aboutsummaryrefslogtreecommitdiff
diff options
context:
space:
mode:
authorPaul Duncan <pabs@pablotron.org>2024-02-24 14:53:57 -0500
committerPaul Duncan <pabs@pablotron.org>2024-02-24 14:53:57 -0500
commit1bae8a7d0816dbcca37d89ee85bcd96d87f1c5b3 (patch)
tree634b3f3a4073cd10371acf96731d6b65e3705330
parentd41372954d22e18a624f9783d90169bc0651b531 (diff)
downloadsha3-1bae8a7d0816dbcca37d89ee85bcd96d87f1c5b3.tar.bz2
sha3-1bae8a7d0816dbcca37d89ee85bcd96d87f1c5b3.zip
sha3.c: xof_squeeze_raw(): improve speed by squeezing in rate-sized chunks
-rw-r--r--sha3.c42
1 files changed, 36 insertions, 6 deletions
diff --git a/sha3.c b/sha3.c
index 34c1664..b6d4f9c 100644
--- a/sha3.c
+++ b/sha3.c
@@ -993,12 +993,42 @@ static inline void xof_absorb_done(sha3_xof_t * const xof, const size_t rate, co
}
// squeeze data without checking state (used by xof_once())
-static inline void xof_squeeze_raw(sha3_xof_t * const xof, const size_t rate, const size_t num_rounds, uint8_t * const dst, const size_t dst_len) {
- for (size_t i = 0; i < dst_len; i++) {
- dst[i] = xof->a.u8[xof->num_bytes++];
- if (xof->num_bytes == rate) {
- permute(xof->a.u64, num_rounds);
- xof->num_bytes = 0;
+static inline void xof_squeeze_raw(sha3_xof_t * const xof, const size_t rate, const size_t num_rounds, uint8_t *dst, size_t dst_len) {
+ if (!xof->num_bytes) {
+ // num_bytes is zero, so we are reading from the start of the
+ // internal state buffer. while `dst_len` is greater than rate,
+ // copy `rate` sized chunks directly from the internal state buffer
+ // to the destination, then permute the internal state. squeeze
+ // rate-sized chunks to destination
+ while (dst_len >= rate) {
+ memcpy(dst, xof->a.u8, rate); // copy rate-sized chunk
+ permute(xof->a.u64, num_rounds); // permute state
+
+ // update destination pointer and length
+ dst += rate;
+ dst_len -= rate;
+ }
+
+ if (dst_len > 0) {
+ // the remaining destination length is less than `rate`, so copy a
+ // `dst_len`-sized chunk from the internal state to the
+ // destination buffer, then update the read byte count.
+
+ // squeeze dst_len-sized block to destination
+ memcpy(dst, xof->a.u8, dst_len); // copy dst_len-sized chunk
+ xof->num_bytes = dst_len; // update read byte count
+ }
+ } else {
+ // fall back to squeezing one byte at a time
+
+ // squeeze bytes to destination
+ for (size_t i = 0; i < dst_len; i++) {
+ dst[i] = xof->a.u8[xof->num_bytes++]; // squeeze byte to destination
+
+ if (xof->num_bytes == rate) {
+ permute(xof->a.u64, num_rounds); // permute state
+ xof->num_bytes = 0; // clear read bytes count
+ }
}
}
}