Peak Performance in Matched Filtering: Fused Correlation and Statistically Controlled Screening on CPUs and GPUs

The 30-second take

• The conjugate product is folded into the first inverse-transform butterflies and the peak reduction into the last, so only candidate peaks leave the registers.
• Two-dimensional tiling lets K squared correlations reuse 2K cached spectra, which moves the job from memory bandwidth to on-chip cache.
• Gravitational-wave, seismic, and radio-transient searches get an open-source library with up to 26 times speedup on a fiducial gravitational-wave screen.

What the paper actually did

Batched matched filtering and cross-correlation over large template banks dominate the compute in gravitational-wave astronomy, seismology, and radio transient searches. Standard pipelines pair a general-purpose inverse fast Fourier transform with separate multiply and peak-finding steps, and they write intermediate spectra and full correlation series to memory for every template. This work fuses the conjugate product into the first inverse-transform butterflies and the peak reduction into the last, so only candidate peaks leave processor registers. Batches are tiled in two dimensions so that K squared correlations reuse 2K cached spectra. The filter moves from a regime bound by DRAM bandwidth to one bound by on-chip cache bandwidth. Where the signal-to-noise ratio sits toward lower frequencies while higher frequencies dominate the transform length, a hierarchical screen runs a sequence of short low-frequency transforms before the full band. Stage thresholds come from the spectral power profile and the matched-filter noise autocovariance, and they target a specified false-dismissal probability. The methods are in the open-source matchedfilter library, exercised on eight CPU architectures and two integrated GPUs, an AMD Radeon 8060S through Vulkan and an Apple M2 through Metal. On an AMD Ryzen AI Max+ 395 at transform length 4096, fused peak-only filtering is 1.4 times faster than a standalone AOCL-FFTW inverse transform on one CPU core, 6.2 times faster than rocFFT on the Radeon 8060S, and 3.1 times faster than MLX FFT on the M2. For a fiducial gravitational-wave search with a signal-to-noise threshold of 5.5 and a false-dismissal target of one in a hundred, hierarchical screening is 26 times faster than AOCL-FFTW on CPU, 21 times faster than rocFFT on the 8060S, and 19 times faster than MLX FFT on the M2.

What makes this disruptive

The arithmetic of a correlation is no longer the scarce resource. The scarce resource is writing every frequency bin back to memory for a peak test that needed only a few samples. Fusion keeps the peak inside the transform. The hierarchical screen then refuses to pay for the high-frequency bins until a cheap low-frequency stage says the template is worth it, with the refusal rate tied to a stated false-dismissal target rather than to an uncalibrated cutoff.

Why it matters (outside the lab)

Template banks in gravitational-wave searches, and the analogous banks in seismology and radio transients, are large enough that a constant-factor cut changes how much sky or how many events a given machine can scan. The open-source library already spans eight CPU architectures and two integrated GPUs. The 26-times, 21-times, and 19-times figures are for one fiducial search setting, which is the number a group should remeasure on its own threshold before planning a campaign around it.

Limitations & open questions

The speedups are for the named chips and libraries: Ryzen AI Max+ 395, AOCL-FFTW, rocFFT on the 8060S, and MLX FFT on the M2, at length 4096 for the fusion result. The 26, 21, and 19 times gains are for a signal-to-noise threshold of 5.5 and a designed false-dismissal target of 10^-2, not for a measured astrophysical detection rate. The screen assumes signal-to-noise concentrated at lower frequencies. The abstract does not say the library is already inside a LIGO production pipeline.