| #!/bin/bash |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| set -uo pipefail |
| MODE=${1:-ncu}; shift |
|
|
| if [ "$MODE" = "nsys" ]; then |
| O=/tmp/_nsys.$$ |
| nsys profile -o $O --force-overwrite true --trace=cuda,nvtx,osrt "$@" >/dev/null 2>&1 |
| echo "== per-kernel totals ==" |
| nsys stats --report cuda_gpu_kern_sum $O.nsys-rep 2>/dev/null | sed -n '/Time (%)/,$p' | head -14 |
| echo |
| echo "Read this first: if the summed kernel time is far below wall time, you are HOST-BOUND --" |
| echo "launch gaps, dispatch, or H2D/D2H. Fix that before opening ncu." |
| echo "Report kept at $O.nsys-rep -- open the timeline for gaps and missing overlap." |
| exit 0 |
| fi |
|
|
| REP=/tmp/_ncu.$$; LOG=/tmp/_ncu.log.$$ |
| ncu --target-processes all ${KERNEL:+--kernel-name regex:$KERNEL} --launch-skip 3 --launch-count 3 \ |
| --section SpeedOfLight --section MemoryWorkloadAnalysis \ |
| --section WarpStateStats --section Occupancy \ |
| --metrics sm__throughput.avg.pct_of_peak_sustained_elapsed,\ |
| gpu__dram_throughput.avg.pct_of_peak_sustained_elapsed,\ |
| l1tex__t_sectors_per_request.avg,\ |
| l1tex__data_bank_conflicts_pipe_lsu_mem_shared.sum,\ |
| sm__warps_active.avg.pct_of_peak_sustained_active,\ |
| launch__registers_per_thread \ |
| -o $REP --force-overwrite "$@" >$LOG 2>&1 |
|
|
| if grep -q ERR_NVGPUCTRPERM $LOG; then |
| echo "ERR_NVGPUCTRPERM -- no access to GPU performance counters." |
| echo " fix: run the container with --cap-add SYS_ADMIN, or set" |
| echo " NVreg_RestrictProfilingToAdminUsers=0 on the host." |
| exit 1 |
| fi |
| if [ ! -f $REP.ncu-rep ]; then |
| echo "ncu produced no report. Last lines:"; tail -5 $LOG; exit 1 |
| fi |
|
|
| ncu -i $REP.ncu-rep --csv 2>/dev/null > /tmp/_ncu.csv.$$ |
| python3 - /tmp/_ncu.csv.$$ <<'PY' |
| import csv, sys |
| rows = list(csv.DictReader(open(sys.argv[1]))) |
| if not rows: |
| print("no kernels matched." |
| + (" Loosen or drop KERNEL=." if __import__("os").environ.get("KERNEL") else "")) |
| raise SystemExit(1) |
| kern = rows[0].get("Kernel Name", "?") |
| def get(name): |
| for r in rows: |
| if r.get("Metric Name", "").strip() == name: |
| try: return float(r["Metric Value"].replace(",", "")) |
| except Exception: return None |
| return None |
| sm = get("sm__throughput.avg.pct_of_peak_sustained_elapsed") |
| dram = get("gpu__dram_throughput.avg.pct_of_peak_sustained_elapsed") |
| sect = get("l1tex__t_sectors_per_request.avg") |
| bank = get("l1tex__data_bank_conflicts_pipe_lsu_mem_shared.sum") |
| occ = get("sm__warps_active.avg.pct_of_peak_sustained_active") |
| regs = get("launch__registers_per_thread") |
| f = lambda v, u="": "n/a" if v is None else f"{v:.1f}{u}" |
| print(f"kernel: {kern[:70]}") |
| print(f"SM {f(sm,'%')} DRAM {f(dram,'%')} occupancy {f(occ,'%')} " |
| f"sectors/req {f(sect)} bank-conflicts {f(bank)} regs/thread {f(regs)}") |
| print("\nDIAGNOSIS") |
| if sm is not None and dram is not None: |
| if sm > 60 and dram < 40: print(" COMPUTE-BOUND -> tensor cores, better instruction mix, less redundant work") |
| elif dram > 60 and sm < 40: print(" BANDWIDTH-BOUND -> cut traffic, fuse passes, improve reuse") |
| elif sm < 40 and dram < 40: print(" LATENCY-BOUND -> read WarpStateStats; more loads in flight (cp.async/TMA), more ILP") |
| else: print(" MIXED -> attack whichever of SM/DRAM is higher first") |
| else: |
| print(" SpeedOfLight metrics unavailable -- open the report by hand.") |
| if sect and sect > 4.5: print(f" UNCOALESCED: {sect:.1f} sectors/request (4 is ideal) -> fix layout/access pattern") |
| if bank: print(f" SHARED-MEM BANK CONFLICTS: {bank:.0f} -> swizzle (do NOT pad 2-byte types)") |
| if occ and occ < 30: print(f" LOW OCCUPANCY {occ:.0f}% -> run tools/occupancy.py to find the limiter") |
| if regs and regs >= 168: print(f" {regs:.0f} regs/thread -> spills likely; check SASS for LDL/STL") |
| print("\nReminder: never take a TIMING from a run under ncu -- it serializes and replays kernels.") |
| PY |
| echo "full report: ncu -i $REP.ncu-rep --page details" |
| rm -f /tmp/_ncu.csv.$$ $LOG |
|
|