Skip to content

Make CUDACachingAllocator::recordStream() a no-op on null ptrs #20658

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
mrshenli wants to merge 2 commits into from
Closed

Make CUDACachingAllocator::recordStream() a no-op on null ptrs #20658

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
06eb014
Make CUDACachingAllocator::recordStream() a no-op on null ptrs
mrshenli May 17, 2019
5de2414
address test failures
mrshenli May 17, 2019
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
24 changes: 14 additions & 10 deletions c10/cuda/CUDACachingAllocator.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -378,17 +378,21 @@ struct THCCachingAllocator

void recordStream(void* ptr, cuda::CUDAStream stream)
{
std::lock_guard<std::recursive_mutex> lock(mutex);
Block* block = find_allocated_block(ptr);
if (!block) {
AT_ERROR("invalid device pointer: ", ptr);
}
if (stream.stream() == block->stream) {
// ignore uses on the allocation stream, since those don't require any
// special synchronization
return;
// Empty tensor's storage().data() might be a null ptr. As there is no
// blocks associated with those tensors, it is fine to do nothing here.
if (ptr) {
std::lock_guard<std::recursive_mutex> lock(mutex);
Block* block = find_allocated_block(ptr);
if (!block) {
AT_ERROR("invalid device pointer: ", ptr);
}
if (stream.stream() == block->stream) {
// ignore uses on the allocation stream, since those don't require any
// special synchronization
return;
}
block->stream_uses.insert(stream);
}
block->stream_uses.insert(stream);
}

/** moves a block into a pool of cached free blocks */
Expand Down
32 changes: 32 additions & 0 deletions test/test_c10d.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -579,6 +579,14 @@ def opts(self, threads=2):
opts.threads = threads
return opts

def test_empty_tensors(self):
store = c10d.FileStore(self.file.name, self.world_size)
pg = c10d.ProcessGroupGloo(store, self.rank, self.world_size, self.opts())

xs = [torch.FloatTensor([])]
pg.broadcast(xs).wait()
self.assertEqual(0, xs[0].numel())

def test_broadcast_checks(self):
store = c10d.FileStore(self.file.name, self.world_size)
pg = c10d.ProcessGroupGloo(store, self.rank, self.world_size, self.opts())
Expand Down Expand Up @@ -1344,6 +1352,30 @@ def setUp(self):
def tearDown(self):
pass

def test_empty_tensors(self):
store = c10d.FileStore(self.file.name, self.world_size)
pg = c10d.ProcessGroupNCCL(store, self.rank, self.world_size)

xs = [torch.cuda.FloatTensor([])]
pg.broadcast(xs).wait()
self.assertEqual(0, xs[0].numel())

pg.allreduce(xs).wait()
self.assertEqual(0, xs[0].numel())

pg.reduce(xs).wait()
self.assertEqual(0, xs[0].numel())

ys = [[torch.cuda.FloatTensor([]) for _ in range(self.world_size)]]
pg.allgather(ys, xs).wait()
for y in ys[0]:
self.assertEqual(0, y.numel())

ys = [torch.cuda.FloatTensor([])]
xs = [[torch.cuda.FloatTensor([]) for _ in range(self.world_size)]]
pg.reduce_scatter(ys, xs).wait()
self.assertEqual(0, ys[0].numel())

def test_broadcast_ops(self):
store = c10d.FileStore(self.file.name, self.world_size)
pg = c10d.ProcessGroupNCCL(store, self.rank, self.world_size)
Expand Down