset + 2 * sizeof(__m128i)]);
+  out[9] = loadu_128(&inputs[1][block_offset + 2 * sizeof(__m128i)]);
+  out[10] = loadu_128(&inputs[2][block_offset + 2 * sizeof(__m128i)]);
+  out[11] = loadu_128(&inputs[3][block_offset + 2 * sizeof(__m128i)]);
+  out[12] = loadu_128(&inputs[0][block_offset + 3 * sizeof(__m128i)]);
+  out[13] = loadu_128(&inputs[1][block_offset + 3 * sizeof(__m128i)]);
+  out[14] = loadu_128(&inputs[2][block_offset + 3 * sizeof(__m128i)]);
+  out[15] = loadu_128(&inputs[3][block_offset + 3 * sizeof(__m128i)]);
+  for (size_t i = 0; i < 4; ++i) {
+    _mm_prefetch((const void *)&inputs[i][block_offset + 256], _MM_HINT_T0);
+  }
+  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,
+                           __m128i *out_lo, __m128i *out_hi) {
+  uint64_t mask = (increment_counter ? ~0 : 0);
+  __m256i mask_vec = _mm256_set1_epi64x(mask);
+  __m256i deltas = _mm256_setr_epi64x(0, 1, 2, 3);
+  deltas = _mm256_and_si256(mask_vec, deltas);
+  __m256i counters =
+      _mm256_add_epi64(_mm256_set1_epi64x((int64_t)counter), deltas);
+  *out_lo = _mm256_cvtepi64_epi32(counters);
+  *out_hi = _mm256_cvtepi64_epi32(_mm256_srli_epi64(counters, 32));
+}
+
+static
+void blake3_hash4_avx512(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) {
+  __m128i 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]),
+  };
+  __m128i 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;
+    }
+    __m128i block_len_vec = set1_128(BLAKE3_BLOCK_LEN);
+    __m128i block_flags_vec = set1_128(block_flags);
+    __m128i msg_vecs[16];
+    transpose_msg_vecs4(inputs, block * BLAKE3_BLOCK_LEN, msg_vecs);
+
+    __m128i 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(__m128i)]);
+  storeu_128(h_vecs[4], &out[1 * sizeof(__m128i)]);
+  storeu_128(h_vecs[1], &out[2 * sizeof(__m128i)]);
+  storeu_128(h_vecs[5], &out[3 * sizeof(__m128i)]);
+  storeu_128(h_vecs[2], &out[4 * sizeof(__m128i)]);
+  storeu_128(h_vecs[6], &out[5 * sizeof(__m128i)]);
+  storeu_128(h_vecs[3], &out[6 * sizeof(__m128i)]);
+  storeu_128(h_vecs[7], &out[7 * sizeof(__m128i)]);
+}
+
+/*
+ * ----------------------------------------------------------------------------
+ * hash8_avx512
+ * ----------------------------------------------------------------------------
+ */
+
+INLINE void round_fn8(__m256i v[16], __m256i m[16], size_t r) {
+  v[0] = add_256(v[0], m[(size_t)MSG_SCHEDULE[r][0]]);
+  v[1] = add_256(v[1], m[(size_t)MSG_SCHEDULE[r][2]]);
+  v[2] = add_256(v[2], m[(size_t)MSG_SCHEDULE[r][4]]);
+  v[3] = add_256(v[3], m[(size_t)MSG_SCHEDULE[r][6]]);
+  v[0] = add_256(v[0], v[4]);
+  v[1] = add_256(v[1], v[5]);
+  v[2] = add_256(v[2], v[6]);
+  v[3] = add_256(v[3], v[7]);
+  v[12] = xor_256(v[12], v[0]);
+  v[13] = xor_256(v[13], v[1]);
+  v[14] = xor_256(v[14], v[2]);
+  v[15] = xor_256(v[15], v[3]);
+  v[12] = rot16_256(v[12]);
+  v[13] = rot16_256(v[13]);
+  v[14] = rot16_256(v[14]);
+  v[15] = rot16_256(v[15]);
+  v[8] = add_256(v[8], v[12]);
+  v[9] = add_256(v[9], v[13]);
+  v[10] = add_256(v[10], v[14]);
+  v[11] = add_256(v[11], v[15]);
+  v[4] = xor_256(v[4], v[8]);
+  v[5] = xor_256(v[5], v[9]);
+  v[6] = xor_256(v[6], v[10]);
+  v[7] = xor_256(v[7], v[11]);
+  v[4] = rot12_256(v[4]);
+  v[5] = rot12_256(v[5]);
+  v[6] = rot12_256(v[6]);
+  v[7] = rot12_256(v[7]);
+  v[0] = add_256(v[0], m[(size_t)MSG_SCHEDULE[r][1]]);
+  v[1] = add_256(v[1], m[(size_t)MSG_SCHEDULE[r][3]]);
+  v[2] = add_256(v[2], m[(size_t)MSG_SCHEDULE[r][5]]);
+  v[3] = add_256(v[3], m[(size_t)MSG_SCHEDULE[r][7]]);
+  v[0] = add_256(v[0], v[4]);
+  v[1] = add_256(v[1], v[5]);
+  v[2] = add_256(v[2], v[6]);
+  v[3] = add_256(v[3], v[7]);
+  v[12] = xor_256(v[12], v[0]);
+  v[13] = xor_256(v[13], v[1]);
+  v[14] = xor_256(v[14], v[2]);
+  v[15] = xor_256(v[15], v[3]);
+  v[12] = rot8_256(v[12]);
+  v[13] = rot8_256(v[13]);
+  v[14] = rot8_256(v[14]);
+  v[15] = rot8_256(v[15]);
+  v[8] = add_256(v[8], v[12]);
+  v[9] = add_256(v[9], v[13]);
+  v[10] = add_256(v[10], v[14]);
+  v[11] = add_256(v[11], v[15]);
+  v[4] = xor_256(v[4], v[8]);
+  v[5] = xor_256(v[5], v[9]);
+  v[6] = xor_256(v[6], v[10]);
+  v[7] = xor_256(v[7], v[11]);
+  v[4] = rot7_256(v[4]);
+  v[5] = rot7_256(v[5]);
+  v[6] = rot7_256(v[6]);
+  v[7] = rot7_256(v[7]);
+
+  v[0] = add_256(v[0], m[(size_t)MSG_SCHEDULE[r][8]]);
+  v[1] = add_256(v[1], m[(size_t)MSG_SCHEDULE[r][10]]);
+  v[2] = add_256(v[2], m[(size_t)MSG_SCHEDULE[r][12]]);
+  v[3] = add_256(v[3], m[(size_t)MSG_SCHEDULE[r][14]]);
+  v[0] = add_256(v[0], v[5]);
+  v[1] = add_256(v[1], v[6]);
+  v[2] = add_256(v[2], v[7]);
+  v[3] = add_256(v[3], v[4]);
+  v[15] = xor_256(v[15], v[0]);
+  v[12] = xor_256(v[12], v[1]);
+  v[13] = xor_256(v[13], v[2]);
+  v[14] = xor_256(v[14], v[3]);
+  v[15] = rot16_256(v[15]);
+  v[12] = rot16_256(v[12]);
+  v[13] = rot16_256(v[13]);
+  v[14] = rot16_256(v[14]);
+  v[10] = add_256(v[10], v[15]);
+  v[11] = add_256(v[11], v[12]);
+  v[8] = add_256(v[8], v[13]);
+  v[9] = add_256(v[9], v[14]);
+  v[5] = xor_256(v[5], v[10]);
+  v[6] = xor_256(v[6], v[11]);
+  v[7] = xor_256(v[7], v[8]);
+  v[4] = xor_256(v[4], v[9]);
+  v[5] = rot12_256(v[5]);
+  v[6] = rot12_256(v[6]);
+  v[7] = rot12_256(v[7]);
+  v[4] = rot12_256(v[4]);
+  v[0] = add_256(v[0], m[(size_t)MSG_SCHEDULE[r][9]]);
+  v[1] = add_256(v[1], m[(size_t)MSG_SCHEDULE[r][11]]);
+  v[2] = add_256(v[2], m[(size_t)MSG_SCHEDULE[r][13]]);
+  v[3] = add_256(v[3], m[(size_t)MSG_SCHEDULE[r][15]]);
+  v[0] = add_256(v[0], v[5]);
+  v[1] = add_256(v[1], v[6]);
+  v[2] = add_256(v[2], v[7]);
+  v[3] = add_256(v[3], v[4]);
+  v[15] = xor_256(v[15], v[0]);
+  v[12] = xor_256(v[12], v[1]);
+  v[13] = xor_256(v[13], v[2]);
+  v[14] = xor_256(v[14], v[3]);
+  v[15] = rot8_256(v[15]);
+  v[12] = rot8_256(v[12]);
+  v[13] = rot8_256(v[13]);
+  v[14] = rot8_256(v[14]);
+  v[10] = add_256(v[10], v[15]);
+  v[11] = add_256(v[11], v[12]);
+  v[8] = add_256(v[8], v[13]);
+  v[9] = add_256(v[9], v[14]);
+  v[5] = xor_256(v[5], v[10]);
+  v[6] = xor_256(v[6], v[11]);
+  v[7] = xor_256(v[7], v[8]);
+  v[4] = xor_256(v[4], v[9]);
+  v[5] = rot7_256(v[5]);
+  v[6] = rot7_256(v[6]);
+  v[7] = rot7_256(v[7]);
+  v[4] = rot7_256(v[4]);
+}
+
+INLINE void transpose_vecs_256(__m256i vecs[8]) {
+  // Interleave 32-bit lanes. The low unpack is lanes 00/11/44/55, and the high
+  // is 22/33/66/77.
+  __m256i ab_0145 = _mm256_unpacklo_epi32(vecs[0], vecs[1]);
+  __m256i ab_2367 = _mm256_unpackhi_epi32(vecs[0], vecs[1]);
+  __m256i cd_0145 = _mm256_unpacklo_epi32(vecs[2], vecs[3]);
+  __m256i cd_2367 = _mm256_unpackhi_epi32(vecs[2], vecs[3]);
+  __m256i ef_0145 = _mm256_unpacklo_epi32(vecs[4], vecs[5]);
+  __m256i ef_2367 = _mm256_unpackhi_epi32(vecs[4], vecs[5]);
+  __m256i gh_0145 = _mm256_unpacklo_epi32(vecs[6], vecs[7]);
+  __m256i gh_2367 = _mm256_unpackhi_epi32(vecs[6], vecs[7]);
+
+  // Interleave 64-bit lanes. The low unpack is lanes 00/22 and the high is
+  // 11/33.
+  __m256i abcd_04 = _mm256_unpacklo_epi64(ab_0145, cd_0145);
+  __m256i abcd_15 = _mm256_unpackhi_epi64(ab_0145, cd_0145);
+  __m256i abcd_26 = _mm256_unpacklo_epi64(ab_2367, cd_2367);
+  __m256i abcd_37 = _mm256_unpackhi_epi64(ab_2367, cd_2367);
+  __m256i efgh_04 = _mm256_unpacklo_epi64(ef_0145, gh_0145);
+  __m256i efgh_15 = _mm256_unpackhi_epi64(ef_0145, gh_0145);
+  __m256i efgh_26 = _mm256_unpacklo_epi64(ef_2367, gh_2367);
+  __m256i efgh_37 = _mm256_unpackhi_epi64(ef_2367, gh_2367);
+
+  // Interleave 128-bit lanes.
+  vecs[0] = _mm256_permute2x128_si256(abcd_04, efgh_04, 0x20);
+  vecs[1] = _mm256_permute2x128_si256(abcd_15, efgh_15, 0x20);
+  vecs[2] = _mm256_permute2x128_si256(abcd_26, efgh_26, 0x20);
+  vecs[3] = _mm256_permute2x128_si256(abcd_37, efgh_37, 0x20);
+  vecs[4] = _mm256_permute2x128_si256(abcd_04, efgh_04, 0x31);
+  vecs[5] = _mm256_permute2x128_si256(abcd_15, efgh_15, 0x31);
+  vecs[6] = _mm256_permute2x128_si256(abcd_26, efgh_26, 0x31);
+  vecs[7] = _mm256_permute2x128_si256(abcd_37, efgh_37, 0x31);
+}
+
+INLINE void transpose_msg_vecs8(const uint8_t *const *inputs,
+                                size_t block_offset, __m256i out[16]) {
+  out[0] = loadu_256(&inputs[0][block_offset + 0 * sizeof(__m256i)]);
+  out[1] = loadu_256(&inputs[1][block_offset + 0 * sizeof(__m256i)]);
+  out[2] = loadu_256(&inputs[2][block_offset + 0 * sizeof(__m256i)]);
+  out[3] = loadu_256(&inputs[3][block_offset + 0 * sizeof(__m256i)]);
+  out[4] = loadu_256(&inputs[4][block_offset + 0 * sizeof(__m256i)]);
+  out[5] = loadu_256(&inputs[5][block_offset + 0 * sizeof(__m256i)]);
+  out[6] = loadu_256(&inputs[6][block_offset + 0 * sizeof(__m256i)]);
+  out[7] = loadu_256(&inputs[7][block_offset + 0 * sizeof(__m256i)]);
+  out[8] = loadu_256(&inputs[0][block_offset + 1 * sizeof(__m256i)]);
+  out[9] = loadu_256(&inputs[1][block_offset + 1 * sizeof(__m256i)]);
+  out[10] = loadu_256(&inputs[2][block_offset + 1 * sizeof(__m256i)]);
+  out[11] = loadu_256(&inputs[3][block_offset + 1 * sizeof(__m256i)]);
+  out[12] = loadu_256(&inputs[4][block_offset + 1 * sizeof(__m256i)]);
+  out[13] = loadu_256(&inputs[5][block_offset + 1 * sizeof(__m256i)]);
+  out[14] = loadu_256(&inputs[6][block_offset + 1 * sizeof(__m256i)]);
+  out[15] = loadu_256(&inputs[7][block_offset + 1 * sizeof(__m256i)]);
+  for (size_t i = 0; i < 8; ++i) {
+    _mm_prefetch((const void *)&inputs[i][block_offset + 256], _MM_HINT_T0);
+  }
+  transpose_vecs_256(&out[0]);
+  transpose_vecs_256(&out[8]);
+}
+
+INLINE void load_counters8(uint64_t counter, bool increment_counter,
+                           __m256i *out_lo, __m256i *out_hi) {
+  uint64_t mask = (increment_counter ? ~0 : 0);
+  __m512i mask_vec = _mm512_set1_epi64(mask);
+  __m512i deltas = _mm512_setr_epi64(0, 1, 2, 3, 4, 5, 6, 7);
+  deltas = _mm512_and_si512(mask_vec, deltas);
+  __m512i counters =
+      _mm512_add_epi64(_mm512_set1_epi64((int64_t)counter), deltas);
+  *out_lo = _mm512_cvtepi64_epi32(counters);
+  *out_hi = _mm512_cvtepi64_epi32(_mm512_srli_epi64(counters, 32));
+}
+
+static
+void blake3_hash8_avx512(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) {
+  __m256i h_vecs[8] = {
+      set1_256(key[0]), set1_256(key[1]), set1_256(key[2]), set1_256(key[3]),
+      set1_256(key[4]), set1_256(key[5]), set1_256(key[6]), set1_256(key[7]),
+  };
+  __m256i counter_low_vec, counter_high_vec;
+  load_counters8(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;
+    }
+    __m256i block_len_vec = set1_256(BLAKE3_BLOCK_LEN);
+    __m256i block_flags_vec = set1_256(block_flags);
+    __m256i msg_vecs[16];
+    transpose_msg_vecs8(inputs, block * BLAKE3_BLOCK_LEN, msg_vecs);
+
+    __m256i 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_256(IV[0]), set1_256(IV[1]),  set1_256(IV[2]), set1_256(IV[3]),
+        counter_low_vec, counter_high_vec, block_len_vec,   block_flags_vec,
+    };
+    round_fn8(v, msg_vecs, 0);
+    round_fn8(v, msg_vecs, 1);
+    round_fn8(v, msg_vecs, 2);
+    round_fn8(v, msg_vecs, 3);
+    round_fn8(v, msg_vecs, 4);
+    round_fn8(v, msg_vecs, 5);
+    round_fn8(v, msg_vecs, 6);
+    h_vecs[0] = xor_256(v[0], v[8]);
+    h_vecs[1] = xor_256(v[1], v[9]);
+    h_vecs[2] = xor_256(v[2], v[10]);
+    h_vecs[3] = xor_256(v[3], v[11]);
+    h_vecs[4] = xor_256(v[4], v[12]);
+    h_vecs[5] = xor_256(v[5], v[13]);
+    h_vecs[6] = xor_256(v[6], v[14]);
+    h_vecs[7] = xor_256(v[7], v[15]);
+
+    block_flags = flags;
+  }
+
+  transpose_vecs_256(h_vecs);
+  storeu_256(h_vecs[0], &out[0 * sizeof(__m256i)]);
+  storeu_256(h_vecs[1], &out[1 * sizeof(__m256i)]);
+  storeu_256(h_vecs[2], &out[2 * sizeof(__m256i)]);
+  storeu_256(h_vecs[3], &out[3 * sizeof(__m256i)]);
+  storeu_256(h_vecs[4], &out[4 * sizeof(__m256i)]);
+  storeu_256(h_vecs[5], &out[5 * sizeof(__m256i)]);
+  storeu_256(h_vecs[6], &out[6 * sizeof(__m256i)]);
+  storeu_256(h_vecs[7], &out[7 * sizeof(__m256i)]);
+}
+
+/*
+ * ----------------------------------------------------------------------------
+ * hash16_avx512
+ * ----------------------------------------------------------------------------
+ */
+
+INLINE void round_fn16(__m512i v[16], __m512i m[16], size_t r) {
+  v[0] = add_512(v[0], m[(size_t)MSG_SCHEDULE[r][0]]);
+  v[1] = add_512(v[1], m[(size_t)MSG_SCHEDULE[r][2]]);
+  v[2] = add_512(v[2], m[(size_t)MSG_SCHEDULE[r][4]]);
+  v[3] = add_512(v[3], m[(size_t)MSG_SCHEDULE[r][6]]);
+  v[0] = add_512(v[0], v[4]);
+  v[1] = add_512(v[1], v[5]);
+  v[2] = add_512(v[2], v[6]);
+  v[3] = add_512(v[3], v[7]);
+  v[12] = xor_512(v[12], v[0]);
+  v[13] = xor_512(v[13], v[1]);
+  v[14] = xor_512(v[14], v[2]);
+  v[15] = xor_512(v[15], v[3]);
+  v[12] = rot16_512(v[12]);
+  v[13] = rot16_512(v[13]);
+  v[14] = rot16_512(v[14]);
+  v[15] = rot16_512(v[15]);
+  v[8] = add_512(v[8], v[12]);
+  v[9] = add_512(v[9], v[13]);
+  v[10] = add_512(v[10], v[14]);
+  v[11] = add_512(v[11], v[15]);
+  v[4] = xor_512(v[4], v[8]);
+  v[5] = xor_512(v[5], v[9]);
+  v[6] = xor_512(v[6], v[10]);
+  v[7] = xor_512(v[7], v[11]);
+  v[4] = rot12_512(v[4]);
+  v[5] = rot12_512(v[5]);
+  v[6] = rot12_512(v[6]);
+  v[7] = rot12_512(v[7]);
+  v[0] = add_512(v[0], m[(size_t)MSG_SCHEDULE[r][1]]);
+  v[1] = add_512(v[1], m[(size_t)MSG_SCHEDULE[r][3]]);
+  v[2] = add_512(v[2], m[(size_t)MSG_SCHEDULE[r][5]]);
+  v[3] = add_512(v[3], m[(size_t)MSG_SCHEDULE[r][7]]);
+  v[0] = add_512(v[0], v[4]);
+  v[1] = add_512(v[1], v[5]);
+  v[2] = add_512(v[2], v[6]);
+  v[3] = add_512(v[3], v[7]);
+  v[12] = xor_512(v[12], v[0]);
+  v[13] = xor_512(v[13], v[1]);
+  v[14] = xor_512(v[14], v[2]);
+  v[15] = xor_512(v[15], v[3]);
+  v[12] = rot8_512(v[12]);
+  v[13] = rot8_512(v[13]);
+  v[14] = rot8_512(v[14]);
+  v[15] = rot8_512(v[15]);
+  v[8] = add_512(v[8], v[12]);
+  v[9] = add_512(v[9], v[13]);
+  v[10] = add_512(v[10], v[14]);
+  v[11] = add_512(v[11], v[15]);
+  v[4] = xor_512(v[4], v[8]);
+  v[5] = xor_512(v[5], v[9]);
+  v[6] = xor_512(v[6], v[10]);
+  v[7] = xor_512(v[7], v[11]);
+  v[4] = rot7_512(v[4]);
+  v[5] = rot7_512(v[5]);
+  v[6] = rot7_512(v[6]);
+  v[7] = rot7_512(v[7]);
+
+  v[0] = add_512(v[0], m[(size_t)MSG_SCHEDULE[r][8]]);
+  v[1] = add_512(v[1], m[(size_t)MSG_SCHEDULE[r][10]]);
+  v[2] = add_512(v[2], m[(size_t)MSG_SCHEDULE[r][12]]);
+  v[3] = add_512(v[3], m[(size_t)MSG_SCHEDULE[r][14]]);
+  v[0] = add_512(v[0], v[5]);
+  v[1] = add_512(v[1], v[6]);
+  v[2] = add_512(v[2], v[7]);
+  v[3] = add_512(v[3], v[4]);
+  v[15] = xor_512(v[15], v[0]);
+  v[12] = xor_512(v[12], v[1]);
+  v[13] = xor_512(v[13], v[2]);
+  v[14] = xor_512(v[14], v[3]);
+  v[15] = rot16_512(v[15]);
+  v[12] = rot16_512(v[12]);
+  v[13] = rot16_512(v[13]);
+  v[14] = rot16_512(v[14]);
+  v[10] = add_512(v[10], v[15]);
+  v[11] = add_512(v[11], v[12]);
+  v[8] = add_512(v[8], v[13]);
+  v[9] = add_512(v[9], v[14]);
+  v[5] = xor_512(v[5], v[10]);
+  v[6] = xor_512(v[6], v[11]);
+  v[7] = xor_512(v[7], v[8]);
+  v[4] = xor_512(v[4], v[9]);
+  v[5] = rot12_512(v[5]);
+  v[6] = rot12_512(v[6]);
+  v[7] = rot12_512(v[7]);
+  v[4] = rot12_512(v[4]);
+  v[0] = add_512(v[0], m[(size_t)MSG_SCHEDULE[r][9]]);
+  v[1] = add_512(v[1], m[(size_t)MSG_SCHEDULE[r][11]]);
+  v[2] = add_512(v[2], m[(size_t)MSG_SCHEDULE[r][13]]);
+  v[3] = add_512(v[3], m[(size_t)MSG_SCHEDULE[r][15]]);
+  v[0] = add_512(v[0], v[5]);
+  v[1] = add_512(v[1], v[6]);
+  v[2] = add_512(v[2], v[7]);
+  v[3] = add_512(v[3], v[4]);
+  v[15] = xor_512(v[15], v[0]);
+  v[12] = xor_512(v[12], v[1]);
+  v[13] = xor_512(v[13], v[2]);
+  v[14] = xor_512(v[14], v[3]);
+  v[15] = rot8_512(v[15]);
+  v[12] = rot8_512(v[12]);
+  v[13] = rot8_512(v[13]);
+  v[14] = rot8_512(v[14]);
+  v[10] = add_512(v[10], v[15]);
+  v[11] = add_512(v[11], v[12]);
+  v[8] = add_512(v[8], v[13]);
+  v[9] = add_512(v[9], v[14]);
+  v[5] = xor_512(v[5], v[10]);
+  v[6] = xor_512(v[6], v[11]);
+  v[7] = xor_512(v[7], v[8]);
+  v[4] = xor_512(v[4], v[9]);
+  v[5] = rot7_512(v[5]);
+  v[6] = rot7_512(v[6]);
+  v[7] = rot7_512(v[7]);
+  v[4] = rot7_512(v[4]);
+}
+
+// 0b10001000, or lanes a0/a2/b0/b2 in little-endian order
+#define LO_IMM8 0x88
+
+INLINE __m512i unpack_lo_128(__m512i a, __m512i b) {
+  return _mm512_shuffle_i32x4(a, b, LO_IMM8);
+}
+
+// 0b11011101, or lanes a1/a3/b1/b3 in little-endian order
+#define HI_IMM8 0xdd
+
+INLINE __m512i unpack_hi_128(__m512i a, __m512i b) {
+  return _mm512_shuffle_i32x4(a, b, HI_IMM8);
+}
+
+INLINE void transpose_vecs_512(__m512i vecs[16]) {
+  // Interleave 32-bit lanes. The _0 unpack is lanes
+  // 0/0/1/1/4/4/5/5/8/8/9/9/12/12/13/13, and the _2 unpack is lanes
+  // 2/2/3/3/6/6/7/7/10/10/11/11/14/14/15/15.
+  __m512i ab_0 = _mm512_unpacklo_epi32(vecs[0], vecs[1]);
+  __m512i ab_2 = _mm512_unpackhi_epi32(vecs[0], vecs[1]);
+  __m512i cd_0 = _mm512_unpacklo_epi32(vecs[2], vecs[3]);
+  __m512i cd_2 = _mm512_unpackhi_epi32(vecs[2], vecs[3]);
+  __m512i ef_0 = _mm512_unpacklo_epi32(vecs[4], vecs[5]);
+  __m512i ef_2 = _mm512_unpackhi_epi32(vecs[4], vecs[5]);
+  __m512i gh_0 = _mm512_unpacklo_epi32(vecs[6], vecs[7]);
+  __m512i gh_2 = _mm512_unpackhi_epi32(vecs[6], vecs[7]);
+  __m512i ij_0 = _mm512_unpacklo_epi32(vecs[8], vecs[9]);
+  __m512i ij_2 = _mm512_unpackhi_epi32(vecs[8], vecs[9]);
+  __m512i kl_0 = _mm512_unpacklo_epi32(vecs[10], vecs[11]);
+  __m512i kl_2 = _mm512_unpackhi_epi32(vecs[10], vecs[11]);
+  __m512i mn_0 = _mm512_unpacklo_epi32(vecs[12], vecs[13]);
+  __m512i mn_2 = _mm512_unpackhi_epi32(vecs[12], vecs[13]);
+  __m512i op_0 = _mm512_unpacklo_epi32(vecs[14], vecs[15]);
+  __m512i op_2 = _mm512_unpackhi_epi32(vecs[14], vecs[15]);
+
+  // Interleave 64-bit lanes. The _0 unpack is lanes
+  // 0/0/0/0/4/4/4/4/8/8/8/8/12/12/12/12, the _1 unpack is lanes
+  // 1/1/1/1/5/5/5/5/9/9/9/9/13/13/13/13, the _2 unpack is lanes
+  // 2/2/2/2/6/6/6/6/10/10/10/10/14/14/14/14, and the _3 unpack is lanes
+  // 3/3/3/3/7/7/7/7/11/11/11/11/15/15/15/15.
+  __m512i abcd_0 = _mm512_unpacklo_epi64(ab_0, cd_0);
+  __m512i abcd_1 = _mm512_unpackhi_epi64(ab_0, cd_0);
+  __m512i abcd_2 = _mm512_unpacklo_epi64(ab_2, cd_2);
+  __m512i abcd_3 = _mm512_unpackhi_epi64(ab_2, cd_2);
+  __m512i efgh_0 = _mm512_unpacklo_epi64(ef_0, gh_0);
+  __m512i efgh_1 = _mm512_unpackhi_epi64(ef_0, gh_0);
+  __m512i efgh_2 = _mm512_unpacklo_epi64(ef_2, gh_2);
+  __m512i efgh_3 = _mm512_unpackhi_epi64(ef_2, gh_2);
+  __m512i ijkl_0 = _mm512_unpacklo_epi64(ij_0, kl_0);
+  __m512i ijkl_1 = _mm512_unpackhi_epi64(ij_0, kl_0);
+  __m512i ijkl_2 = _mm512_unpacklo_epi64(ij_2, kl_2);
+  __m512i ijkl_3 = _mm512_unpackhi_epi64(ij_2, kl_2);
+  __m512i mnop_0 = _mm512_unpacklo_epi64(mn_0, op_0);
+  __m512i mnop_1 = _mm512_unpackhi_epi64(mn_0, op_0);
+  __m512i mnop_2 = _mm512_unpacklo_epi64(mn_2, op_2);
+  __m512i mnop_3 = _mm512_unpackhi_epi64(mn_2, op_2);
+
+  // Interleave 128-bit lanes. The _0 unpack is
+  // 0/0/0/0/8/8/8/8/0/0/0/0/8/8/8/8, the _1 unpack is
+  // 1/1/1/1/9/9/9/9/1/1/1/1/9/9/9/9, and so on.
+  __m512i abcdefgh_0 = unpack_lo_128(abcd_0, efgh_0);
+  __m512i abcdefgh_1 = unpack_lo_128(abcd_1, efgh_1);
+  __m512i abcdefgh_2 = unpack_lo_128(abcd_2, efgh_2);
+  __m512i abcdefgh_3 = unpack_lo_128(abcd_3, efgh_3);
+  __m512i abcdefgh_4 = unpack_hi_128(abcd_0, efgh_0);
+  __m512i abcdefgh_5 = unpack_hi_128(abcd_1, efgh_1);
+  __m512i abcdefgh_6 = unpack_hi_128(abcd_2, efgh_2);
+  __m512i abcdefgh_7 = unpack_hi_128(abcd_3, efgh_3);
+  __m512i ijklmnop_0 = unpack_lo_128(ijkl_0, mnop_0);
+  __m512i ijklmnop_1 = unpack_lo_128(ijkl_1, mnop_1);
+  __m512i ijklmnop_2 = unpack_lo_128(ijkl_2, mnop_2);
+  __m512i ijklmnop_3 = unpack_lo_128(ijkl_3, mnop_3);
+  __m512i ijklmnop_4 = unpack_hi_128(ijkl_0, mnop_0);
+  __m512i ijklmnop_5 = unpack_hi_128(ijkl_1, mnop_1);
+  __m512i ijklmnop_6 = unpack_hi_128(ijkl_2, mnop_2);
+  __m512i ijklmnop_7 = unpack_hi_128(ijkl_3, mnop_3);
+
+  // Interleave 128-bit lanes again for the final outputs.
+  vecs[0] = unpack_lo_128(abcdefgh_0, ijklmnop_0);
+  vecs[1] = unpack_lo_128(abcdefgh_1, ijklmnop_1);
+  vecs[2] = unpack_lo_128(abcdefgh_2, ijklmnop_2);
+  vecs[3] = unpack_lo_128(abcdefgh_3, ijklmnop_3);
+  vecs[4] = unpack_lo_128(abcdefgh_4, ijklmnop_4);
+  vecs[5] = unpack_lo_128(abcdefgh_5, ijklmnop_5);
+  vecs[6] = unpack_lo_128(abcdefgh_6, ijklmnop_6);
+  vecs[7] = unpack_lo_128(abcdefgh_7, ijklmnop_7);
+  vecs[8] = unpack_hi_128(abcdefgh_0, ijklmnop_0);
+  vecs[9] = unpack_hi_128(abcdefgh_1, ijklmnop_1);
+  vecs[10] = unpack_hi_128(abcdefgh_2, ijklmnop_2);
+  vecs[11] = unpack_hi_128(abcdefgh_3, ijklmnop_3);
+  vecs[12] = unpack_hi_128(abcdefgh_4, ijklmnop_4);
+  vecs[13] = unpack_hi_128(abcdefgh_5, ijklmnop_5);
+  vecs[14] = unpack_hi_128(abcdefgh_6, ijklmnop_6);
+  vecs[15] = unpack_hi_128(abcdefgh_7, ijklmnop_7);
+}
+
+INLINE void transpose_msg_vecs16(const uint8_t *const *inputs,
+                                 size_t block_offset, __m512i out[16]) {
+  out[0] = loadu_512(&inputs[0][block_offset]);
+  out[1] = loadu_512(&inputs[1][block_offset]);
+  out[2] = loadu_512(&inputs[2][block_offset]);
+  out[3] = loadu_512(&inputs[3][block_offset]);
+  out[4] = loadu_512(&inputs[4][block_offset]);
+  out[5] = loadu_512(&inputs[5][block_offset]);
+  out[6] = loadu_512(&inputs[6][block_offset]);
+  out[7] = loadu_512(&inputs[7][block_offset]);
+  out[8] = loadu_512(&inputs[8][block_offset]);
+  out[9] = loadu_512(&inputs[9][block_offset]);
+  out[10] = loadu_512(&inputs[10][block_offset]);
+  out[11] = loadu_512(&inputs[11][block_offset]);
+  out[12] = loadu_512(&inputs[12][block_offset]);
+  out[13] = loadu_512(&inputs[13][block_offset]);
+  out[14] = loadu_512(&inputs[14][block_offset]);
+  out[15] = loadu_512(&inputs[15][block_offset]);
+  for (size_t i = 0; i < 16; ++i) {
+    _mm_prefetch((const void *)&inputs[i][block_offset + 256], _MM_HINT_T0);
+  }
+  transpose_vecs_512(out);
+}
+
+INLINE void load_counters16(uint64_t counter, bool increment_counter,
+                            __m512i *out_lo, __m512i *out_hi) {
+  const __m512i mask = _mm512_set1_epi32(-(int32_t)increment_counter);
+  const __m512i deltas = _mm512_set_epi32(15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0);
+  const __m512i masked_deltas = _mm512_and_si512(deltas, mask);
+  const __m512i low_words = _mm512_add_epi32(
+    _mm512_set1_epi32((int32_t)counter),
+    masked_deltas);
+  // The carry bit is 1 if the high bit of the word was 1 before addition and is
+  // 0 after.
+  // NOTE: It would be a bit more natural to use _mm512_cmp_epu32_mask to
+  // compute the carry bits here, and originally we did, but that intrinsic is
+  // broken under GCC 5.4. See https://github.com/BLAKE3-team/BLAKE3/issues/271.
+  const __m512i carries = _mm512_srli_epi32(
+    _mm512_andnot_si512(
+        low_words, // 0 after (gets inverted by andnot)
+        _mm512_set1_epi32((int32_t)counter)), // and 1 before
+    31);
+  const __m512i high_words = _mm512_add_epi32(
+    _mm512_set1_epi32((int32_t)(counter >> 32)),
+    carries);
+  *out_lo = low_words;
+  *out_hi = high_words;
+}
+
+static
+void blake3_hash16_avx512(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) {
+  __m512i h_vecs[8] = {
+      set1_512(key[0]), set1_512(key[1]), set1_512(key[2]), set1_512(key[3]),
+      set1_512(key[4]), set1_512(key[5]), set1_512(key[6]), set1_512(key[7]),
+  };
+  __m512i counter_low_vec, counter_high_vec;
+  load_counters16(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;
+    }
+    __m512i block_len_vec = set1_512(BLAKE3_BLOCK_LEN);
+    __m512i block_flags_vec = set1_512(block_flags);
+    __m512i msg_vecs[16];
+    transpose_msg_vecs16(inputs, block * BLAKE3_BLOCK_LEN, msg_vecs);
+
+    __m512i 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_512(IV[0]), set1_512(IV[1]),  set1_512(IV[2]), set1_512(IV[3]),
+        counter_low_vec, counter_high_vec, block_len_vec,   block_flags_vec,
+    };
+    round_fn16(v, msg_vecs, 0);
+    round_fn16(v, msg_vecs, 1);
+    round_fn16(v, msg_vecs, 2);
+    round_fn16(v, msg_vecs, 3);
+    round_fn16(v, msg_vecs, 4);
+    round_fn16(v, msg_vecs, 5);
+    round_fn16(v, msg_vecs, 6);
+    h_vecs[0] = xor_512(v[0], v[8]);
+    h_vecs[1] = xor_512(v[1], v[9]);
+    h_vecs[2] = xor_512(v[2], v[10]);
+    h_vecs[3] = xor_512(v[3], v[11]);
+    h_vecs[4] = xor_512(v[4], v[12]);
+    h_vecs[5] = xor_512(v[5], v[13]);
+    h_vecs[6] = xor_512(v[6], v[14]);
+    h_vecs[7] = xor_512(v[7], v[15]);
+
+    block_flags = flags;
+  }
+
+  // transpose_vecs_512 operates on a 16x16 matrix of words, but we only have 8
+  // state vectors. Pad the matrix with zeros. After transposition, store the
+  // lower half of each vector.
+  __m512i padded[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_512(0), set1_512(0), set1_512(0), set1_512(0),
+      set1_512(0), set1_512(0), set1_512(0), set1_512(0),
+  };
+  transpose_vecs_512(padded);
+  _mm256_mask_storeu_epi32(&out[0 * sizeof(__m256i)], (__mmask8)-1, _mm512_castsi512_si256(padded[0]));
+  _mm256_mask_storeu_epi32(&out[1 * sizeof(__m256i)], (__mmask8)-1, _mm512_castsi512_si256(padded[1]));
+  _mm256_mask_storeu_epi32(&out[2 * sizeof(__m256i)], (__mmask8)-1, _mm512_castsi512_si256(padded[2]));
+  _mm256_mask_storeu_epi32(&out[3 * sizeof(__m256i)], (__mmask8)-1, _mm512_castsi512_si256(padded[3]));
+  _mm256_mask_storeu_epi32(&out[4 * sizeof(__m256i)], (__mmask8)-1, _mm512_castsi512_si256(padded[4]));
+  _mm256_mask_storeu_epi32(&out[5 * sizeof(__m256i)], (__mmask8)-1, _mm512_castsi512_si256(padded[5]));
+  _mm256_mask_storeu_epi32(&out[6 * sizeof(__m256i)], (__mmask8)-1, _mm512_castsi512_si256(padded[6]));
+  _mm256_mask_storeu_epi32(&out[7 * sizeof(__m256i)], (__mmask8)-1, _mm512_castsi512_si256(padded[7]));
+  _mm256_mask_storeu_epi32(&out[8 * sizeof(__m256i)], (__mmask8)-1, _mm512_castsi512_si256(padded[8]));
+  _mm256_mask_storeu_epi32(&out[9 * sizeof(__m256i)], (__mmask8)-1, _mm512_castsi512_si256(padded[9]));
+  _mm256_mask_storeu_epi32(&out[10 * sizeof(__m256i)], (__mmask8)-1, _mm512_castsi512_si256(padded[10]));
+  _mm256_mask_storeu_epi32(&out[11 * sizeof(__m256i)], (__mmask8)-1, _mm512_castsi512_si256(padded[11]));
+  _mm256_mask_storeu_epi32(&out[12 * sizeof(__m256i)], (__mmask8)-1, _mm512_castsi512_si256(padded[12]));
+  _mm256_mask_storeu_epi32(&out[13 * sizeof(__m256i)], (__mmask8)-1, _mm512_castsi512_si256(padded[13]));
+  _mm256_mask_storeu_epi32(&out[14 * sizeof(__m256i)], (__mmask8)-1, _mm512_castsi512_si256(padded[14]));
+  _mm256_mask_storeu_epi32(&out[15 * sizeof(__m256i)], (__mmask8)-1, _mm512_castsi512_si256(padded[15]));
+}
+
+/*
+ * ----------------------------------------------------------------------------
+ * hash_many_avx512
+ * ----------------------------------------------------------------------------
+ */
+
+INLINE void hash_one_avx512(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;
+    }
+    blake3_compress_in_place_avx512(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_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) {
+  while (num_inputs >= 16) {
+    blake3_hash16_avx512(inputs, blocks, key, counter, increment_counter, flags,
+                         flags_start, flags_end, out);
+    if (increment_counter) {
+      counter += 16;
+    }
+    inputs += 16;
+    num_inputs -= 16;
+    out = &out[16 * BLAKE3_OUT_LEN];
+  }
+  while (num_inputs >= 8) {
+    blake3_hash8_avx512(inputs, blocks, key, counter, increment_counter, flags,
+                        flags_start, flags_end, out);
+    if (increment_counter) {
+      counter += 8;
+    }
+    inputs += 8;
+    num_inputs -= 8;
+    out = &out[8 * BLAKE3_OUT_LEN];
+  }
+  while (num_inputs >= 4) {
+    blake3_hash4_avx512(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_avx512(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];
+  }
+}
diff --git a/src/util/blake3/blake3_avx512_x86-64_windows_gnu.S b/src/util/blake3/blake3_avx512_x86-64_windows_gnu.S
index e10b9f36cbc..ba4fc5fa984 100644
--- a/src/util/blake3/blake3_avx512_x86-64_windows_gnu.S
+++ b/src/util/blake3/blake3_avx512_x86-64_windows_gnu.S
@@ -2587,7 +2587,7 @@ blake3_compress_xof_avx512:
         add     rsp, 72
         ret
 
-.section .rodata
+.section .rdata
 .p2align  6
 INDEX0:
         .long    0,  1,  2,  3, 16, 17, 18, 19
diff --git a/src/util/blake3/blake3_dispatch.c b/src/util/blake3/blake3_dispatch.c
index 0f348efafca..af6c3dadc7b 100644
--- a/src/util/blake3/blake3_dispatch.c
+++ b/src/util/blake3/blake3_dispatch.c
@@ -6,6 +6,7 @@
 
 #if defined(IS_X86)
 #if defined(_MSC_VER)
+#include <Windows.h>
 #include <intrin.h>
 #elif defined(__GNUC__)
 #include <immintrin.h>
@@ -14,6 +15,32 @@
 #endif
 #endif
 
+#if !defined(BLAKE3_ATOMICS)
+#if defined(__has_include)
+#if __has_include(<stdatomic.h>) && !defined(_MSC_VER)
+#define BLAKE3_ATOMICS 1
+#else
+#define BLAKE3_ATOMICS 0
+#endif /* __has_include(<stdatomic.h>) && !defined(_MSC_VER) */
+#else
+#define BLAKE3_ATOMICS 0
+#endif /* defined(__has_include) */
+#endif /* BLAKE3_ATOMICS */
+
+#if BLAKE3_ATOMICS
+#define ATOMIC_INT _Atomic int
+#define ATOMIC_LOAD(x) x
+#define ATOMIC_STORE(x, y) x = y
+#elif defined(_MSC_VER)
+#define ATOMIC_INT LONG
+#define ATOMIC_LOAD(x) InterlockedOr(&x, 0)
+#define ATOMIC_STORE(x, y) InterlockedExchange(&x, y)
+#else
+#define ATOMIC_INT int
+#define ATOMIC_LOAD(x) x
+#define ATOMIC_STORE(x, y) x = y
+#endif
+
 #define MAYBE_UNUSED(x) (void)((x))
 
 #if defined(IS_X86)
@@ -76,7 +103,7 @@ enum cpu_feature {
 #if !defined(BLAKE3_TESTING)
 static /* Allow the variable to be controlled manually for testing */
 #endif
-    enum cpu_feature g_cpu_features = UNDEFINED;
+    ATOMIC_INT g_cpu_features = UNDEFINED;
 
 #if !defined(BLAKE3_TESTING)
 static
@@ -84,14 +111,16 @@ static
     enum cpu_feature
     get_cpu_features(void) {
 
-  if (g_cpu_features != UNDEFINED) {
-    return g_cpu_features;
+  /* If TSAN detects a data race here, try compiling with -DBLAKE3_ATOMICS=1 */
+  enum cpu_feature features = ATOMIC_LOAD(g_cpu_features);
+  if (features != UNDEFINED) {
+    return features;
   } else {
 #if defined(IS_X86)
     uint32_t regs[4] = {0};
     uint32_t *eax = &regs[0], *ebx = &regs[1], *ecx = &regs[2], *edx = &regs[3];
     (void)edx;
-    enum cpu_feature features = 0;
+    features = 0;
     cpuid(regs, 0);
     const int max_id = *eax;
     cpuid(regs, 1);
@@ -101,7 +130,7 @@ static
     if (*edx & (1UL << 26))
       features |= SSE2;
 #endif
-    if (*ecx & (1UL << 0))
+    if (*ecx & (1UL << 9))
       features |= SSSE3;
     if (*ecx & (1UL << 19))
       features |= SSE41;
@@ -124,7 +153,7 @@ static
         }
       }
     }
-    g_cpu_features = features;
+    ATOMIC_STORE(g_cpu_features, features);
     return features;
 #else
     /* How to detect NEON? */
diff --git a/src/util/blake3/blake3_impl.h b/src/util/blake3/blake3_impl.h
index b0ce5c9999e..beab5cf53c6 100644
--- a/src/util/blake3/blake3_impl.h
+++ b/src/util/blake3/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)) && !defined(_M_ARM64EC)
+#if defined(__x86_64__) || defined(_M_X64) 
 #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) || defined(_M_ARM64EC)
+#if defined(__aarch64__) || defined(_M_ARM64)
 #define IS_AARCH64
 #endif
 
@@ -51,7 +51,11 @@ enum blake3_flags {
 #if !defined(BLAKE3_USE_NEON) 
   // If BLAKE3_USE_NEON not manually set, autodetect based on AArch64ness
   #if defined(IS_AARCH64)
-    #define BLAKE3_USE_NEON 1
+    #if defined(__ARM_BIG_ENDIAN)
+      #define BLAKE3_USE_NEON 0
+    #else
+      #define BLAKE3_USE_NEON 1
+    #endif
   #else
     #define BLAKE3_USE_NEON 0
   #endif
@@ -87,7 +91,7 @@ static const uint8_t MSG_SCHEDULE[7][16] = {
 /* x is assumed to be nonzero.       */
 static unsigned int highest_one(uint64_t x) {
 #if defined(__GNUC__) || defined(__clang__)
-  return 63 ^ __builtin_clzll(x);
+  return 63 ^ (unsigned int)__builtin_clzll(x);
 #elif defined(_MSC_VER) && defined(IS_X86_64)
   unsigned long index;
   _BitScanReverse64(&index, x);
@@ -117,7 +121,7 @@ static unsigned int highest_one(uint64_t x) {
 // Count the number of 1 bits.
 INLINE unsigned int popcnt(uint64_t x) {
 #if defined(__GNUC__) || defined(__clang__)
-  return __builtin_popcountll(x);
+  return (unsigned int)__builtin_popcountll(x);
 #else
   unsigned int count = 0;
   while (x != 0) {
diff --git a/src/util/blake3/blake3_neon.c b/src/util/blake3/blake3_neon.c
index 689e3259df7..50a0a40fabf 100644
--- a/src/util/blake3/blake3_neon.c
+++ b/src/util/blake3/blake3_neon.c
@@ -10,14 +10,12 @@
 
 INLINE uint32x4_t loadu_128(const uint8_t src[16]) {
   // vld1q_u32 has alignment requirements. Don't use it.
-  uint32x4_t x;
-  memcpy(&x, src, 16);
-  return x;
+  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.
-  memcpy(dest, &src, 16);
+  vst1q_u8(dest, vreinterpretq_u8_u32(src));
 }
 
 INLINE uint32x4_t add_128(uint32x4_t a, uint32x4_t b) {
@@ -36,19 +34,36 @@ INLINE uint32x4_t set4(uint32_t a, uint32_t b, uint32_t c, uint32_t d) {
 }
 
 INLINE uint32x4_t rot16_128(uint32x4_t x) {
-  return vorrq_u32(vshrq_n_u32(x, 16), vshlq_n_u32(x, 32 - 16));
+  // The straightfoward 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) {
-  return vorrq_u32(vshrq_n_u32(x, 12), vshlq_n_u32(x, 32 - 12));
+  // 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) {
-  return vorrq_u32(vshrq_n_u32(x, 8), vshlq_n_u32(x, 32 - 8));
+  // 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) {
-  return vorrq_u32(vshrq_n_u32(x, 7), vshlq_n_u32(x, 32 - 7));
+  // 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
@@ -229,9 +244,9 @@ INLINE void load_counters4(uint64_t counter, bool increment_counter,
 }
 
 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) {
+                       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]),
diff --git a/src/util/blake3/blake3_sse2.c b/src/util/blake3/blake3_sse2.c
new file mode 100644
index 00000000000..691e1c6806c
--- /dev/null
+++ b/src/util/blake3/blake3_sse2.c
@@ -0,0 +1,566 @@
+#include "blake3_impl.h"
+
+#include <immintrin.h>
+
+#define DEGREE 4
+
+#define _mm_shuffle_ps2(a, b, c)                                               \
+  (_mm_castps_si128(                                                           \
+      _mm_shuffle_ps(_mm_castsi128_ps(a), _mm_castsi128_ps(b), (c))))
+
+INLINE __m128i loadu(const uint8_t src[16]) {
+  return _mm_loadu_si128((const __m128i *)src);
+}
+
+INLINE void storeu(__m128i src, uint8_t dest[16]) {
+  _mm_storeu_si128((__m128i *)dest, src);
+}
+
+INLINE __m128i addv(__m128i a, __m128i b) { return _mm_add_epi32(a, b); }
+
+// Note that clang-format doesn't like the name "xor" for some reason.
+INLINE __m128i xorv(__m128i a, __m128i b) { return _mm_xor_si128(a, b); }
+
+INLINE __m128i set1(uint32_t x) { return _mm_set1_epi32((int32_t)x); }
+
+INLINE __m128i set4(uint32_t a, uint32_t b, uint32_t c, uint32_t d) {
+  return _mm_setr_epi32((int32_t)a, (int32_t)b, (int32_t)c, (int32_t)d);
+}
+
+INLINE __m128i rot16(__m128i x) {
+  return _mm_shufflehi_epi16(_mm_shufflelo_epi16(x, 0xB1), 0xB1);
+}
+
+INLINE __m128i rot12(__m128i x) {
+  return xorv(_mm_srli_epi32(x, 12), _mm_slli_epi32(x, 32 - 12));
+}
+
+INLINE __m128i rot8(__m128i x) {
+  return xorv(_mm_srli_epi32(x, 8), _mm_slli_epi32(x, 32 - 8));
+}
+
+INLINE __m128i rot7(__m128i x) {
+  return xorv(_mm_srli_epi32(x, 7), _mm_slli_epi32(x, 32 - 7));
+}
+
+INLINE void g1(__m128i *row0, __m128i *row1, __m128i *row2, __m128i *row3,
+               __m128i m) {
+  *row0 = addv(addv(*row0, m), *row1);
+  *row3 = xorv(*row3, *row0);
+  *row3 = rot16(*row3);
+  *row2 = addv(*row2, *row3);
+  *row1 = xorv(*row1, *row2);
+  *row1 = rot12(*row1);
+}
+
+INLINE void g2(__m128i *row0, __m128i *row1, __m128i *row2, __m128i *row3,
+               __m128i m) {
+  *row0 = addv(addv(*row0, m), *row1);
+  *row3 = xorv(*row3, *row0);
+  *row3 = rot8(*row3);
+  *row2 = addv(*row2, *row3);
+  *row1 = xorv(*row1, *row2);
+  *row1 = rot7(*row1);
+}
+
+// Note the optimization here of leaving row1 as the unrotated row, rather than
+// row0. All the message loads below are adjusted to compensate for this. See
+// discussion at https://github.com/sneves/blake2-avx2/pull/4
+INLINE void diagonalize(__m128i *row0, __m128i *row2, __m128i *row3) {
+  *row0 = _mm_shuffle_epi32(*row0, _MM_SHUFFLE(2, 1, 0, 3));
+  *row3 = _mm_shuffle_epi32(*row3, _MM_SHUFFLE(1, 0, 3, 2));
+  *row2 = _mm_shuffle_epi32(*row2, _MM_SHUFFLE(0, 3, 2, 1));
+}
+
+INLINE void undiagonalize(__m128i *row0, __m128i *row2, __m128i *row3) {
+  *row0 = _mm_shuffle_epi32(*row0, _MM_SHUFFLE(0, 3, 2, 1));
+  *row3 = _mm_shuffle_epi32(*row3, _MM_SHUFFLE(1, 0, 3, 2));
+  *row2 = _mm_shuffle_epi32(*row2, _MM_SHUFFLE(2, 1, 0, 3));
+}
+
+INLINE __m128i blend_epi16(__m128i a, __m128i b, const int16_t imm8) {
+  const __m128i bits = _mm_set_epi16(0x80, 0x40, 0x20, 0x10, 0x08, 0x04, 0x02, 0x01);
+  __m128i mask = _mm_set1_epi16(imm8);
+  mask = _mm_and_si128(mask, bits);
+  mask = _mm_cmpeq_epi16(mask, bits);
+  return _mm_or_si128(_mm_and_si128(mask, b), _mm_andnot_si128(mask, a));
+}
+
+INLINE void compress_pre(__m128i 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((uint8_t *)&cv[0]);
+  rows[1] = loadu((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);
+
+  __m128i m0 = loadu(&block[sizeof(__m128i) * 0]);
+  __m128i m1 = loadu(&block[sizeof(__m128i) * 1]);
+  __m128i m2 = loadu(&block[sizeof(__m128i) * 2]);
+  __m128i m3 = loadu(&block[sizeof(__m128i) * 3]);
+
+  __m128i 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 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(2, 0, 2, 0)); //  6  4  2  0
+  g1(&rows[0], &rows[1], &rows[2], &rows[3], t0);
+  t1 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(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 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(2, 0, 2, 0)); // 14 12 10  8
+  t2 = _mm_shuffle_epi32(t2, _MM_SHUFFLE(2, 1, 0, 3));   // 12 10  8 14
+  g1(&rows[0], &rows[1], &rows[2], &rows[3], t2);
+  t3 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 1, 3, 1)); // 15 13 11  9
+  t3 = _mm_shuffle_epi32(t3, _MM_SHUFFLE(2, 1, 0, 3));   // 13 11  9 15
+  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 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2));
+  t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1));
+  g1(&rows[0], &rows[1], &rows[2], &rows[3], t0);
+  t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2));
+  tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3));
+  t1 = blend_epi16(tt, t1, 0xCC);
+  g2(&rows[0], &rows[1], &rows[2], &rows[3], t1);
+  diagonalize(&rows[0], &rows[2], &rows[3]);
+  t2 = _mm_unpacklo_epi64(m3, m1);
+  tt = blend_epi16(t2, m2, 0xC0);
+  t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0));
+  g1(&rows[0], &rows[1], &rows[2], &rows[3], t2);
+  t3 = _mm_unpackhi_epi32(m1, m3);
+  tt = _mm_unpacklo_epi32(m2, t3);
+  t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(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 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2));
+  t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1));
+  g1(&rows[0], &rows[1], &rows[2], &rows[3], t0);
+  t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2));
+  tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3));
+  t1 = blend_epi16(tt, t1, 0xCC);
+  g2(&rows[0], &rows[1], &rows[2], &rows[3], t1);
+  diagonalize(&rows[0], &rows[2], &rows[3]);
+  t2 = _mm_unpacklo_epi64(m3, m1);
+  tt = blend_epi16(t2, m2, 0xC0);
+  t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0));
+  g1(&rows[0], &rows[1], &rows[2], &rows[3], t2);
+  t3 = _mm_unpackhi_epi32(m1, m3);
+  tt = _mm_unpacklo_epi32(m2, t3);
+  t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(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 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2));
+  t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1));
+  g1(&rows[0], &rows[1], &rows[2], &rows[3], t0);
+  t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2));
+  tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3));
+  t1 = blend_epi16(tt, t1, 0xCC);
+  g2(&rows[0], &rows[1], &rows[2], &rows[3], t1);
+  diagonalize(&rows[0], &rows[2], &rows[3]);
+  t2 = _mm_unpacklo_epi64(m3, m1);
+  tt = blend_epi16(t2, m2, 0xC0);
+  t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0));
+  g1(&rows[0], &rows[1], &rows[2], &rows[3], t2);
+  t3 = _mm_unpackhi_epi32(m1, m3);
+  tt = _mm_unpacklo_epi32(m2, t3);
+  t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(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 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2));
+  t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1));
+  g1(&rows[0], &rows[1], &rows[2], &rows[3], t0);
+  t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2));
+  tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3));
+  t1 = blend_epi16(tt, t1, 0xCC);
+  g2(&rows[0], &rows[1], &rows[2], &rows[3], t1);
+  diagonalize(&rows[0], &rows[2], &rows[3]);
+  t2 = _mm_unpacklo_epi64(m3, m1);
+  tt = blend_epi16(t2, m2, 0xC0);
+  t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0));
+  g1(&rows[0], &rows[1], &rows[2], &rows[3], t2);
+  t3 = _mm_unpackhi_epi32(m1, m3);
+  tt = _mm_unpacklo_epi32(m2, t3);
+  t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(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 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2));
+  t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1));
+  g1(&rows[0], &rows[1], &rows[2], &rows[3], t0);
+  t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2));
+  tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3));
+  t1 = blend_epi16(tt, t1, 0xCC);
+  g2(&rows[0], &rows[1], &rows[2], &rows[3], t1);
+  diagonalize(&rows[0], &rows[2], &rows[3]);
+  t2 = _mm_unpacklo_epi64(m3, m1);
+  tt = blend_epi16(t2, m2, 0xC0);
+  t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0));
+  g1(&rows[0], &rows[1], &rows[2], &rows[3], t2);
+  t3 = _mm_unpackhi_epi32(m1, m3);
+  tt = _mm_unpacklo_epi32(m2, t3);
+  t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(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 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2));
+  t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1));
+  g1(&rows[0], &rows[1], &rows[2], &rows[3], t0);
+  t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2));
+  tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3));
+  t1 = blend_epi16(tt, t1, 0xCC);
+  g2(&rows[0], &rows[1], &rows[2], &rows[3], t1);
+  diagonalize(&rows[0], &rows[2], &rows[3]);
+  t2 = _mm_unpacklo_epi64(m3, m1);
+  tt = blend_epi16(t2, m2, 0xC0);
+  t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0));
+  g1(&rows[0], &rows[1], &rows[2], &rows[3], t2);
+  t3 = _mm_unpackhi_epi32(m1, m3);
+  tt = _mm_unpacklo_epi32(m2, t3);
+  t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(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_sse2(uint32_t cv[8],
+                                   const uint8_t block[BLAKE3_BLOCK_LEN],
+                                   uint8_t block_len, uint64_t counter,
+                                   uint8_t flags) {
+  __m128i rows[4];
+  compress_pre(rows, cv, block, block_len, counter, flags);
+  storeu(xorv(rows[0], rows[2]), (uint8_t *)&cv[0]);
+  storeu(xorv(rows[1], rows[3]), (uint8_t *)&cv[4]);
+}
+
+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]) {
+  __m128i rows[4];
+  compress_pre(rows, cv, block, block_len, counter, flags);
+  storeu(xorv(rows[0], rows[2]), &out[0]);
+  storeu(xorv(rows[1], rows[3]), &out[16]);
+  storeu(xorv(rows[2], loadu((uint8_t *)&cv[0])), &out[32]);
+  storeu(xorv(rows[3], loadu((uint8_t *)&cv[4])), &out[48]);
+}
+
+INLINE void round_fn(__m128i v[16], __m128i m[16], size_t r) {
+  v[0] = addv(v[0], m[(size_t)MSG_SCHEDULE[r][0]]);
+  v[1] = addv(v[1], m[(size_t)MSG_SCHEDULE[r][2]]);
+  v[2] = addv(v[2], m[(size_t)MSG_SCHEDULE[r][4]]);
+  v[3] = addv(v[3], m[(size_t)MSG_SCHEDULE[r][6]]);
+  v[0] = addv(v[0], v[4]);
+  v[1] = addv(v[1], v[5]);
+  v[2] = addv(v[2], v[6]);
+  v[3] = addv(v[3], v[7]);
+  v[12] = xorv(v[12], v[0]);
+  v[13] = xorv(v[13], v[1]);
+  v[14] = xorv(v[14], v[2]);
+  v[15] = xorv(v[15], v[3]);
+  v[12] = rot16(v[12]);
+  v[13] = rot16(v[13]);
+  v[14] = rot16(v[14]);
+  v[15] = rot16(v[15]);
+  v[8] = addv(v[8], v[12]);
+  v[9] = addv(v[9], v[13]);
+  v[10] = addv(v[10], v[14]);
+  v[11] = addv(v[11], v[15]);
+  v[4] = xorv(v[4], v[8]);
+  v[5] = xorv(v[5], v[9]);
+  v[6] = xorv(v[6], v[10]);
+  v[7] = xorv(v[7], v[11]);
+  v[4] = rot12(v[4]);
+  v[5] = rot12(v[5]);
+  v[6] = rot12(v[6]);
+  v[7] = rot12(v[7]);
+  v[0] = addv(v[0], m[(size_t)MSG_SCHEDULE[r][1]]);
+  v[1] = addv(v[1], m[(size_t)MSG_SCHEDULE[r][3]]);
+  v[2] = addv(v[2], m[(size_t)MSG_SCHEDULE[r][5]]);
+  v[3] = addv(v[3], m[(size_t)MSG_SCHEDULE[r][7]]);
+  v[0] = addv(v[0], v[4]);
+  v[1] = addv(v[1], v[5]);
+  v[2] = addv(v[2], v[6]);
+  v[3] = addv(v[3], v[7]);
+  v[12] = xorv(v[12], v[0]);
+  v[13] = xorv(v[13], v[1]);
+  v[14] = xorv(v[14], v[2]);
+  v[15] = xorv(v[15], v[3]);
+  v[12] = rot8(v[12]);
+  v[13] = rot8(v[13]);
+  v[14] = rot8(v[14]);
+  v[15] = rot8(v[15]);
+  v[8] = addv(v[8], v[12]);
+  v[9] = addv(v[9], v[13]);
+  v[10] = addv(v[10], v[14]);
+  v[11] = addv(v[11], v[15]);
+  v[4] = xorv(v[4], v[8]);
+  v[5] = xorv(v[5], v[9]);
+  v[6] = xorv(v[6], v[10]);
+  v[7] = xorv(v[7], v[11]);
+  v[4] = rot7(v[4]);
+  v[5] = rot7(v[5]);
+  v[6] = rot7(v[6]);
+  v[7] = rot7(v[7]);
+
+  v[0] = addv(v[0], m[(size_t)MSG_SCHEDULE[r][8]]);
+  v[1] = addv(v[1], m[(size_t)MSG_SCHEDULE[r][10]]);
+  v[2] = addv(v[2], m[(size_t)MSG_SCHEDULE[r][12]]);
+  v[3] = addv(v[3], m[(size_t)MSG_SCHEDULE[r][14]]);
+  v[0] = addv(v[0], v[5]);
+  v[1] = addv(v[1], v[6]);
+  v[2] = addv(v[2], v[7]);
+  v[3] = addv(v[3], v[4]);
+  v[15] = xorv(v[15], v[0]);
+  v[12] = xorv(v[12], v[1]);
+  v[13] = xorv(v[13], v[2]);
+  v[14] = xorv(v[14], v[3]);
+  v[15] = rot16(v[15]);
+  v[12] = rot16(v[12]);
+  v[13] = rot16(v[13]);
+  v[14] = rot16(v[14]);
+  v[10] = addv(v[10], v[15]);
+  v[11] = addv(v[11], v[12]);
+  v[8] = addv(v[8], v[13]);
+  v[9] = addv(v[9], v[14]);
+  v[5] = xorv(v[5], v[10]);
+  v[6] = xorv(v[6], v[11]);
+  v[7] = xorv(v[7], v[8]);
+  v[4] = xorv(v[4], v[9]);
+  v[5] = rot12(v[5]);
+  v[6] = rot12(v[6]);
+  v[7] = rot12(v[7]);
+  v[4] = rot12(v[4]);
+  v[0] = addv(v[0], m[(size_t)MSG_SCHEDULE[r][9]]);
+  v[1] = addv(v[1], m[(size_t)MSG_SCHEDULE[r][11]]);
+  v[2] = addv(v[2], m[(size_t)MSG_SCHEDULE[r][13]]);
+  v[3] = addv(v[3], m[(size_t)MSG_SCHEDULE[r][15]]);
+  v[0] = addv(v[0], v[5]);
+  v[1] = addv(v[1], v[6]);
+  v[2] = addv(v[2], v[7]);
+  v[3] = addv(v[3], v[4]);
+  v[15] = xorv(v[15], v[0]);
+  v[12] = xorv(v[12], v[1]);
+  v[13] = xorv(v[13], v[2]);
+  v[14] = xorv(v[14], v[3]);
+  v[15] = rot8(v[15]);
+  v[12] = rot8(v[12]);
+  v[13] = rot8(v[13]);
+  v[14] = rot8(v[14]);
+  v[10] = addv(v[10], v[15]);
+  v[11] = addv(v[11], v[12]);
+  v[8] = addv(v[8], v[13]);
+  v[9] = addv(v[9], v[14]);
+  v[5] = xorv(v[5], v[10]);
+  v[6] = xorv(v[6], v[11]);
+  v[7] = xorv(v[7], v[8]);
+  v[4] = xorv(v[4], v[9]);
+  v[5] = rot7(v[5]);
+  v[6] = rot7(v[6]);
+  v[7] = rot7(v[7]);
+  v[4] = rot7(v[4]);
+}
+
+INLINE void transpose_vecs(__m128i vecs[DEGREE]) {
+  // Interleave 32-bit lanes. The low unpack is lanes 00/11 and the high is
+  // 22/33. Note that this doesn't split the vector into two lanes, as the
+  // AVX2 counterparts do.
+  __m128i ab_01 = _mm_unpacklo_epi32(vecs[0], vecs[1]);
+  __m128i ab_23 = _mm_unpackhi_epi32(vecs[0], vecs[1]);
+  __m128i cd_01 = _mm_unpacklo_epi32(vecs[2], vecs[3]);
+  __m128i cd_23 = _mm_unpackhi_epi32(vecs[2], vecs[3]);
+
+  // Interleave 64-bit lanes.
+  __m128i abcd_0 = _mm_unpacklo_epi64(ab_01, cd_01);
+  __m128i abcd_1 = _mm_unpackhi_epi64(ab_01, cd_01);
+  __m128i abcd_2 = _mm_unpacklo_epi64(ab_23, cd_23);
+  __m128i abcd_3 = _mm_unpackhi_epi64(ab_23, cd_23);
+
+  vecs[0] = abcd_0;
+  vecs[1] = abcd_1;
+  vecs[2] = abcd_2;
+  vecs[3] = abcd_3;
+}
+
+INLINE void transpose_msg_vecs(const uint8_t *const *inputs,
+                               size_t block_offset, __m128i out[16]) {
+  out[0] = loadu(&inputs[0][block_offset + 0 * sizeof(__m128i)]);
+  out[1] = loadu(&inputs[1][block_offset + 0 * sizeof(__m128i)]);
+  out[2] = loadu(&inputs[2][block_offset + 0 * sizeof(__m128i)]);
+  out[3] = loadu(&inputs[3][block_offset + 0 * sizeof(__m128i)]);
+  out[4] = loadu(&inputs[0][block_offset + 1 * sizeof(__m128i)]);
+  out[5] = loadu(&inputs[1][block_offset + 1 * sizeof(__m128i)]);
+  out[6] = loadu(&inputs[2][block_offset + 1 * sizeof(__m128i)]);
+  out[7] = loadu(&inputs[3][block_offset + 1 * sizeof(__m128i)]);
+  out[8] = loadu(&inputs[0][block_offset + 2 * sizeof(__m128i)]);
+  out[9] = loadu(&inputs[1][block_offset + 2 * sizeof(__m128i)]);
+  out[10] = loadu(&inputs[2][block_offset + 2 * sizeof(__m128i)]);
+  out[11] = loadu(&inputs[3][block_offset + 2 * sizeof(__m128i)]);
+  out[12] = loadu(&inputs[0][block_offset + 3 * sizeof(__m128i)]);
+  out[13] = loadu(&inputs[1][block_offset + 3 * sizeof(__m128i)]);
+  out[14] = loadu(&inputs[2][block_offset + 3 * sizeof(__m128i)]);
+  out[15] = loadu(&inputs[3][block_offset + 3 * sizeof(__m128i)]);
+  for (size_t i = 0; i < 4; ++i) {
+    _mm_prefetch((const void *)&inputs[i][block_offset + 256], _MM_HINT_T0);
+  }
+  transpose_vecs(&out[0]);
+  transpose_vecs(&out[4]);
+  transpose_vecs(&out[8]);
+  transpose_vecs(&out[12]);
+}
+
+INLINE void load_counters(uint64_t counter, bool increment_counter,
+                          __m128i *out_lo, __m128i *out_hi) {
+  const __m128i mask = _mm_set1_epi32(-(int32_t)increment_counter);
+  const __m128i add0 = _mm_set_epi32(3, 2, 1, 0);
+  const __m128i add1 = _mm_and_si128(mask, add0);
+  __m128i l = _mm_add_epi32(_mm_set1_epi32((int32_t)counter), add1);
+  __m128i carry = _mm_cmpgt_epi32(_mm_xor_si128(add1, _mm_set1_epi32(0x80000000)), 
+                                  _mm_xor_si128(   l, _mm_set1_epi32(0x80000000)));
+  __m128i h = _mm_sub_epi32(_mm_set1_epi32((int32_t)(counter >> 32)), carry);
+  *out_lo = l;
+  *out_hi = h;
+}
+
+static
+void blake3_hash4_sse2(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) {
+  __m128i h_vecs[8] = {
+      set1(key[0]), set1(key[1]), set1(key[2]), set1(key[3]),
+      set1(key[4]), set1(key[5]), set1(key[6]), set1(key[7]),
+  };
+  __m128i counter_low_vec, counter_high_vec;
+  load_counters(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;
+    }
+    __m128i block_len_vec = set1(BLAKE3_BLOCK_LEN);
+    __m128i block_flags_vec = set1(block_flags);
+    __m128i msg_vecs[16];
+    transpose_msg_vecs(inputs, block * BLAKE3_BLOCK_LEN, msg_vecs);
+
+    __m128i 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(IV[0]),     set1(IV[1]),      set1(IV[2]),   set1(IV[3]),
+        counter_low_vec, counter_high_vec, block_len_vec, block_flags_vec,
+    };
+    round_fn(v, msg_vecs, 0);
+    round_fn(v, msg_vecs, 1);
+    round_fn(v, msg_vecs, 2);
+    round_fn(v, msg_vecs, 3);
+    round_fn(v, msg_vecs, 4);
+    round_fn(v, msg_vecs, 5);
+    round_fn(v, msg_vecs, 6);
+    h_vecs[0] = xorv(v[0], v[8]);
+    h_vecs[1] = xorv(v[1], v[9]);
+    h_vecs[2] = xorv(v[2], v[10]);
+    h_vecs[3] = xorv(v[3], v[11]);
+    h_vecs[4] = xorv(v[4], v[12]);
+    h_vecs[5] = xorv(v[5], v[13]);
+    h_vecs[6] = xorv(v[6], v[14]);
+    h_vecs[7] = xorv(v[7], v[15]);
+
+    block_flags = flags;
+  }
+
+  transpose_vecs(&h_vecs[0]);
+  transpose_vecs(&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(h_vecs[0], &out[0 * sizeof(__m128i)]);
+  storeu(h_vecs[4], &out[1 * sizeof(__m128i)]);
+  storeu(h_vecs[1], &out[2 * sizeof(__m128i)]);
+  storeu(h_vecs[5], &out[3 * sizeof(__m128i)]);
+  storeu(h_vecs[2], &out[4 * sizeof(__m128i)]);
+  storeu(h_vecs[6], &out[5 * sizeof(__m128i)]);
+  storeu(h_vecs[3], &out[6 * sizeof(__m128i)]);
+  storeu(h_vecs[7], &out[7 * sizeof(__m128i)]);
+}
+
+INLINE void hash_one_sse2(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;
+    }
+    blake3_compress_in_place_sse2(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_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) {
+  while (num_inputs >= DEGREE) {
+    blake3_hash4_sse2(inputs, blocks, key, counter, increment_counter, flags,
+                      flags_start, flags_end, out);
+    if (increment_counter) {
+      counter += DEGREE;
+    }
+    inputs += DEGREE;
+    num_inputs -= DEGREE;
+    out = &out[DEGREE * BLAKE3_OUT_LEN];
+  }
+  while (num_inputs > 0) {
+    hash_one_sse2(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];
+  }
+}
diff --git a/src/util/blake3/blake3_sse2_x86-64_windows_gnu.S b/src/util/blake3/blake3_sse2_x86-64_windows_gnu.S
index 8852ba5976e..4facb50e75b 100644
--- a/src/util/blake3/blake3_sse2_x86-64_windows_gnu.S
+++ b/src/util/blake3/blake3_sse2_x86-64_windows_gnu.S
@@ -2301,7 +2301,7 @@ blake3_compress_xof_sse2:
         ret
 
 
-.section .rodata
+.section .rdata
 .p2align  6
 BLAKE3_IV:
         .long  0x6A09E667, 0xBB67AE85
diff --git a/src/util/blake3/blake3_sse41.c b/src/util/blake3/blake3_sse41.c
new file mode 100644
index 00000000000..4653a856fe6
--- /dev/null
+++ b/src/util/blake3/blake3_sse41.c
@@ -0,0 +1,560 @@
+#include "blake3_impl.h"
+
+#include <immintrin.h>
+
+#define DEGREE 4
+
+#define _mm_shuffle_ps2(a, b, c)                                               \
+  (_mm_castps_si128(                                                           \
+      _mm_shuffle_ps(_mm_castsi128_ps(a), _mm_castsi128_ps(b), (c))))
+
+INLINE __m128i loadu(const uint8_t src[16]) {
+  return _mm_loadu_si128((const __m128i *)src);
+}
+
+INLINE void storeu(__m128i src, uint8_t dest[16]) {
+  _mm_storeu_si128((__m128i *)dest, src);
+}
+
+INLINE __m128i addv(__m128i a, __m128i b) { return _mm_add_epi32(a, b); }
+
+// Note that clang-format doesn't like the name "xor" for some reason.
+INLINE __m128i xorv(__m128i a, __m128i b) { return _mm_xor_si128(a, b); }
+
+INLINE __m128i set1(uint32_t x) { return _mm_set1_epi32((int32_t)x); }
+
+INLINE __m128i set4(uint32_t a, uint32_t b, uint32_t c, uint32_t d) {
+  return _mm_setr_epi32((int32_t)a, (int32_t)b, (int32_t)c, (int32_t)d);
+}
+
+INLINE __m128i rot16(__m128i x) {
+  return _mm_shuffle_epi8(
+      x, _mm_set_epi8(13, 12, 15, 14, 9, 8, 11, 10, 5, 4, 7, 6, 1, 0, 3, 2));
+}
+
+INLINE __m128i rot12(__m128i x) {
+  return xorv(_mm_srli_epi32(x, 12), _mm_slli_epi32(x, 32 - 12));
+}
+
+INLINE __m128i rot8(__m128i x) {
+  return _mm_shuffle_epi8(
+      x, _mm_set_epi8(12, 15, 14, 13, 8, 11, 10, 9, 4, 7, 6, 5, 0, 3, 2, 1));
+}
+
+INLINE __m128i rot7(__m128i x) {
+  return xorv(_mm_srli_epi32(x, 7), _mm_slli_epi32(x, 32 - 7));
+}
+
+INLINE void g1(__m128i *row0, __m128i *row1, __m128i *row2, __m128i *row3,
+               __m128i m) {
+  *row0 = addv(addv(*row0, m), *row1);
+  *row3 = xorv(*row3, *row0);
+  *row3 = rot16(*row3);
+  *row2 = addv(*row2, *row3);
+  *row1 = xorv(*row1, *row2);
+  *row1 = rot12(*row1);
+}
+
+INLINE void g2(__m128i *row0, __m128i *row1, __m128i *row2, __m128i *row3,
+               __m128i m) {
+  *row0 = addv(addv(*row0, m), *row1);
+  *row3 = xorv(*row3, *row0);
+  *row3 = rot8(*row3);
+  *row2 = addv(*row2, *row3);
+  *row1 = xorv(*row1, *row2);
+  *row1 = rot7(*row1);
+}
+
+// Note the optimization here of leaving row1 as the unrotated row, rather than
+// row0. All the message loads below are adjusted to compensate for this. See
+// discussion at https://github.com/sneves/blake2-avx2/pull/4
+INLINE void diagonalize(__m128i *row0, __m128i *row2, __m128i *row3) {
+  *row0 = _mm_shuffle_epi32(*row0, _MM_SHUFFLE(2, 1, 0, 3));
+  *row3 = _mm_shuffle_epi32(*row3, _MM_SHUFFLE(1, 0, 3, 2));
+  *row2 = _mm_shuffle_epi32(*row2, _MM_SHUFFLE(0, 3, 2, 1));
+}
+
+INLINE void undiagonalize(__m128i *row0, __m128i *row2, __m128i *row3) {
+  *row0 = _mm_shuffle_epi32(*row0, _MM_SHUFFLE(0, 3, 2, 1));
+  *row3 = _mm_shuffle_epi32(*row3, _MM_SHUFFLE(1, 0, 3, 2));
+  *row2 = _mm_shuffle_epi32(*row2, _MM_SHUFFLE(2, 1, 0, 3));
+}
+
+INLINE void compress_pre(__m128i 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((uint8_t *)&cv[0]);
+  rows[1] = loadu((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);
+
+  __m128i m0 = loadu(&block[sizeof(__m128i) * 0]);
+  __m128i m1 = loadu(&block[sizeof(__m128i) * 1]);
+  __m128i m2 = loadu(&block[sizeof(__m128i) * 2]);
+  __m128i m3 = loadu(&block[sizeof(__m128i) * 3]);
+
+  __m128i 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 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(2, 0, 2, 0)); //  6  4  2  0
+  g1(&rows[0], &rows[1], &rows[2], &rows[3], t0);
+  t1 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(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 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(2, 0, 2, 0)); // 14 12 10  8
+  t2 = _mm_shuffle_epi32(t2, _MM_SHUFFLE(2, 1, 0, 3));   // 12 10  8 14
+  g1(&rows[0], &rows[1], &rows[2], &rows[3], t2);
+  t3 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 1, 3, 1)); // 15 13 11  9
+  t3 = _mm_shuffle_epi32(t3, _MM_SHUFFLE(2, 1, 0, 3));   // 13 11  9 15
+  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 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2));
+  t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1));
+  g1(&rows[0], &rows[1], &rows[2], &rows[3], t0);
+  t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2));
+  tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3));
+  t1 = _mm_blend_epi16(tt, t1, 0xCC);
+  g2(&rows[0], &rows[1], &rows[2], &rows[3], t1);
+  diagonalize(&rows[0], &rows[2], &rows[3]);
+  t2 = _mm_unpacklo_epi64(m3, m1);
+  tt = _mm_blend_epi16(t2, m2, 0xC0);
+  t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0));
+  g1(&rows[0], &rows[1], &rows[2], &rows[3], t2);
+  t3 = _mm_unpackhi_epi32(m1, m3);
+  tt = _mm_unpacklo_epi32(m2, t3);
+  t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(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 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2));
+  t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1));
+  g1(&rows[0], &rows[1], &rows[2], &rows[3], t0);
+  t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2));
+  tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3));
+  t1 = _mm_blend_epi16(tt, t1, 0xCC);
+  g2(&rows[0], &rows[1], &rows[2], &rows[3], t1);
+  diagonalize(&rows[0], &rows[2], &rows[3]);
+  t2 = _mm_unpacklo_epi64(m3, m1);
+  tt = _mm_blend_epi16(t2, m2, 0xC0);
+  t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0));
+  g1(&rows[0], &rows[1], &rows[2], &rows[3], t2);
+  t3 = _mm_unpackhi_epi32(m1, m3);
+  tt = _mm_unpacklo_epi32(m2, t3);
+  t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(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 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2));
+  t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1));
+  g1(&rows[0], &rows[1], &rows[2], &rows[3], t0);
+  t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2));
+  tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3));
+  t1 = _mm_blend_epi16(tt, t1, 0xCC);
+  g2(&rows[0], &rows[1], &rows[2], &rows[3], t1);
+  diagonalize(&rows[0], &rows[2], &rows[3]);
+  t2 = _mm_unpacklo_epi64(m3, m1);
+  tt = _mm_blend_epi16(t2, m2, 0xC0);
+  t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0));
+  g1(&rows[0], &rows[1], &rows[2], &rows[3], t2);
+  t3 = _mm_unpackhi_epi32(m1, m3);
+  tt = _mm_unpacklo_epi32(m2, t3);
+  t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(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 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2));
+  t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1));
+  g1(&rows[0], &rows[1], &rows[2], &rows[3], t0);
+  t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2));
+  tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3));
+  t1 = _mm_blend_epi16(tt, t1, 0xCC);
+  g2(&rows[0], &rows[1], &rows[2], &rows[3], t1);
+  diagonalize(&rows[0], &rows[2], &rows[3]);
+  t2 = _mm_unpacklo_epi64(m3, m1);
+  tt = _mm_blend_epi16(t2, m2, 0xC0);
+  t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0));
+  g1(&rows[0], &rows[1], &rows[2], &rows[3], t2);
+  t3 = _mm_unpackhi_epi32(m1, m3);
+  tt = _mm_unpacklo_epi32(m2, t3);
+  t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(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 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2));
+  t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1));
+  g1(&rows[0], &rows[1], &rows[2], &rows[3], t0);
+  t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2));
+  tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3));
+  t1 = _mm_blend_epi16(tt, t1, 0xCC);
+  g2(&rows[0], &rows[1], &rows[2], &rows[3], t1);
+  diagonalize(&rows[0], &rows[2], &rows[3]);
+  t2 = _mm_unpacklo_epi64(m3, m1);
+  tt = _mm_blend_epi16(t2, m2, 0xC0);
+  t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0));
+  g1(&rows[0], &rows[1], &rows[2], &rows[3], t2);
+  t3 = _mm_unpackhi_epi32(m1, m3);
+  tt = _mm_unpacklo_epi32(m2, t3);
+  t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(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 = _mm_shuffle_ps2(m0, m1, _MM_SHUFFLE(3, 1, 1, 2));
+  t0 = _mm_shuffle_epi32(t0, _MM_SHUFFLE(0, 3, 2, 1));
+  g1(&rows[0], &rows[1], &rows[2], &rows[3], t0);
+  t1 = _mm_shuffle_ps2(m2, m3, _MM_SHUFFLE(3, 3, 2, 2));
+  tt = _mm_shuffle_epi32(m0, _MM_SHUFFLE(0, 0, 3, 3));
+  t1 = _mm_blend_epi16(tt, t1, 0xCC);
+  g2(&rows[0], &rows[1], &rows[2], &rows[3], t1);
+  diagonalize(&rows[0], &rows[2], &rows[3]);
+  t2 = _mm_unpacklo_epi64(m3, m1);
+  tt = _mm_blend_epi16(t2, m2, 0xC0);
+  t2 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(1, 3, 2, 0));
+  g1(&rows[0], &rows[1], &rows[2], &rows[3], t2);
+  t3 = _mm_unpackhi_epi32(m1, m3);
+  tt = _mm_unpacklo_epi32(m2, t3);
+  t3 = _mm_shuffle_epi32(tt, _MM_SHUFFLE(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_sse41(uint32_t cv[8],
+                                    const uint8_t block[BLAKE3_BLOCK_LEN],
+                                    uint8_t block_len, uint64_t counter,
+                                    uint8_t flags) {
+  __m128i rows[4];
+  compress_pre(rows, cv, block, block_len, counter, flags);
+  storeu(xorv(rows[0], rows[2]), (uint8_t *)&cv[0]);
+  storeu(xorv(rows[1], rows[3]), (uint8_t *)&cv[4]);
+}
+
+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]) {
+  __m128i rows[4];
+  compress_pre(rows, cv, block, block_len, counter, flags);
+  storeu(xorv(rows[0], rows[2]), &out[0]);
+  storeu(xorv(rows[1], rows[3]), &out[16]);
+  storeu(xorv(rows[2], loadu((uint8_t *)&cv[0])), &out[32]);
+  storeu(xorv(rows[3], loadu((uint8_t *)&cv[4])), &out[48]);
+}
+
+INLINE void round_fn(__m128i v[16], __m128i m[16], size_t r) {
+  v[0] = addv(v[0], m[(size_t)MSG_SCHEDULE[r][0]]);
+  v[1] = addv(v[1], m[(size_t)MSG_SCHEDULE[r][2]]);
+  v[2] = addv(v[2], m[(size_t)MSG_SCHEDULE[r][4]]);
+  v[3] = addv(v[3], m[(size_t)MSG_SCHEDULE[r][6]]);
+  v[0] = addv(v[0], v[4]);
+  v[1] = addv(v[1], v[5]);
+  v[2] = addv(v[2], v[6]);
+  v[3] = addv(v[3], v[7]);
+  v[12] = xorv(v[12], v[0]);
+  v[13] = xorv(v[13], v[1]);
+  v[14] = xorv(v[14], v[2]);
+  v[15] = xorv(v[15], v[3]);
+  v[12] = rot16(v[12]);
+  v[13] = rot16(v[13]);
+  v[14] = rot16(v[14]);
+  v[15] = rot16(v[15]);
+  v[8] = addv(v[8], v[12]);
+  v[9] = addv(v[9], v[13]);
+  v[10] = addv(v[10], v[14]);
+  v[11] = addv(v[11], v[15]);
+  v[4] = xorv(v[4], v[8]);
+  v[5] = xorv(v[5], v[9]);
+  v[6] = xorv(v[6], v[10]);
+  v[7] = xorv(v[7], v[11]);
+  v[4] = rot12(v[4]);
+  v[5] = rot12(v[5]);
+  v[6] = rot12(v[6]);
+  v[7] = rot12(v[7]);
+  v[0] = addv(v[0], m[(size_t)MSG_SCHEDULE[r][1]]);
+  v[1] = addv(v[1], m[(size_t)MSG_SCHEDULE[r][3]]);
+  v[2] = addv(v[2], m[(size_t)MSG_SCHEDULE[r][5]]);
+  v[3] = addv(v[3], m[(size_t)MSG_SCHEDULE[r][7]]);
+  v[0] = addv(v[0], v[4]);
+  v[1] = addv(v[1], v[5]);
+  v[2] = addv(v[2], v[6]);
+  v[3] = addv(v[3], v[7]);
+  v[12] = xorv(v[12], v[0]);
+  v[13] = xorv(v[13], v[1]);
+  v[14] = xorv(v[14], v[2]);
+  v[15] = xorv(v[15], v[3]);
+  v[12] = rot8(v[12]);
+  v[13] = rot8(v[13]);
+  v[14] = rot8(v[14]);
+  v[15] = rot8(v[15]);
+  v[8] = addv(v[8], v[12]);
+  v[9] = addv(v[9], v[13]);
+  v[10] = addv(v[10], v[14]);
+  v[11] = addv(v[11], v[15]);
+  v[4] = xorv(v[4], v[8]);
+  v[5] = xorv(v[5], v[9]);
+  v[6] = xorv(v[6], v[10]);
+  v[7] = xorv(v[7], v[11]);
+  v[4] = rot7(v[4]);
+  v[5] = rot7(v[5]);
+  v[6] = rot7(v[6]);
+  v[7] = rot7(v[7]);
+
+  v[0] = addv(v[0], m[(size_t)MSG_SCHEDULE[r][8]]);
+  v[1] = addv(v[1], m[(size_t)MSG_SCHEDULE[r][10]]);
+  v[2] = addv(v[2], m[(size_t)MSG_SCHEDULE[r][12]]);
+  v[3] = addv(v[3], m[(size_t)MSG_SCHEDULE[r][14]]);
+  v[0] = addv(v[0], v[5]);
+  v[1] = addv(v[1], v[6]);
+  v[2] = addv(v[2], v[7]);
+  v[3] = addv(v[3], v[4]);
+  v[15] = xorv(v[15], v[0]);
+  v[12] = xorv(v[12], v[1]);
+  v[13] = xorv(v[13], v[2]);
+  v[14] = xorv(v[14], v[3]);
+  v[15] = rot16(v[15]);
+  v[12] = rot16(v[12]);
+  v[13] = rot16(v[13]);
+  v[14] = rot16(v[14]);
+  v[10] = addv(v[10], v[15]);
+  v[11] = addv(v[11], v[12]);
+  v[8] = addv(v[8], v[13]);
+  v[9] = addv(v[9], v[14]);
+  v[5] = xorv(v[5], v[10]);
+  v[6] = xorv(v[6], v[11]);
+  v[7] = xorv(v[7], v[8]);
+  v[4] = xorv(v[4], v[9]);
+  v[5] = rot12(v[5]);
+  v[6] = rot12(v[6]);
+  v[7] = rot12(v[7]);
+  v[4] = rot12(v[4]);
+  v[0] = addv(v[0], m[(size_t)MSG_SCHEDULE[r][9]]);
+  v[1] = addv(v[1], m[(size_t)MSG_SCHEDULE[r][11]]);
+  v[2] = addv(v[2], m[(size_t)MSG_SCHEDULE[r][13]]);
+  v[3] = addv(v[3], m[(size_t)MSG_SCHEDULE[r][15]]);
+  v[0] = addv(v[0], v[5]);
+  v[1] = addv(v[1], v[6]);
+  v[2] = addv(v[2], v[7]);
+  v[3] = addv(v[3], v[4]);
+  v[15] = xorv(v[15], v[0]);
+  v[12] = xorv(v[12], v[1]);
+  v[13] = xorv(v[13], v[2]);
+  v[14] = xorv(v[14], v[3]);
+  v[15] = rot8(v[15]);
+  v[12] = rot8(v[12]);
+  v[13] = rot8(v[13]);
+  v[14] = rot8(v[14]);
+  v[10] = addv(v[10], v[15]);
+  v[11] = addv(v[11], v[12]);
+  v[8] = addv(v[8], v[13]);
+  v[9] = addv(v[9], v[14]);
+  v[5] = xorv(v[5], v[10]);
+  v[6] = xorv(v[6], v[11]);
+  v[7] = xorv(v[7], v[8]);
+  v[4] = xorv(v[4], v[9]);
+  v[5] = rot7(v[5]);
+  v[6] = rot7(v[6]);
+  v[7] = rot7(v[7]);
+  v[4] = rot7(v[4]);
+}
+
+INLINE void transpose_vecs(__m128i vecs[DEGREE]) {
+  // Interleave 32-bit lanes. The low unpack is lanes 00/11 and the high is
+  // 22/33. Note that this doesn't split the vector into two lanes, as the
+  // AVX2 counterparts do.
+  __m128i ab_01 = _mm_unpacklo_epi32(vecs[0], vecs[1]);
+  __m128i ab_23 = _mm_unpackhi_epi32(vecs[0], vecs[1]);
+  __m128i cd_01 = _mm_unpacklo_epi32(vecs[2], vecs[3]);
+  __m128i cd_23 = _mm_unpackhi_epi32(vecs[2], vecs[3]);
+
+  // Interleave 64-bit lanes.
+  __m128i abcd_0 = _mm_unpacklo_epi64(ab_01, cd_01);
+  __m128i abcd_1 = _mm_unpackhi_epi64(ab_01, cd_01);
+  __m128i abcd_2 = _mm_unpacklo_epi64(ab_23, cd_23);
+  __m128i abcd_3 = _mm_unpackhi_epi64(ab_23, cd_23);
+
+  vecs[0] = abcd_0;
+  vecs[1] = abcd_1;
+  vecs[2] = abcd_2;
+  vecs[3] = abcd_3;
+}
+
+INLINE void transpose_msg_vecs(const uint8_t *const *inputs,
+                               size_t block_offset, __m128i out[16]) {
+  out[0] = loadu(&inputs[0][block_offset + 0 * sizeof(__m128i)]);
+  out[1] = loadu(&inputs[1][block_offset + 0 * sizeof(__m128i)]);
+  out[2] = loadu(&inputs[2][block_offset + 0 * sizeof(__m128i)]);
+  out[3] = loadu(&inputs[3][block_offset + 0 * sizeof(__m128i)]);
+  out[4] = loadu(&inputs[0][block_offset + 1 * sizeof(__m128i)]);
+  out[5] = loadu(&inputs[1][block_offset + 1 * sizeof(__m128i)]);
+  out[6] = loadu(&inputs[2][block_offset + 1 * sizeof(__m128i)]);
+  out[7] = loadu(&inputs[3][block_offset + 1 * sizeof(__m128i)]);
+  out[8] = loadu(&inputs[0][block_offset + 2 * sizeof(__m128i)]);
+  out[9] = loadu(&inputs[1][block_offset + 2 * sizeof(__m128i)]);
+  out[10] = loadu(&inputs[2][block_offset + 2 * sizeof(__m128i)]);
+  out[11] = loadu(&inputs[3][block_offset + 2 * sizeof(__m128i)]);
+  out[12] = loadu(&inputs[0][block_offset + 3 * sizeof(__m128i)]);
+  out[13] = loadu(&inputs[1][block_offset + 3 * sizeof(__m128i)]);
+  out[14] = loadu(&inputs[2][block_offset + 3 * sizeof(__m128i)]);
+  out[15] = loadu(&inputs[3][block_offset + 3 * sizeof(__m128i)]);
+  for (size_t i = 0; i < 4; ++i) {
+    _mm_prefetch((const void *)&inputs[i][block_offset + 256], _MM_HINT_T0);
+  }
+  transpose_vecs(&out[0]);
+  transpose_vecs(&out[4]);
+  transpose_vecs(&out[8]);
+  transpose_vecs(&out[12]);
+}
+
+INLINE void load_counters(uint64_t counter, bool increment_counter,
+                          __m128i *out_lo, __m128i *out_hi) {
+  const __m128i mask = _mm_set1_epi32(-(int32_t)increment_counter);
+  const __m128i add0 = _mm_set_epi32(3, 2, 1, 0);
+  const __m128i add1 = _mm_and_si128(mask, add0);
+  __m128i l = _mm_add_epi32(_mm_set1_epi32((int32_t)counter), add1);
+  __m128i carry = _mm_cmpgt_epi32(_mm_xor_si128(add1, _mm_set1_epi32(0x80000000)), 
+                                  _mm_xor_si128(   l, _mm_set1_epi32(0x80000000)));
+  __m128i h = _mm_sub_epi32(_mm_set1_epi32((int32_t)(counter >> 32)), carry);
+  *out_lo = l;
+  *out_hi = h;
+}
+
+static
+void blake3_hash4_sse41(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) {
+  __m128i h_vecs[8] = {
+      set1(key[0]), set1(key[1]), set1(key[2]), set1(key[3]),
+      set1(key[4]), set1(key[5]), set1(key[6]), set1(key[7]),
+  };
+  __m128i counter_low_vec, counter_high_vec;
+  load_counters(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;
+    }
+    __m128i block_len_vec = set1(BLAKE3_BLOCK_LEN);
+    __m128i block_flags_vec = set1(block_flags);
+    __m128i msg_vecs[16];
+    transpose_msg_vecs(inputs, block * BLAKE3_BLOCK_LEN, msg_vecs);
+
+    __m128i 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(IV[0]),     set1(IV[1]),      set1(IV[2]),   set1(IV[3]),
+        counter_low_vec, counter_high_vec, block_len_vec, block_flags_vec,
+    };
+    round_fn(v, msg_vecs, 0);
+    round_fn(v, msg_vecs, 1);
+    round_fn(v, msg_vecs, 2);
+    round_fn(v, msg_vecs, 3);
+    round_fn(v, msg_vecs, 4);
+    round_fn(v, msg_vecs, 5);
+    round_fn(v, msg_vecs, 6);
+    h_vecs[0] = xorv(v[0], v[8]);
+    h_vecs[1] = xorv(v[1], v[9]);
+    h_vecs[2] = xorv(v[2], v[10]);
+    h_vecs[3] = xorv(v[3], v[11]);
+    h_vecs[4] = xorv(v[4], v[12]);
+    h_vecs[5] = xorv(v[5], v[13]);
+    h_vecs[6] = xorv(v[6], v[14]);
+    h_vecs[7] = xorv(v[7], v[15]);
+
+    block_flags = flags;
+  }
+
+  transpose_vecs(&h_vecs[0]);
+  transpose_vecs(&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(h_vecs[0], &out[0 * sizeof(__m128i)]);
+  storeu(h_vecs[4], &out[1 * sizeof(__m128i)]);
+  storeu(h_vecs[1], &out[2 * sizeof(__m128i)]);
+  storeu(h_vecs[5], &out[3 * sizeof(__m128i)]);
+  storeu(h_vecs[2], &out[4 * sizeof(__m128i)]);
+  storeu(h_vecs[6], &out[5 * sizeof(__m128i)]);
+  storeu(h_vecs[3], &out[6 * sizeof(__m128i)]);
+  storeu(h_vecs[7], &out[7 * sizeof(__m128i)]);
+}
+
+INLINE void hash_one_sse41(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;
+    }
+    blake3_compress_in_place_sse41(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_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) {
+  while (num_inputs >= DEGREE) {
+    blake3_hash4_sse41(inputs, blocks, key, counter, increment_counter, flags,
+                       flags_start, flags_end, out);
+    if (increment_counter) {
+      counter += DEGREE;
+    }
+    inputs += DEGREE;
+    num_inputs -= DEGREE;
+    out = &out[DEGREE * BLAKE3_OUT_LEN];
+  }
+  while (num_inputs > 0) {
+    hash_one_sse41(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];
+  }
+}
diff --git a/src/util/blake3/blake3_sse41_x86-64_windows_gnu.S b/src/util/blake3/blake3_sse41_x86-64_windows_gnu.S
index 60d0a4042e7..02083f9d592 100644
--- a/src/util/blake3/blake3_sse41_x86-64_windows_gnu.S
+++ b/src/util/blake3/blake3_sse41_x86-64_windows_gnu.S
@@ -2042,7 +2042,7 @@ blake3_compress_xof_sse41:
         ret
 
 
-.section .rodata
+.section .rdata
 .p2align  6
 BLAKE3_IV:
         .long  0x6A09E667, 0xBB67AE85