580 California St., Suite 400
San Francisco, CA, 94104
Key research themes
1. How can memory and storage optimization be achieved effectively for large-scale and irregularly accessed arrays?
This research area focuses on optimizing the memory footprint and access patterns of large arrays, especially in embedded and high-performance computing systems where memory size directly impacts cost, energy consumption, and performance. The challenge lies in computing the minimum storage size needed for arrays under complex access patterns including overlapping reads/writes and irregular indexing, while keeping analysis computationally feasible. This theme is critical for designing efficient hardware-software systems and scalable array-based applications.
Key finding: Introduces a novel methodology that efficiently computes the near-minimal storage size for arrays under overlapping and irregular memory access patterns. This approach balances exploration time and precision better than... Read more
Key finding: Presents Cash, an innovative method utilizing segmentation hardware in x86 architectures to perform array bound checking at virtually zero per-access overhead. By assigning arrays to separate segments, the CPU’s native... Read more
2. What computational strategies enable effective source enumeration and subspace estimation in array signal processing under challenging noise and sample conditions?
Source enumeration—the problem of determining the number of signal sources impinging on sensor arrays—is fundamental in array signal processing with applications in radar, communications, and seismology. This theme involves designing robust algorithms that can estimate the signal subspace dimension despite non-ideal noise (e.g., non-iid) and limited sample sizes, improving accuracy in the small-sample regime and high-dimensional settings. Approaches using subspace averaging, matrix completion, and Toeplitz structure exploitation are prominent, providing new ways to tackle traditional and modern array processing challenges.
Key finding: Introduces a robust subspace averaging (SA) technique which estimates source number by averaging projection matrices derived from overlapping subarrays and examining the eigenvalues of the average. The method avoids... Read more
Key finding: Develops a method for source enumeration in non-iid noise environments by reconstructing the low-rank Toeplitz-structured signal covariance matrix via matrix completion. By removing the diagonal (noise) terms and performing... Read more
Key finding: Extends the classical theory of arrays by incorporating set and copy operations to model region-based memory updates such as memset and memcpy in programs. Though primarily theoretical, this work supports accurate reasoning... Read more
3. How can large array computations be efficiently mapped and executed on modern high-performance hardware architectures such as GPUs and massive MIMO antenna arrays?
This research area explores efficient code generation, memory management, and algorithm design for large multidimensional array operations on hardware platforms like GPUs and massive multiple-antenna arrays. Tackling challenges like device memory limitations, the need for streaming data, gracefully handling non-uniform arrays, and reducing computational complexity, this theme aims to improve throughput and scalability for scientific, communication, and signal processing applications.
Key finding: Presents automated strategies to generate CUDA streaming code for map-like and iterative stencil array operations that can handle datasets exceeding GPU memory capacity (out-of-core). The proposed parameterized streaming... Read more
Key finding: Introduces a novel design of a 77 GHz FMCW MIMO radar frontend utilizing non-uniform planar antenna arrays and substrate integrated waveguides that reduce sidelobe levels and mitigate microstrip antenna drawbacks. The convex... Read more
Key finding: Proposes a Kronecker-separable beamforming method tailored for planar massive MIMO arrays that decomposes the high-dimensional filter into two low-dimensional factors optimized via alternating minimization and regularized... Read more