{"id":"a7ac6eef-e28e-4542-a8ed-66da8a2e49f2","arxiv_id":"2611.17229","paper_version":1,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Fused FFT correlation with peak-only output, 2D cache tiling, and covariance-calibrated hierarchical screening gives 1.4-6.2x kernel speedups and 19-26x end-to-end matched-filter search speedups across eight CPUs and two integrated GPUs.","lead":"This paper makes matched filtering, the standard calculation for finding faint gravitational-wave signals and other buried correlations, much cheaper by fusing the FFT steps so only the winning peak is written to memory. On tested CPUs and integrated GPUs the fused kernels run 1.4 to 6.2 times faster than vendor FFT libraries, and a screening stage that checks short low-frequency transforms first delivers up to 26 times total search speedups.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek/deepseek-v4-flash","headline":"Load-bearing risk: FDR control outside the matched-reference Gaussian regime.","rationale":"The central claim decomposes into (a) fused peak-only throughput gains and (b) hierarchical screening speedups at a statistically controlled false-dismissal rate. I checked the arithmetic of (a) and the internal consistency of (b). Table 2 is self-consistent: 32.84/22.78 = 1.44x, 32.84/1.24 = 26.5x, 22.78/1.24 = 18.4x, matching the abstract; the Figure 3(b) throughput values (8192 pairs over 1.24 ms = 6.6 Mpairs/s) agree; the traffic model Eqs. (3)-(5) evaluates correctly (R ~ 95x at K=32, N=4096). Re-deriving Eqs. (7)-(8) under the paper's unit-variance coarse normalization, the coarse response peak E[z_c[L]] = rho*sqrt(mu) and autocovariance A(x) = m*F^{-1}{q_hat} follow, and the cross-tier covariance sqrt(mu_i/mu_j)*A_bi(x) matches the shared-band overlap, so the screening model is coherent in its stated regime. The Figure 2(c) matched-reference validation (three decades of alpha, >2.5e5 injections per profile, disclosed Wilson intervals) is genuine supporting evidence. The baselines are honest: FFTW_PATIENT on CPU, rocFFT and MLX on GPU, with the standalone inverse transform timed without the conjugate product or peak search, making the fused comparison conservative. The Section 4.3 stress test that fails (27% FDR at a 1e-2 target) is disclosed rather than hidden, which further supports good faith. The load-bearing risk is the transfer of the FDR guarantee: reference representativeness is the controlling variable, the grouping mitigation is validated only on a synthetic f^gamma bank with no specified procedure, and the b95 direction claim is unproven - high-frequency-concentrated non-Gaussianity would inflate FDR while shrinking the speedup. This does not refute the claims as stated for the tested regime, and the main limitation is acknowledged by the authors, so the reader's ACCEPT stands unchanged; the concrete test above defines the condition under which the concern would land.","tokens_in":18438,"tokens_out":27285,"duration_ms":186441,"concrete_test":"Measure the Section 4.3 conditional FDR on real data: inject ~1e5 signals at SNR 5.5-8 at random lags into 16 public LIGO O4 noise segments (N=4096, Delta f=0.5 Hz) and run the (512, 1024, 2048) chain at rho_th=5.5, alpha=1e-2 under two reference configurations: the single BNS-matched envelope and a slope-partitioned multi-reference grouping. If the measured FDR exceeds ~2e-2 in either configuration, Section 6's guarantee fails in the paper's fiducial search regime and the mitigation needs a specified grouping procedure plus real-data calibration; if the FDR stays within the Wilson bounds of the Figure 2(c) targets - and deviations show up, per b95, as reduced speedup rather than increased FDR - the concern does not land.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The screening speedups (26x/21x/19x) and the closing guarantee (b144: 'without sacrificing signal recovery beyond a user-specified false-dismissal rate') both rest on the FDR bound holding when the method is deployed. Within the tested regime the model is internally consistent: re-deriving Eqs. (7)-(8) with the paper's unit-variance coarse normalization gives E[z_c[L]] = rho*sqrt(mu) and A(x) = m*F^{-1}{q_hat}, and the Section 4.2 cross-tier covariance sqrt(mu_i/mu_j)*A_bi(x) matches the shared-band overlap, so the Figure 2(c) matched-reference validation across three decades of alpha with >2.5e5 injections per profile is genuine evidence. The fragile pillar is reference representativeness, and Section 4.3 demonstrates the failure: a single edge reference over an f^gamma bank, gamma in [-2.33, -1.33], yields ~27% bank-wide FDR against a 1e-2 target (27x), and even the bank-mean reference gives ~3.5%. The proposed mitigation ('partition the bank into four reference groups ordered by spectral slope') is validated only on this synthetic bank; no algorithm or tolerance for choosing the groups is specified, so a user cannot determine when a real bank is 'similar enough'. Second, the failure-direction claim (b95) that rejection rates 'would be more permissive where these assumptions do not hold' is asserted without argument and is not directionally guaranteed: non-Gaussian noise concentrated in the high-frequency part of the band (glitches, scattered light, spectral lines above 256 Hz) raises the full-band peak without raising the coarse-stage response, which would inflate the FDR above target while simultaneously reducing the screening speedup - the worst case for the advertised claim. All validation is simulated whitened Gaussian noise; real-data FDR is entirely untested.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek/deepseek-v4-flash","summary":"The paper presents two complementary techniques for accelerating batched matched filtering: (1) fusing the conjugate product and peak search into the inverse FFT butterflies, eliminating intermediate memory traffic, and (2) a hierarchical screening method that uses short low-frequency transforms to reject noise-dominated pairs before full-bandwidth evaluation, with thresholds calibrated from the matched-filter noise autocovariance to control the false-dismissal rate. The methods are implemented in the open-source matchedfilter library and benchmarked on eight CPU architectures and two integrated GPUs. The central claim is that fused peak-only filtering is 1.4–6.2× faster than vendor FFT libraries on a single core/GPU, and hierarchical screening yields 19–26× speedups for a fiducial gravitational-wave search at ρth=5.5 and α=10^{-2}, without sacrificing signal recovery beyond the specified false-dismissal rate.","tokens_in":18913,"tokens_out":1862,"duration_ms":46788,"significance":"If the claims hold, this work provides a practical, open-source acceleration of a core computational kernel in gravitational-wave astronomy, seismology, and radio transient searches. The operator-fusion and 2D tiling approach is a clean application of known IO-aware optimization principles to the matched-filtering context, and the statistical screening model with covariance-calibrated thresholds is a principled alternative to ad-hoc hierarchical search thresholds. The validation includes extensive Monte Carlo testing (2.5×10^5 injections per profile) and cross-platform benchmarks. The paper is clearly written and the code is released under an open-source license, supporting reproducibility.","major_comments":[{"comment":"The heterogeneous-bank stress test (right side of Fig. 2(c)) shows that a single reference envelope can produce false-dismissal rates far exceeding the target (∼27% bank-wide for an edge reference at α=10^{-2}). The authors propose partitioning the bank into reference groups ordered by spectral slope, but provide no algorithm or tolerance for determining when a real bank is 'similar enough' to use a single reference. This is a load-bearing limitation because the speedups claimed for hierarchical screening (26×, 21×, 19×) and the closing guarantee in b144 ('without sacrificing signal recovery beyond a user-specified false-dismissal rate') both depend on the FDR bound holding in deployment. The paper should either specify a concrete procedure for grouping templates (e.g., a distance metric on q[k] with a tolerance bound) or state explicitly that the method is only validated for banks where a single reference profile is representative, and that users must validate FDR for their specific bank.","section":"§4.3, Fig. 2(c), sentence b95","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The covariance model assumes stationary Gaussian noise and independent whitened frequency bins. The authors acknowledge in b95 that 'rejection rates would be more permissive where these assumptions do not hold', but the direction of the failure is not guaranteed: non-Gaussian noise or non-stationary features could produce either more or fewer false dismissals than the model predicts. The paper should discuss whether the screening thresholds are conservative or anti-conservative under realistic noise deviations (e.g., spectral lines, glitches, scattered light), and whether the method can be adapted to estimate the noise distribution empirically from the data rather than relying solely on the Gaussian model.","section":"§4.1, Eqs. (7)–(9), and §4.2","authors_acknowledge":"The rejection rates hold for noise-dominated data which is approximately Gaussian; rejection rates would be more permissive where these assumptions do not hold. 4.3.","authors_acknowledge_p":0.68},{"comment":"The fused peak-only speedups are reported relative to a standalone inverse FFT (AOCL-FFTW or rocFFT) that does not include the conjugate product or peak search. The authors note this in the caption of Fig. 3(a) and in the text, but the headline speedups (1.4×, 6.2×, 3.1×) compare against the inverse FFT only, not against a complete unfused pipeline that includes multiplication and peak finding. For a fair comparison, the baseline should include the full pipeline cost. The paper does provide some comparisons against the complete unfused pipeline on GPU (6–13× for fused peak-only vs. unfused pipeline on Radeon 8060S, §5.2), but the CPU comparisons are only against the standalone FFT. The authors should either provide CPU benchmarks of the complete unfused pipeline or clarify that the speedups are relative to the transform-only baseline and that the advantage over a full pipeline would be larger.","section":"§5, Tables 1 and 2, Fig. 3","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The chain selection minimizes expected cost using modeled cumulative noise survival fractions π_i. However, the π_i depend on the detection threshold ρ_th and the false-dismissal budget α, and the model assumes Gaussian noise. The paper does not report how sensitive the optimal chain is to deviations from the Gaussian assumption or to errors in the estimated survival fractions. If the model underestimates π_i, the chain may be too aggressive and cause excessive false dismissals; if it overestimates π_i, the chain may be too conservative and reduce the speedup. The authors should discuss the robustness of the chain selection to model misspecification.","section":"§4.2, Eq. (10) and chain selection","authors_acknowledge":null,"authors_acknowledge_p":null}],"minor_comments":[{"comment":"The abstract states 'fused peak-only filtering is 1.4× faster than a standalone AOCL-FFTW inverse transform on one CPU core and 6.2× faster than rocFFT on its Radeon 8060S GPU (3.1× over MLX FFT on the M2 GPU).' The phrase 'standalone AOCL-FFTW inverse transform' is clear, but the GPU comparisons should also specify that the baseline is the inverse transform only, not the full pipeline, for consistency.","section":"Abstract and §1","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The definition of A(x) as m F^{-1}(q_hat)[x] is given, but the normalization of the inverse transform is not explicitly stated. It would be helpful to note that the 1/m normalization of the length-m inverse transform cancels the factor m, so that A(0)=1.","section":"§4.1, Eq. (7)","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The notation P_FD(ρ_c) uses ρ_c as a vector of thresholds, but the equation shows a union over tiers. Clarify that ρ_c is the set of tier thresholds (ρ_{c,1}, ρ_{c,2}, ...).","section":"§4.2, Eq. (9)","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The shaded bands span the 10th–90th percentile of timing blocks, but the number of timing blocks is not stated. Adding this information would help assess the variability.","section":"§5.1, Fig. 3(a)","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The 'effective problem throughput' is defined as 48N bytes divided by execution time for the fused and hierarchical cases, but the hierarchical case skips operations and memory reads on rejected pairs, so the throughput metric is not directly comparable to the memory bandwidth ceilings. The authors note this in the caption, but it would be clearer to state in the text that the hierarchical curves are not physical memory throughputs.","section":"§5.2, Fig. 4(c)–(d)","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The suggestion that reduced-precision formats (float16, bfloat16, fp8) could be used for early screening tiers is interesting but speculative. The paper does not test this, so it should be clearly marked as future work rather than presented as a near-term possibility.","section":"§6, last paragraph","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"Reference [40] (Google Highway) is listed as 2024, but the URL points to a GitHub repository. The reference should include the version or commit hash used for reproducibility.","section":"References","authors_acknowledge":null,"authors_acknowledge_p":null}],"recommendation":"major_revision"},"author_rebuttal":null,"desk_editor":null,"pith_extraction":null,"created_at":"2026-11-13T09:06:38.999020+00:00","model_set":{"reader":"deepseek/deepseek-v4-flash"},"pipeline_version":"pith-pipeline@v1.4.1-daily-deepseek-or-medium","input_kind":null,"input_chars":null,"input_total_chars":null,"passes_missing":null,"published_standing":null,"published_confidence":null,"objection_audit":null,"run_record":null,"basis":{"kind":"not_recorded","source":null,"chars_given":null,"chars_total":null},"receipt":null,"falsifier":null,"supporting_citations":[],"lane_accuracy":{"status":"not yet measured","lane":{"model":"deepseek/deepseek-v4-flash","input_kind":null},"url":"https://pith.science/accuracy"},"review_version":1}