engine-release.yml

# SKaiNET Compute Engine — release benchmark manifest.
#
# Every public benchmark run is parameterized by this file: scenario list,
# warmup/measured-run counts, JDK requirements, JVM flags, and the schema
# version the JSON artifacts are validated against. Bumping any value here
# is a public methodology change — bump the manifest's `manifest_version`
# and call it out in the release notes.

manifest_version: "1.0.0"
schema_version: "1.0.0"

# Toolchain pinning. Both jobs (ubuntu-latest smoke + self-hosted full)
# install the same major JDK; the build itself enforces JDK 21+ via the
# root subprojects block, but we record the version actually used in the
# resulting BenchmarkRecord.runtime.* fields.
toolchain:
  jdk_major: 21
  jvm_args:
    - "--enable-preview"
    - "--add-modules"
    - "jdk.incubator.vector"

# CoV gate — runs above this are marked unstable=true in the JSON record
# and SHOULD NOT be promoted into public leaderboards. Mirrors the PRD
# success metric of 3% CoV on stable CPU lanes.
stability:
  cov_limit_percent: 3.0

# Run shape. Smoke is for CI on ubuntu-latest (≤2 min total); full is for
# the self-hosted i7-9750H lane and matches the PRD recommendation of
# 3 warmup + 5 measured runs.
modes:
  smoke:
    warmup_runs: 1
    measured_runs: 1
    seed: 42
  full:
    warmup_runs: 8
    measured_runs: 5
    seed: 42

# Scenario list. Provider names map to KernelProvider.name returned by
# sk.ainet.backend.api.kernel.KernelRegistry.
scenarios:
  - id: engine-fp32-gemm
    provider: panama
  - id: engine-q4-gemm
    provider: panama
  - id: engine-kernel-matmul
    provider: panama
  - id: engine-kernel-matmul
    provider: scalar
  - id: engine-bf16-matmul
    provider: panama
  - id: engine-bf16-matmul
    provider: scalar
  - id: engine-q8-matmul
    provider: panama
  - id: engine-q8-matmul
    provider: scalar
  - id: engine-elementwise-add
    provider: panama
  - id: engine-reductions-sum
    provider: panama
  - id: engine-reductions-mean
    provider: panama