experiments/STT-tensorflow
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

add benchmarks_test option to capture profiles

Browse Source 
PiperOrigin-RevId: 322256578
Change-Id: I895261966df5fcec4104299e4697e889624e08db
This commit is contained in:
Yanhua Sun 2020-07-20 17:02:38 -07:00 committed by TensorFlower Gardener
parent 0a67d25653
commit d5097840b9
1 changed files with 45 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

46
tensorflow/python/eager/benchmarks_test_base.py
View File

@ -18,11 +18,37 @@ from __future__ import absolute_import as _absolute_import
from __future__ import division as _division
from __future__ import print_function as _print_function
import os
import uuid
from tensorflow.python.eager import test
from tensorflow.python.platform import flags
from tensorflow.python.profiler import profiler_v2 as profiler
flags.DEFINE_bool("xprof", False, "Run and report benchmarks with xprof on")
flags.DEFINE_string("logdir", "/tmp/xprof/", "Directory to store xprof data")
class MicroBenchmarksBase(test.Benchmark):
"""Run and report benchmark results."""
"""Run and report benchmark results.
The first run is without any profilng.
Second run is with xprof and python trace. Third run is with xprof without
python trace. Note: xprof runs are with fewer iterations.
"""
def run_with_xprof(self, enable_python_trace, run_benchmark, func,
num_iters_xprof, execution_mode, suid):
if enable_python_trace:
options = profiler.ProfilerOptions(python_tracer_level=1)
logdir = os.path.join(flags.FLAGS.logdir, suid + "_with_python")
else:
options = profiler.ProfilerOptions(python_tracer_level=0)
logdir = os.path.join(flags.FLAGS.logdir, suid)
with profiler.Profile(logdir, options):
total_time = run_benchmark(func, num_iters_xprof, execution_mode)
us_per_example = float("{0:.3f}".format(total_time * 1e6 / num_iters_xprof))
return logdir, us_per_example
def run_report(self, run_benchmark, func, num_iters, execution_mode=None):
"""Run and report benchmark results."""
@ -32,6 +58,24 @@ class MicroBenchmarksBase(test.Benchmark):
"examples_per_sec": float("{0:.3f}".format(num_iters / total_time)),
"us_per_example": float("{0:.3f}".format(total_time * 1e6 / num_iters))
}
if flags.FLAGS.xprof:
suid = str(uuid.uuid4())
# Re-run with xprof and python trace.
num_iters_xprof = min(100, num_iters)
xprof_link, us_per_example = self.run_with_xprof(True, run_benchmark,
func, num_iters_xprof,
execution_mode, suid)
extras["xprof link with python trace"] = xprof_link
extras["us_per_example with xprof and python"] = us_per_example
# Re-run with xprof but no python trace.
xprof_link, us_per_example = self.run_with_xprof(False, run_benchmark,
func, num_iters_xprof,
execution_mode, suid)
extras["xprof link"] = xprof_link
extras["us_per_example with xprof"] = us_per_example
benchmark_name = self._get_benchmark_name()
self.report_benchmark(
iters=num_iters, wall_time=mean_us, extras=extras, name=benchmark_name)