Avoid large device allocation in UMAP with nndescent (#6292)

Currently `NNDescent` returns two arrays:

- `graph.graph()`: (n x graph_degree) on host
- `graph.distances()`: (n x graph_degree) on device

Downstream, the rest of UMAP wants both of these to be device arrays of shape (n x n_neighbors).

Currently we copy `graph.graph()` to a temporary device array, then slice and and copy it to the output array `out.knn_indices`.

Ideally we'd force `graph_degree = n_neighbors` to avoid the slicing entirely (and reduce the size of the intermediate results). However, it seems like currently there's a bug in `NNDescent` where reducing `graph_degree` to `n_neighbors` causes a significant decrease in result quality. So for now we need to keep the slicing around.

We can avoid allocating the temporary device array though, instead doing the slicing on host. Doing this avoids allocating a (n x graph_degree) device array entirely; for large `n` this can be a significant savings (47 GiB on one test problem I was trying).

We still should fix the `graph_degree` issue, but for now this should help unblock running UMAP on very large datasets.

Authors:
  - Jim Crist-Harif (https://github.com/jcrist)

Approvers:
  - Divye Gala (https://github.com/divyegala)
  - William Hicks (https://github.com/wphicks)

URL: #6292

cpp/src/umap/knn_graph/algo.cuh

@@ -118,31 +118,42 @@ inline void launcher(const raft::handle_t& handle,
     // TODO:  use nndescent from cuvs
     RAFT_EXPECTS(static_cast<size_t>(n_neighbors) <= params->nn_descent_params.graph_degree,
                  "n_neighbors should be smaller than the graph degree computed by nn descent");
+    RAFT_EXPECTS(params->nn_descent_params.return_distances,
+                 "return_distances for nn descent should be set to true to be used for UMAP");
 
     auto graph = get_graph_nnd(handle, inputsA, params);
 
-    auto indices_d = raft::make_device_matrix<int64_t, int64_t>(
-      handle, inputsA.n, params->nn_descent_params.graph_degree);
-
-    raft::copy(indices_d.data_handle(),
-               graph.graph().data_handle(),
-               inputsA.n * params->nn_descent_params.graph_degree,
-               stream);
-
+    // `graph.graph()` is a host array (n x graph_degree).
+    // Slice and copy to a temporary host array (n x n_neighbors), then copy
+    // that to the output device array `out.knn_indices` (n x n_neighbors).
+    // TODO: force graph_degree = n_neighbors so the temporary host array and
+    // slice isn't necessary.
+    auto temp_indices_h = raft::make_host_matrix<int64_t, int64_t>(inputsA.n, n_neighbors);
+    size_t graph_degree = params->nn_descent_params.graph_degree;
+#pragma omp parallel for
+    for (size_t i = 0; i < static_cast<size_t>(inputsA.n); i++) {
+      for (int j = 0; j < n_neighbors; j++) {
+        auto target                 = temp_indices_h.data_handle();
+        auto source                 = graph.graph().data_handle();
+        target[i * n_neighbors + j] = source[i * graph_degree + j];
+      }
+    }
+    raft::copy(handle,
+               raft::make_device_matrix_view(out.knn_indices, inputsA.n, n_neighbors),
+               temp_indices_h.view());
+
+    // `graph.distances()` is a device array (n x graph_degree).
+    // Slice and copy to the output device array `out.knn_dists` (n x n_neighbors).
+    // TODO: force graph_degree = n_neighbors so this slice isn't necessary.
     raft::matrix::slice_coordinates coords{static_cast<int64_t>(0),
                                            static_cast<int64_t>(0),
                                            static_cast<int64_t>(inputsA.n),
                                            static_cast<int64_t>(n_neighbors)};
-
-    RAFT_EXPECTS(graph.distances().has_value(),
-                 "return_distances for nn descent should be set to true to be used for UMAP");
-    auto out_knn_dists_view = raft::make_device_matrix_view(out.knn_dists, inputsA.n, n_neighbors);
     raft::matrix::slice<float, int64_t, raft::row_major>(
-      handle, raft::make_const_mdspan(graph.distances().value()), out_knn_dists_view, coords);
-    auto out_knn_indices_view =
-      raft::make_device_matrix_view(out.knn_indices, inputsA.n, n_neighbors);
-    raft::matrix::slice<int64_t, int64_t, raft::row_major>(
-      handle, raft::make_const_mdspan(indices_d.view()), out_knn_indices_view, coords);
+      handle,
+      raft::make_const_mdspan(graph.distances().value()),
+      raft::make_device_matrix_view(out.knn_dists, inputsA.n, n_neighbors),
+      coords);
   }
 }