saymrwulf/pytorch
1
0
Fork 0
mirror of https://github.com/saymrwulf/pytorch.git synced 2026-05-14 20:57:59 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Strengthening unit test cases for quueue depth calculation

Browse source 
Pull Request resolved: https://github.com/pytorch/pytorch/pull/79974

Approved by: https://github.com/robieta
This commit is contained in:
David Chen 2022-06-22 19:30:46 +00:00 committed by PyTorch MergeBot
parent 14eadf937b
commit 38ee99b382
1 changed files with 18 additions and 3 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

21
test/test_profiler.py
View file

@ -1131,7 +1131,7 @@ class TestProfiler(TestCase):
@unittest.skipIf(not torch.cuda.is_available(), "CUDA is required")
def test_utils_compute_queue_depth(self):
x = torch.ones((4096, 4096), device="cuda")
x = torch.ones((8096, 8096), device="cuda")
with profile() as prof:
# First half we want it to be compute bound
for _ in range(5):
@ -1143,8 +1143,23 @@ class TestProfiler(TestCase):
y[0] += 1
time.sleep(0.1)
basic_evaluation = _utils.BasicEvaluation(prof.profiler)
for entry in basic_evaluation.compute_queue_depth():
self.assertTrue(entry.queue_depth >= 0)
# We can assume golden because mm is compute intensive,
# so kernel will queued up.
# But later tensor indexing is overhead bound, and there
# is sleep to make sure kernel finished before next dispatch.
golden_queue_depth_list = [1, 2, 3, 4, 5, 1, 1, 1]
for entry, golden in zip(basic_evaluation.compute_queue_depth(),
golden_queue_depth_list):
self.assertTrue(entry.queue_depth == golden)
def test_utils_compute_queue_depth_when_no_cuda_events(self):
# For traces with only cpu events, we expect empty queue depth list
x = torch.ones((1024, 1024))
with profile() as prof:
for _ in range(5):
x = x @ x
basic_evaluation = _utils.BasicEvaluation(prof.profiler)
self.assertFalse(basic_evaluation.compute_queue_depth())
def test_extra_fields(self):