File size: 2,381 Bytes
1ab6c33
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
"""Occupancy / register / shared-memory budget, and persistent-grid co-residency.

Two questions this answers that cost the most time when guessed:
  1. which resource is limiting occupancy -- registers, shared memory, or block size
  2. whether a persistent kernel's grid actually fits (if it does not, a grid-wide barrier DEADLOCKS)

    python3 occupancy.py --threads 256 --regs 96 --smem 49152 [--blocks 264]
"""
import argparse
import torch


def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("--threads", type=int, required=True)
    ap.add_argument("--regs", type=int, required=True, help="registers per thread (ncu: launch__registers_per_thread)")
    ap.add_argument("--smem", type=int, default=0, help="dynamic shared memory per block, bytes")
    ap.add_argument("--blocks", type=int, default=0, help="grid size, to check persistent co-residency")
    a = ap.parse_args()

    p = torch.cuda.get_device_properties(0)
    sm = p.multi_processor_count
    max_thr = p.max_threads_per_multi_processor
    regs_sm = p.regs_per_multiprocessor
    smem_sm = getattr(p, "shared_memory_per_multiprocessor", p.shared_memory_per_block_optin)
    warps_blk = (a.threads + 31) // 32

    by_thread = max_thr // a.threads
    by_regs = regs_sm // (a.regs * warps_blk * 32) if a.regs else 999
    by_smem = (smem_sm // a.smem) if a.smem else 999
    blocks_sm = max(0, min(by_thread, by_regs, by_smem))
    limiter = min((by_thread, "block size"), (by_regs, "registers"), (by_smem, "shared memory"))[1]
    occ = blocks_sm * warps_blk * 32 / max_thr if max_thr else 0

    print(f"device {p.name}  SMs={sm}  max_threads/SM={max_thr}  regs/SM={regs_sm}  smem/SM={smem_sm//1024}KB")
    print(f"blocks/SM: by-threads {by_thread}, by-registers {by_regs}, by-smem {by_smem}"
          f"  -> {blocks_sm}  (limiter: {limiter})")
    print(f"achieved occupancy  {occ:.0%}   resident blocks total {blocks_sm*sm}")
    if a.regs >= 168:
        print("  !! >=168 regs/thread: spills are likely; check SASS for LDL/STL")
    if a.blocks:
        cap = blocks_sm * sm
        ok = a.blocks <= cap
        print(f"\npersistent grid {a.blocks} vs co-resident capacity {cap}: {'FITS' if ok else 'DOES NOT FIT'}")
        if not ok:
            print("  !! a grid-wide barrier WILL DEADLOCK: blocks that never get scheduled cannot arrive.")


if __name__ == "__main__":
    main()