Skip to content

Don't duplicate a device-resident dataset in CAGRA sort_knn_graph #2834

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
maxwbuckley wants to merge 2 commits into
base: main
Choose a base branch
from
Open

Don't duplicate a device-resident dataset in CAGRA sort_knn_graph #2834

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
9e52aea
cagra: do not duplicate a device-resident dataset in sort_knn_graph
maxwbuckley Oct 11, 2026
4121201
cagra: keep staging datasets owned by another GPU in sort_knn_graph
maxwbuckley Oct 11, 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
27 changes: 23 additions & 4 deletions cpp/src/neighbors/detail/cagra/graph_core.cuh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -14,6 +14,8 @@
#include <raft/core/host_mdspan.hpp>
#include <raft/core/mdspan.hpp>
#include <raft/core/resource/cuda_stream.hpp>
#include <raft/core/resource/device_id.hpp>
#include <raft/util/cudart_utils.hpp>
#include <raft/core/resources.hpp>
#include <raft/matrix/init.cuh>

Expand All @@ -38,6 +40,7 @@
#include <climits>
#include <iostream>
#include <memory>
#include <optional>
#include <random>
#include <type_traits>

Expand Down Expand Up @@ -895,16 +898,32 @@ void sort_knn_graph(
const double time_sort_start = cur_time();
RAFT_LOG_DEBUG("# Sorting kNN Graph on GPUs ");

auto d_dataset = raft::make_device_mdarray<DataT>(
res, large_tmp_mr, raft::make_extents<int64_t>(dataset_size, dataset_dim));
raft::copy(res, d_dataset.view(), dataset);
// A dataset in device memory of the resource's GPU (or in managed memory) is read in place
// instead of being duplicated into a second full-size device buffer. Host datasets, and device
// allocations owned by another GPU (which the sort kernel cannot read without peer access), are
// staged as before.
bool stage_dataset = !d_accessor::is_device_accessible;
if constexpr (d_accessor::is_device_accessible) {
const int dataset_device = raft::get_device_for_address(dataset_ptr);
stage_dataset = dataset_device >= 0 && dataset_device != raft::resource::get_device_id(res);
}
using d_dataset_t = decltype(raft::make_device_mdarray<DataT>(
res, large_tmp_mr, raft::make_extents<int64_t>(dataset_size, dataset_dim)));
std::optional<d_dataset_t> d_dataset;
const DataT* d_dataset_ptr = dataset_ptr;
if (stage_dataset) {
d_dataset.emplace(raft::make_device_mdarray<DataT>(
res, large_tmp_mr, raft::make_extents<int64_t>(dataset_size, dataset_dim)));
raft::copy(res, d_dataset->view(), dataset);
d_dataset_ptr = d_dataset->data_handle();
}

raft::copy(res, d_input_graph.view(), knn_graph);

RAFT_LOG_DEBUG(".");
launch_sort_knn_graph(res,
metric,
d_dataset.data_handle(),
d_dataset_ptr,
static_cast<uint32_t>(dataset_size),
static_cast<uint32_t>(dataset_dim),
d_input_graph.data_handle(),
Expand Down
Loading