Siebel School Speaker Series Master Calendar

Speaker: Chamika Sudusinghe

Title: COMPASS: Guiding Dataset Selection for Cost Model Fine-Tuning to Optimize Sparse Kernels on Emerging Hardware Accelerators

Abstract: Sparse matrix kernels are central to modern machine learning and graph analytics workloads, yet optimizing them remains challenging due to complexities in sparsity patterns and large compiler schedule search spaces. Learned cost models have emerged as an effective approach for navigating these spaces by predicting performance and guiding optimization. However, training accurate cost models typically requires a large number of performance measurements, which becomes prohibitive on emerging accelerators where evaluation relies on expensive, simulator-based execution. While prior work has shown that transfer learning can reduce the data requirements of cost model fine-tuning, existing approaches rely on manually curated fine-tuning datasets and carefully constrained schedule search spaces. In this paper, we argue that dataset selection for fine-tuning is itself a critical optimization problem and that the lack of automation in this process is a key limitation of existing approaches. To address this challenge, we propose an automated, architecture-aware strategy that jointly guides matrix selection and schedule selection during fine-tuning, eliminating reliance on expert-curated datasets while efficiently allocating a fixed evaluation budget. We evaluate the proposed approach on an emerging sparse accelerator and modern NVIDIA GPUs. Our method achieves competitive average speedups of 1.50× (up to 5.46×) for SpMM and 1.40× (up to 4.22×) for SDDMM, while remaining data-efficient and scaling to larger schedule search spaces that were previously infeasible.

Speaker: Siyuan Brant Qian

Title: Thinking Fast and Correct: Automated Rewriting of Numerical Code through Compiler Augmentation

Conference: CGO 2026

Abstract: Floating-point numbers are finite-precision approximations to real numbers and are ubiquitous in computer applications in nearly every field. Selecting the right floating-point representation that balances performance and numerical accuracy is a difficult task – one that has become even more critical as hardware trends toward high-performance, low-precision operations. Although the common wisdom around changing floating-point precision implies that accuracy and performance are inversely correlated, more advanced techniques can often circumvent this tradeoff. Applying complex numerical optimizations to real-world code, however, is an arduous engineering task that requires expertise in numerical analysis and performance engineering, and application-specific numerical context. While there is a plethora of existing tools that partially automate this process, they are limited in the scope of optimization techniques or still require substantial human intervention. We present Poseidon, a modular and extensible framework that fully automates floating-point optimizations for real-world applications within a production compiler. Our key insight is that a small surrogate profile often reveals sufficient numerical context to drive effective rewrites. Poseidon operates as a two-phase compiler: the first compilation instruments the program to capture numerical context; the second compilation consumes profiled data, generates and evaluates candidate rewrites, and solves for optimal performance/accuracy tradeoffs. Poseidon’s interoperability with standard compiler analyses and optimizations grants it analysis and optimization advantages unavailable to existing source- and binary-level approaches. On multiple large-scale applications, Poseidon leads to outsized benefits in performance without substantially changing accuracy, and outsized accuracy benefits without diminishing performance. On a quaternion differentiator, Poseidon enables a 1.46× speedup with a relative error of 10−7. On DOE’s LULESH hydrodynamics application, Poseidon improves program accuracy to exactly match a 512-bit simulation run without substantially reducing performance.