kevin/libsodium
1
Fork 0
mirror of https://github.com/jedisct1/libsodium.git synced 2024-12-23 20:15:19 -07:00
Code Releases Wiki Activity

AES-256-GCM (AES-NI): prefetch the next blocks

Browse Source 
...while computing the GHASH of the previous blocks.

For AMD CPUs with disabled hardware prefetchers, the gain may
be significant.
This commit is contained in:
Frank Denis 2023-01-14 00:01:42 +01:00
parent 67f1b568c5
commit 48af322b7a
1 changed files with 9 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

9
src/libsodium/crypto_aead/aes256gcm/aesni/aead_aes256gcm_aesni.c
View File

@ -444,6 +444,11 @@ aes_gcm_encrypt_generic(const State *st, GHash *sth, unsigned char mac[ABYTES],
counter = incr_counters(rev_counters, counter, PARALLEL_BLOCKS); counter = incr_counters(rev_counters, counter, PARALLEL_BLOCKS);
encrypt_xor_wide(st, dst + i, src + i, rev_counters); encrypt_xor_wide(st, dst + i, src + i, rev_counters);
PREFETCH_READ(src + i + PARALLEL_BLOCKS * 16);
#if PARALLEL_BLOCKS >= 64 / 16
PREFETCH_READ(src + i + PARALLEL_BLOCKS * 16 + 64);
#endif
pi = i - PARALLEL_BLOCKS * 16; pi = i - PARALLEL_BLOCKS * 16;
u = gh_update0(sth, dst + pi, st->hx[2 * PARALLEL_BLOCKS - 1 - 0]); u = gh_update0(sth, dst + pi, st->hx[2 * PARALLEL_BLOCKS - 1 - 0]);
for (j = 1; j < PARALLEL_BLOCKS; j += 1) { for (j = 1; j < PARALLEL_BLOCKS; j += 1) {
@ -454,6 +459,10 @@ aes_gcm_encrypt_generic(const State *st, GHash *sth, unsigned char mac[ABYTES],
encrypt_xor_wide(st, dst + i + PARALLEL_BLOCKS * 16, src + i + PARALLEL_BLOCKS * 16, encrypt_xor_wide(st, dst + i + PARALLEL_BLOCKS * 16, src + i + PARALLEL_BLOCKS * 16,
rev_counters); rev_counters);
PREFETCH_READ(src + i + 2 * PARALLEL_BLOCKS * 16);
#if PARALLEL_BLOCKS >= 64 / 16
PREFETCH_READ(src + i + 2 * PARALLEL_BLOCKS * 16 + 64);
#endif
pi = i; pi = i;
for (j = 0; j < PARALLEL_BLOCKS; j += 1) { for (j = 0; j < PARALLEL_BLOCKS; j += 1) {
gh_update(&u, dst + pi + j * 16, st->hx[PARALLEL_BLOCKS - 1 - j]); gh_update(&u, dst + pi + j * 16, st->hx[PARALLEL_BLOCKS - 1 - j]);