Header grinding for locality on the frozen objectc3d32437(igneum-pow identical at base5e412177). BOUND: ten accepted class v4 programs over 2.6e7 header-chosen hashes sit on the windowed random baseline at mean, min and the 1e-3/1e-4/1e-5 tails for 2 KiB rows, 8 KiB rows, 64 B lines and items, per hash and per unit, closed form and live dataset alike; one-bit header flips move 100 percent of the 4,096 unit addresses; 2 to 4 of 16 load sites are header-predictable, none after iteration 0. Price: the best 1-in-15 header groups run 0.09 percent faster on an RTX A6000 (measured) and cost 15 hashes each; no grind nets 1 percent. One small FINDING: drawn program 0x5d7cc2b09fc6922a repeats a word across two load sites in 1.55 percent of hashes (a register-amount rotate as the only write between two loads from one register), 0.1 percent of loads, header-independent. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
20 lines
1.7 KiB
Text
20 lines
1.7 KiB
Text
[2026-10-07T18:51:07Z] rows: prog=devnet id=0xa785001687d8688a attempt=1 sites=16 plant=none live=false hashes=10000000 shard=0/1
|
|
[2026-10-07T18:57:24Z] done 312500 units (10000000 lane-hashes) in 376.9s
|
|
sites win (0 = whole dataset, 1 = half, 2 = quarter): [2, 2, 2, 1, 0, 0, 0, 0, 2, 0, 2, 0, 2, 1, 0, 2]
|
|
== PER HASH (128 loads) prog=devnet plant=none live=false n=10000000
|
|
metric mean min q1e-3 q1e-4 q1e-5 | windowed baseline mean min q1e-3 q1e-4 q1e-5 | uniform mean min
|
|
rows2KiB 127.981 125 127 126 126 | 127.981 125 127 126 126 | 127.984 125
|
|
rows8KiB 127.924 124 126 126 125 | 127.924 124 126 126 125 | 127.938 124
|
|
lines64B 127.999 126 128 127 127 | 127.999 126 128 127 127 | 128.000 126
|
|
items64B 127.999 126 128 127 127 | 127.999 126 128 127 127 | 128.000 126
|
|
== PER UNIT (4096 loads) prog=devnet plant=none live=false n=312500
|
|
metric mean min q1e-3 q1e-4 q1e-5 | windowed baseline mean min q1e-3 q1e-4 q1e-5 | uniform mean min
|
|
rows2KiB 4076.401 4054 4062 4058 4055 | 4076.387 4054 4062 4058 4055 | 4080.051 4060
|
|
rows8KiB 4018.434 3978 3991 3985 3981 | 4018.409 3974 3991 3985 3978 | 4032.667 3994
|
|
lines64B 4095.384 4089 4092 4091 4090 | 4095.385 4089 4092 4091 4090 | 4095.499 4090
|
|
items64B 4095.384 4089 4092 4091 4090 | 4095.385 4089 4092 4091 4090 | 4095.499 4090
|
|
== SAME-INSTRUCTION COALESCING per unit (lane pairs of one load site sharing a row or line; 128 sites x 496 pairs)
|
|
metric mean max (expected uniform: 2KiB 0.1211, 8KiB 0.4844, 64B 0.00378 per unit)
|
|
pairs row2KiB 0.2973 5
|
|
pairs row8KiB 1.1827 9
|
|
pairs line64B 0.0093 2
|