Pr90 testland

.github/workflows/release.yml

@@ -90,33 +90,49 @@ jobs:
       fail-fast: false
       matrix:
         include:
+          # Linux targets are built with profile-guided optimization (`pgo:
+          # true`): a fresh synthetic genome+reads fixture (scripts/../
+          # src/bin/gen_pgo_fixture.rs) trains an instrumented build, which is
+          # then rebuilt using that profile. ~10-20% faster alignment in
+          # practice (see docs-old perf notes), output-identical to a plain
+          # build. Restricted to Linux because these are the only matrix
+          # entries that run NATIVELY on their own build runner — PGO needs to
+          # execute the instrumented binary to collect a profile, which is not
+          # possible for the cross-compiled targets below (e.g. macOS x86_64
+          # built on an Apple Silicon runner).
+
           # Linux x86_64 (baseline)
           - name: linux-x86_64
             os: ubuntu-latest
             target: x86_64-unknown-linux-gnu
+            pgo: true
 
           # Linux x86_64 with AVX2 (x86-64-v3)
           - name: linux-x86_64-v3
             os: ubuntu-latest
             target: x86_64-unknown-linux-gnu
             target_cpu: x86-64-v3
+            pgo: true
 
           # Linux x86_64 with AVX-512 (x86-64-v4)
           - name: linux-x86_64-v4
             os: ubuntu-latest
             target: x86_64-unknown-linux-gnu
             target_cpu: x86-64-v4
+            pgo: true
 
           # Linux aarch64 (native arm64 runner, baseline)
           - name: linux-aarch64
             os: ubuntu-24.04-arm
             target: aarch64-unknown-linux-gnu
+            pgo: true
 
           # Linux aarch64 with SVE (Graviton 3+, Axion, Grace)
           - name: linux-aarch64-neoverse-v1
             os: ubuntu-24.04-arm
             target: aarch64-unknown-linux-gnu
             target_cpu: neoverse-v1
+            pgo: true
 
           # macOS x86_64 (cross-compiled on Apple Silicon runner)
           - name: macos-x86_64
@@ -147,8 +163,31 @@ jobs:
         uses: dtolnay/rust-toolchain@631a55b12751854ce901bb631d5902ceb48146f7 # ratchet:dtolnay/rust-toolchain@stable
         with:
           targets: ${{ matrix.target }}
+          components: ${{ matrix.pgo && 'llvm-tools' || '' }}
+
+      - name: Build (PGO)
+        if: matrix.pgo
+        shell: bash
+        run: |
+          # Build a plain (non-instrumented, host-default-target) binary just to
+          # generate the tiny synthetic training fixture and its genome index —
+          # scripts/pgo-build.sh does the actual instrumented + profile-use
+          # builds of the SHIPPED binary below.
+          cargo build --release --bin gen_pgo_fixture --bin rustar-aligner
+          FIXTURE_DIR="$(mktemp -d)"
+          ./target/release/gen_pgo_fixture --out-dir "$FIXTURE_DIR" --n-reads 50000
+          IDX_DIR="$(mktemp -d)"
+          ./target/release/rustar-aligner --runMode genomeGenerate \
+            --genomeDir "$IDX_DIR" --genomeFastaFiles "$FIXTURE_DIR/genome.fa" \
+            --genomeSAindexNbases 8 --runThreadN 2
+          PGO_TARGET="${{ matrix.target }}" \
+          PGO_EXTRA_RUSTFLAGS="${TARGET_CPU:+-Ctarget-cpu=$TARGET_CPU}" \
+            bash scripts/pgo-build.sh "$IDX_DIR" "$FIXTURE_DIR/reads.fastq" 50000
+        env:
+          TARGET_CPU: ${{ matrix.target_cpu || '' }}
 
       - name: Build
+        if: ${{ !matrix.pgo }}
         shell: bash
         run: >
           cargo build --release --target ${{ matrix.target }}

.gitignore

@@ -36,4 +36,13 @@ target
 *.out
 *.log
 *.tab
-*.sam
+*.sam
+
+# Linux build dir used by the solo Docker benchmark/diff (CARGO_TARGET_DIR)
+/target-linux/
+
+# amd64 Linux build dir for the benchmark container
+/target-amd64/
+
+# macOS Finder metadata
+.DS_Store

CHANGELOG.md

@@ -13,6 +13,27 @@ Sections commonly used: Features, Bug fixes, Other changes.
 
 ### Features
 
+- **STARsolo single-cell quantification (`--soloType`)** — the 10x
+  Chromium / plate-based count-matrix pipeline, ported from STAR and
+  verified against real STARsolo (#90).
+
+  - Chemistries: `CB_UMI_Simple` (10x 3'/5'), `CB_UMI_Complex`
+    (multi-segment CB), and `SmartSeq` (manifest-driven, SE read counts
+    / PE fragment counts).
+  - Features: `Gene`, `GeneFull` (pre-mRNA), `SJ` (splice-junction
+    counts), and `Velocyto` (spliced/unspliced/ambiguous matrices for
+    RNA velocity).
+  - Barcode correction: all `--soloCBmatchWLtype` modes
+    (`Exact`/`1MM`/`1MM_multi`/`1MM_multi_pseudocounts`/
+    `1MM_multi_Nbase_pseudocounts`); all `--soloUMIdedup` and
+    `--soloUMIfiltering` modes; real-valued `--soloMultiMappers`
+    (`Uniform`/`PropUnique`/`Rescue`/`EM`).
+  - Output: `raw/` + `filtered/` matrices, `EmptyDrops_CR` cell calling,
+    CellRanger-style `Summary.csv`, and `--soloOutGzip`.
+
+- **`--outSAMattributes GX GN`** — per-read gene-id / gene-name SAM tags
+  for the solo `Gene` assignment.
+
 - **`genomeGenerate` peak RSS cut from ~113 GB → ~11 GB** on the human
   genome (GRCh38, 32 threads). The construction pipeline no longer
   materialises three large intermediates that were dominating the peak:
@@ -76,6 +97,20 @@ Sections commonly used: Features, Bug fixes, Other changes.
   glibc's malloc and per-thread heaps that return whole segments to
   the OS when abandoned, so allocator cache size stays bounded.
 
+### Bug fixes
+
+- **STARsolo `Gene` assignment now requires exon concordance**, matching
+  STARsolo: a read counts toward a gene only when every aligned block
+  lies within the gene's exons, rather than merely overlapping one. This
+  removes over-assignment of reads that extend past an exon boundary into
+  an intron. On the mouse-chr19 differential test the `Gene` count matrix
+  now matches STAR 2.7.11b (raw matrix identical apart from the
+  multimapper tie-break tail that even a byte-faithful reimplementation
+  cannot reproduce).
+- **STARsolo `features.tsv` column 2 now emits the GTF `gene_name`**
+  (symbol), with the STAR gene_id fallback, instead of duplicating the
+  gene_id.
+
 ### Bumps
 
 - `caps-sa` → `0.5` (adds `build_ext_mem_for_filter*`; see the caps-sa

Cargo.toml

@@ -35,8 +35,12 @@ memmap2 = "0.9"
 byteorder = "1"
 noodles = { version = "0.111", features = ["fastq", "sam", "bam", "bgzf"] }
 bstr = "1"
-flate2 = "1"
+# Use the pure-Rust `zlib-rs` backend instead of flate2's default `miniz_oxide`:
+# ~2-3x faster inflate/deflate on the FASTQ decode + BGZF (BAM) paths, with no C
+# toolchain dependency. Backend is chosen at build time; API is unchanged.
+flate2 = { version = "1", default-features = false, features = ["zlib-rs"] }
 rayon = "1"
+rustc-hash = "2"
 dashmap = "6"
 chrono = "0.4"
 rand = "0.10"
@@ -59,6 +63,9 @@ caps-sa = "0.6"
 # extra cost in release, but the default features add a few hundred
 # kB of fast-path stubs we don't need).
 mimalloc = { version = "0.1", default-features = false }
+libmimalloc-sys = { version = "0.1.49", features = ["extended"] } # mi_option_set (purge_delay); see main.rs
+libdeflater = "1.25.2"
+noodles-bgzf = { version = "0.47", features = ["libdeflate"] }
 
 [dev-dependencies]
 assert_cmd = "2"