Skip to content

Add workaround for ACPP allocation tracker destruction order #343

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

Merged
GagaLP merged 1 commit into from
Oct 21, 2026
Merged

Add workaround for ACPP allocation tracker destruction order #343

Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
8 changes: 8 additions & 0 deletions src/backend/sycl_backend.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -181,6 +181,14 @@ struct sycl_backend::impl {
};

sycl_backend::sycl_backend(const std::vector<sycl::device>& devices, const configuration& config) : m_impl(new impl(devices, config)) {
#if CELERITY_WORKAROUND(ACPP)
// AdaptiveCpp tracks USM allocations in a function-local static that is created by the first allocation. The runtime (and with it, all backend
// allocations) is torn down from an atexit handler that is registered right after the backend has been constructed. Unless the tracker already exists
// at that point it will be destroyed before that handler runs, and every sycl::free() during shutdown then operates on a dead object, which corrupts the
// heap (observed as a hang in release builds and a segfault in debug builds). Allocating once here establishes the correct destruction order.
sycl::free(sycl::aligned_alloc_host(alignof(std::max_align_t), 1, m_impl->host.sycl_context), m_impl->host.sycl_context);
#endif

// Initialize a submission thread with hydrator for each device, if they are enabled
if(m_impl->config.per_device_submission_threads) {
for(device_id did = 0; did < m_impl->system.devices.size(); ++did) {
Expand Down
Loading