amethyst

Files

T

Claude b71641a15f fix: fair C-to-C benchmark — ACINQ sign must include keypair_create

The previous benchmark unfairly gave ACINQ a cached keypair (created
once outside the loop) while our sign derived the pubkey each call.

Fixed to test both patterns:
- "sign (full)": both derive pubkey each call
  ACINQ: keypair_create + sign32 = 36.3µs
  Ours: ecmult_gen + sign_internal = 33.6µs → 1.08x faster

- "sign (cached)": both reuse precomputed pubkey
  ACINQ: sign32 with cached keypair = 18.8µs
  Ours: signXOnly with cached xonly = 17.4µs → 1.08x faster

Fair native C-to-C results (x86_64, BMI2):
  pubkeyCreate:          ACINQ 18.0  Ours 16.3  1.10x faster ✓
  sign (full):           ACINQ 36.3  Ours 33.6  1.08x faster ✓
  sign (cached):         ACINQ 18.8  Ours 17.4  1.08x faster ✓
  verify (BIP-340):      ACINQ 36.7  Ours 40.4  0.91x slower ✗
  verifyFast (Nostr):    ACINQ 36.7  Ours 34.5  1.06x faster ✓
  ECDH (cached):         ACINQ 37.2  Ours 35.4  1.05x faster ✓
  batch(200):            ACINQ 42.4  Ours  8.3  5.1x faster  ✓

We beat ACINQ on 5 of 6 operations. The only loss is full BIP-340
verify (0.91x) due to ACINQ's 5x52+ADCX/ADOX field assembly.

https://claude.ai/code/session_011KVZhDcV2G7idNWEBz12GY

2026-04-11 12:38:03 +00:00

.idea

…

benchmarks

feat: upgrade negentropy-kmp to v1.0.2, enable macosArm64 target

2026-04-11 00:44:33 +00:00

src

fix: fair C-to-C benchmark — ACINQ sign must include keypair_create

2026-04-11 12:38:03 +00:00

.gitignore

…

build.gradle.kts