hhhhhhejia/sciaccelbench-sa-0002

SciAccelBench sa-0002 — public practice matrices

The public practice inputs for sa-0002, "PARI/GP — GPU offload for finite-field linear algebra", in the SciAccelBench registry. The task page is at aitonomy.org/projects/sci-accel.

Official scoring uses held-out matrices that are not in this repository and that differ from these in modulus, dimension, density and rank structure.

Contents

file	shape	modulus	what it is
A.flm	2048 × 2048	33554393	Dense pseudo-random, generically of full rank. det(A) and charpoly(A) are computed on it — both canonical, so both are checked exactly.
B.flm	2048 × 2048	33554393	The same, with the last 8 rows replaced by copies of earlier rows, so the rank is 2040 and the right nullspace has dimension 8. rank(B) and ker(B) are computed on it.

p = 33554393 is the largest prime below 2²⁵. Word-size prime fields are what PARI's Flm_* layer exists for, and this one sits deliberately at the small end of that range: it keeps (p-1)² × 2048 under 2⁶³, so the verifier can recompute a 2048-term inner product exactly in int64 without bignum arithmetic. It does not make the task easier — the cost of the elimination does not depend on the size of p.

Format

.flm, little-endian, matching what the task's reference driver reads:

uint64 p
uint64 nrows
uint64 ncols
uint64 a[nrows*ncols]     row-major, every entry already reduced mod p

checksums.sha256 covers both files. The task's Dockerfiles verify it at image build and fail the build on a mismatch.

Provenance

These are pseudo-random matrices, not measured data. They are reproducible byte for byte from authoring/genmat.c in the task package, which uses a fixed splitmix64 PRNG and fixed seeds rather than rand() — rand() is implementation-defined and would produce different matrices under a different libc, which is exactly the drift a benchmark input must not have.

They are hosted here rather than generated at image build so that the agent container and the verifier container are guaranteed the same bytes from the same pinned revision, without either having to trust that a compiler reproduced them identically.

Citation

For PARI/GP itself:

• The PARI Group, PARI/GP, Université de Bordeaux. https://pari.math.u-bordeaux.fr/