Skip to content

SYCL: skip grid-stride loop when not needed for MDRange #9250

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

Draft
masterleinad wants to merge 2 commits into
base: develop
Choose a base branch
from
Draft

SYCL: skip grid-stride loop when not needed for MDRange #9250

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
95ef08e
SYCL: skip grid-stride loop when not needed for MDRange (align with P…
masterleinad Jun 5, 2026
fdf6370
Clean-up
masterleinad Jun 5, 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
81 changes: 72 additions & 9 deletions core/src/SYCL/Kokkos_SYCL_ParallelFor_MDRange.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -35,9 +35,10 @@ class Kokkos::Impl::ParallelFor<FunctorType, Kokkos::MDRangePolicy<Traits...>,
array_type m_upper;
array_type m_extent; // tile_size * num_tiles

template <typename FunctorWrapper>
template <bool UseStride, typename FunctorWrapper>
sycl::event sycl_direct_launch(const FunctorWrapper& functor_wrapper,
const sycl::event& memcpy_event) const {
const sycl::event& memcpy_event,
const sycl::nd_range<3>& range) const {
// Convenience references
const Kokkos::SYCL& space = m_policy.space();
sycl::queue& q = space.sycl_queue();
Expand All @@ -50,9 +51,6 @@ class Kokkos::Impl::ParallelFor<FunctorType, Kokkos::MDRangePolicy<Traits...>,

desul::ensure_sycl_lock_arrays_on_device(q);

const auto range =
Kokkos::Impl::compute_device_launch_params(m_policy, m_max_grid_size);

auto cgh_lambda = [&, range](sycl::handler& cgh) {
const sycl::range<3> global_range = range.get_global_range();
const sycl::range<3> local_range = range.get_local_range();
Expand Down Expand Up @@ -83,8 +81,8 @@ class Kokkos::Impl::ParallelFor<FunctorType, Kokkos::MDRangePolicy<Traits...>,
const index_type n_global_z = item.get_group_range(0);

Kokkos::Impl::DeviceIterate<Policy::rank, array_index_type, index_type,
Policy::inner_direction, true, FunctorType,
typename Policy::work_tag>(
Policy::inner_direction, UseStride,
FunctorType, typename Policy::work_tag>(
lower_bound, upper_bound, extent, functor_wrapper.get_functor(),
{n_global_x, n_global_y, n_global_z},
{n_local_x, n_local_y, n_local_z}, {global_x, global_y, global_z},
Expand Down Expand Up @@ -150,8 +148,73 @@ class Kokkos::Impl::ParallelFor<FunctorType, Kokkos::MDRangePolicy<Traits...>,

auto functor_wrapper =
Impl::make_sycl_function_wrapper(m_functor, indirectKernelMem);
sycl::event event =
sycl_direct_launch(functor_wrapper, functor_wrapper.get_copy_event());

// Compute launch range once and determine if a grid-stride loop is needed
const auto range =
Kokkos::Impl::compute_device_launch_params(m_policy, m_max_grid_size);
const sycl::range<3> local_range = range.get_local_range();

bool need_grid_stride = true;

if constexpr (Policy::rank == 1) {
if ((m_max_grid_size[0] * local_range[0]) >=
static_cast<std::size_t>(m_extent[0])) {
need_grid_stride = false;
}
} else if constexpr (Policy::rank == 2) {
if ((m_max_grid_size[0] * local_range[0]) >=
static_cast<std::size_t>(m_extent[0]) &&
(m_max_grid_size[1] * local_range[1]) >=
static_cast<std::size_t>(m_extent[1])) {
need_grid_stride = false;
}
} else if constexpr (Policy::rank == 3) {
if ((m_max_grid_size[0] * local_range[0]) >=
static_cast<std::size_t>(m_extent[0]) &&
(m_max_grid_size[1] * local_range[1]) >=
static_cast<std::size_t>(m_extent[1]) &&
(m_max_grid_size[2] * local_range[2]) >=
static_cast<std::size_t>(m_extent[2])) {
need_grid_stride = false;
}
} else if constexpr (Policy::rank == 4) {
if ((m_max_grid_size[0] * local_range[0]) >=
static_cast<std::size_t>(m_extent[0] * m_extent[1]) &&
(m_max_grid_size[1] * local_range[1]) >=
static_cast<std::size_t>(m_extent[2]) &&
(m_max_grid_size[2] * local_range[2]) >=
static_cast<std::size_t>(m_extent[3])) {
need_grid_stride = false;
}
} else if constexpr (Policy::rank == 5) {
if ((m_max_grid_size[0] * local_range[0]) >=
static_cast<std::size_t>(m_extent[0] * m_extent[1]) &&
(m_max_grid_size[1] * local_range[1]) >=
static_cast<std::size_t>(m_extent[2] * m_extent[3]) &&
(m_max_grid_size[2] * local_range[2]) >=
static_cast<std::size_t>(m_extent[4])) {
need_grid_stride = false;
}
} else if constexpr (Policy::rank == 6) {
if ((m_max_grid_size[0] * local_range[0]) >=
static_cast<std::size_t>(m_extent[0] * m_extent[1]) &&
(m_max_grid_size[1] * local_range[1]) >=
static_cast<std::size_t>(m_extent[2] * m_extent[3]) &&
(m_max_grid_size[2] * local_range[2]) >=
static_cast<std::size_t>(m_extent[4] * m_extent[5])) {
need_grid_stride = false;
}
}

sycl::event event;
if (need_grid_stride) {
event = sycl_direct_launch<true>(functor_wrapper,
functor_wrapper.get_copy_event(), range);
} else {
event = sycl_direct_launch<false>(
functor_wrapper, functor_wrapper.get_copy_event(), range);
}

functor_wrapper.register_event(event);
}

Expand Down
Loading