BLAS full-overwrite: matched-provider instruction evidence

Local candidate, not a native-speedup claim or merge-ready dependency chain.

Results

Same migrated baseline, strided revision, compiler, OpenBLAS and 1T settings; only tenferro's linked-BLAS full-overwrite implementation differs.

Workload	Baseline Ir / operation	Candidate Ir / operation	Paired reduction
b32 × 128 × 128 × 128	43,895,074	39,696,305	9.560–9.569%
GEMM 1024	533,358,520	524,962,594	1.573–1.574%
LM sentence 3, 12d	1,009,589,326.5	968,164,607.75	4.103%, one exploratory pair

Binary cases report medians of three independent N3/N1 pairs. Every pair passes the predeclared instruction gate: BMM reduction >=5%, GEMM control regression <=1%. LM was an additional check, not a predeclared acceptance workload; its single pair is not a repeated-measurement performance claim. GEMM kernel self-Ir is identical in every baseline/candidate pair (33,273,536 for BMM, 512,386,632 for GEMM, and path-mean 444,650,842 for LM).

The candidate removes redundant Rust output initialization when BLAS beta=0 fully overwrites the result. It does not change the GEMM math kernel, add a batched BLAS API, remove canonical packing, or eliminate eager input copies. Empty contractions still initialize zeros; nonzero beta stays initialized. Injected providers do not advertise the new witness: their existing unsafe registration contract promises ABI compatibility rather than full overwrite.

Julia workloads were active, including in CPU16's L3 domain (CPUs16–23). No native timings, bandwidth figures, time shares or predicted speedups are reported. Timing fields in diagnostic/profiler stdout are intentionally ignored.

PyTorch and batched GEMM findings

Exact wheel source: PyTorch 0d62256a2b23365f8e1604297eb23a6545102aa8:

LinearAlgebra.cpp: CPU bmm_out_or_baddbmm_ chooses the native small kernel when M*N*K < 400.
CPUBlas.cpp: eligible float64 batches can use MKL batching; the non-MKL implementation loops over ordinary GEMM. This source-built OpenBLAS wheel is non-MKL.
Installed OpenBLAS 0.3.26 exports no gemm_batch symbol. This is a finding about this exact provider, not all newer OpenBLAS releases. Tenferro already contains feature-gated grouped batching; adding a batch-shaped wrapper around the current system-OpenBLAS loop is not vendor batching acceleration.
raw/torch-shapes.jsonl records both LM paths at actual mm/bmm boundaries. Batches of 1100 include 12×12×12, 11×1×11, and 1×1×11 products. The latter two are below PyTorch's threshold. Observed BMM operands have nonzero batch strides, not shared-operand broadcast strides; they cannot simply be folded into one ordinary GEMM without extra work.
raw/rust-gemm.log.gz contains 61,688 calls for the entire diagnostic invocation, not a claimed count per contraction. This older shape trace uses tenferro d8759f / strided fbd10f; it is not the candidate performance measurement.

This justifies investigating small independent products separately, but not copying PyTorch's threshold into Rust without matching-kernel evidence.

Provenance and validation

Benchmark source: 976c4b9a459f94086c714ec0f6836978a7bbdfc0.
Baseline tenferro: 03ec9804beca51dce6355d1751e10251ff2972d4.
Candidate tenferro: 57bee76de1a752782bdd985f332050422f885372.
Fixed strided: 17e05ffb168d0f826529ea2c3aceeb4ec851448b.
OpenBLAS 0.3.26, LP64/pthreads; Rust 1.98.1 / LLVM22.1.8 for Docker builds; Valgrind3.22. CPU16 affinity and explicitly configured 1T, not inferred from affinity. verify_threads.c separately queried openblas_get_num_threads() inside every intercepted dgemm call, aborting unless it was 1. It was never preloaded into measurements. Environment, image/library/binary hashes and Cargo features are in raw/environment.txt and raw/final-build-features.log.gz.
The preserved pre-edit baseline diff matches committed 03ec980 exactly.
Combined faer/BLAS: 564 library tests passed. Faer-only: 543 passed.
BLAS-only optimized build: 6 focused tests passed, covering four scalar types, alpha=0, beta rejection, nonzero products, transpose, views, empty contractions/extents, batches, invalid storage, and untouched unsupported output.
Inject-only build checked; its no-uninit-witness regression test passed.
Optimized focused tests under Memcheck: zero errors; leak checking disabled. Earlier debug Memcheck also reported zero errors. This is not a leak-free claim.
Existing pooled-output/nonzero-beta regression tests remain in the library run. Benchmark fixture runs themselves are zero-input smoke checks, not the nonzero numerical oracle; the focused tests provide that oracle.
Normal dependency pins still need the unpublished upstream strided commits. Tests/builds use the retained local overrides. No push or PR was performed. Quiet-host native validation and the broader integration gate remain pending; pre-existing strict-Clippy failures from the migration have not been resolved.

Reproduction and retained evidence

Use the existing OpenBLAS image identified in raw/environment.txt, mount the workspace at the same absolute path (the harness embeds CARGO_MANIFEST_DIR), and point extern/tenferro-rs at separate baseline/candidate worktrees and extern/strided-rs at the fixed strided worktree. Do not overwrite a dirty checkout. raw/blas-bench-patch.toml records all six local strided overrides; raw/benchmark-Cargo.lock records dependency resolution. Sequential build command:

RUSTC_WRAPPER= CARGO_PROFILE_RELEASE_DEBUG=1 cargo build -j 16 \
  --config /tmp/local-patch.toml --release --no-default-features \
  --features system-openblas --bin tenferro-einsum-benchmark

Preserve each binary before rebuilding. The build container used CARGO_HOME=/tmp/bench-cargo (mounted host Cargo cache), RUSTUP_HOME=/home/vscode/.rustup, and HOME=/tmp/bench-home. The archived patch is mounted as /tmp/local-patch.toml.

Run reproduce.sh in the benchmark worktree inside the image, supplying the two binaries and output directory. Baseline then candidate, never concurrent. It uses configure_cpu_thread_env 1, explicit eager/BLAS, CPU16, one fixed warmup, and N1/N3. The whole-process difference cancels fixed preparation; small per-repeat harness/statistics overhead remains, so these are warmed eager operation counts, not kernel-only windows. Identical binary paths share the harness's strategy cache: divide by 2. LM has two distinct paths: divide by 4.

python3 result/amd-cpu/blas-full-overwrite/summarize.py
# To inspect a compressed raw profile:
gzip -dc raw/final/PROFILE.callgrind.gz > /tmp/profile.callgrind
callgrind_annotate --auto=no --inclusive=no --threshold=100 /tmp/profile.callgrind

Raw profiles, annotations, logs and the source patch are gzip-compressed without changing their bytes (including original tool-output whitespace). summary.json is generated from the raw profiles; the parser was checked against both uncompressed and compressed evidence with identical output. Raw logs include the initial candidate 133ca79 trial (not used for the final table) and a rejected missing-fixtures launch. The final table uses only raw/final/. Non-einsum JSON fixture warnings are loader skips, identical across variants. Auto source annotation produced a Valgrind/Perl warning; final annotations disable automatic source annotation and retain complete function self-cost tables.