NumPy

A SciPy-stack probe for BLAS and LAPACK through NumPy — from opensolvers/benchmarks/numpy (bench_blas.py).

Kernel	NumPy call	Backend	Metric
DGEMM	A @ B	dgemm	GFLOP/s
EIGH	np.linalg.eigvalsh(S)	dsyevd	seconds

Unlike raw OpenBLAS verification, eigvalsh exercises BLAS-2 tridiagonalization — a correctness gate for the same gemv_n bug that breaks HPL and ELPA.

module load SciPy-bundle/2025.07-gfbf-2025b
OMP_NUM_THREADS=8 OPENBLAS_NUM_THREADS=8 python3 bench_blas.py [dgemm_N] [eigh_N]
# defaults: dgemm_N=4096  eigh_N=2048

Swap backends per run: OPENBLAS_CORETYPE=RISCV64_GENERIC or FLEXIBLAS=/path/to/libopenblas.so.

Orange Pi RV2 — re-verify (2026-08-22)

8 threads, SciPy-bundle 2025.07 (numpy 2.3.2), N_dgemm=4096, N_eig=2048:

Tag	DGEMM GFLOP/s	EIGH s	finite
scalar (RISCV64_GENERIC)	4.87	10.61	yes
stock (default FlexiBLAS OpenBLAS)	10.06	—	eig fails (LinAlgError)
patched RVV	15.24	6.57	yes

Patched vs scalar: DGEMM 3.13×, EIGH 1.61×. Stock RVV dgemm is fast but eigvalsh aborts (same gemv_n NaN path as HPL).

Earlier RV2 A/B (same sizes): scalar 4.77 → patched 11.52 GFLOP/s (2.4×); EIGH 1.6×. See LAPACK.

Banana Pi F3 (cross-board, 8 threads)

Kernel	Scalar	Patched RVV	Speedup
DGEMM N=4096	4.91 GFLOP/s	17.51 GFLOP/s	3.6×
EIGH N=2048	9.59 s	5.94 s	1.6×

Stock unpatched RVV: eigvalsh raises LinAlgError: Eigenvalues did not converge; DGEMM alone stays finite at 10.94 GFLOP/s (only gemv is broken — see difftest).

Eigensolver speedup (~1.6×) is lower than DGEMM (~2.4–3.6×) because eigvalsh mixes BLAS-3 with latency-bound BLAS-2 — same pattern as ELPA.