ck-profilelisted

Profile a Composable Kernel (CK) build target two ways: static compile-time resource analysis (registers, occupancy ceiling, spills, scratch, LDS) and dynamic runtime profiling with rocprofv3 (timing, HBM traffic, cache, occupancy) averaged over N runs and an argument sweep, with a roofline-lite compute/memory/latency verdict. Use when the user wants to profile, benchmark, or find the bottleneck of a CK example/test binary; says "profile <target>", "is <target> compute- or memory-bound", "check register spills / occupancy", "benchmark fp16 vs bf16", "rocprof <target>", "static/dynamic profile", or "sweep -B / -prec and compare". The user selects one or more of six modes: static, dynamic, trace, cfg, depgraph, compute. Not for non-CK or CPU-only targets, or for correctness testing.
steven112163/dotharness · ★ 0 · Data & Documents · score 65

Install: claude install-skill steven112163/dotharness

# CK profiling (ck-profile) Six independent modes; the user picks one or several: - **static** — compile-time only, no GPU run. Builds the target with `-Rpass-analysis=kernel-resource-usage` and parses the remarks for register/ occupancy/spill/scratch/LDS. Answers "what each kernel reserves and its occupancy ceiling." - **dynamic** — runs the target under rocprofv3 (kernel trace + PMC multipass), aggregates over runs/variants, and classifies the bottleneck. Answers "what actually happened at runtime and what is the limiter." - **trace** — rocprofv3 `--sys-trace` → an offline **HTML timeline** (host + device lanes, opens in Live Preview) plus a perfetto `.pftrace` (+ sys-trace CSVs); optional best-effort PC sampling. Answers "what is the dispatch timeline / where on the host-GPU timeline does time go." - **cfg** — extracts the amdgcn code object and disassembles it (`llvm-objdump`), emitting one per-kernel ISA control-flow graph as Graphviz `.dot`. Answers "what does each kernel's basic-block / branch structure look like." - **depgraph** — two dependency graphs as `.dot`: a logical data-dependency DAG (workspace producer→consumer, from the launch source) and a runtime dispatch graph (from a kernel trace). Answers "how do the kernels depend on / follow each other." - **compute** — `rocprof-compute` (ROCm Compute Profiler): roofline + speed-of-light + memory-hierarchy panels. Answers "where in the microarchitecture is the kernel limited." Requires