ELPA dense eigensolver on Orange Pi RV2: stock EESSI OpenBLAS RVV finishes in 42.17 s but ev0=nan (finite=0). Patched gemv_n → 34.81 s, finite=1, 1.58× scalar. Same FlexiBLAS swap as HPL and QE — no ELPA rebuild.
Two RVV BLAS libraries on SpaceMiT X60 — BLIS vs patched OpenBLAS. Square DGEMM 1T N=4096: BLIS 1.29× OpenBLAS; 8T OpenBLAS scales better. verify_ctrsm 2400/2400. HPL on static libblis.a is PASSED but only 0.35–0.53× OpenBLAS-RVV.
RVV FFTW is real on SpaceMiT X60 (1.60× at N=256 under MEASURE), but the same lib in Quantum ESPRESSO yields 1.002× whole SCF. Planner trap: ESTIMATE→MEASURE is worth 3–5×; QE uses ESTIMATE across ~3260 small FFTs.
Stock EESSI OpenBLAS aborts Quantum ESPRESSO in inverse_s (MPI_ABORT) — the same gemv_n bug that NaN’d HPL. Patched RVV: −14.57861334 Ry, then a 64-atom SCF 144.8 → 110.3 s (1.31×). calbec ~2×; fftw ~1.01×. Measured on Banana Pi F3.
ONNX Runtime int4 MatMulNBits on Orange Pi RV2 spent ~32 s until accuracy_level=4 selected the X60 IME path — 9.1× (1T) and 10.3× (8T). Config beat kernels; a hand RVV “fast path” was −28% and reverted.
RVV FFT wins ~1.23× on PME 3D-FFT but Force is ~90% of the run — whole-app barely moves. A hand-written impl_riscv_rvv/ Force backend delivers 3.31× whole-app (0.380 → 1.256 ns/day) on Orange Pi RV2.
Stock EESSI OpenBLAS on an Orange Pi RV2 reports ~8.5 GFLOP/s on HPL — then residual nan, FAILED. We isolate 192 NaNs in RVV dgemv with difftest, and fix it via OpenBLAS#5408 / easyconfigs#26444 through FlexiBLAS (10.53 GFLOP/s PASSED).
How we lifted HPL on a StarFive VisionFive 2 from 3.13 to 5.28 GFLOP/s without rebuilding HPL — by swapping in a SiFive U74-tuned OpenBLAS kernel through EESSI and FlexiBLAS.