Benchmarking and Selecting State-of-the-Art Modern Fourier Transformation Methods
Fourier transformation remains a core primitive in scientific computing, signal processing, and ML systems, but practical performance and accuracy depend heavily on algorithm and hardware choices. This project will build a reproducible benchmark framework to compare modern Fourier transformation implementations across representative workloads (1D/2D/3D, real/complex, varying sizes, and precision modes). We will evaluate both numerical fidelity and system-level efficiency to identify configuration-specific SOTA candidates rather than a one-size-fits-all winner. The expected impact is a decision guide and open benchmark suite that helps researchers and engineers select the best Fourier pipeline for their constraints.
Problem Workspace
Problem Statement
Scope: This submission targets rigorous, application-relevant comparison of modern Fourier transformation approaches, with emphasis on practical SOTA selection by scenario. The work will cover direct transform performance, inverse transform stability, batched throughput, and behavior under precision and size variation. The output is not a new transform theory, but a validated methodology and recommendation matrix for current best-performing approaches. Constraints: Results must be reproducible, hardware-annotated, and comparable across methods using consistent preprocessing, normalization, and measurement procedures. The benchmark should explicitly separate algorithmic effects from implementation/runtime effects. Claims will be limited to tested environments and clearly labeled by workload class to avoid overgeneralization. Success criteria: (1) A complete benchmark protocol with fixed data generation and split policy, (2) statistically stable performance and…Read more
Read less
Execution plan
Metrics: latency (p50/p95), throughput (transforms/sec), peak memory, numerical error versus high-precision reference (L2 relative error, max absolute error), and stability under repeated forward-inverse cycles. Baselines: naive DFT implementation, standard FFT baseline configuration, and currently deployed/default pipeline in the target stack. Data/splits: synthetic signal/image volumes with controlled frequency content plus a held-out stress suite; split into tuning (20%), validation (20%), and final reporting (60%) with fixed random seeds and size-stratified buckets. Acceptance criteria: at least one candidate must exceed baseline throughput by >=20% at equal or better error tolerance, or reduce error by >=25% at no worse than 10% latency cost, with results reproduced across at least two independent runs per setting.