Minor follow-up fixes for ivf-flat

cpp/include/raft/spatial/knn/detail/ivf_flat_search.cuh

@@ -696,8 +696,8 @@ __global__ void __launch_bounds__(kThreadsPerBlock)
   copy_vectorized(query_shared, query, std::min(dim, query_smem_elems));
   __syncthreads();
 
-  topk::block_sort<topk::warp_sort_filtered, Capacity, Ascending, float, IdxT> queue(
-    k, interleaved_scan_kernel_smem + query_smem_elems * sizeof(T));
+  using block_sort_t = topk::block_sort<topk::warp_sort_filtered, Capacity, Ascending, float, IdxT>;
+  block_sort_t queue(k, interleaved_scan_kernel_smem + query_smem_elems * sizeof(T));
 
   {
     using align_warp  = Pow2<WarpSize>;
@@ -766,8 +766,7 @@ __global__ void __launch_bounds__(kThreadsPerBlock)
         }
 
         // Enqueue one element per thread
-        constexpr float kDummy = Ascending ? upper_bound<float>() : lower_bound<float>();
-        const float val        = valid ? static_cast<float>(dist) : kDummy;
+        const float val  = valid ? static_cast<float>(dist) : block_sort_t::queue_t::kDummy;
         const size_t idx = valid ? static_cast<size_t>(list_indices[list_offset + vec_id]) : 0;
         queue.add(val, idx);
       }
@@ -826,7 +825,7 @@ void launch_kernel(Lambda lambda,
     std::min<int>(max_query_smem / sizeof(T), Pow2<Veclen * WarpSize>::roundUp(index.dim()));
   int smem_size              = query_smem_elems * sizeof(T);
   constexpr int kSubwarpSize = std::min<int>(Capacity, WarpSize);
-  smem_size += raft::spatial::knn::detail::topk::calc_smem_size_for_block_wide<AccT, size_t>(
+  smem_size += raft::spatial::knn::detail::topk::calc_smem_size_for_block_wide<AccT, IdxT>(
     kThreadsPerBlock / kSubwarpSize, k);
 
   // power-of-two less than cuda limit (for better addr alignment)

cpp/include/raft/spatial/knn/detail/topk/warpsort_topk.cuh

@@ -135,6 +135,7 @@ constexpr auto calc_capacity(int k) -> int
 template <int Capacity, bool Ascending, typename T, typename IdxT>
 class warp_sort {
   static_assert(isPo2(Capacity));
+  static_assert(std::is_default_constructible_v<IdxT>);
 
  public:
   /**
@@ -158,6 +159,7 @@ class warp_sort {
 #pragma unroll
     for (int i = 0; i < kMaxArrLen; i++) {
       val_arr_[i] = kDummy;
+      idx_arr_[i] = IdxT{};
     }
   }
 
@@ -280,6 +282,7 @@ class warp_sort_filtered : public warp_sort<Capacity, Ascending, T, IdxT> {
 #pragma unroll
     for (int i = 0; i < kMaxBufLen; i++) {
       val_buf_[i] = kDummy;
+      idx_buf_[i] = IdxT{};
     }
   }
 
@@ -371,6 +374,7 @@ class warp_sort_immediate : public warp_sort<Capacity, Ascending, T, IdxT> {
 #pragma unroll
     for (int i = 0; i < kMaxArrLen; i++) {
       val_buf_[i] = kDummy;
+      idx_buf_[i] = IdxT{};
     }
   }
 
@@ -429,9 +433,9 @@ template <template <int, bool, typename, typename> class WarpSortWarpWide,
           typename T,
           typename IdxT>
 class block_sort {
+ public:
   using queue_t = WarpSortWarpWide<Capacity, Ascending, T, IdxT>;
 
- public:
   __device__ block_sort(int k, uint8_t* smem_buf) : queue_(k)
   {
     val_smem_             = reinterpret_cast<T*>(smem_buf);

cpp/test/spatial/ann_ivf_flat.cu

@@ -20,7 +20,6 @@
 #include <raft/core/logger.hpp>
 #include <raft/distance/distance_type.hpp>
 #include <raft/random/rng.cuh>
-#include <raft/sparse/detail/utils.h>
 #include <raft/spatial/knn/ann.cuh>
 #include <raft/spatial/knn/ivf_flat.cuh>
 #include <raft/spatial/knn/knn.cuh>
@@ -30,6 +29,8 @@
 
 #include <gtest/gtest.h>
 
+#include <thrust/sequence.h>
+
 #include <cstddef>
 #include <iostream>
 #include <vector>
@@ -209,7 +210,9 @@ class AnnIVFFlatTest : public ::testing::TestWithParam<AnnIvfFlatInputs> {
           ivf_flat::build(handle_, index_params, database.data(), int64_t(ps.num_db_vecs), ps.dim);
 
         rmm::device_uvector<int64_t> vector_indices(ps.num_db_vecs, stream_);
-        sparse::iota_fill(vector_indices.data(), int64_t(ps.num_db_vecs), int64_t(1), stream_);
+        thrust::sequence(handle_.get_thrust_policy(),
+                         thrust::device_pointer_cast(vector_indices.data()),
+                         thrust::device_pointer_cast(vector_indices.data() + ps.num_db_vecs));
         handle_.sync_stream(stream_);
 
         int64_t half_of_data = ps.num_db_vecs / 2;