Every native kernel competes with a vendor library, and the choice between them is a measured window. These pages record the measurements behind each window: the grid that sets each boundary, the alternatives that were rejected, and the open debts. Read an op's page before widening its window or adding a tier.

Since flat selection (../design/flat-kernel-selection.md), each op ranks its kernels from tuned/<op>.<device>.txt. Where a page quotes a preferred() or supports() predicate with its source line number, that predicate is deleted. Its window lives on as table rows.

Page	Ops	Does anything route natively by default?
dispatch.md	the BATCHLAS_<OP>_ROUTE words, the vendor gate, the coverage instrument, level-3 boundaries	n/a (the mechanism)
gemm.md	gemm	yes: double broadly, float NN squares at max_dim <= 32; complex never
level3.md	symm hemm syrk herk syr2k her2k trmm	symm, syrk, syr2k, trmm: flat selection tables; hemm, herk, her2k: hand-written in cublas.cc
trsm.md	trsm	yes, broadly; read its open debts before trusting a ratio
potrf.md	potrf	yes, per cell; measured tables; five native kernels
qr.md	geqrf orgqr ormqr	yes: ormqr native-first; geqrf above a per-type order floor plus a tall-panel clause; orgqr to n = 512
lu.md	getrf getrs getri	yes: four windows, all float/cfloat-leaning
gemv.md	gemv	yes: one complex<double> transposed window
spmm.md	spmm	yes: the NoTrans gather; the transposed scatter stays vendor-first
steqr.md	steqr_cta (also the fused small-n syev and the stedc leaves)	n/a; no vendor arm
syev.md	syev (Jacobi CTA, fused CTA, blocked, two-stage; vendor)	yes: native small-n and blocked windows per type
syevx.md	syevx (Direct, DirectSubset, Filtered, LOBPCG, stebz/stein)	n/a; no vendor arm for the selected-range solve
stedc.md	stedc (divide and conquer, merge kernels)	n/a; no vendor arm
sytrd.md	sytrd_blocked + latrd, sytrd_sy2sb, sytrd_sb2st_hh and its Q2 back-transform	n/a; internal reductions under syev
ortho.md	ortho (the Gram product and the algorithm rules)	n/a; the Gram product goes through syrk (real, small k) or gemm; host devices force Householder
iluk.md	iluk (ILU(k) numeric phase and apply)	n/a; host or device chosen on batch size
tuning.md	the tuning_params.hh constants and BATCHLAS_TUNE_* overrides	n/a; kernel parameters, not a routing choice
gesvd.md	gesvd	yes: jacobi for real general max(m,n) <= 32 and complex general input up to its ceiling; blocked above for real general; cta for square Hermitian to n = 32
sycl-implementations.md	gemm, potrf, getrf, geqrf, the CTA eigensolvers, stedc and syev built with DPC++ and with AdaptiveCpp: performance and accuracy A/B	n/a; both builds read the same tables and take the same Auto routes

perf index: two rules

The shipped code decides what ships; the notes explain why. Several windows were narrowed or widened after their note was written. Read the op's table and its can_run in src/ops/<op>/<op>.cc. Use the select trace (BATCHLAS_SELECT_TRACE) or the coverage reached row to see which kernel a shape takes.

can_run and table rows are different kinds of false. can_run is correctness: false means the kernel would refuse or return a wrong answer. A table row is speed: a candidate ranked last is slower but stays eligible. A speed threshold in can_run makes the shape unservable vendor-free and turns a pin into a throw.

perf index: measurement rules

At saturation. Below saturation a ratio measures overhead. Where an arm is below saturation, its page says so beside the number.
One harness on the box at a time. The two RTX 4090s share one NUMA node and one UVM driver. A sweep on device 1 alongside one on device 0 read a cell 5.5x slow, while nvidia-smi --query-compute-apps showed no foreign process and rel_sd stayed under 0.02. The effect is cell-specific and intermittent; see lu.md.
JIT warmed. Medians of interleaved A/B. Cells over a relative-sd gate are discarded.
Vendor-free means the build (-DBATCHLAS_ENABLE_VENDOR_BLAS=OFF), never an environment variable inside a build that still links the vendor. There, a pin the shape cannot run throws, and the class word native with nothing runnable falls back to Auto (with a warning), which may pick the vendor, so an A/B can be vendor against vendor.
Every boundary needs a bracketing non-winner. A window edge with no measured loss just outside it is a guess. Where one is missing, the page says so.

perf index: the raw data

The raw data behind the older pages (about 4,800 files and 13 MB of CSV, logs, profiler captures and harnesses) is preserved at the tag perf-evidence/vendor-independence. No build or test reads it, so it is out of the working tree. Every experiments/... path on these pages is an archive path, and resolves only against the tag:

git show perf-evidence/vendor-independence:experiments/wp6_lu/bench/README.md
git show perf-evidence/vendor-independence:experiments/sparse_spmm/verdict.txt
git worktree add /tmp/perf-evidence perf-evidence/vendor-independence

Each page's final section maps its claims to the paths that hold their data.

New raw data lives in benchmarks/results/, in Git LFS

New grids go under benchmarks/results/. .gitattributes routes that directory through Git LFS, so the repository stores a three-line pointer per file. A PR that adds a grid costs about three lines per file in its diff. File names, paths, citations and scripts are unchanged.

git lfs install # once per machine

git lfs pull # fetch the content for the current checkout

Committing without git-lfs installed stores the raw file silently. .github/ci/check_lfs_pointers.py catches this: it fails CI if an LFS-tracked blob is not a pointer, and .github/ci/run_local_checks.sh runs it against the index before the commit. The fix is git lfs install, then git add --renormalize benchmarks/results. No data is lost.

perf index: related pages

../design/vendor-independence.md: how dispatch works and how to add an op to it.
../design/vendor-free-status.md: what the vendor-free build does and does not do.
../design/known-defects.md: located, unfixed defects, with line numbers.