Utilities

To enable seamless integration with target applications, AccelProf is built as a shared library and injected at runtime using the LD_PRELOAD mechanism. This setup allows AccelProf to intercept both deep learning framework calls and accelerator runtime calls.

---

Event Capture Mechanism

Once loaded, AccelProf activates the underlying event capture infrastructure through vendor-specific APIs, such as:

• sanitizerEnableDomain (Compute Sanitizer)

• nvbit_at_cuda_event (NVIDIA NVBit)

• rocprofiler_configure_callback (AMD ROCm ROCProfiler)

These APIs initialize and register the profiling hooks required to monitor accelerator behavior.

---

Callback Integration

For every intercepted event, AccelProf uses a dedicated callback implementation to route data into the event handler system. This consistent routing mechanism enables modular and extensible event processing.

---

Developer Introspection & Customization

AccelProf supports a rich set of features to enable developer-level introspection and cross-language integration:

• 🧩 pybind11: Enables Python-side annotations and tool development

• 🐍 PyFrame (CPython): Captures Python-level call stacks

• ⚙️ libbacktrace: Extracts symbolic stack traces for C/C++ applications

These integrations allow users to analyze runtime behavior with high fidelity, offering both framework-level and instruction-level views. The result is a powerful profiling platform for debugging, performance tuning, and workload characterization.