Speaker
Description
Reliable floating-point behavior is increasingly difficult to ensure as HEP applications adopt heterogeneous architectures, multiple GPU vendors, and aggressive compiler optimizations such as fast-math. We introduce a non-intrusive workflow that enables detailed floating-point error analysis of GPU kernels without modifying application code. The method records SYCL kernel executions on Intel GPUs using the OpenCL Intercept Layer, capturing SPIR-V kernels and their input/output buffers. These kernels are then replayed on CPUs through PyOpenCL and the PoCL runtime, where they are instrumented by Verificarlo to explore IEEE-compliant behavior, stochastic arithmetic, and reduced-precision formats.
By isolating kernel execution from the surrounding application, the workflow accommodates the additional overhead of floating-point instrumentation while preserving a realistic production workload. This targeted replay model supports the validation process without architecture-specific code paths or rebuilding the full application. Demonstrated on a gravity kernel from HACC (Hardware/Hybrid Accelerated Cosmology Code), used for extreme-scale cosmological simulations, and currently being evaluated on tracking kernels from the ACTS/traccc project, a performance-portable particle tracking framework, it delivers reproducible cross-runtime behavior, well-structured ULP-error distributions, and clear quantification of reduced-precision and fast-math effects.
The workflow offers a sustainable mechanism for HEP developers to assess numerical robustness as software evolves and hardware diversifies, supporting long-term maintainability and reproducibility.