Skip to content

Fix scanning out-of-bounds items in OpenMP scan #9759

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

Open
bernhardmgruber wants to merge 2 commits into
base: main
Choose a base branch
from
+33 −37
Open

Fix scanning out-of-bounds items in OpenMP scan #9759

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
d5fc08b
Fix scanning out-of-bounds items in OpenMP scan
bernhardmgruber Jul 9, 2026
3078353
Review
bernhardmgruber Jul 9, 2026
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
70 changes: 33 additions & 37 deletions thrust/thrust/system/omp/detail/scan.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -98,22 +98,22 @@ OutputIterator scan_impl(

_CCCL_ASSERT(num_threads > 1, "Parallel scan requires multiple threads");

temporary_array<accum_t, DerivedPolicy> block_sums(exec, num_threads);
// The chosen block size may leave some threads without work
const Size block_size = ::cuda::ceil_div(n, static_cast<Size>(num_threads));
const int active_threads = static_cast<int>(::cuda::ceil_div(n, block_size));

temporary_array<accum_t, DerivedPolicy> block_sums(exec, active_threads);

// Step 1: Reduce each block (N reads)
THRUST_PRAGMA_OMP(parallel num_threads(num_threads))
THRUST_PRAGMA_OMP(parallel num_threads(active_threads))
{
const int tid = omp_get_thread_num();
const Size block_size = ::cuda::ceil_div(n, num_threads);
const Size start = tid * block_size;
const Size end = (::cuda::std::min) (start + block_size, n);
const int tid = omp_get_thread_num();
const Size start = tid * block_size;
const Size end = (::cuda::std::min) (start + block_size, n);

if (start < n)
{
// For both has_init and no-init cases: reduce each block using first element as init
accum_t first_elem = *(first + start);
block_sums[tid] = ::cuda::std::reduce(first + start + 1, first + end, first_elem, wrapped_binary_op);
}
// For both has_init and no-init cases: reduce each block using first element as init
const accum_t first_elem = *(first + start);
block_sums[tid] = ::cuda::std::reduce(first + start + 1, first + end, first_elem, wrapped_binary_op);
}

// Step 2: Scan block sums
Expand All @@ -130,41 +130,37 @@ OutputIterator scan_impl(
}

// Step 3: Scan each block with offset (N reads/writes)
THRUST_PRAGMA_OMP(parallel num_threads(num_threads))
THRUST_PRAGMA_OMP(parallel num_threads(active_threads))
{
const int tid = omp_get_thread_num();
const Size block_size = ::cuda::ceil_div(n, num_threads);
const Size start = tid * block_size;
const Size end = (::cuda::std::min) (start + block_size, n);
const int tid = omp_get_thread_num();
const Size start = tid * block_size;
const Size end = (::cuda::std::min) (start + block_size, n);

if (start < n)
if constexpr (IsInclusive)
{
if constexpr (IsInclusive)
if constexpr (has_init)
{
if constexpr (has_init)
const accum_t prefix = block_sums[tid];
::cuda::std::inclusive_scan(first + start, first + end, result + start, wrapped_binary_op, prefix);
}
else
{
// For no init: thread 0 has no prefix, others use block_sums
if (tid == 0)
{
const accum_t prefix = block_sums[tid];
::cuda::std::inclusive_scan(first + start, first + end, result + start, wrapped_binary_op, prefix);
::cuda::std::inclusive_scan(first + start, first + end, result + start, wrapped_binary_op);
}
else
{
// For no init: thread 0 has no prefix, others use block_sums
if (tid == 0)
{
::cuda::std::inclusive_scan(first + start, first + end, result + start, wrapped_binary_op);
}
else
{
const accum_t prefix = block_sums[tid];
::cuda::std::inclusive_scan(first + start, first + end, result + start, wrapped_binary_op, prefix);
}
const accum_t prefix = block_sums[tid];
::cuda::std::inclusive_scan(first + start, first + end, result + start, wrapped_binary_op, prefix);
}
}
else
{
const accum_t prefix = block_sums[tid];
::cuda::std::exclusive_scan(first + start, first + end, result + start, prefix, wrapped_binary_op);
}
}
else
{
const accum_t prefix = block_sums[tid];
::cuda::std::exclusive_scan(first + start, first + end, result + start, prefix, wrapped_binary_op);
}
}

Expand Down
Loading