|
19 | 19 | > results are for THIS workload, THIS machine, ONE corpus — see "Scope" at |
20 | 20 | > the end. |
21 | 21 | > |
22 | | -> **Lane S provenance.** Lane S (SWAR) is NOT in this PR's diff — it ships in |
23 | | -> the companion PR (`onebrc-probe: lane S`). Its rows below were measured with |
24 | | -> that PR applied; on **this branch alone**, the runnable lanes are |
25 | | -> `a c r f t8 t` and the `lane_s_agrees_with_lane_a` parity test lives in the |
26 | | -> companion PR. S is kept in the ladder because the report's whole point is the |
27 | | -> full RAM-table-vs-trie comparison, and S is the fastest RAM-table method. |
| 22 | +> **Lane S provenance.** Lane S (SWAR) shipped separately (PR #637, merged to |
| 23 | +> `main`); this branch is rebased on top of it, so `lane_s` and its |
| 24 | +> `lane_s_agrees_with_lane_a` parity test are present here and all lanes in the |
| 25 | +> ladder — `a c r f t8 t s` — are runnable. S is kept in the ladder because the |
| 26 | +> report's whole point is the full RAM-table-vs-trie comparison, and S is the |
| 27 | +> fastest RAM-table method. |
28 | 28 |
|
29 | 29 | ## The methods (group-by-aggregate, min/max/sum/count per station) |
30 | 30 |
|
@@ -155,7 +155,7 @@ count (10M). Not a claim about other CPUs, other cardinalities, or the full |
155 | 155 | cd crates/onebrc-probe |
156 | 156 | cargo build --release |
157 | 157 | target/release/onebrc-probe gen /tmp/brc10m.txt 10000000 42 # if absent |
158 | | -for lane in a c r f t8 t; do # add `s` once the companion lane-S PR is merged |
| 158 | +for lane in a c r f t8 t s; do |
159 | 159 | for i in $(seq 1 11); do target/release/onebrc-probe run /tmp/brc10m.txt $lane 4; done |
160 | 160 | done # take median + min/max/sd per lane, not best-of-N |
161 | 161 | # native: RUSTFLAGS="-C target-cpu=native" cargo build --release --target-dir target-native |
|
0 commit comments