Concepts

• Kernel: A GPU function executed in parallel by many threads.
• Warp: A group of 32 threads that execute in lockstep.
• SASS / Opcode: NVIDIA GPU assembly instruction and its mnemonic.
• Clock instruction: SASS CS2R SR_CLOCK* reads the GPU clock. CUTracer uses these as region markers.
• Region (clock-to-clock): A segment of dynamic instruction execution delimited by consecutive clock reads.
• Histogram (instruction): Count of instruction mnemonics within a region per warp.
• Instrumentation: Injecting lightweight device calls into SASS via NVBit to emit data.
• Analysis: Host-side processing of emitted data into logs/CSVs.
• Kernel filter: Substring match on kernel names (mangled/unmangled) to limit instrumentation.

Limitations to note:

• CUTracer's instruction histogram analysis does not support nested regions; it treats consecutive clock reads as a flat sequence of start/stop markers.
• Graph/stream-capture launches are supported but have distinct synchronization paths.