Architecture and design

Design records: decisions and their rationale. Measured evidence is in Routing and performance evidence.

Kernel selection

Page	What you find there
Flat kernel selection	Families, per-device tuned tables, select::choose, pins, rules R1-R8. API: Kernel selection (developer); inventory: Kernel selection tables.
Tiered tuning	The tuning engine behind the tables: tiers, racing, ledgers, workers.
Tuning data layout	The two files per (op, device): the ledger, the generated table, kernel hashes and the CI check.
Extending BatchLAS	Adding an op or entry point: src/ops/<op>/, tables, tuner spec, tests.
Vendor independence	The vendor seam, the vendor-free build, the coverage instrument; the replaced RouteTable layer as history.
Vendor-free status	Current state of the vendor-free build.

Core model and API

Page	What you find there
Matrix model	Column-major storage, ld/stride/batch, owning matrices vs views, the device-side view.
API conventions	Option structs vs positional overloads, owning arguments, argument checks, per-item info.
Workspace	The per-Queue arena, leases, BumpAllocator sizing.
Error model	Exception hierarchy, convergence reporting.
Environment	Every BATCHLAS_* variable the library reads.
Symbol visibility	What BATCHLAS_API must cover.
Runtime internals	Queue implementation, settings snapshot, embedded tables, template instantiation.
Device-side group BLAS	In-kernel work-group and sub-group BLAS templates.
Verification library	batchlas::verify: the inputs, references, residuals and tolerances shared by tests, benchmarks and the tuner.

Operations

Page	What you find there
syevx: algorithm survey	Selected-eigenpair solvers considered, verdicts.
syevx: range selection	The range API and its hard cases.
gesvd: design	One-sided Jacobi, the bidiagonal pipeline, kernel design rules.

Build

Page	What you find there
CPU target detection	How the build decides whether CPU SYCL kernels exist, and which tests depend on it.
Build performance	Header-structure decisions that keep consumer TUs cheap (e.g. why the umbrella header excludes device code).
Two SYCL implementations	Building with DPC++ or AdaptiveCpp from one tree: gaps, interop seam, A/B protocol, phases, decisions. Measurements: DPC++ vs AdaptiveCpp: performance and accuracy A/B.

Defects and history

Page	What you find there
Known defects	Located-but-unfixed defects.
Flat selection, phase 3 plan (history)	Execution plan for posv, trsm and gemm.
Small-n factorization campaign (history)	Campaign plan and comment-density burn-down; kept for its evidence and cited anchors.