File size: 2,727 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 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 | """Honest GPU timing. Import `bench` rather than hand-rolling a timer.
The failure this exists to prevent: an eager reference in this repo measured 185 ms wall against
2.06 ms of GPU-busy time. Wall-clock around a python call measures dispatch, not the kernel.
Rules enforced here: CUDA events (not time.time), warmup excluded, FRESH inputs per rep so a warm L2
does not masquerade as bandwidth, min-of-N (not mean, which drifts with clocks), and an optional
GPU-busy cross-check so host-bound code is caught rather than reported as kernel time.
"""
import statistics
import torch
def bench(fn, make_args=None, args=None, reps=10, warmup=5, return_all=False):
"""Time `fn`. Pass make_args(i)->tuple for fresh inputs per rep, or a fixed `args` tuple."""
if make_args is None and args is None:
raise ValueError("pass make_args or args")
get = make_args if make_args is not None else (lambda i: args)
for i in range(warmup):
fn(*get(-1 - i))
torch.cuda.synchronize()
times = []
for i in range(reps):
a = get(i)
torch.cuda.synchronize()
s = torch.cuda.Event(enable_timing=True)
e = torch.cuda.Event(enable_timing=True)
s.record()
out = fn(*a)
e.record()
torch.cuda.synchronize()
times.append(s.elapsed_time(e) / 1e3)
del a, out
return (min(times), times) if return_all else min(times)
def gpu_busy(fn, args, reps=3):
"""Sum of kernel time from the profiler. If this is far below `bench`, you are HOST-bound and the
kernel is not your problem yet -- go look at nsys, not ncu."""
from torch.profiler import profile, ProfilerActivity
for _ in range(3):
fn(*args)
torch.cuda.synchronize()
with profile(activities=[ProfilerActivity.CUDA]) as p:
for _ in range(reps):
fn(*args)
torch.cuda.synchronize()
ks = [k for k in p.key_averages() if k.self_device_time_total > 0]
return sum(k.self_device_time_total for k in ks) / reps / 1e6, sum(k.count for k in ks) / reps
def report(fn, make_args=None, args=None, work=None, unit="TFLOP/s", **kw):
"""Time, cross-check against GPU-busy, and convert to an achieved metric."""
wall = bench(fn, make_args=make_args, args=args, **kw)
a = (make_args(0) if make_args else args)
busy, nk = gpu_busy(fn, a)
print(f"wall {wall*1e6:9.1f} us gpu-busy {busy*1e6:9.1f} us kernels/call {nk:.0f}")
if busy > 0 and wall / busy > 1.5:
print(f" !! wall is {wall/busy:.1f}x gpu-busy -- HOST-BOUND. Profile with nsys, not ncu.")
if work:
scale = 1e12 if unit == "TFLOP/s" else 2 ** 30
print(f" achieved {work/wall/scale:.4g} {unit}")
return wall
|