Instrumentation Modes

opcode_only (lightweight) 🪶

• Emits: opcode ID, warp ID, PC, kernel_launch_id, CTA IDs
• Use for: Proton instruction histogram; lowest overhead
• May be auto-enabled by analyses (see below)

reg_trace (medium) 🔬

• Emits: per-thread register values (plus unified registers), opcode ID, PC
• Use for: register value tracing and dataflow inspection

mem_trace (heavy) 🧠

• Emits: 32-lane memory addresses for memory-reference instructions
• Use for: memory access pattern analysis

Combining modes ➕

• CUTRACER_INSTRUMENT accepts comma-separated values. Analyses may also enable required modes implicitly.
• proton_instr_histogram auto-enables opcode_only.
• deadlock_detection auto-enables reg_trace.

Notes 📝

• When an analysis auto-enables a mode, you do not need to repeat it in CUTRACER_INSTRUMENT.
• Enabling additional modes increases overhead and output volume; prefer the minimal set that satisfies your analysis.