manually prefetch message blocks

1 files changed, 9 insertions, 0 deletions

diff --git a/c/blake3_avx512.c b/c/blake3_avx512.c
index fc754e2..3e07ecb 100644
--- a/c/blake3_avx512.c
+++ b/c/blake3_avx512.c
@@ -467,6 +467,9 @@ INLINE void transpose_msg_vecs4(const uint8_t *const *inputs,
   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(&inputs[i][block_offset + 256], _MM_HINT_T0);
+  }
   transpose_vecs_128(&out[0]);
   transpose_vecs_128(&out[4]);
   transpose_vecs_128(&out[8]);
@@ -720,6 +723,9 @@ INLINE void transpose_msg_vecs8(const uint8_t *const *inputs,
   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(&inputs[i][block_offset + 256], _MM_HINT_T0);
+  }
   transpose_vecs_256(&out[0]);
   transpose_vecs_256(&out[8]);
 }
@@ -1030,6 +1036,9 @@ INLINE void transpose_msg_vecs16(const uint8_t *const *inputs,
   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(&inputs[i][block_offset + 256], _MM_HINT_T0);
+  }
   transpose_vecs_512(out);
 }