[ENH] [3/5] Header structure: force explicit instantiation in tests and benchmarks

README.md

@@ -198,7 +198,7 @@ RAFT itself can be installed through conda, [CMake Package Manager (CPM)](https:
 
 The easiest way to install RAFT is through conda and several packages are provided.
 - `libraft-headers` RAFT headers
-- `libraft` (optional) shared library of pre-compiled template specializations and runtime APIs.
+- `libraft` (optional) shared library of pre-compiled template instantiations and runtime APIs.
 - `pylibraft` (optional) Python wrappers around RAFT algorithms and primitives.
 - `raft-dask` (optional) enables deployment of multi-node multi-GPU algorithms that use RAFT `raft::comms` in Dask clusters.
 
@@ -231,11 +231,11 @@ You can find an [example RAFT](cpp/template/README.md) project template in the `
 
 Additional CMake targets can be made available by adding components in the table below to the `RAFT_COMPONENTS` list above, separated by spaces. The `raft::raft` target will always be available. RAFT headers require, at a minimum, the CUDA toolkit libraries and RMM dependencies.
 
-| Component   | Target              | Description                                               | Base Dependencies                     |
-|-------------|---------------------|-----------------------------------------------------------|---------------------------------------|
-| n/a         | `raft::raft`        | Full RAFT header library                                  | CUDA toolkit, RMM, NVTX, CCCL, CUTLASS |
-| compiled    | `raft::compiled`    | Pre-compiled template specializations and runtime library | raft::raft                            |
-| distributed | `raft::distributed` | Dependencies for `raft::comms` APIs                       | raft::raft, UCX, NCCL                 |
+| Component   | Target              | Description                                              | Base Dependencies                      |
+|-------------|---------------------|----------------------------------------------------------|----------------------------------------|
+| n/a         | `raft::raft`        | Full RAFT header library                                 | CUDA toolkit, RMM, NVTX, CCCL, CUTLASS |
+| compiled    | `raft::compiled`    | Pre-compiled template instantiations and runtime library | raft::raft                             |
+| distributed | `raft::distributed` | Dependencies for `raft::comms` APIs                      | raft::raft, UCX, NCCL                  |
 
 ### Source
 
@@ -282,7 +282,7 @@ The folder structure mirrors other RAPIDS repos, with the following folders:
     - `util`: Various reusable tools and utilities for accelerated algorithm development
   - `internal`: A private header-only component that hosts the code shared between benchmarks and tests.
   - `scripts`: Helpful scripts for development
-  - `src`: Compiled APIs and template specializations for the shared libraries
+  - `src`: Compiled APIs and template instantiations for the shared libraries
   - `template`: A skeleton template containing the bare-bones file structure and cmake configuration for writing applications with RAFT.
   - `test`: Googletests source code
 - `docs`: Source code and scripts for building library documentation (Uses breath, doxygen, & pydocs)

cpp/bench/ann/src/raft/raft_benchmark.cu

@@ -22,10 +22,6 @@
 #include <type_traits>
 #include <utility>
 
-#ifdef RAFT_COMPILED
-#include <raft/neighbors/specializations.cuh>
-#endif
-
 #include "../common/ann_types.hpp"
 #include "../common/benchmark_util.hpp"
 #undef WARP_SIZE

cpp/bench/ann/src/raft/raft_ivf_flat.cu

@@ -15,12 +15,8 @@
  */
 #include "raft_ivf_flat_wrapper.h"
 
-#ifdef RAFT_COMPILED
-#include <raft/neighbors/specializations.cuh>
-#endif
-
 namespace raft::bench::ann {
 template class RaftIvfFlatGpu<float, int64_t>;
 template class RaftIvfFlatGpu<uint8_t, int64_t>;
 template class RaftIvfFlatGpu<int8_t, int64_t>;
-}  // namespace raft::bench::ann
+}  // namespace raft::bench::ann

cpp/bench/ann/src/raft/raft_ivf_flat_wrapper.h

@@ -29,6 +29,7 @@
 #include <raft/neighbors/ivf_flat_types.hpp>
 #include <raft/util/cudart_utils.hpp>
 #include <rmm/device_uvector.hpp>
+#include <rmm/mr/device/pool_memory_resource.hpp>
 #include <stdexcept>
 #include <string>
 #include <type_traits>

cpp/bench/ann/src/raft/raft_ivf_pq.cu

@@ -15,10 +15,6 @@
  */
 #include "raft_ivf_pq_wrapper.h"
 
-#ifdef RAFT_COMPILED
-#include <raft/neighbors/specializations.cuh>
-#endif
-
 namespace raft::bench::ann {
 template class RaftIvfPQ<float, int64_t>;
 template class RaftIvfPQ<uint8_t, int64_t>;

cpp/bench/prims/cluster/kmeans.cu

@@ -18,10 +18,6 @@
 #include <raft/cluster/kmeans.cuh>
 #include <raft/cluster/kmeans_types.hpp>
 
-#if defined RAFT_COMPILED
-#include <raft/cluster/specializations.cuh>
-#endif
-
 namespace raft::bench::cluster {
 
 struct KMeansBenchParams {

cpp/bench/prims/cluster/kmeans_balanced.cu

@@ -18,10 +18,6 @@
 #include <raft/cluster/kmeans_balanced.cuh>
 #include <raft/random/rng.cuh>
 
-#if defined RAFT_COMPILED
-#include <raft/cluster/specializations.cuh>
-#endif
-
 namespace raft::bench::cluster {
 
 struct KMeansBalancedBenchParams {

cpp/bench/prims/distance/distance_common.cuh

@@ -17,9 +17,6 @@
 #include <common/benchmark.hpp>
 #include <raft/distance/distance.cuh>
 #include <raft/util/cudart_utils.hpp>
-#if defined RAFT_COMPILED
-#include <raft/distance/specializations.cuh>
-#endif
 #include <rmm/device_uvector.hpp>
 
 namespace raft::bench::distance {

cpp/bench/prims/distance/fused_l2_nn.cu

@@ -16,10 +16,8 @@
 
 #include <common/benchmark.hpp>
 #include <raft/distance/fused_l2_nn.cuh>
+#include <raft/linalg/norm.cuh>
 #include <raft/util/cudart_utils.hpp>
-#if defined RAFT_COMPILED
-#include <raft/distance/specializations.cuh>
-#endif
 #include <rmm/device_uvector.hpp>
 
 namespace raft::bench::distance {

cpp/bench/prims/distance/kernels.cu

@@ -13,10 +13,6 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  */
-#if defined RAFT_COMPILED
-#include <raft/distance/specializations.cuh>
-#endif
-
 #include <common/benchmark.hpp>
 #include <memory>
 #include <raft/core/device_resources.hpp>

cpp/bench/prims/distance/masked_nn.cu

@@ -30,10 +30,6 @@
 #include <raft/random/rng.cuh>
 #include <raft/util/cudart_utils.hpp>
 
-#ifdef RAFT_COMPILED
-#include <raft/distance/specializations.cuh>
-#endif
-
 namespace raft::bench::distance::masked_nn {
 
 // Introduce various sparsity patterns

cpp/bench/prims/matrix/select_k.cu

@@ -23,10 +23,6 @@
 #include <raft/sparse/detail/utils.h>
 #include <raft/util/cudart_utils.hpp>
 
-#if defined RAFT_COMPILED
-#include <raft/matrix/specializations.cuh>
-#endif
-
 #include <raft/matrix/detail/select_radix.cuh>
 #include <raft/matrix/detail/select_warpsort.cuh>
 #include <raft/matrix/select_k.cuh>

cpp/bench/prims/neighbors/knn.cuh

@@ -24,10 +24,6 @@
 #include <raft/neighbors/ivf_pq.cuh>
 #include <raft/spatial/knn/knn.cuh>
 
-#if defined RAFT_COMPILED
-#include <raft/neighbors/specializations.cuh>
-#endif
-
 #include <rmm/mr/device/managed_memory_resource.hpp>
 #include <rmm/mr/device/per_device_resource.hpp>
 

cpp/bench/prims/neighbors/refine_float_int64_t.cu

@@ -17,11 +17,6 @@
 #include "refine.cuh"
 #include <common/benchmark.hpp>
 
-#if defined RAFT_COMPILED
-#include <raft/neighbors/specializations/refine.cuh>
-#include <raft/spatial/knn/specializations.cuh>
-#endif
-
 using namespace raft::neighbors;
 
 namespace raft::bench::neighbors {

cpp/bench/prims/neighbors/refine_uint8_t_int64_t.cu

@@ -17,10 +17,6 @@
 #include "refine.cuh"
 #include <common/benchmark.hpp>
 
-#if defined RAFT_COMPILED
-#include <raft/neighbors/specializations.cuh>
-#endif
-
 using namespace raft::neighbors;
 
 namespace raft::bench::neighbors {

cpp/doxygen/Doxyfile

@@ -918,6 +918,7 @@ EXCLUDE_SYMLINKS       = NO
 # Note that the wildcards are matched against the file with absolute path, so to
 # exclude all test directories for example use the pattern */test/*
 
+# TODO: remove specializations from exclude patterns when headers have been removed.
 EXCLUDE_PATTERNS       = */detail/* \
                          */specializations/* \
                          */thirdparty/*

cpp/include/raft/cluster/detail/kmeans_common.cuh

@@ -38,6 +38,7 @@
 #include <raft/distance/distance.cuh>
 #include <raft/distance/distance_types.hpp>
 #include <raft/distance/fused_l2_nn.cuh>
+#include <raft/linalg/norm.cuh>
 #include <raft/linalg/reduce_rows_by_key.cuh>
 #include <raft/linalg/unary_op.cuh>
 #include <raft/matrix/gather.cuh>

cpp/include/raft/cluster/specializations.cuh

@@ -13,12 +13,10 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  */
-#ifndef __CLUSTER_SPECIALIZATIONS_H
-#define __CLUSTER_SPECIALIZATIONS_H
-
 #pragma once
 
-#include <raft/distance/specializations.cuh>
-#include <raft/neighbors/specializations.cuh>
-
-#endif
+#pragma message(                                          \
+  __FILE__                                                \
+  " is deprecated and will be removed."                   \
+  " Including specializations is not necessary any more." \
+  " For more information, see: https://docs.rapids.ai/api/raft/nightly/using_libraft.html")

cpp/include/raft/core/mdarray.hpp

@@ -25,6 +25,7 @@
 #include <stddef.h>
 
 #include <raft/core/detail/macros.hpp>
+#include <raft/core/device_resources.hpp>
 #include <raft/core/host_device_accessor.hpp>
 #include <raft/core/mdspan.hpp>
 #include <raft/core/mdspan_types.hpp>

cpp/include/raft/core/resource/device_memory_resource.hpp

@@ -18,6 +18,7 @@
 #include <raft/core/resource/resource_types.hpp>
 #include <raft/core/resources.hpp>
 #include <rmm/mr/device/device_memory_resource.hpp>
+#include <rmm/mr/device/per_device_resource.hpp>
 
 namespace raft::resource {
 class device_memory_resource : public resource {
@@ -72,4 +73,4 @@ inline void set_workspace_resource(resources const& res, rmm::mr::device_memory_
 {
   res.add_resource_factory(std::make_shared<workspace_resource_factory>(mr));
 };
-}  // namespace raft::resource
+}  // namespace raft::resource

cpp/include/raft/core/resources.hpp

@@ -18,6 +18,7 @@
 #include "resource/resource_types.hpp"
 #include <algorithm>
 #include <mutex>
+#include <raft/core/error.hpp>  // RAFT_EXPECTS
 #include <raft/core/logger.hpp>
 #include <string>
 #include <vector>
@@ -128,4 +129,4 @@ class resources {
   mutable std::vector<pair_res_factory> factories_;
   mutable std::vector<pair_resource> resources_;
 };
-}  // namespace raft
+}  // namespace raft

cpp/include/raft/distance/detail/kernels/kernel_matrices.cuh

@@ -17,10 +17,11 @@
 #pragma once
 
 #include "gram_matrix.cuh"
-#include <raft/util/cuda_utils.cuh>
 
+#include <raft/distance/detail/kernels/rbf_fin_op.cuh>
 #include <raft/distance/distance.cuh>
 #include <raft/linalg/gemm.cuh>
+#include <raft/util/cuda_utils.cuh>
 
 namespace raft::distance::kernels::detail {
 
@@ -353,7 +354,7 @@ class RBFKernel : public GramMatrixBase<math_t> {
     math_t gain   = this->gain;
     using index_t = int64_t;
 
-    auto fin_op = [gain] __device__(math_t d_val, index_t idx) { return exp(-gain * d_val); };
+    rbf_fin_op fin_op{gain};
     raft::distance::distance<raft::distance::DistanceType::L2Unexpanded,
                              math_t,
                              math_t,

cpp/include/raft/distance/detail/kernels/rbf_fin_op.cuh

@@ -0,0 +1,51 @@
+/*
+ * Copyright (c) 2019-2023, NVIDIA CORPORATION.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#pragma once
+
+/*
+ * This file defines rbf_fin_op, which is used in GramMatrixBase.
+ *
+ * This struct has been moved to a separate file, so that it is cheap to include
+ * in distance/distance-ext.cuh, where an instance of raft::distance::distance
+ * with the rbf_fin_op is instantiated.
+ *
+ */
+
+#include <raft/core/math.hpp>                 // raft::exp
+#include <raft/util/cuda_dev_essentials.cuh>  // HD
+
+namespace raft::distance::kernels::detail {
+
+/** @brief: Final op for Gram matrix with RBF kernel.
+ *
+ * Calculates output = e^(-gain * in)
+ *
+ */
+template <typename OutT>
+struct rbf_fin_op {
+  OutT gain;
+
+  explicit HD rbf_fin_op(OutT gain_) noexcept : gain(gain_) {}
+
+  template <typename... Args>
+  HDI OutT operator()(OutT d_val, Args... unused_args)
+  {
+    return raft::exp(-gain * d_val);
+  }
+};  // struct rbf_fin_op
+
+}  // namespace raft::distance::kernels::detail