v2.3.3

salmon 2.3.3

A features-and-fixes release on 2.3.2. The default salmon quant output is
unchanged — the new inference accelerators are opt-in (default off), the
profiling additions are pure instrumentation, and the one bug fix is
behavior-neutral. No index rebuild is required.

Optional EM/VBEM acceleration (--emAccel)

salmon's abundance estimation is a fixed-point EM (or VBEM under --useVBOpt)
that converges only linearly; on high-ambiguity / slow-mixing problems it can
crawl, and every bootstrap replicate pays the cost again. Two off-the-shelf
accelerators are now available (thanks
@BenjaminDEMAILLE, #1051, #1052):

• --emAccel squarem — SQUAREM (Varadhan & Roland 2008): a squared
  extrapolation from the last two EM maps plus a stabilizing step. Near-zero
  overhead; a small, safe speedup on typical runs and a large one on
  slow-mixing problems.
• --emAccel daarem — DAAREM (Henderson & Varadhan 2019): damped Anderson
  acceleration over a window of residuals (a multi-secant quasi-Newton step),
  with a self-contained eigensolver (no new dependency). Wins big on hard,
  ill-conditioned problems; because of its per-iteration overhead it is best
  reserved for those cases (it can be slower than plain EM when the base EM
  already converges quickly).

Both reach the same fixed point as plain EM/VBEM (validated: TPM Pearson
≈ 0.9998–0.99999 vs. the default, total mass conserved, identical accuracy vs.
simulation ground truth), and both accelerate the point estimate and every
bootstrap replicate (Gibbs is unaffected). The default is none, so
existing results and --deterministic output are unchanged; the accelerators
are not byte-identical to none (differences are at the convergence-tolerance
level), which is why they are opt-in.

Profiling harness + per-phase timing

• A reproducible profiling harness (#1051, @BenjaminDEMAILLE): a profiling
  build profile, a criterion EM/VBEM M-step benchmark, and scripts/profile.sh
  (samply).
• Per-phase wall-clock timing on the salmon::timing tracing target
  (index_load / mapping / eff_length_collapse / em_bias / posterior / output),
  available inline at info or in isolation via RUST_LOG=salmon::timing=info.
  It now covers all quantification paths — reads (online and
  --deterministic), alignment (-a), and --rad — so the EM phase is visible
  regardless of input mode. Pure instrumentation; no effect on output or
  determinism.

Fixes

• Radix-sort scratch fill (#1055): the (transcript, orientation) uni-MEM
  radix grouping added in 2.3.2 constructed a zero-length Mem as a
  never-read scratch placeholder, which tripped a debug assertion and panicked
  debug builds on real reads. Release builds (all shipped binaries) were
  unaffected — the assertion is compiled out and the placeholder was always
  overwritten — so this changes no results (verified byte-identical quant.sf).
  The fix restores usable debug builds and closes the invariant hole.

Other

• Tidier salmon --help: the removed alevin subcommand is hidden and the
  (Rust port) parenthetical is dropped (#1053, @BenjaminDEMAILLE).

Upgrading

Drop-in from 2.3.2: no index rebuild, and default quantification results are
unchanged. Try --emAccel squarem (or daarem for slow-mixing / large
bootstrap workloads) if you want faster inference.

---

Install salmon-cli 2.3.3

Install prebuilt binaries via shell script

curl --proto '=https' --tlsv1.2 -LsSf https://github.com/COMBINE-lab/salmon/releases/download/v2.3.3/salmon-cli-installer.sh | sh

Download salmon-cli 2.3.3

File	Platform	Checksum
salmon-cli-aarch64-apple-darwin.tar.xz	Apple Silicon macOS	checksum
salmon-cli-x86_64-apple-darwin.tar.xz	Intel macOS	checksum
salmon-cli-aarch64-unknown-linux-gnu.tar.xz	ARM64 Linux	checksum
salmon-cli-x86_64-unknown-linux-gnu.tar.xz	x64 Linux	checksum