|
Everett
|
I keep source, fixture, command and result names consistent with the current Everett API throughout pre-release development.
Timing and allocation samples retain their measured values. Recorded measurement hashes identify the original source and binary bytes, before name normalization; they do not claim that renamed source was remeasured. The artifact manifest records original and normalized artifact hashes separately.
The snapshot helper extracts a pinned revision into the current include layout and records both original and normalized source hashes. Use the current runners with the revisions named in each report. Reproduction makes new measurements with normalized source names; it does not replace the recorded observations.
New measurements of Everett should record their own source hashes and results. The implementation ledger links the accepted measurements and describes the current library.
For current sources, build the C++26 CMake experiments using the module toolchain. They select an explicit CPU architecture and record the configured compiler. The historical Python comparison runners retain the compiler settings for their recorded revisions; supply the candidate revision named in the report when replaying one.
The bounded fixed-key search comparison measures scalar and SIMD windows on NEON and AVX2, with protected-page correctness checks and conservative automatic choices.
The byte-profile GPU driver merges mapped KV02 files directly, including GPU frame parsing and Elias–Fano construction. It compares complete output bytes against the ordinary CPU writer. Its timing boundary includes fresh output mappings, checksums and cleanup; fractional indexes and durable publication remain outside it. Its M2 Max measurements retain all trials at three input sizes, including small-input losses and process variability.
The GPU merge experiment reads compressed mapped inputs and emits a complete compressed output file. Its measurements include parsing, merging, navigation construction and checksums. The collision-rank comparison explores temporary cancellation bitmaps; the production rank formats stay unchanged. The calibration guide describes a measured CPU/GPU choice using headers and file sizes. The M2 Max results retain the complete 50-case calibration and its five successful held-out GPU selections. The Vulkan qualification records 150 successful native correctness runs on an RTX 4090; it contains no performance claim. The tiled-prefix comparison reduces tree-construction dispatches but has mixed complete-path results, so the tiled path remains an explicit experiment. The output-plan comparison removes one host wait, with mixed complete-path results. It also remains opt-in.
This is a separate opt-in Metal program. It does not publish durable catalog updates, and the ordinary library has no shader-toolchain dependency.
The fixed-key experiment separately constructs native .ff and .fv candidates, including scheduled cancellations, actual payload copying and complete native Elias–Fano output. It uses an experimental envelope and has its own correctness and timing boundaries. Its M2 Max measurements retain both random-select and stronger forward-decoding CPU baselines, payload-size and cancellation sweeps, skewed inputs and scenario-specific crossover brackets.
The matched format comparison gives both Metal paths identical logical keys and values, with fresh output mappings and checksums included. It reports complete latency and encoded size for structured and hash-like keys, preserving all trials and their variability.
The matched CPU profile comparison uses those same logical records in raw and typed bit/byte files. It measures complete native construction and reports exact file sizes, before/after byte transport results and the bit-path controls.
The current matched lookup comparison uses identical logical tables and queries in byte and bit formats, including the reservoir decoder. It reports complete-file space alongside whole-query latency over three table sizes and retains CPU/wall diagnostics for every trial.
The select comparison gives every candidate the same sampled offsets from native and fractional-index fixtures. It separates alternatives to Elias–Fano from changes to its internal selector, and reports construction, array and resident space, query patterns, forward scans and larger replayed-gap cases. Production navigation remains unchanged.
1.9.8