Project Description
Develop and maintain an open-sourced profiler tool for AMD GPUs. The main components of profiler are:
ROC profiler library, https://github.com/ROCm-Developer-Tools/rocprofiler
ROCm Tracer Callback/Activity library, https://github.com/ROCm-Developer-Tools/roctracer
Responsibilities
Examples of tasks:
- To implement shader-engine counters indexing with two dimensional indexing of shader memory block counters
- Optimizing of data generation mechanism
- Dumping SQLite DB directly from profiler, C++
- Optimizing post-processing, Python
- Multi-node time synchronization
- Profiler overhead tracing
- Configurable trace format generator
Skills
Must have
- C/C++, assembly, python
- System/Linux programming, resource management
- Generic data processing and algorithms
- CPU/GPU architecture knowledge
- Strong self-management/motivated
- Good written and verbal English
Nice to have
- HW/SW profiling experience, perf-counters, performance metrics
- Experience in HW/SW performance analysis and optimization
- Experience with Linux perf, VTune, CUPTI/CUDA