MCPcopy Create free account
hub / github.com/RightNow-AI/autokernel / _benchmark_fn

Function _benchmark_fn

prepare.py:60–82  ·  view source on GitHub ↗

Benchmark *fn* using CUDA events. Returns median latency in microseconds.

(fn, *args, warmup: int = _WARMUP_ITERS, iters: int = _BENCH_ITERS)

Source from the content-addressed store, hash-verified

58
59
60def _benchmark_fn(fn, *args, warmup: int = _WARMUP_ITERS, iters: int = _BENCH_ITERS):
61 """
62 Benchmark *fn* using CUDA events. Returns median latency in microseconds.
63 """
64 # Warmup
65 for _ in range(warmup):
66 fn(*args)
67 torch.cuda.synchronize()
68
69 start_events = [torch.cuda.Event(enable_timing=True) for _ in range(iters)]
70 end_events = [torch.cuda.Event(enable_timing=True) for _ in range(iters)]
71
72 torch.cuda.synchronize()
73 for i in range(iters):
74 start_events[i].record()
75 fn(*args)
76 end_events[i].record()
77 torch.cuda.synchronize()
78
79 times_ms = [s.elapsed_time(e) for s, e in zip(start_events, end_events)]
80 times_ms.sort()
81 median_ms = times_ms[len(times_ms) // 2]
82 return median_ms * 1000.0 # convert to microseconds
83
84
85# ---------------------------------------------------------------------------

Callers 1

benchmark_baselinesFunction · 0.85

Calls

no outgoing calls

Tested by

no test coverage detected