Detailed benchmark reports

These reports retain the full tables, failed comparisons, uncertainty and reproduction details behind the summary. Each describes its measured source and environment; filenames such as current and main identify historical baselines, not the current checkout. Existing report URLs remain unchanged.

2026-10-18

• Same-code calibration — eight process blocks per side and all uncertainty
• CPU dense decision — fixed-work rejection, same-call GEMV attribution and storage observations
• CUDA completion evaluation — 336 primary rows, overall inconclusive; baseline retained

2026-10-17

• Complete mode baseline — standard/explicit, fusion and CPU/CUDA scopes (compiled 445275a)
• Complete fixed A/B evaluation — candidate decisions and all native/framework controls
• CUDA same-call diagnostics — fresh API, Session, prepared and resource costs
• CPU GEMV diagnostics — independent oracles and portable code generation
• Restart-aware optimization decisions — rejected variants, allocation ownership and launch evaluation
• Restart-aware baseline — all CPU/CUDA scopes, remaining losses and seventeen diagnostics

2026-10-16

• Execution reuse and launches — GEMV changes, allocation scopes and remaining losses
• cuBLAS API lifecycle — which construction and synchronization calls cost time?
• CPU/CUDA optimization evaluation — which candidates meet the comparison gate?
• Main baseline — performance before the optimization campaign

2026-10-15

• Cleanup baseline — dense, CSR and sparse strategies before later changes
• CPU implementation results — C++20 comparison and native versus BLAS-backed references
• Sparse solver results — CG/PCG versus official SciPy and SuperLU
• CUDA results — native/cuBLAS versus eager Torch on resident inputs

CUB adoption decision

The CUB report was published on 2026-10-15; it does not record a measurement date. Its source revisions, hardware and sampling protocol identify the comparison. Correctness passed, but the complete adoption gate was not met.

• CUB block reduction — should it replace the existing implementation?

For fresh measurements, use Methodology and reproduction. Raw JSON, frozen binaries and profiler traces referenced by historical reports are local evidence under build/; a clean clone reruns the public commands.