diff --git a/.github/workflows/build.yaml b/.github/workflows/build.yaml index db20bdbc1..b73a8eea9 100644 --- a/.github/workflows/build.yaml +++ b/.github/workflows/build.yaml @@ -28,7 +28,7 @@ concurrency: jobs: cpp-build: secrets: inherit - uses: rapidsai/shared-workflows/.github/workflows/conda-cpp-build.yaml@branch-24.10 + uses: rapidsai/shared-workflows/.github/workflows/conda-cpp-build.yaml@python-3.12 with: build_type: ${{ inputs.build_type || 'branch' }} branch: ${{ inputs.branch }} @@ -37,7 +37,7 @@ jobs: rust-build: needs: cpp-build secrets: inherit - uses: rapidsai/shared-workflows/.github/workflows/custom-job.yaml@branch-24.10 + uses: rapidsai/shared-workflows/.github/workflows/custom-job.yaml@python-3.12 with: build_type: ${{ inputs.build_type || 'branch' }} branch: ${{ inputs.branch }} @@ -50,7 +50,7 @@ jobs: python-build: needs: [cpp-build] secrets: inherit - uses: rapidsai/shared-workflows/.github/workflows/conda-python-build.yaml@branch-24.10 + uses: rapidsai/shared-workflows/.github/workflows/conda-python-build.yaml@python-3.12 with: build_type: ${{ inputs.build_type || 'branch' }} branch: ${{ inputs.branch }} @@ -59,7 +59,7 @@ jobs: upload-conda: needs: [cpp-build, python-build] secrets: inherit - uses: rapidsai/shared-workflows/.github/workflows/conda-upload-packages.yaml@branch-24.10 + uses: rapidsai/shared-workflows/.github/workflows/conda-upload-packages.yaml@python-3.12 with: build_type: ${{ inputs.build_type || 'branch' }} branch: ${{ inputs.branch }} @@ -70,7 +70,7 @@ jobs: if: github.ref_type == 'branch' needs: python-build secrets: inherit - uses: rapidsai/shared-workflows/.github/workflows/custom-job.yaml@branch-24.10 + uses: rapidsai/shared-workflows/.github/workflows/custom-job.yaml@python-3.12 with: arch: "amd64" branch: ${{ inputs.branch }} @@ -82,7 +82,7 @@ jobs: sha: ${{ inputs.sha }} wheel-build-cuvs: secrets: inherit - uses: rapidsai/shared-workflows/.github/workflows/wheels-build.yaml@branch-24.10 + uses: rapidsai/shared-workflows/.github/workflows/wheels-build.yaml@python-3.12 with: build_type: ${{ inputs.build_type || 'branch' }} branch: ${{ inputs.branch }} @@ -92,7 +92,7 @@ jobs: wheel-publish-cuvs: needs: wheel-build-cuvs secrets: inherit - uses: rapidsai/shared-workflows/.github/workflows/wheels-publish.yaml@branch-24.10 + uses: rapidsai/shared-workflows/.github/workflows/wheels-publish.yaml@python-3.12 with: build_type: ${{ inputs.build_type || 'branch' }} branch: ${{ inputs.branch }} diff --git a/.github/workflows/pr.yaml b/.github/workflows/pr.yaml index 07b10e85a..d34f74062 100644 --- a/.github/workflows/pr.yaml +++ b/.github/workflows/pr.yaml @@ -24,29 +24,29 @@ jobs: - wheel-tests-cuvs - devcontainer secrets: inherit - uses: rapidsai/shared-workflows/.github/workflows/pr-builder.yaml@branch-24.10 + uses: rapidsai/shared-workflows/.github/workflows/pr-builder.yaml@python-3.12 checks: secrets: inherit - uses: rapidsai/shared-workflows/.github/workflows/checks.yaml@branch-24.10 + uses: rapidsai/shared-workflows/.github/workflows/checks.yaml@python-3.12 with: enable_check_generated_files: false conda-cpp-build: needs: checks secrets: inherit - uses: rapidsai/shared-workflows/.github/workflows/conda-cpp-build.yaml@branch-24.10 + uses: rapidsai/shared-workflows/.github/workflows/conda-cpp-build.yaml@python-3.12 with: build_type: pull-request node_type: cpu16 conda-cpp-tests: needs: conda-cpp-build secrets: inherit - uses: rapidsai/shared-workflows/.github/workflows/conda-cpp-tests.yaml@branch-24.10 + uses: rapidsai/shared-workflows/.github/workflows/conda-cpp-tests.yaml@python-3.12 with: build_type: pull-request conda-cpp-checks: needs: conda-cpp-build secrets: inherit - uses: rapidsai/shared-workflows/.github/workflows/conda-cpp-post-build-checks.yaml@branch-24.10 + uses: rapidsai/shared-workflows/.github/workflows/conda-cpp-post-build-checks.yaml@python-3.12 with: build_type: pull-request enable_check_symbols: true @@ -54,19 +54,19 @@ jobs: conda-python-build: needs: conda-cpp-build secrets: inherit - uses: rapidsai/shared-workflows/.github/workflows/conda-python-build.yaml@branch-24.10 + uses: rapidsai/shared-workflows/.github/workflows/conda-python-build.yaml@python-3.12 with: build_type: pull-request conda-python-tests: needs: conda-python-build secrets: inherit - uses: rapidsai/shared-workflows/.github/workflows/conda-python-tests.yaml@branch-24.10 + uses: rapidsai/shared-workflows/.github/workflows/conda-python-tests.yaml@python-3.12 with: build_type: pull-request docs-build: needs: conda-python-build secrets: inherit - uses: rapidsai/shared-workflows/.github/workflows/custom-job.yaml@branch-24.10 + uses: rapidsai/shared-workflows/.github/workflows/custom-job.yaml@python-3.12 with: build_type: pull-request node_type: "gpu-v100-latest-1" @@ -76,7 +76,7 @@ jobs: rust-build: needs: conda-cpp-build secrets: inherit - uses: rapidsai/shared-workflows/.github/workflows/custom-job.yaml@branch-24.10 + uses: rapidsai/shared-workflows/.github/workflows/custom-job.yaml@python-3.12 with: build_type: pull-request node_type: "gpu-v100-latest-1" @@ -86,20 +86,20 @@ jobs: wheel-build-cuvs: needs: checks secrets: inherit - uses: rapidsai/shared-workflows/.github/workflows/wheels-build.yaml@branch-24.10 + uses: rapidsai/shared-workflows/.github/workflows/wheels-build.yaml@python-3.12 with: build_type: pull-request script: ci/build_wheel_cuvs.sh wheel-tests-cuvs: needs: wheel-build-cuvs secrets: inherit - uses: rapidsai/shared-workflows/.github/workflows/wheels-test.yaml@branch-24.10 + uses: rapidsai/shared-workflows/.github/workflows/wheels-test.yaml@python-3.12 with: build_type: pull-request script: ci/test_wheel_cuvs.sh devcontainer: secrets: inherit - uses: rapidsai/shared-workflows/.github/workflows/build-in-devcontainer.yaml@branch-24.10 + uses: rapidsai/shared-workflows/.github/workflows/build-in-devcontainer.yaml@python-3.12 with: arch: '["amd64"]' cuda: '["12.5"]' diff --git a/.github/workflows/test.yaml b/.github/workflows/test.yaml index 0821233a1..f2daecbec 100644 --- a/.github/workflows/test.yaml +++ b/.github/workflows/test.yaml @@ -16,7 +16,7 @@ on: jobs: conda-cpp-checks: secrets: inherit - uses: rapidsai/shared-workflows/.github/workflows/conda-cpp-post-build-checks.yaml@branch-24.10 + uses: rapidsai/shared-workflows/.github/workflows/conda-cpp-post-build-checks.yaml@python-3.12 with: build_type: nightly branch: ${{ inputs.branch }} @@ -26,7 +26,7 @@ jobs: symbol_exclusions: (void (thrust::|cub::)|raft_cutlass) conda-cpp-tests: secrets: inherit - uses: rapidsai/shared-workflows/.github/workflows/conda-cpp-tests.yaml@branch-24.10 + uses: rapidsai/shared-workflows/.github/workflows/conda-cpp-tests.yaml@python-3.12 with: build_type: nightly branch: ${{ inputs.branch }} @@ -34,7 +34,7 @@ jobs: sha: ${{ inputs.sha }} conda-python-tests: secrets: inherit - uses: rapidsai/shared-workflows/.github/workflows/conda-python-tests.yaml@branch-24.10 + uses: rapidsai/shared-workflows/.github/workflows/conda-python-tests.yaml@python-3.12 with: build_type: nightly branch: ${{ inputs.branch }} @@ -42,7 +42,7 @@ jobs: sha: ${{ inputs.sha }} wheel-tests-cuvs: secrets: inherit - uses: rapidsai/shared-workflows/.github/workflows/wheels-test.yaml@branch-24.10 + uses: rapidsai/shared-workflows/.github/workflows/wheels-test.yaml@python-3.12 with: build_type: nightly branch: ${{ inputs.branch }} diff --git a/ci/release/update-version.sh b/ci/release/update-version.sh index 6d7d022c2..feb0a400c 100755 --- a/ci/release/update-version.sh +++ b/ci/release/update-version.sh @@ -45,6 +45,7 @@ DEPENDENCIES=( dask-cuda cuvs pylibraft + librmm rmm rapids-dask-dependency ) diff --git a/conda/environments/all_cuda-118_arch-aarch64.yaml b/conda/environments/all_cuda-118_arch-aarch64.yaml index 4bbdc3650..cfcb56225 100644 --- a/conda/environments/all_cuda-118_arch-aarch64.yaml +++ b/conda/environments/all_cuda-118_arch-aarch64.yaml @@ -35,6 +35,7 @@ dependencies: - libcusolver=11.4.1.48 - libcusparse-dev=11.7.5.86 - libcusparse=11.7.5.86 +- librmm==24.10.*,>=0.0.0a0 - make - nccl>=2.9.9 - ninja @@ -49,7 +50,6 @@ dependencies: - pytest==7.* - rapids-build-backend>=0.3.0,<0.4.0.dev0 - recommonmark -- rmm==24.10.*,>=0.0.0a0 - rust - scikit-build-core>=0.10.0 - scikit-learn diff --git a/conda/environments/all_cuda-118_arch-x86_64.yaml b/conda/environments/all_cuda-118_arch-x86_64.yaml index 908421d08..dc519d1b5 100644 --- a/conda/environments/all_cuda-118_arch-x86_64.yaml +++ b/conda/environments/all_cuda-118_arch-x86_64.yaml @@ -35,6 +35,7 @@ dependencies: - libcusolver=11.4.1.48 - libcusparse-dev=11.7.5.86 - libcusparse=11.7.5.86 +- librmm==24.10.*,>=0.0.0a0 - make - nccl>=2.9.9 - ninja @@ -49,7 +50,6 @@ dependencies: - pytest==7.* - rapids-build-backend>=0.3.0,<0.4.0.dev0 - recommonmark -- rmm==24.10.*,>=0.0.0a0 - rust - scikit-build-core>=0.10.0 - scikit-learn diff --git a/conda/environments/all_cuda-125_arch-aarch64.yaml b/conda/environments/all_cuda-125_arch-aarch64.yaml index 3131d0b77..b32650e44 100644 --- a/conda/environments/all_cuda-125_arch-aarch64.yaml +++ b/conda/environments/all_cuda-125_arch-aarch64.yaml @@ -32,6 +32,7 @@ dependencies: - libcurand-dev - libcusolver-dev - libcusparse-dev +- librmm==24.10.*,>=0.0.0a0 - make - nccl>=2.9.9 - ninja @@ -45,7 +46,6 @@ dependencies: - pytest==7.* - rapids-build-backend>=0.3.0,<0.4.0.dev0 - recommonmark -- rmm==24.10.*,>=0.0.0a0 - rust - scikit-build-core>=0.10.0 - scikit-learn diff --git a/conda/environments/all_cuda-125_arch-x86_64.yaml b/conda/environments/all_cuda-125_arch-x86_64.yaml index 2f107c4fb..d40fc3b99 100644 --- a/conda/environments/all_cuda-125_arch-x86_64.yaml +++ b/conda/environments/all_cuda-125_arch-x86_64.yaml @@ -32,6 +32,7 @@ dependencies: - libcurand-dev - libcusolver-dev - libcusparse-dev +- librmm==24.10.*,>=0.0.0a0 - make - nccl>=2.9.9 - ninja @@ -45,7 +46,6 @@ dependencies: - pytest==7.* - rapids-build-backend>=0.3.0,<0.4.0.dev0 - recommonmark -- rmm==24.10.*,>=0.0.0a0 - rust - scikit-build-core>=0.10.0 - scikit-learn diff --git a/conda/environments/bench_ann_cuda-118_arch-aarch64.yaml b/conda/environments/bench_ann_cuda-118_arch-aarch64.yaml new file mode 100644 index 000000000..c6e8b05a2 --- /dev/null +++ b/conda/environments/bench_ann_cuda-118_arch-aarch64.yaml @@ -0,0 +1,47 @@ +# This file is generated by `rapids-dependency-file-generator`. +# To make changes, edit ../../dependencies.yaml and run `rapids-dependency-file-generator`. +channels: +- rapidsai +- rapidsai-nightly +- dask/label/dev +- conda-forge +- nvidia +dependencies: +- benchmark>=1.8.2 +- c-compiler +- clang-tools=16.0.6 +- clang==16.0.6 +- click +- cmake>=3.26.4,!=3.30.0 +- cuda-nvtx=11.8 +- cuda-profiler-api=11.8.86 +- cuda-python>=11.7.1,<12.0a0 +- cuda-version=11.8 +- cudatoolkit +- cxx-compiler +- cython>=3.0.0 +- dlpack>=0.8,<1.0 +- gcc_linux-aarch64=11.* +- glog>=0.6.0 +- h5py>=3.8.0 +- hnswlib=0.6.2 +- libcublas-dev=11.11.3.6 +- libcublas=11.11.3.6 +- libcurand-dev=10.3.0.86 +- libcurand=10.3.0.86 +- libcusolver-dev=11.4.1.48 +- libcusolver=11.4.1.48 +- libcusparse-dev=11.7.5.86 +- libcusparse=11.7.5.86 +- librmm==24.10.*,>=0.0.0a0 +- matplotlib +- nccl>=2.9.9 +- ninja +- nlohmann_json>=3.11.2 +- nvcc_linux-aarch64=11.8 +- openblas +- pandas +- pylibraft==24.10.*,>=0.0.0a0 +- pyyaml +- sysroot_linux-aarch64==2.17 +name: bench_ann_cuda-118_arch-aarch64 diff --git a/conda/environments/bench_ann_cuda-118_arch-x86_64.yaml b/conda/environments/bench_ann_cuda-118_arch-x86_64.yaml new file mode 100644 index 000000000..d6c023ae9 --- /dev/null +++ b/conda/environments/bench_ann_cuda-118_arch-x86_64.yaml @@ -0,0 +1,47 @@ +# This file is generated by `rapids-dependency-file-generator`. +# To make changes, edit ../../dependencies.yaml and run `rapids-dependency-file-generator`. +channels: +- rapidsai +- rapidsai-nightly +- dask/label/dev +- conda-forge +- nvidia +dependencies: +- benchmark>=1.8.2 +- c-compiler +- clang-tools=16.0.6 +- clang==16.0.6 +- click +- cmake>=3.26.4,!=3.30.0 +- cuda-nvtx=11.8 +- cuda-profiler-api=11.8.86 +- cuda-python>=11.7.1,<12.0a0 +- cuda-version=11.8 +- cudatoolkit +- cxx-compiler +- cython>=3.0.0 +- dlpack>=0.8,<1.0 +- gcc_linux-64=11.* +- glog>=0.6.0 +- h5py>=3.8.0 +- hnswlib=0.6.2 +- libcublas-dev=11.11.3.6 +- libcublas=11.11.3.6 +- libcurand-dev=10.3.0.86 +- libcurand=10.3.0.86 +- libcusolver-dev=11.4.1.48 +- libcusolver=11.4.1.48 +- libcusparse-dev=11.7.5.86 +- libcusparse=11.7.5.86 +- librmm==24.10.*,>=0.0.0a0 +- matplotlib +- nccl>=2.9.9 +- ninja +- nlohmann_json>=3.11.2 +- nvcc_linux-64=11.8 +- openblas +- pandas +- pylibraft==24.10.*,>=0.0.0a0 +- pyyaml +- sysroot_linux-64==2.17 +name: bench_ann_cuda-118_arch-x86_64 diff --git a/conda/environments/bench_ann_cuda-125_arch-aarch64.yaml b/conda/environments/bench_ann_cuda-125_arch-aarch64.yaml new file mode 100644 index 000000000..4d0ca9496 --- /dev/null +++ b/conda/environments/bench_ann_cuda-125_arch-aarch64.yaml @@ -0,0 +1,43 @@ +# This file is generated by `rapids-dependency-file-generator`. +# To make changes, edit ../../dependencies.yaml and run `rapids-dependency-file-generator`. +channels: +- rapidsai +- rapidsai-nightly +- dask/label/dev +- conda-forge +- nvidia +dependencies: +- benchmark>=1.8.2 +- c-compiler +- clang-tools=16.0.6 +- clang==16.0.6 +- click +- cmake>=3.26.4,!=3.30.0 +- cuda-cudart-dev +- cuda-nvcc +- cuda-nvtx-dev +- cuda-profiler-api +- cuda-python>=12.0,<13.0a0 +- cuda-version=12.5 +- cxx-compiler +- cython>=3.0.0 +- dlpack>=0.8,<1.0 +- gcc_linux-aarch64=11.* +- glog>=0.6.0 +- h5py>=3.8.0 +- hnswlib=0.6.2 +- libcublas-dev +- libcurand-dev +- libcusolver-dev +- libcusparse-dev +- librmm==24.10.*,>=0.0.0a0 +- matplotlib +- nccl>=2.9.9 +- ninja +- nlohmann_json>=3.11.2 +- openblas +- pandas +- pylibraft==24.10.*,>=0.0.0a0 +- pyyaml +- sysroot_linux-aarch64==2.17 +name: bench_ann_cuda-125_arch-aarch64 diff --git a/conda/environments/bench_ann_cuda-125_arch-x86_64.yaml b/conda/environments/bench_ann_cuda-125_arch-x86_64.yaml new file mode 100644 index 000000000..7dd67ab5e --- /dev/null +++ b/conda/environments/bench_ann_cuda-125_arch-x86_64.yaml @@ -0,0 +1,43 @@ +# This file is generated by `rapids-dependency-file-generator`. +# To make changes, edit ../../dependencies.yaml and run `rapids-dependency-file-generator`. +channels: +- rapidsai +- rapidsai-nightly +- dask/label/dev +- conda-forge +- nvidia +dependencies: +- benchmark>=1.8.2 +- c-compiler +- clang-tools=16.0.6 +- clang==16.0.6 +- click +- cmake>=3.26.4,!=3.30.0 +- cuda-cudart-dev +- cuda-nvcc +- cuda-nvtx-dev +- cuda-profiler-api +- cuda-python>=12.0,<13.0a0 +- cuda-version=12.5 +- cxx-compiler +- cython>=3.0.0 +- dlpack>=0.8,<1.0 +- gcc_linux-64=11.* +- glog>=0.6.0 +- h5py>=3.8.0 +- hnswlib=0.6.2 +- libcublas-dev +- libcurand-dev +- libcusolver-dev +- libcusparse-dev +- librmm==24.10.*,>=0.0.0a0 +- matplotlib +- nccl>=2.9.9 +- ninja +- nlohmann_json>=3.11.2 +- openblas +- pandas +- pylibraft==24.10.*,>=0.0.0a0 +- pyyaml +- sysroot_linux-64==2.17 +name: bench_ann_cuda-125_arch-x86_64 diff --git a/cpp/CMakeLists.txt b/cpp/CMakeLists.txt index 90ed0b160..ba46e60b4 100644 --- a/cpp/CMakeLists.txt +++ b/cpp/CMakeLists.txt @@ -551,6 +551,7 @@ if(BUILD_C_LIBRARY) src/neighbors/ivf_flat_c.cpp src/neighbors/ivf_pq_c.cpp src/neighbors/cagra_c.cpp + src/neighbors/hnsw_c.cpp src/neighbors/refine/refine_c.cpp src/distance/pairwise_distance_c.cpp ) diff --git a/cpp/include/cuvs/neighbors/cagra.h b/cpp/include/cuvs/neighbors/cagra.h index 87541f7f0..241f5d8b0 100644 --- a/cpp/include/cuvs/neighbors/cagra.h +++ b/cpp/include/cuvs/neighbors/cagra.h @@ -337,7 +337,10 @@ cuvsError_t cuvsCagraBuild(cuvsResources_t res, * It is also important to note that the CAGRA Index must have been built * with the same type of `queries`, such that `index.dtype.code == * queries.dl_tensor.dtype.code` Types for input are: - * 1. `queries`: `kDLDataType.code == kDLFloat` and `kDLDataType.bits = 32` + * 1. `queries`: + *` a. kDLDataType.code == kDLFloat` and `kDLDataType.bits = 32` + * b. `kDLDataType.code == kDLInt` and `kDLDataType.bits = 8` + * c. `kDLDataType.code == kDLUInt` and `kDLDataType.bits = 8` * 2. `neighbors`: `kDLDataType.code == kDLUInt` and `kDLDataType.bits = 32` * 3. `distances`: `kDLDataType.code == kDLFloat` and `kDLDataType.bits = 32` * @@ -394,7 +397,7 @@ cuvsError_t cuvsCagraSearch(cuvsResources_t res, * * Experimental, both the API and the serialization format are subject to change. * - * @code{.cpp} + * @code{.c} * #include * * // Create cuvsResources_t @@ -416,6 +419,34 @@ cuvsError_t cuvsCagraSerialize(cuvsResources_t res, cuvsCagraIndex_t index, bool include_dataset); +/** + * Save the CAGRA index to file in hnswlib format. + * NOTE: The saved index can only be read by the hnswlib wrapper in cuVS, + * as the serialization format is not compatible with the original hnswlib. + * + * Experimental, both the API and the serialization format are subject to change. + * + * @code{.c} + * #include + * #include + * + * // Create cuvsResources_t + * cuvsResources_t res; + * cuvsError_t res_create_status = cuvsResourcesCreate(&res); + * + * // create an index with `cuvsCagraBuild` + * cuvsCagraSerializeHnswlib(res, "/path/to/index", index); + * @endcode + * + * @param[in] res cuvsResources_t opaque C handle + * @param[in] filename the file name for saving the index + * @param[in] index CAGRA index + * + */ +cuvsError_t cuvsCagraSerializeToHnswlib(cuvsResources_t res, + const char* filename, + cuvsCagraIndex_t index); + /** * Load index from file. * diff --git a/cpp/include/cuvs/neighbors/cagra.hpp b/cpp/include/cuvs/neighbors/cagra.hpp index f74eac711..5f77eb8a3 100644 --- a/cpp/include/cuvs/neighbors/cagra.hpp +++ b/cpp/include/cuvs/neighbors/cagra.hpp @@ -1345,6 +1345,8 @@ void deserialize(raft::resources const& handle, /** * Write the CAGRA built index as a base layer HNSW index to an output stream + * NOTE: The saved index can only be read by the hnswlib wrapper in cuVS, + * as the serialization format is not compatible with the original hnswlib. * * Experimental, both the API and the serialization format are subject to change. * @@ -1371,6 +1373,8 @@ void serialize_to_hnswlib(raft::resources const& handle, /** * Save a CAGRA build index in hnswlib base-layer-only serialized format + * NOTE: The saved index can only be read by the hnswlib wrapper in cuVS, + * as the serialization format is not compatible with the original hnswlib. * * Experimental, both the API and the serialization format are subject to change. * @@ -1398,6 +1402,8 @@ void serialize_to_hnswlib(raft::resources const& handle, /** * Write the CAGRA built index as a base layer HNSW index to an output stream + * NOTE: The saved index can only be read by the hnswlib wrapper in cuVS, + * as the serialization format is not compatible with the original hnswlib. * * Experimental, both the API and the serialization format are subject to change. * @@ -1424,6 +1430,8 @@ void serialize_to_hnswlib(raft::resources const& handle, /** * Save a CAGRA build index in hnswlib base-layer-only serialized format + * NOTE: The saved index can only be read by the hnswlib wrapper in cuVS, + * as the serialization format is not compatible with the original hnswlib. * * Experimental, both the API and the serialization format are subject to change. * @@ -1451,6 +1459,8 @@ void serialize_to_hnswlib(raft::resources const& handle, /** * Write the CAGRA built index as a base layer HNSW index to an output stream + * NOTE: The saved index can only be read by the hnswlib wrapper in cuVS, + * as the serialization format is not compatible with the original hnswlib. * * Experimental, both the API and the serialization format are subject to change. * @@ -1477,6 +1487,8 @@ void serialize_to_hnswlib(raft::resources const& handle, /** * Save a CAGRA build index in hnswlib base-layer-only serialized format + * NOTE: The saved index can only be read by the hnswlib wrapper in cuVS, + * as the serialization format is not compatible with the original hnswlib. * * Experimental, both the API and the serialization format are subject to change. * diff --git a/cpp/include/cuvs/neighbors/hnsw.h b/cpp/include/cuvs/neighbors/hnsw.h new file mode 100644 index 000000000..5e94de60a --- /dev/null +++ b/cpp/include/cuvs/neighbors/hnsw.h @@ -0,0 +1,207 @@ +/* + * Copyright (c) 2024, 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 + +#include +#include +#include +#include +#include + +#ifdef __cplusplus +extern "C" { +#endif + +/** + * @defgroup hnsw_c_search_params C API for hnswlib wrapper search params + * @{ + */ + +struct cuvsHnswSearchParams { + int32_t ef; + int32_t numThreads; +}; + +typedef struct cuvsHnswSearchParams* cuvsHnswSearchParams_t; + +/** + * @brief Allocate HNSW search params, and populate with default values + * + * @param[in] params cuvsHnswSearchParams_t to allocate + * @return cuvsError_t + */ +cuvsError_t cuvsHnswSearchParamsCreate(cuvsHnswSearchParams_t* params); + +/** + * @brief De-allocate HNSW search params + * + * @param[in] params cuvsHnswSearchParams_t to de-allocate + * @return cuvsError_t + */ +cuvsError_t cuvsHnswSearchParamsDestroy(cuvsHnswSearchParams_t params); + +/** + * @} + */ + +/** + * @defgroup hnsw_c_index C API for hnswlib wrapper index + * @{ + */ + +/** + * @brief Struct to hold address of cuvs::neighbors::Hnsw::index and its active trained dtype + * + */ +typedef struct { + uintptr_t addr; + DLDataType dtype; + +} cuvsHnswIndex; + +typedef cuvsHnswIndex* cuvsHnswIndex_t; + +/** + * @brief Allocate HNSW index + * + * @param[in] index cuvsHnswIndex_t to allocate + * @return HnswError_t + */ +cuvsError_t cuvsHnswIndexCreate(cuvsHnswIndex_t* index); + +/** + * @brief De-allocate HNSW index + * + * @param[in] index cuvsHnswIndex_t to de-allocate + */ +cuvsError_t cuvsHnswIndexDestroy(cuvsHnswIndex_t index); + +/** + * @} + */ + +/** + * @defgroup hnsw_c_index_search C API for CUDA ANN Graph-based nearest neighbor search + * @{ + */ +/** + * @brief Search a HNSW index with a `DLManagedTensor` which has underlying + * `DLDeviceType` equal to `kDLCPU`, `kDLCUDAHost`, or `kDLCUDAManaged`. + * It is also important to note that the HNSW Index must have been built + * with the same type of `queries`, such that `index.dtype.code == + * queries.dl_tensor.dtype.code` + * Supported types for input are: + * 1. `queries`: `kDLDataType.code == kDLFloat` or `kDLDataType.code == kDLInt` and + * `kDLDataType.bits = 32` + * 2. `neighbors`: `kDLDataType.code == kDLUInt` and `kDLDataType.bits = 64` + * 3. `distances`: `kDLDataType.code == kDLFloat` and `kDLDataType.bits = 32` + * NOTE: The HNSW index can only be searched by the hnswlib wrapper in cuVS, + * as the format is not compatible with the original hnswlib. + * + * @code {.c} + * #include + * #include + * + * // Create cuvsResources_t + * cuvsResources_t res; + * cuvsError_t res_create_status = cuvsResourcesCreate(&res); + * + * // Assume a populated `DLManagedTensor` type here + * DLManagedTensor dataset; + * DLManagedTensor queries; + * DLManagedTensor neighbors; + * + * // Create default search params + * cuvsHnswSearchParams_t params; + * cuvsError_t params_create_status = cuvsHnswSearchParamsCreate(¶ms); + * + * // Search the `index` built using `cuvsHnswBuild` + * cuvsError_t search_status = cuvsHnswSearch(res, params, index, &queries, &neighbors, + * &distances); + * + * // de-allocate `params` and `res` + * cuvsError_t params_destroy_status = cuvsHnswSearchParamsDestroy(params); + * cuvsError_t res_destroy_status = cuvsResourcesDestroy(res); + * @endcode + * + * @param[in] res cuvsResources_t opaque C handle + * @param[in] params cuvsHnswSearchParams_t used to search Hnsw index + * @param[in] index cuvsHnswIndex which has been returned by `cuvsHnswBuild` + * @param[in] queries DLManagedTensor* queries dataset to search + * @param[out] neighbors DLManagedTensor* output `k` neighbors for queries + * @param[out] distances DLManagedTensor* output `k` distances for queries + */ +cuvsError_t cuvsHnswSearch(cuvsResources_t res, + cuvsHnswSearchParams_t params, + cuvsHnswIndex_t index, + DLManagedTensor* queries, + DLManagedTensor* neighbors, + DLManagedTensor* distances); + +/** + * @} + */ + +/** + * @defgroup hnsw_c_serialize HNSW C-API serialize functions + * @{ + */ + +/** + * Load hnswlib index from file which was serialized from a HNSW index. + * NOTE: The loaded hnswlib index is immutable, and only be read by the + * hnswlib wrapper in cuVS, as the serialization format is not compatible with the original hnswlib. + * Experimental, both the API and the serialization format are subject to change. + * + * @code{.c} + * #include + * #include + * #include + * + * // Create cuvsResources_t + * cuvsResources_t res; + * cuvsError_t res_create_status = cuvsResourcesCreate(&res); + * + * // create an index with `cuvsCagraBuild` + * cuvsCagraSerializeHnswlib(res, "/path/to/index", index); + * + * // Load the serialized CAGRA index from file as an hnswlib index + * // The index should have the same dtype as the one used to build CAGRA the index + * cuvsHnswIndex_t hnsw_index; + * cuvsHnswIndexCreate(&hnsw_index); + * hnsw_index->dtype = index->dtype; + * cuvsCagraDeserialize(res, "/path/to/index", hnsw_index); + * @endcode + * + * @param[in] res cuvsResources_t opaque C handle + * @param[in] filename the name of the file that stores the index + * @param[in] dim the dimension of the vectors in the index + * @param[in] metric the distance metric used to build the index + * @param[out] index HNSW index loaded disk + */ +cuvsError_t cuvsHnswDeserialize(cuvsResources_t res, + const char* filename, + int dim, + cuvsDistanceType metric, + cuvsHnswIndex_t index); +/** + * @} + */ + +#ifdef __cplusplus +} +#endif diff --git a/cpp/include/cuvs/neighbors/hnsw.hpp b/cpp/include/cuvs/neighbors/hnsw.hpp index 86f321564..007adef0d 100644 --- a/cpp/include/cuvs/neighbors/hnsw.hpp +++ b/cpp/include/cuvs/neighbors/hnsw.hpp @@ -34,7 +34,7 @@ namespace cuvs::neighbors::hnsw { /** - * @defgroup hnsw Build CAGRA index and search with hnswlib + * @defgroup hnsw_cpp_search_params Build CAGRA index and search with hnswlib * @{ */ @@ -44,6 +44,13 @@ struct search_params : cuvs::neighbors::search_params { // automatically maximizes parallelism }; +/**@}*/ + +/** + * @defgroup hnsw_cpp_index hnswlib index wrapper + * @{ + */ + template struct index : cuvs::neighbors::index { public: @@ -58,6 +65,8 @@ struct index : cuvs::neighbors::index { */ index(int dim, cuvs::distance::DistanceType metric) : dim_{dim}, metric_{metric} {} + virtual ~index() {} + /** @brief Get underlying index */ @@ -77,11 +86,19 @@ struct index : cuvs::neighbors::index { cuvs::distance::DistanceType metric_; }; +/**@}*/ + +/** + * @defgroup hnsw_cpp_index_load Load CAGRA index as hnswlib index + * @{ + */ + /** - * @brief Construct an hnswlib base-layer-only index from a CAGRA index - * NOTE: 1. This method uses the filesystem to write the CAGRA index in `/tmp/.bin` - * before reading it as an hnswlib index, then deleting the temporary file. - * 2. This function is only offered as a compiled symbol in `libraft.so` + * @brief Construct an immutable hnswlib base-layer-only index from a CAGRA index + * NOTE: This method uses the filesystem to write the CAGRA index in `/tmp/.bin` + * before reading it as an hnswlib index, then deleting the temporary file. The returned index + * is immutable and can only be searched by the hnswlib wrapper in cuVS, as the format is not + * compatible with the original hnswlib. * * @param[in] res raft resources * @param[in] cagra_index cagra index @@ -103,10 +120,11 @@ std::unique_ptr> from_cagra( raft::resources const& res, const cuvs::neighbors::cagra::index& cagra_index); /** - * @brief Construct an hnswlib base-layer-only index from a CAGRA index - * NOTE: 1. This method uses the filesystem to write the CAGRA index in `/tmp/.bin` - * before reading it as an hnswlib index, then deleting the temporary file. - * 2. This function is only offered as a compiled symbol in `libraft.so` + * @brief Construct an immutable hnswlib base-layer-only index from a CAGRA index + * NOTE: This method uses the filesystem to write the CAGRA index in `/tmp/.bin` + * before reading it as an hnswlib index, then deleting the temporary file. The returned index + * is immutable and can only be searched by the hnswlib wrapper in cuVS, as the format is not + * compatible with the original hnswlib. * * @param[in] res raft resources * @param[in] cagra_index cagra index @@ -128,10 +146,11 @@ std::unique_ptr> from_cagra( raft::resources const& res, const cuvs::neighbors::cagra::index& cagra_index); /** - * @brief Construct an hnswlib base-layer-only index from a CAGRA index - * NOTE: 1. This method uses the filesystem to write the CAGRA index in `/tmp/.bin` - * before reading it as an hnswlib index, then deleting the temporary file. - * 2. This function is only offered as a compiled symbol in `libraft.so` + * @brief Construct an immutable hnswlib base-layer-only index from a CAGRA index + * NOTE: This method uses the filesystem to write the CAGRA index in `/tmp/.bin` + * before reading it as an hnswlib index, then deleting the temporary file. The returned index + * is immutable and can only be searched by the hnswlib wrapper in cuVS, as the format is not + * compatible with the original hnswlib. * * @param[in] res raft resources * @param[in] cagra_index cagra index @@ -152,8 +171,17 @@ std::unique_ptr> from_cagra( std::unique_ptr> from_cagra( raft::resources const& res, const cuvs::neighbors::cagra::index& cagra_index); +/**@}*/ + +/** + * @defgroup hnsw_cpp_index_search Search hnswlib index + * @{ + */ + /** * @brief Search hnswlib base-layer-only index constructed from a CAGRA index + * NOTE: The HNSW index can only be searched by the hnswlib wrapper in cuVS, + * as the format is not compatible with the original hnswlib. * * @param[in] res raft resources * @param[in] params configure the search @@ -195,6 +223,8 @@ void search(raft::resources const& res, /** * @brief Search hnswlib base-layer-only index constructed from a CAGRA index + * NOTE: The HNSW index can only be searched by the hnswlib wrapper in cuVS, + * as the format is not compatible with the original hnswlib. * * @param[in] res raft resources * @param[in] params configure the search @@ -236,6 +266,8 @@ void search(raft::resources const& res, /** * @brief Search hnswlib base-layer-only index constructed from a CAGRA index + * NOTE: The HNSW index can only be searched by the hnswlib wrapper in cuVS, + * as the format is not compatible with the original hnswlib. * * @param[in] res raft resources * @param[in] params configure the search @@ -275,8 +307,17 @@ void search(raft::resources const& res, raft::host_matrix_view neighbors, raft::host_matrix_view distances); +/**@}*/ + +/** + * @defgroup hnsw_cpp_index_deserialize Deserialize CAGRA index as hnswlib index + * @{ + */ + /** - * @brief De-serialize a CAGRA index saved to a file as an hnsw index + * @brief De-serialize a CAGRA index saved to a file as an hnswlib index + * NOTE: The loaded hnswlib index is immutable, and only be read by the + * hnswlib wrapper in cuVS, as the serialization format is not compatible with the original hnswlib. * * @param[in] res raft resources * @param[in] filename path to the file containing the serialized CAGRA index @@ -310,7 +351,9 @@ void deserialize(raft::resources const& res, index** index); /** - * @brief De-serialize a CAGRA index saved to a file as an hnsw index + * @brief De-serialize a CAGRA index saved to a file as an hnswlib index + * NOTE: The loaded hnswlib index is immutable, and only be read by the + * hnswlib wrapper in cuVS, as the serialization format is not compatible with the original hnswlib. * * @param[in] res raft resources * @param[in] filename path to the file containing the serialized CAGRA index @@ -344,7 +387,9 @@ void deserialize(raft::resources const& res, index** index); /** - * @brief De-serialize a CAGRA index saved to a file as an hnsw index + * @brief De-serialize a CAGRA index saved to a file as an hnswlib index + * NOTE: The loaded hnswlib index is immutable, and only be read by the + * hnswlib wrapper in cuVS, as the serialization format is not compatible with the original hnswlib. * * @param[in] res raft resources * @param[in] filename path to the file containing the serialized CAGRA index diff --git a/cpp/src/neighbors/cagra_c.cpp b/cpp/src/neighbors/cagra_c.cpp index 868b3dec0..164448f2c 100644 --- a/cpp/src/neighbors/cagra_c.cpp +++ b/cpp/src/neighbors/cagra_c.cpp @@ -130,6 +130,14 @@ void _serialize(cuvsResources_t res, cuvs::neighbors::cagra::serialize(*res_ptr, std::string(filename), *index_ptr, include_dataset); } +template +void _serialize_to_hnswlib(cuvsResources_t res, const char* filename, cuvsCagraIndex_t index) +{ + auto res_ptr = reinterpret_cast(res); + auto index_ptr = reinterpret_cast*>(index->addr); + cuvs::neighbors::cagra::serialize_to_hnswlib(*res_ptr, std::string(filename), *index_ptr); +} + template void* _deserialize(cuvsResources_t res, const char* filename) { @@ -326,3 +334,20 @@ extern "C" cuvsError_t cuvsCagraSerialize(cuvsResources_t res, } }); } + +extern "C" cuvsError_t cuvsCagraSerializeToHnswlib(cuvsResources_t res, + const char* filename, + cuvsCagraIndex_t index) +{ + return cuvs::core::translate_exceptions([=] { + if (index->dtype.code == kDLFloat && index->dtype.bits == 32) { + _serialize_to_hnswlib(res, filename, index); + } else if (index->dtype.code == kDLInt && index->dtype.bits == 8) { + _serialize_to_hnswlib(res, filename, index); + } else if (index->dtype.code == kDLUInt && index->dtype.bits == 8) { + _serialize_to_hnswlib(res, filename, index); + } else { + RAFT_FAIL("Unsupported index dtype: %d and bits: %d", index->dtype.code, index->dtype.bits); + } + }); +} diff --git a/cpp/src/neighbors/hnsw_c.cpp b/cpp/src/neighbors/hnsw_c.cpp new file mode 100644 index 000000000..ab5268a6d --- /dev/null +++ b/cpp/src/neighbors/hnsw_c.cpp @@ -0,0 +1,166 @@ + +/* + * Copyright (c) 2024, 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. + */ + +#include "cuvs/distance/distance.h" +#include +#include + +#include +#include +#include +#include + +#include +#include +#include +#include +#include + +namespace { +template +void _search(cuvsResources_t res, + cuvsHnswSearchParams params, + cuvsHnswIndex index, + DLManagedTensor* queries_tensor, + DLManagedTensor* neighbors_tensor, + DLManagedTensor* distances_tensor) +{ + auto res_ptr = reinterpret_cast(res); + auto index_ptr = reinterpret_cast*>(index.addr); + + auto search_params = cuvs::neighbors::hnsw::search_params(); + search_params.ef = params.ef; + search_params.num_threads = params.numThreads; + + using queries_mdspan_type = raft::host_matrix_view; + using neighbors_mdspan_type = raft::host_matrix_view; + using distances_mdspan_type = raft::host_matrix_view; + auto queries_mds = cuvs::core::from_dlpack(queries_tensor); + auto neighbors_mds = cuvs::core::from_dlpack(neighbors_tensor); + auto distances_mds = cuvs::core::from_dlpack(distances_tensor); + cuvs::neighbors::hnsw::search( + *res_ptr, search_params, *index_ptr, queries_mds, neighbors_mds, distances_mds); +} + +template +void* _deserialize(cuvsResources_t res, const char* filename, int dim, cuvsDistanceType metric) +{ + auto res_ptr = reinterpret_cast(res); + cuvs::neighbors::hnsw::index* index = nullptr; + cuvs::neighbors::hnsw::deserialize(*res_ptr, std::string(filename), dim, metric, &index); + return index; +} +} // namespace + +extern "C" cuvsError_t cuvsHnswSearchParamsCreate(cuvsHnswSearchParams_t* params) +{ + return cuvs::core::translate_exceptions( + [=] { *params = new cuvsHnswSearchParams{.ef = 200, .numThreads = 0}; }); +} + +extern "C" cuvsError_t cuvsHnswSearchParamsDestroy(cuvsHnswSearchParams_t params) +{ + return cuvs::core::translate_exceptions([=] { delete params; }); +} + +extern "C" cuvsError_t cuvsHnswIndexCreate(cuvsHnswIndex_t* index) +{ + return cuvs::core::translate_exceptions([=] { *index = new cuvsHnswIndex{}; }); +} + +extern "C" cuvsError_t cuvsHnswIndexDestroy(cuvsHnswIndex_t index_c_ptr) +{ + return cuvs::core::translate_exceptions([=] { + auto index = *index_c_ptr; + + if (index.dtype.code == kDLFloat) { + auto index_ptr = reinterpret_cast*>(index.addr); + delete index_ptr; + } else if (index.dtype.code == kDLInt) { + auto index_ptr = reinterpret_cast*>(index.addr); + delete index_ptr; + } else if (index.dtype.code == kDLUInt) { + auto index_ptr = reinterpret_cast*>(index.addr); + delete index_ptr; + } + delete index_c_ptr; + }); +} + +extern "C" cuvsError_t cuvsHnswSearch(cuvsResources_t res, + cuvsHnswSearchParams_t params, + cuvsHnswIndex_t index_c_ptr, + DLManagedTensor* queries_tensor, + DLManagedTensor* neighbors_tensor, + DLManagedTensor* distances_tensor) +{ + return cuvs::core::translate_exceptions([=] { + auto queries = queries_tensor->dl_tensor; + auto neighbors = neighbors_tensor->dl_tensor; + auto distances = distances_tensor->dl_tensor; + + RAFT_EXPECTS(cuvs::core::is_dlpack_host_compatible(queries), + "queries should have host compatible memory"); + RAFT_EXPECTS(cuvs::core::is_dlpack_host_compatible(neighbors), + "neighbors should have host compatible memory"); + RAFT_EXPECTS(cuvs::core::is_dlpack_host_compatible(distances), + "distances should have host compatible memory"); + + RAFT_EXPECTS(neighbors.dtype.code == kDLUInt && neighbors.dtype.bits == 64, + "neighbors should be of type uint64_t"); + RAFT_EXPECTS(distances.dtype.code == kDLFloat && distances.dtype.bits == 32, + "distances should be of type float32"); + + auto index = *index_c_ptr; + RAFT_EXPECTS(queries.dtype.code == index.dtype.code, "type mismatch between index and queries"); + RAFT_EXPECTS(queries.dtype.bits == 32, "number of bits in queries dtype should be 32"); + + if (index.dtype.code == kDLFloat) { + _search( + res, *params, index, queries_tensor, neighbors_tensor, distances_tensor); + } else if (index.dtype.code == kDLUInt) { + _search( + res, *params, index, queries_tensor, neighbors_tensor, distances_tensor); + } else if (index.dtype.code == kDLInt) { + _search(res, *params, index, queries_tensor, neighbors_tensor, distances_tensor); + } else { + RAFT_FAIL("Unsupported index dtype: %d and bits: %d", queries.dtype.code, queries.dtype.bits); + } + }); +} + +extern "C" cuvsError_t cuvsHnswDeserialize(cuvsResources_t res, + const char* filename, + int dim, + cuvsDistanceType metric, + cuvsHnswIndex_t index) +{ + return cuvs::core::translate_exceptions([=] { + if (index->dtype.code == kDLFloat && index->dtype.bits == 32) { + index->addr = reinterpret_cast(_deserialize(res, filename, dim, metric)); + index->dtype.code = kDLFloat; + } else if (index->dtype.code == kDLUInt && index->dtype.bits == 8) { + index->addr = reinterpret_cast(_deserialize(res, filename, dim, metric)); + index->dtype.code = kDLInt; + } else if (index->dtype.code == kDLInt && index->dtype.bits == 8) { + index->addr = reinterpret_cast(_deserialize(res, filename, dim, metric)); + index->dtype.code = kDLUInt; + } else { + RAFT_FAIL("Unsupported dtype in file %s", filename); + } + }); +} diff --git a/cpp/test/CMakeLists.txt b/cpp/test/CMakeLists.txt index 570220084..e04c39318 100644 --- a/cpp/test/CMakeLists.txt +++ b/cpp/test/CMakeLists.txt @@ -212,6 +212,10 @@ if(BUILD_C_TESTS) ConfigureTest(NAME IVF_PQ_C_TEST PATH neighbors/run_ivf_pq_c.c neighbors/ann_ivf_pq_c.cu C_LIB) ConfigureTest(NAME CAGRA_C_TEST PATH neighbors/ann_cagra_c.cu C_LIB) + + if(BUILD_CAGRA_HNSWLIB) + ConfigureTest(NAME HNSW_C_TEST PATH neighbors/ann_hnsw_c.cu C_LIB) + endif() endif() # ################################################################################################## diff --git a/cpp/test/neighbors/ann_hnsw_c.cu b/cpp/test/neighbors/ann_hnsw_c.cu new file mode 100644 index 000000000..fc740b924 --- /dev/null +++ b/cpp/test/neighbors/ann_hnsw_c.cu @@ -0,0 +1,131 @@ +/* + * Copyright (c) 2023-2024, 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. + */ + +#include "../test_utils.cuh" +#include +#include +#include + +#include +#include +#include + +#include +#include +#include +#include + +float dataset[4][2] = {{0.74021935, 0.9209938}, + {0.03902049, 0.9689629}, + {0.92514056, 0.4463501}, + {0.6673192, 0.10993068}}; +float queries[4][2] = {{0.48216683, 0.0428398}, + {0.5084142, 0.6545497}, + {0.51260436, 0.2643005}, + {0.05198065, 0.5789965}}; + +std::vector neighbors_exp = {3, 0, 3, 1}; +std::vector distances_exp = {0.03878258, 0.12472608, 0.04776672, 0.15224178}; + +TEST(CagraHnswC, BuildSearch) +{ + // create cuvsResources_t + cuvsResources_t res; + cuvsResourcesCreate(&res); + + // create dataset DLTensor + DLManagedTensor dataset_tensor; + dataset_tensor.dl_tensor.data = dataset; + dataset_tensor.dl_tensor.device.device_type = kDLCPU; + dataset_tensor.dl_tensor.ndim = 2; + dataset_tensor.dl_tensor.dtype.code = kDLFloat; + dataset_tensor.dl_tensor.dtype.bits = 32; + dataset_tensor.dl_tensor.dtype.lanes = 1; + int64_t dataset_shape[2] = {4, 2}; + dataset_tensor.dl_tensor.shape = dataset_shape; + dataset_tensor.dl_tensor.strides = nullptr; + + // create index + cuvsCagraIndex_t index; + cuvsCagraIndexCreate(&index); + + // build index + cuvsCagraIndexParams_t build_params; + cuvsCagraIndexParamsCreate(&build_params); + cuvsCagraBuild(res, build_params, &dataset_tensor, index); + cuvsCagraSerializeToHnswlib(res, "/tmp/cagra_hnswlib.index", index); + + DLManagedTensor queries_tensor; + queries_tensor.dl_tensor.data = queries; + queries_tensor.dl_tensor.device.device_type = kDLCPU; + queries_tensor.dl_tensor.ndim = 2; + queries_tensor.dl_tensor.dtype.code = kDLFloat; + queries_tensor.dl_tensor.dtype.bits = 32; + queries_tensor.dl_tensor.dtype.lanes = 1; + int64_t queries_shape[2] = {4, 2}; + queries_tensor.dl_tensor.shape = queries_shape; + queries_tensor.dl_tensor.strides = nullptr; + + // create neighbors DLTensor + std::vector neighbors(4); + + DLManagedTensor neighbors_tensor; + neighbors_tensor.dl_tensor.data = neighbors.data(); + neighbors_tensor.dl_tensor.device.device_type = kDLCPU; + neighbors_tensor.dl_tensor.ndim = 2; + neighbors_tensor.dl_tensor.dtype.code = kDLUInt; + neighbors_tensor.dl_tensor.dtype.bits = 64; + neighbors_tensor.dl_tensor.dtype.lanes = 1; + int64_t neighbors_shape[2] = {4, 1}; + neighbors_tensor.dl_tensor.shape = neighbors_shape; + neighbors_tensor.dl_tensor.strides = nullptr; + + // create distances DLTensor + std::vector distances(4); + + DLManagedTensor distances_tensor; + distances_tensor.dl_tensor.data = distances.data(); + distances_tensor.dl_tensor.device.device_type = kDLCPU; + distances_tensor.dl_tensor.ndim = 2; + distances_tensor.dl_tensor.dtype.code = kDLFloat; + distances_tensor.dl_tensor.dtype.bits = 32; + distances_tensor.dl_tensor.dtype.lanes = 1; + int64_t distances_shape[2] = {4, 1}; + distances_tensor.dl_tensor.shape = distances_shape; + distances_tensor.dl_tensor.strides = nullptr; + + // create hnsw index + cuvsHnswIndex_t hnsw_index; + cuvsHnswIndexCreate(&hnsw_index); + hnsw_index->dtype = index->dtype; + cuvsHnswDeserialize(res, "/tmp/cagra_hnswlib.index", 2, L2Expanded, hnsw_index); + + // search index + cuvsHnswSearchParams_t search_params; + cuvsHnswSearchParamsCreate(&search_params); + cuvsHnswSearch( + res, search_params, hnsw_index, &queries_tensor, &neighbors_tensor, &distances_tensor); + + // verify output + ASSERT_TRUE(cuvs::hostVecMatch(neighbors_exp, neighbors, cuvs::Compare())); + ASSERT_TRUE(cuvs::hostVecMatch(distances_exp, distances, cuvs::CompareApprox(0.001f))); + + cuvsCagraIndexParamsDestroy(build_params); + cuvsCagraIndexDestroy(index); + cuvsHnswSearchParamsDestroy(search_params); + cuvsHnswIndexDestroy(hnsw_index); + cuvsResourcesDestroy(res); +} diff --git a/dependencies.yaml b/dependencies.yaml index d8763e076..9fcbeaae2 100644 --- a/dependencies.yaml +++ b/dependencies.yaml @@ -11,6 +11,8 @@ files: - build_py_cuvs - cuda - cuda_version + - depends_on_pylibraft + - depends_on_librmm - develop - checks - build_wheels @@ -21,6 +23,21 @@ files: - test_py_cuvs - cupy - rust + bench_ann: + output: conda + matrix: + cuda: ["11.8", "12.5"] + arch: [x86_64, aarch64] + includes: + - rapids_build + - build_py_cuvs + - cuda + - cuda_version + - depends_on_pylibraft + - depends_on_librmm + - develop + - bench + - bench_python test_cpp: output: none includes: @@ -80,6 +97,7 @@ files: includes: - cuda_wheels - run_py_cuvs + - depends_on_pylibraft py_test_py_cuvs: output: pyproject pyproject_dir: python/cuvs @@ -90,6 +108,20 @@ files: - test_python_common - test_py_cuvs - cupy + py_build_cuvs_bench: + output: pyproject + pyproject_dir: python/cuvs_bench + extras: + table: build-system + includes: + - rapids_build_setuptools + py_run_cuvs_bench: + output: pyproject + pyproject_dir: python/cuvs_bench + extras: + table: project + includes: + - bench_python channels: - rapidsai - rapidsai-nightly @@ -154,19 +186,19 @@ dependencies: - matrix: {cuda: "11.2", arch: aarch64} packages: [nvcc_linux-aarch64=11.2] + rapids_build_setuptools: + common: + - output_types: [requirements, pyproject] + packages: + - &rapids_build_backend rapids-build-backend>=0.3.0,<0.4.0.dev0 + - setuptools + - wheel + build_py_cuvs: common: - output_types: [conda] packages: - - &rmm_unsuffixed rmm==24.10.*,>=0.0.0a0 - - &pylibraft_unsuffixed pylibraft==24.10.*,>=0.0.0a0 - dlpack>=0.8,<1.0 - - output_types: requirements - packages: - # pip recognizes the index as a global option for the requirements.txt file - # This index is needed for rmm-cu{11,12}. - - --extra-index-url=https://pypi.nvidia.com - - --extra-index-url=https://pypi.anaconda.org/rapidsai-wheels-nightly/simple specific: - output_types: [conda, requirements, pyproject] matrices: @@ -181,21 +213,6 @@ dependencies: - matrix: packages: - &cuda_python cuda-python - - output_types: [requirements, pyproject] - matrices: - - matrix: - cuda: "12.*" - cuda_suffixed: "true" - packages: - - &rmm_cu12 rmm-cu12==24.10.*,>=0.0.0a0 - - &pylibraft_cu12 pylibraft-cu12==24.10.*,>=0.0.0a0 - - matrix: - cuda: "11.*" - cuda_suffixed: "true" - packages: - - &rmm_cu11 rmm-cu11==24.10.*,>=0.0.0a0 - - &pylibraft_cu11 pylibraft-cu11==24.10.*,>=0.0.0a0 - - {matrix: null, packages: [*rmm_unsuffixed, *pylibraft_unsuffixed] } checks: common: - output_types: [conda, requirements] @@ -404,24 +421,18 @@ dependencies: py: "3.11" packages: - python=3.11 + - matrix: + py: "3.12" + packages: + - python=3.12 - matrix: packages: - - python>=3.10,<3.12 + - python>=3.10,<3.13 run_py_cuvs: common: - - output_types: [conda, pyproject] + - output_types: [conda, requirements, pyproject] packages: - &numpy numpy>=1.23,<3.0a0 - - output_types: [conda] - packages: - - *rmm_unsuffixed - - *pylibraft_unsuffixed - - output_types: requirements - packages: - # pip recognizes the index as a global option for the requirements.txt file - # This index is needed for cudf and rmm. - - --extra-index-url=https://pypi.nvidia.com - - --extra-index-url=https://pypi.anaconda.org/rapidsai-wheels-nightly/simple specific: - output_types: [conda, requirements, pyproject] matrices: @@ -436,15 +447,6 @@ dependencies: - matrix: packages: - *cuda_python - - output_types: [requirements, pyproject] - matrices: - - matrix: {cuda: "12.*"} - packages: - - *pylibraft_cu12 - - matrix: {cuda: "11.*"} - packages: - - *pylibraft_cu11 - - {matrix: null, packages: [*pylibraft_unsuffixed]} test_python_common: common: - output_types: [conda, requirements, pyproject] @@ -456,3 +458,70 @@ dependencies: - output_types: [conda, requirements, pyproject] packages: - scikit-learn + bench: + common: + - output_types: [conda, pyproject, requirements] + packages: + - hnswlib=0.6.2 + - nlohmann_json>=3.11.2 + - glog>=0.6.0 + - h5py>=3.8.0 + - benchmark>=1.8.2 + - openblas + bench_python: + common: + - output_types: [conda] + packages: + - matplotlib + - pandas + - pyyaml + - pandas + - click + depends_on_librmm: + common: + - output_types: conda + packages: + - &librmm_unsuffixed librmm==24.10.*,>=0.0.0a0 + - output_types: requirements + packages: + # pip recognizes the index as a global option for the requirements.txt file + - --extra-index-url=https://pypi.nvidia.com + - --extra-index-url=https://pypi.anaconda.org/rapidsai-wheels-nightly/simple + specific: + - output_types: [requirements, pyproject] + matrices: + - matrix: + cuda: "12.*" + cuda_suffixed: "true" + packages: + - librmm-cu12==24.10.*,>=0.0.0a0 + - matrix: + cuda: "11.*" + cuda_suffixed: "true" + packages: + - librmm-cu11==24.10.*,>=0.0.0a0 + - {matrix: null, packages: [*librmm_unsuffixed]} + depends_on_pylibraft: + common: + - output_types: conda + packages: + - &pylibraft_unsuffixed pylibraft==24.10.*,>=0.0.0a0 + - output_types: requirements + packages: + # pip recognizes the index as a global option for the requirements.txt file + - --extra-index-url=https://pypi.nvidia.com + - --extra-index-url=https://pypi.anaconda.org/rapidsai-wheels-nightly/simple + specific: + - output_types: [requirements, pyproject] + matrices: + - matrix: + cuda: "12.*" + cuda_suffixed: "true" + packages: + - pylibraft-cu12==24.10.*,>=0.0.0a0 + - matrix: + cuda: "11.*" + cuda_suffixed: "true" + packages: + - pylibraft-cu11==24.10.*,>=0.0.0a0 + - {matrix: null, packages: [*pylibraft_unsuffixed]} diff --git a/docs/source/c_api/neighbors_cagra_c.rst b/docs/source/c_api/neighbors_cagra_c.rst index eb40d5578..a5ffc45b9 100644 --- a/docs/source/c_api/neighbors_cagra_c.rst +++ b/docs/source/c_api/neighbors_cagra_c.rst @@ -50,4 +50,10 @@ Index search :members: :content-only: +Index serialize +------------ +.. doxygengroup:: cagra_c_index_serialize + :project: cuvs + :members: + :content-only: diff --git a/docs/source/c_api/neighbors_hnsw_c.rst b/docs/source/c_api/neighbors_hnsw_c.rst new file mode 100644 index 000000000..4d83cd3e3 --- /dev/null +++ b/docs/source/c_api/neighbors_hnsw_c.rst @@ -0,0 +1,43 @@ +HNSW +==== + +This is a wrapper for hnswlib, to load a CAGRA index as an immutable HNSW index. The loaded HNSW index is only compatible in cuVS, and can be searched using wrapper functions. + + +.. role:: py(code) + :language: c + :class: highlight + +``#include `` + +Index search parameters +----------------------- + +.. doxygengroup:: hnsw_c_search_params + :project: cuvs + :members: + :content-only: + +Index +----- + +.. doxygengroup:: hnsw_c_index + :project: cuvs + :members: + :content-only: + +Index search +------------ + +.. doxygengroup:: cagra_c_index_search + :project: cuvs + :members: + :content-only: + +Index serialize +------------ + +.. doxygengroup:: hnsw_c_index_serialize + :project: cuvs + :members: + :content-only: diff --git a/docs/source/cpp_api/neighbors_cagra.rst b/docs/source/cpp_api/neighbors_cagra.rst index 3a1ed861a..d9f503871 100644 --- a/docs/source/cpp_api/neighbors_cagra.rst +++ b/docs/source/cpp_api/neighbors_cagra.rst @@ -75,3 +75,10 @@ Index extend :members: :content-only: +Index serialize +--------------- + +.. doxygengroup:: cagra_cpp_serialize + :project: cuvs + :members: + :content-only: diff --git a/docs/source/cpp_api/neighbors_hnsw.rst b/docs/source/cpp_api/neighbors_hnsw.rst new file mode 100644 index 000000000..b0af88af0 --- /dev/null +++ b/docs/source/cpp_api/neighbors_hnsw.rst @@ -0,0 +1,52 @@ +HNSW +==== + +This is a wrapper for hnswlib, to load a CAGRA index as an immutable HNSW index. The loaded HNSW index is only compatible in cuVS, and can be searched using wrapper functions. + +.. role:: py(code) + :language: c++ + :class: highlight + +``#include `` + +namespace *cuvs::neighbors::hnsw* + +Index search parameters +----------------------- + +.. doxygengroup:: hnsw_cpp_search_params + :project: cuvs + :members: + :content-only: + +Index +----- + +.. doxygengroup:: hnsw_cpp_index + :project: cuvs + :members: + :content-only: + +Index load +------------ + +.. doxygengroup:: hnsw_cpp_index_search + :project: cuvs + :members: + :content-only: + +Index search +------------ + +.. doxygengroup:: hnsw_cpp_index_search + :project: cuvs + :members: + :content-only: + +Index deserialize +--------------- + +.. doxygengroup:: hnsw_cpp_index_deserialize + :project: cuvs + :members: + :content-only: diff --git a/examples/c/CMakeLists.txt b/examples/c/CMakeLists.txt index fb508728f..d47cd4f1c 100644 --- a/examples/c/CMakeLists.txt +++ b/examples/c/CMakeLists.txt @@ -35,3 +35,7 @@ include(../cmake/thirdparty/get_cuvs.cmake) add_executable(CAGRA_C_EXAMPLE src/cagra_c_example.c) target_include_directories(CAGRA_C_EXAMPLE PUBLIC "$") target_link_libraries(CAGRA_C_EXAMPLE PRIVATE cuvs::c_api $) + +add_executable(L2_C_EXAMPLE src/L2_c_example.c) +target_include_directories(L2_C_EXAMPLE PUBLIC "$") +target_link_libraries(L2_C_EXAMPLE PRIVATE cuvs::c_api $) diff --git a/examples/c/src/L2_c_example.c b/examples/c/src/L2_c_example.c new file mode 100644 index 000000000..73ddf6103 --- /dev/null +++ b/examples/c/src/L2_c_example.c @@ -0,0 +1,123 @@ +/* + * Copyright (c) 2024, 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. + */ + +#include +#include +#include + +#include + +#include +#include +#include + +#define DIM 4 +#define N_ROWS 1 + +float PointA[N_ROWS][DIM] = {1.0,2.0,3.0,4.0}; +float PointB[N_ROWS][DIM] = {2.0,3.0,4.0,5.0}; + +cuvsResources_t res; + +void outputVector(float * Vec) { + printf("Vector is "); + for (int i = 0; i < DIM; ++i){ + printf(" %f",Vec[i]); + } + printf("\n"); +} + +/** + * @brief Initialize Tensor. + * + * @param[in] x_d Pointer to a vector + * @param[in] x_shape[] Two-dimensional array, which stores the number of rows and columns of vectors. + * @param[out] x_tensor Stores the initialized DLManagedTensor. + */ +void tensor_initialize(float* x_d, int64_t x_shape[2], DLManagedTensor* x_tensor) { + x_tensor->dl_tensor.data = x_d; + x_tensor->dl_tensor.device.device_type = kDLCUDA; + x_tensor->dl_tensor.ndim = 2; + x_tensor->dl_tensor.dtype.code = kDLFloat; + x_tensor->dl_tensor.dtype.bits = 32; + x_tensor->dl_tensor.dtype.lanes = 1; + x_tensor->dl_tensor.shape = x_shape; + x_tensor->dl_tensor.strides = NULL; +} + +/** + * @brief Calculate the euclidean distance between two arrays. + * + * @param[in] n_cols array length，also the dimension of the vector + * @param[in] x[] Pointer to a vector + * @param[in] y[] Pointer to another vector + * @param[out] ret will store the result about the euclidean distance + */ +void l2_distance_calc(int64_t n_cols,float x[], float y[], float *ret) { + float *x_d, *y_d; + float *distance_d; + cuvsRMMAlloc(res, (void**) &x_d, sizeof(float) * N_ROWS * n_cols); + cuvsRMMAlloc(res, (void**) &y_d, sizeof(float) * N_ROWS * n_cols); + cuvsRMMAlloc(res, (void**) &distance_d, sizeof(float) * N_ROWS * N_ROWS); + + // Use DLPack to represent x[] and y[] as tensors + cudaMemcpy(x_d, x, sizeof(float) * N_ROWS * n_cols, cudaMemcpyDefault); + cudaMemcpy(y_d, y, sizeof(float) * N_ROWS * n_cols, cudaMemcpyDefault); + + DLManagedTensor x_tensor; + int64_t x_shape[2] = {N_ROWS, n_cols}; + tensor_initialize(x_d, x_shape, &x_tensor); + + DLManagedTensor y_tensor; + int64_t y_shape[2] = {N_ROWS, n_cols}; + tensor_initialize(y_d, y_shape, &y_tensor); + + DLManagedTensor dist_tensor; + int64_t distances_shape[2] = {N_ROWS, N_ROWS}; + tensor_initialize(distance_d, distances_shape, &dist_tensor); + + // metric_arg default value is 2.0,used for Minkowski distance + cuvsPairwiseDistance(res, &x_tensor, &y_tensor, &dist_tensor, L2SqrtUnexpanded, 2.0); + + cudaMemcpy(ret, distance_d, sizeof(float) * N_ROWS * N_ROWS, cudaMemcpyDefault); + + cuvsRMMFree(res, distance_d, sizeof(float) * N_ROWS * N_ROWS); + cuvsRMMFree(res, x_d, sizeof(float) * N_ROWS * n_cols); + cuvsRMMFree(res, y_d, sizeof(float) * N_ROWS * n_cols); + +} + +int euclidean_distance_calculation_example() { + // Create a cuvsResources_t object + cuvsResourcesCreate(&res); + + outputVector((float *)PointA); + outputVector((float *)PointB); + + float ret; + + l2_distance_calc(DIM, (float *)PointA, (float *)PointB, &ret); + printf("L2 distance is %f.\n", ret); + + cuvsResourcesDestroy(res); + + return 0; +} + +int main() { + euclidean_distance_calculation_example(); + return 0; +} diff --git a/python/cuvs/pyproject.toml b/python/cuvs/pyproject.toml index fe0e9022f..68bd9a868 100644 --- a/python/cuvs/pyproject.toml +++ b/python/cuvs/pyproject.toml @@ -44,6 +44,7 @@ classifiers = [ "Programming Language :: Python", "Programming Language :: Python :: 3.10", "Programming Language :: Python :: 3.11", + "Programming Language :: Python :: 3.12", ] [project.optional-dependencies] @@ -127,8 +128,6 @@ requires = [ "cuda-python", "cython>=3.0.0", "ninja", - "pylibraft==24.10.*,>=0.0.0a0", - "rmm==24.10.*,>=0.0.0a0", ] # This list was generated by `rapids-dependency-file-generator`. To make changes, edit ../../dependencies.yaml and run `rapids-dependency-file-generator`. build-backend = "scikit_build_core.build" dependencies-file = "../../dependencies.yaml" diff --git a/python/cuvs_bench/LICENSE b/python/cuvs_bench/LICENSE new file mode 120000 index 000000000..30cff7403 --- /dev/null +++ b/python/cuvs_bench/LICENSE @@ -0,0 +1 @@ +../../LICENSE \ No newline at end of file diff --git a/python/cuvs_bench/cuvs_bench/VERSION b/python/cuvs_bench/cuvs_bench/VERSION new file mode 120000 index 000000000..d62dc733e --- /dev/null +++ b/python/cuvs_bench/cuvs_bench/VERSION @@ -0,0 +1 @@ +../../../VERSION \ No newline at end of file diff --git a/python/cuvs_bench/pyproject.toml b/python/cuvs_bench/pyproject.toml new file mode 100644 index 000000000..7bb9e2f8d --- /dev/null +++ b/python/cuvs_bench/pyproject.toml @@ -0,0 +1,70 @@ +# Copyright (c) 2024, NVIDIA CORPORATION. + +[build-system] +build-backend = "rapids_build_backend.build" +requires = [ + "rapids-build-backend>=0.3.0,<0.4.0.dev0", + "setuptools", + "wheel", +] # This list was generated by `rapids-dependency-file-generator`. To make changes, edit ../../dependencies.yaml and run `rapids-dependency-file-generator`. + +[project] +name = "cuvs-bench" +dynamic = ["version"] +description = "cuVS benchmarks" +authors = [ + { name = "NVIDIA Corporation" }, +] +license = { text = "Apache 2.0" } +requires-python = ">=3.10" +dependencies = [ +] # This list was generated by `rapids-dependency-file-generator`. To make changes, edit ../../dependencies.yaml and run `rapids-dependency-file-generator`. +classifiers = [ + "Intended Audience :: Developers", + "Topic :: Database", + "Topic :: Scientific/Engineering", + "License :: OSI Approved :: Apache Software License", + "Programming Language :: Python", + "Programming Language :: Python :: 3.10", + "Programming Language :: Python :: 3.11", + "Programming Language :: Python :: 3.12", +] + +[project.urls] +Homepage = "https://github.com/rapidsai/raft" + +[tool.setuptools.packages.find] +where = ["src"] + +[tool.setuptools.package-data] +"*" = ["*.*", "VERSION"] + +[tool.isort] +line_length = 79 +multi_line_output = 3 +include_trailing_comma = true +force_grid_wrap = 0 +combine_as_imports = true +order_by_type = true +skip = [ + "thirdparty", + ".eggs", + ".git", + ".hg", + ".mypy_cache", + ".tox", + ".venv", + "_build", + "buck-out", + "build", + "dist", +] + +[tool.setuptools.dynamic] +version = { file = "cuvs_bench/VERSION" } + +[tool.rapids-build-backend] +build-backend = "scikit_build_core.build" +requires = [] +dependencies-file = "../../dependencies.yaml" +matrix-entry = "cuda_suffixed=true"