Repository navigation

perf: widening 32x32 multiply in the vector kernels, extracted into VectorMath - #8

Merged

unsafePtr merged 20 commits into

masterfrom

perf/widening-multiply-32

Aug 7, 2026

Merged

perf: widening 32x32 multiply in the vector kernels, extracted into VectorMath#8
unsafePtr merged 20 commits into
masterfrom
perf/widening-multiply-32

Commits on Jul 27, 2026
- refactor: move vector kernels into VectorMath
  unsafePtr
  committed
  74ba40a
- perf: use widening 32x32 multiply in vector kernels
  unsafePtr
  committed
  9efcd13
- test: cover VectorMath kernels directly
  unsafePtr
  committed
  0b82f2a
- test: cover fast-path rejection of values too large for 32/64 bytes
  unsafePtr
  committed
  210d6d0
- refactor: fast decode paths return bool, drop misleading Try prefix
  unsafePtr
  committed
  f64e22d
- refactor: mark stateless Base58 helpers static
  unsafePtr
  committed
  aaa6951
- ci: add arm64 codegen probe for the vector kernels
  unsafePtr
  committed
  175c4fb
- ci: fix arm64 codegen probe filter and fail on empty dump
  unsafePtr
  committed
  b2b9aec
- perf: use uzp1 and umull for the widening multiply on arm64
  unsafePtr
  committed
  36f4555
- perf: narrow with xtn instead of uzp1 in the arm64 widening multiply
  unsafePtr
  committed
  d1ef6ef
- ci: probe arm64 on neoverse and apple silicon
  unsafePtr
  committed
  dcd9bdc
- ci: add cross-platform benchmark comparison against the branch baseline
  unsafePtr
  committed
  cd141ac
- chore: drop the net9 benchmark jobs
  unsafePtr
  committed
  548fb9c
- Revert "ci: add cross-platform benchmark comparison against the branch baseline"
  unsafePtr
  committed
  8f70bea
Commits on Aug 6, 2026
- docs: link the upstream issue and benchmark behind the safe-rewrite note
  unsafePtr
  committed
  9106039
- perf: hoist the prepend-zero fill out of the fast decode loop
  unsafePtr
  committed
  059b6d3
Commits on Aug 7, 2026

Navigation Menu

Repository navigation

perf: widening 32x32 multiply in the vector kernels, extracted into VectorMath - #8

perf: widening 32x32 multiply in the vector kernels, extracted into VectorMath#8
unsafePtr merged 20 commits into
masterfrom
perf/widening-multiply-32

Commits on Jul 27, 2026

refactor: move vector kernels into VectorMath

perf: use widening 32x32 multiply in vector kernels

test: cover VectorMath kernels directly

test: cover fast-path rejection of values too large for 32/64 bytes

refactor: fast decode paths return bool, drop misleading Try prefix

refactor: mark stateless Base58 helpers static

ci: add arm64 codegen probe for the vector kernels

ci: fix arm64 codegen probe filter and fail on empty dump

perf: use uzp1 and umull for the widening multiply on arm64

perf: narrow with xtn instead of uzp1 in the arm64 widening multiply

ci: probe arm64 on neoverse and apple silicon

ci: add cross-platform benchmark comparison against the branch baseline

chore: drop the net9 benchmark jobs

Revert "ci: add cross-platform benchmark comparison against the branch baseline"

Commits on Aug 6, 2026

docs: link the upstream issue and benchmark behind the safe-rewrite note

perf: hoist the prepend-zero fill out of the fast decode loop

Commits on Aug 7, 2026

refactor: slice to the fixed lengths in the bitcoin encode fast paths

chore: hold SimpleBase at 5.6.2 pending upstream encoder fix

docs: note the nuget cache mount for the big-endian docker run

test: run the vector fallback tests by default