commit bf5a9ac4e089b661c3170e657862949b722d48fd
parent ab624023d522b689e233f881aae97b52a832e181
Author: finwo <finwo@pm.me>
Date: Wed, 14 Oct 2026 19:02:39 +0200
Fix benchmark scalar split
Diffstat:
2 files changed, 76 insertions(+), 23 deletions(-)
diff --git a/docs/bench.md b/docs/bench.md
@@ -77,28 +77,33 @@ in millions of hashes per second on the 9800X3D. Higher is better.
- `glue`: the tweak over the batch API, rebuilding records each step
- `fused`: the chain kernel
-| hash | backend | plain | glue | fused |
-| -------- | --------- | ----- | ----- | ----- |
-| SHA3-224 | scalar | 5.94 | 5.73 | 5.11 |
-| SHA3-224 | avx2 x4 | 11.92 | 10.54 | 13.51 |
-| SHA3-224 | avx512 x8 | 24.44 | 20.36 | 40.10 |
-| SHA3-256 | scalar | 5.92 | 5.67 | 5.12 |
-| SHA3-256 | avx2 x4 | 11.60 | 10.22 | 13.53 |
-| SHA3-256 | avx512 x8 | 24.05 | 19.73 | 40.44 |
-| SHA3-384 | scalar | 5.83 | 5.45 | 5.12 |
-| SHA3-384 | avx2 x4 | 10.93 | 9.57 | 13.54 |
-| SHA3-384 | avx512 x8 | 21.63 | 17.13 | 40.24 |
-| SHA3-512 | scalar | 5.70 | 5.59 | 5.12 |
-| SHA3-512 | avx2 x4 | 10.50 | 10.03 | 13.59 |
-| SHA3-512 | avx512 x8 | 20.24 | 18.76 | 40.91 |
+| hash | backend | plain | glue | fused |
+| -------- | ---------- | ----- | ----- | ----- |
+| SHA3-224 | scalar | 5.14 | 4.98 | 5.11 |
+| SHA3-224 | scalar+bmi | 5.92 | 5.70 | 5.84 |
+| SHA3-224 | avx2 x4 | 11.81 | 10.77 | 13.43 |
+| SHA3-224 | avx512 x8 | 26.13 | 22.13 | 39.83 |
+| SHA3-256 | scalar | 5.14 | 4.97 | 5.10 |
+| SHA3-256 | scalar+bmi | 5.91 | 5.66 | 5.86 |
+| SHA3-256 | avx2 x4 | 11.74 | 10.64 | 13.50 |
+| SHA3-256 | avx512 x8 | 25.68 | 21.42 | 40.35 |
+| SHA3-384 | scalar | 5.11 | 4.83 | 5.13 |
+| SHA3-384 | scalar+bmi | 5.83 | 5.45 | 5.86 |
+| SHA3-384 | avx2 x4 | 11.20 | 9.56 | 13.53 |
+| SHA3-384 | avx512 x8 | 23.99 | 17.35 | 40.21 |
+| SHA3-512 | scalar | 5.02 | 4.94 | 5.14 |
+| SHA3-512 | scalar+bmi | 5.70 | 5.60 | 5.85 |
+| SHA3-512 | avx2 x4 | 10.92 | 10.50 | 13.59 |
+| SHA3-512 | avx512 x8 | 22.33 | 20.79 | 40.90 |
A chain step is one permutation at every hash size, so the fused kernel holds
-about 40 Mh/s on avx512 from SHA3-224 to SHA3-512 where an untweaked chain
-falls from 24.4 to 20.2. Reaching the chain through the batch API instead costs
+about 40 Mh/s on avx512 from SHA3-224 to SHA3-512, where an untweaked chain
+falls from 26.1 to 22.3. Reaching the chain through the batch API instead costs
a transpose and a scalar absorb per step, which is the `glue` column.
-Scalar has no lanes to fill, so it does not recover the larger block and sits
-about 14% below an untweaked chain.
+The scalar rows have one lane, so the batch API has nothing to overlap and its
+`plain` column lands on the same one-permutation loop as `fused`. `scalar+bmi`
+is the fallback on x86 without AVX, about 15% ahead.
## Notes
diff --git a/test/bench-chain.c b/test/bench-chain.c
@@ -9,11 +9,27 @@
#define KF_CHAIN_PROTO(name) \
void name(size_t, uint8_t *, size_t, const uint8_t *, size_t, size_t, uint8_t, uint32_t, size_t, uint8_t)
+#define KF_HASH_PROTO(name) int name(uint8_t *, size_t, const uint8_t *, size_t)
KF_CHAIN_PROTO(kf_scalar_chain);
KF_CHAIN_PROTO(kf_scalar_chain12);
void kf_scalar_batch(uint64_t states[][25], size_t count);
void kf_scalar_batch12(uint64_t states[][25], size_t count);
+KF_HASH_PROTO(kf_scalar_sha3_224);
+KF_HASH_PROTO(kf_scalar_sha3_256);
+KF_HASH_PROTO(kf_scalar_sha3_384);
+KF_HASH_PROTO(kf_scalar_sha3_512);
+
+#if (defined(__x86_64__) || defined(_M_X64) || defined(__i386__)) && !defined(KECCAK_FAST_NO_BMI)
+KF_CHAIN_PROTO(kf_bmi_chain);
+KF_CHAIN_PROTO(kf_bmi_chain12);
+void kf_bmi_batch(uint64_t states[][25], size_t count);
+void kf_bmi_batch12(uint64_t states[][25], size_t count);
+KF_HASH_PROTO(kf_bmi_sha3_224);
+KF_HASH_PROTO(kf_bmi_sha3_256);
+KF_HASH_PROTO(kf_bmi_sha3_384);
+KF_HASH_PROTO(kf_bmi_sha3_512);
+#endif
#if (defined(__x86_64__) || defined(_M_X64) || defined(__i386__)) && !defined(KECCAK_FAST_NO_AVX2)
KF_CHAIN_PROTO(kf_avx2_chain);
@@ -42,6 +58,7 @@ typedef struct {
kf_chain_fn chain;
kf_chain_fn chain12;
int lanes;
+ kf_hash_fn one[4];
} row;
static uint64_t now_ns(void) {
@@ -66,7 +83,7 @@ int main(int argc, char **argv) {
static const sfinx_hash hashes[] = {SFINX_HASH_224, SFINX_HASH_256, SFINX_HASH_384, SFINX_HASH_512};
static const char *sizes[] = {"SHA3-224", "SHA3-256", "SHA3-384", "SHA3-512"};
unsigned reps = 5;
- row rows[4];
+ row rows[5];
int nr = 0;
for (int i = 1; i + 1 < argc; i += 2) {
@@ -78,20 +95,45 @@ int main(int argc, char **argv) {
reps = 1;
}
- rows[nr++] = (row){"scalar", kf_scalar_batch, kf_scalar_batch12, kf_scalar_chain, kf_scalar_chain12, 1};
+ rows[nr++] = (row){"scalar",
+ kf_scalar_batch,
+ kf_scalar_batch12,
+ kf_scalar_chain,
+ kf_scalar_chain12,
+ 1,
+ {kf_scalar_sha3_224, kf_scalar_sha3_256, kf_scalar_sha3_384, kf_scalar_sha3_512}};
+#if (defined(__x86_64__) || defined(_M_X64) || defined(__i386__)) && !defined(KECCAK_FAST_NO_BMI)
+#if defined(__clang__)
+ if (__builtin_cpu_supports("bmi")) {
+#else
+ if (__builtin_cpu_supports("bmi") && __builtin_cpu_supports("bmi2")) {
+#endif
+ rows[nr++] = (row){"scalar+bmi",
+ kf_bmi_batch,
+ kf_bmi_batch12,
+ kf_bmi_chain,
+ kf_bmi_chain12,
+ 1,
+ {kf_bmi_sha3_224, kf_bmi_sha3_256, kf_bmi_sha3_384, kf_bmi_sha3_512}};
+ }
+#endif
#if (defined(__x86_64__) || defined(_M_X64) || defined(__i386__)) && !defined(KECCAK_FAST_NO_AVX2)
if (__builtin_cpu_supports("avx2")) {
- rows[nr++] = (row){"avx2 x4", kf_avx2_batch, kf_avx2_batch12, kf_avx2_chain, kf_avx2_chain12, 4};
+ rows[nr++] =
+ (row){"avx2 x4", kf_avx2_batch, kf_avx2_batch12, kf_avx2_chain, kf_avx2_chain12, 4, {NULL, NULL, NULL, NULL}};
}
#endif
#if (defined(__x86_64__) || defined(_M_X64) || defined(__i386__)) && !defined(KECCAK_FAST_NO_AVX512)
if (__builtin_cpu_supports("avx512f")) {
- rows[nr++] = (row){"avx512 x8", kf_avx512_batch, kf_avx512_batch12, kf_avx512_chain, kf_avx512_chain12, 8};
+ rows[nr++] = (row){"avx512 x8", kf_avx512_batch, kf_avx512_batch12, kf_avx512_chain, kf_avx512_chain12,
+ 8, {NULL, NULL, NULL, NULL}};
}
#endif
#if defined(__aarch64__) && !defined(KECCAK_FAST_NO_SHA3)
if (kf_armsha3_available()) {
- rows[nr++] = (row){"arm-sha3 x2", kf_armsha3_batch, kf_armsha3_batch12, kf_armsha3_chain, kf_armsha3_chain12, 2};
+ rows[nr++] =
+ (row){"arm-sha3 x2", kf_armsha3_batch, kf_armsha3_batch12, kf_armsha3_chain, kf_armsha3_chain12, 2,
+ {NULL, NULL, NULL, NULL}};
}
#endif
@@ -144,6 +186,12 @@ int main(int argc, char **argv) {
kf_chain_kernel = rows[i].chain;
kf_chain_kernel12 = rows[i].chain12;
kf_chain_lanes = rows[i].lanes;
+ if (rows[i].one[0] != NULL) {
+ kf_sha3_224 = rows[i].one[0];
+ kf_sha3_256 = rows[i].one[1];
+ kf_sha3_384 = rows[i].one[2];
+ kf_sha3_512 = rows[i].one[3];
+ }
for (unsigned rep = 0; rep < reps; rep++) {
uint64_t t0;