git:20261029.0fb7586 to git:20261030.e7612a4

# Benchmarks: file map

> Read when: adding, changing or running a .NET benchmark, or a script that turns benchmark results into the README, docs or a release body.

BenchmarkDotNet projects and the Python and shell scripts around them. benchmarks/README.md is the reference for

every option, class and environment variable: read it instead of guessing. The runtimes outside .NET keep their

benchmarks in `runtimes/*/bench`. `audit-code` audits the scripts, `benchmark` runs them, `audit-perf` and `optimize` use them.

```

README.md options of run.sh, the classes, MessagePack / protobuf-net, member by member, machine code

(asm.py, cross JIT), the runtimes' benchmarks

run.sh the published runtime benchmark: v3 3.3.13 then the full v4 suite on one machine (QUICK=1,

FILTER=...); writes benchmarks/results/ (ignored)

Nino.Benchmark/ the BenchmarkDotNet suite: Nino against MemoryPack, MessagePack and protobuf-net per class

(SimpleTest, PolymorphicTest, GenericPolymorphicTest, PooledTest, PooledGraphTest,

ReferenceTest, VersionTolerantTest, FixedSizeTest, CommonTypesTest, CollectionsTest,

BclTypesTest, ScatteredTest, UntrustedTest), CpuProbe (block copies against Memmove),

Verify.cs (--verify: every case reads back right), NinoJobAttribute (the job, quick mode), PayloadColumn; V3/ +

NINO_BENCH_V3=true: the v3 3.3.13 build of SimpleTest; -p:NinoBenchLegacy for older refs

Nino.GeneratorBenchmark/ the source generator on a synthetic workload: generator time, generated size, emit,

analyzer, re-generation after edits

shards.json the classes and the filters of each CI shard: "report" (report.yml, x64 and Arm64) and

"ab" (ab-bench.yml)

ab.py two github reports of the same benchmarks (base, head): head against base and MemoryPack

asm.py the Tier1 machine code of the cases: dump / diff / compare REF (`compare origin/main`

before committing a read or write path change)

jitwarm/ the startup hook asm.py runs Nino.Benchmark with: calls each case until Tier1, then exits

cross-jit.sh installs the JIT for the other 64-bit architecture, so DOTNET_JitDisasm prints x64 and

Arm64 code on one machine

members/ member-by-member v3 against v4 (run.sh: v3, v4, v3 in process; DISASM=1 for listings):

shared/Parts.cs built as v3/P3.csproj (NuGet 3.3.13) and v4/P4.csproj; table.py,

split-asm.py; results/ ignored

bdn.py the one parser of BenchmarkDotNet github reports, units and FLOOR_NS (1 ns), for the

merge_reports.py merges the reports of parallel jobs into one report (report.yml)

compact_report.py shortens a merged report for a release body (125,000-character limit)

compare_v3.py v3 3.3.13 against v4 from two SimpleTest reports, with both MemoryPack runs

readme.py writes the numbers of README.md / README.zh.md, the chart (docs/assets) and the runtime

tables of docs/v4-benchmarks.md from a joined report

release_body.py the release body's benchmark sections (Performance Report, v3 comparison, Arm64) for

report.yml; the docs site finds them by their headings

runtimes_report.py the release body's Go, Rust and TypeScript sections from each benchmark's output

unity_report.py the release body's Unity section from the IL2CPP player's results (unity.yml artifacts)

test_unity_report.py tests of unity_report.py on fixtures/

fixtures/ unity-bench-<version>/ results.xml + build.txt, and unity-bench-empty/, for those tests

.gitignore results/

```

## Rules for these files

- Headings the release body uses (release_body.py, runtimes_report.py, unity_report.py) are read by the docs site

(Nino.Docs `scripts/sync-benchmarks.ts`): change one only together with its parser there.

- A script that reads BenchmarkDotNet reports does it through `bdn.py` (one parser, one unit table, one floor);

the 1 ns floor is the docs site's (`src/lib/benchmarks.ts`) too.

## Running them

How to run, A/B on runners and what may be published: the `benchmark` skill.