Skip to content

Speed up custom reductions #1932

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
oleksandr-pavlyk merged 4 commits into from
Dec 13, 2024
Merged

Speed up custom reductions #1932

Changes from 1 commit
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
4549117
Optimization of custom_reduce_over_group function.
oleksandr-pavlyk Dec 12, 2024
03910f3
Properly set properties of group_load/group_store to striped
oleksandr-pavlyk Dec 12, 2024
e15e3aa
Tweak bounds of cooperative reduction steps
oleksandr-pavlyk Dec 12, 2024
63c1947
Add line for performance improvement in reductions with custom reduce…
oleksandr-pavlyk Dec 12, 2024
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
Prev Previous commit
Next Next commit
Tweak bounds of cooperative reduction steps 
Factor out bounds as constexpr values, reused between power-of-2
branch and not-power-of-two branch.

Lowered lower bounds from 32 to 8 based on pefrormance testing
on PVC and Iris Xe.
  • Loading branch information
@oleksandr-pavlyk
oleksandr-pavlyk committed Dec 12, 2024
commit e15e3aae6e53e045f48d4d10118944d619231766
8 changes: 4 additions & 4 deletions dpctl/tensor/libtensor/include/utils/sycl_utils.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -166,6 +166,8 @@ T custom_reduce_over_group(const GroupT &wg,
const T &local_val,
const OpT &op)
{
constexpr std::uint32_t low_sz = 8u;
constexpr std::uint32_t high_sz = 1024u;
Comment thread
oleksandr-pavlyk marked this conversation as resolved.
const std::uint32_t wgs = wg.get_local_linear_range();
const std::uint32_t lid = wg.get_local_linear_id();

Expand All @@ -176,7 +178,7 @@ T custom_reduce_over_group(const GroupT &wg,
if (wgs & (wgs - 1)) {
// wgs is not a power of 2
#pragma unroll
for (std::uint32_t sz = 1024; sz >= 32; sz >>= 1) {
for (std::uint32_t sz = high_sz; sz >= low_sz; sz >>= 1) {
if (n_witems >= sz) {
const std::uint32_t n_witems_ = (n_witems + 1) >> 1;
_fold(local_mem_acc, lid, n_witems - n_witems_, n_witems_, op);
Expand All @@ -188,7 +190,7 @@ T custom_reduce_over_group(const GroupT &wg,
else {
// wgs is a power of 2
#pragma unroll
for (std::uint32_t sz = 1024; sz >= 32; sz >>= 1) {
for (std::uint32_t sz = high_sz; sz >= low_sz; sz >>= 1) {
if (n_witems >= sz) {
n_witems = (n_witems + 1) >> 1;
Comment thread
oleksandr-pavlyk marked this conversation as resolved.
_fold(local_mem_acc, lid, n_witems, op);
Expand All @@ -204,8 +206,6 @@ T custom_reduce_over_group(const GroupT &wg,
}
}

// sycl::group_barrier(wg, sycl::memory_scope::work_group);

return sycl::group_broadcast(wg, red_val_over_wg, 0);
}

Expand Down