Author: @jurajselep
Date: 2026-09-04
Problem: 16x16 matmul
Cost: 65,084
IR: best_65084.ir
Verifier: best_65084.py
This improves the previous 66,178 record by 1,094 read-cost units (1.65%) without changing the arithmetic circuit.
The computation processes 4x8 output blocks. Selected contraction passes reverse the column order, so each pass consumes the previous partial sums newest first. This LIFO snake order nests accumulator lifetimes instead of crossing them. A values are staged immediately before their first multiply, and the resulting live intervals are assigned to address-cost tiers. The address assignment is locally optimized; no global-optimality claim is made.
| instruction | count | paid reads | read cost |
|---|---|---|---|
mul |
4,096 | 8,192 | 14,300 |
add |
3,840 | 7,680 | 24,897 |
copy |
1,649 | 1,649 | 21,539 |
| output | – | 256 | 4,348 |
| total | 9,585 | 17,777 | 65,084 |
python3 matmul/submissions/best_65084.py
This checks the SHA-256, official score, operation counts, read-cost breakdown, and all 256 outputs as exact noncommutative integer polynomials.