kernel-perf-analysis

To identify performance bottlenecks in AMD GPU kernels, you need to collect timing data, hardware counters, ATT traces, and memory-hierarchy profiles. This automated analysis evaluates MFMA efficiency, memory bandwidth, and latency to generate actionable optimization recommendations.

Profiling GEMM kernels on CDNA3 requires analyzing MFMA efficiency and memory bandwidth across multiple modes. The process collects ATT traces and hardware counters to detect LDS conflicts and generate regression-friendly reports for targeted kernel optimizations.

ATT trace decoding for AMD GPU memory analysis works by capturing memory-hierarchy data to reveal HBM, L1, and L2 bandwidth insights. It requires an ATT decoder and Triton metadata to translate hardware traces into actionable latency and conflict detection reports.

Yes, you need rocprofv3 to analyze AMD GPU MFMA efficiency. The profiling process requires rocprofv3 to collect hardware counters and ATT traces, which are then decoded alongside Triton metadata to produce accurate performance insights and optimization recommendations.

Yes, you can use this AMD GPU profiling approach for general Python kernels. The multi-mode profiling applies to both GEMM-style kernels on CDNA3/4 and general Python kernels, collecting timing and memory-hierarchy data to generate comparative performance reports.

Low memory bandwidth utilization in AMD GPU kernels often indicates bottlenecks detectable through HBM, L1, and L2 bandwidth analysis. Profiling with ATT traces and hardware counters identifies LDS conflicts and memory-hierarchy inefficiencies to guide targeted optimizations.