Drop in
Zero code changes. Preload the library and the covered routines are served by VVE: dsyev · dsyevd · dsyevr · dgesdd · dgesvd · dstev. Python through vvepy. This is the speed path.
LD_PRELOAD=libvve.so ./your_doa_pipelineArray signal processing
Direction of arrival, on hardware that has no room for a BLAS.
MUSIC and ESPRIT are eigendecomposition in a latency budget. The array covariance matrix has to be decomposed, the signal and noise subspaces split, and the spectrum scanned — inside a frame, on a board that has to fit and draw the power it is allowed, producing a result that will later have to be qualified. This is the one workload that needs all four VVE advantages at the same time.
Wilkinson, N=16,383
89.6×
vs single-threaded reference netlib LAPACK dsyev (the default dense path) — wilkinson, N = 16,383
Wilkinson, N = 45,056
5.16×
vs LAPACK's own expert O(N²) routine, single-threaded dstemr (MRRR)
libvve.so
1.4MB
libvve.so, zero external dependencies
x86_64 ≡ aarch64
identical bits, same source
Where the eigensolver sits.
Snapshot averaging, spatial smoothing and the spectral scan are all well-behaved. The eigendecomposition is the block that grows with the array, pulls in a numerical library, and produces the number that has to be identical next year on next year's board.
Four requirements. Four advantages.
| The workload requires | VVE answers with |
|---|---|
| A frame-rate latency budget. | Structure-aware O(N²) routing — 89.6× vs single-threaded reference netlib dsyev at N = 16,383, and up to 5.16× vs single-threaded LAPACK dstemr on clustered spectra; both comparisons are single-threaded on both sides. |
| A board with no room for a tuned BLAS. | 1.4 MB, zero external dependencies. Runtime dependencies are libc, libm and libpthread. x86_64 and aarch64 from the same source. |
| Evidence that survives qualification. | Bit-identical output across machine, architecture, compiler and hardware generation — 96 SHA-256 records across 12 groups, zero mismatches, 541 / 541 across CPU generations. |
| Correlated sources. | Clustered spectra are where VVE's margin is largest and where it holds at every size measured — the wilkinson and schrödinger families in the campaign. |
Why correlated sources favour this solver.
Subspace methods need a clean split between the signal eigenvalues and the noise floor. Correlated and coherent arrivals collapse that separation — the signal eigenvalues bunch, and the classical response is to spend more work resolving them. MRRR resolves a cluster by refining its representation tree, and that refinement is exactly where its cost goes up. VVE's routing does not pay that penalty, which is why the two clustered families in the campaign are the two where it wins at every size measured.
Scoped to this workload.
Wilkinson, N=16,383
89.6×
vs single-threaded reference netlib LAPACK dsyev (the default dense path) — wilkinson, N = 16,383, full decomposition
Full decomposition, N=45,056, Wilkinson spectrum
5.16×
vs LAPACK's own expert O(N²) routine, single-threaded dstemr (MRRR) — wilkinson, N = 45,056, accuracy gated on both sides
Wilkinson, N=45,056, wall clock
171.1s
dstemr vs VVE, wilkinson, N = 45,056, both single-threaded
96 / 96
SHA-256 records identical across x86_64 ≡ aarch64 and GCC ≡ Clang
Two ways in.
Bands-native
One small code change, and the dense buffer never exists. You pass the diagonal and the off-diagonal directly. This is the memory path, and it is what makes large arrays fit in a board's memory budget rather than a server's.
/* pass D and E — no dense N×N is ever allocated */
vve_eigen_values_tridiag(D, E, N, W);Send us your array geometry.
Array size, snapshot count, source correlation, target board and instruction set. We will run your case on the harness and send the numbers back with the build, so the first result you see comes from your problem rather than ours.