TileOPs-Metax/benchmarks/ops/bench_instance_norm.py

71 lines
2.4 KiB
Python

import math
from typing import Optional
import pytest
import torch
import torch.nn.functional as F
from benchmarks.benchmark import BenchmarkBase, BenchmarkReport
from tileops.manifest import eval_roofline, load_workloads
from tileops.ops.norm.instance_norm import InstanceNormFwdOp
from workloads.instance_norm import InstanceNormTest
_OP_NAME = "InstanceNormFwdOp"
class InstanceNormBenchmark(BenchmarkBase):
_roofline_cache: Optional[tuple[float, float]] = None
def _get_roofline(self) -> tuple[float, float]:
if self._roofline_cache is None:
t = self.workload
spatial_size = math.prod(t.spatial)
elem_bytes = torch.tensor([], dtype=t.dtype).element_size()
self._roofline_cache = eval_roofline(
_OP_NAME, N=t.n, C=t.c,
spatial_size=spatial_size, elem_bytes=elem_bytes)
return self._roofline_cache
def calculate_flops(self) -> Optional[float]:
return self._get_roofline()[0]
def calculate_memory(self) -> Optional[float]:
return self._get_roofline()[1]
def _manifest_params():
params = []
for w in load_workloads(_OP_NAME):
shape = w["x_shape"]
n, c, spatial = shape[0], shape[1], tuple(shape[2:])
label = w.get("label", f"{n}x{c}x{'x'.join(map(str, spatial))}")
for dtype_str in w["dtypes"]:
dtype = getattr(torch, dtype_str)
params.append(pytest.param(n, c, spatial, dtype, True,
id=f"{label}-{dtype_str}"))
return params
@pytest.mark.parametrize("n, c, spatial, dtype, tune", _manifest_params())
def test_instance_norm_bench(n: int, c: int, spatial: tuple,
dtype: torch.dtype, tune: bool) -> None:
test = InstanceNormTest(n, c, spatial, dtype)
bm = InstanceNormBenchmark(test)
inputs = test.gen_inputs()
op = InstanceNormFwdOp(N=n, C=c, spatial=spatial, dtype=dtype, tune=tune)
result = bm.profile(op, *inputs)
BenchmarkReport.record(op, locals(), result, tag="tileops")
# Baseline: torch.nn.functional.instance_norm
def baseline_fn(x, weight, bias):
return F.instance_norm(x, weight=weight, bias=bias, eps=1e-5)
result_bl = bm.profile(baseline_fn, *inputs)
BenchmarkReport.record(op, locals(), result_bl, tag="torch")
if __name__ == "__main__":
pytest.main([__file__, "-vvs"])