Add end-to-end CUDA ann-benchmarks to raft (#1304)

New `bench/ann` artifact for comparing (C++ APIs for) GPU-acclerated algorithms end-to-end. Working on this w/ @tfeher but had to squash the original commits into a single commit. Things left to do: - [x] Separate `benchmarks` executables for each different algorithm - [x] Separate build targets for `ggnn` and `hnswlib` - [x] Revise `bench/ann` docs - [x] Break `factory.cuh` abd `benchmark.cu` / `benchmark.cpp` into individual files for each different algorithm to make it easier to plug in new algorithms. - [x] Separate into its own conda package closes #1211 Authors: - Corey J. Nolet (https://github.com/cjnolet) Approvers: - Ray Douglass (https://github.com/raydouglass) - Ben Frederickson (https://github.com/benfred) - Tamas Bela Feher (https://github.com/tfeher) URL: #1304
rapidsai · Mar 28, 2023 · 22ebc72 · 22ebc72
1 parent 0d3bd3d
commit 22ebc72
Show file tree

Hide file tree

Showing 103 changed files with 8,371 additions and 125 deletions.
diff --git a/.pre-commit-config.yaml b/.pre-commit-config.yaml
@@ -62,7 +62,7 @@ repos:
                 entry: ./cpp/scripts/run-cmake-format.sh cmake-format
                 language: python
                 types: [cmake]
-                exclude: .*/thirdparty/.*
+                exclude: .*/thirdparty/.*|.*FindAVX.cmake.*
                 # Note that pre-commit autoupdate does not update the versions
                 # of dependencies, so we'll have to update this manually.
                 additional_dependencies:

diff --git a/build.sh b/build.sh
@@ -2,7 +2,7 @@
 
 # Copyright (c) 2020-2023, NVIDIA CORPORATION.
 
-# raft build script
+# raft build scripts
 
 # This script is used to build the component(s) in this repo from
 # source, and can be called with various options to customize the
@@ -15,11 +15,11 @@ NUMARGS=$#
 ARGS=$*
 
 # NOTE: ensure all dir changes are relative to the location of this
-# script, and that this script resides in the repo dir!
+# scripts, and that this script resides in the repo dir!
 REPODIR=$(cd $(dirname $0); pwd)
 
-VALIDARGS="clean libraft pylibraft raft-dask docs tests bench template clean --uninstall  -v -g -n --compile-lib --allgpuarch --no-nvtx --show_depr_warn -h"
-HELP="$0 [<target> ...] [<flag> ...] [--cmake-args=\"<args>\"] [--cache-tool=<tool>] [--limit-tests=<targets>] [--limit-bench=<targets>]
+VALIDARGS="clean libraft pylibraft raft-dask docs tests template bench-prims bench-ann clean --uninstall  -v -g -n --compile-lib --allgpuarch --no-nvtx --show_depr_warn -h"
+HELP="$0 [<target> ...] [<flag> ...] [--cmake-args=\"<args>\"] [--cache-tool=<tool>] [--limit-tests=<targets>] [--limit-bench-prims=<targets>] [--limit-bench-ann=<targets>]
  where <target> is:
    clean            - remove all existing build artifacts and configuration (start over)
    libraft          - build the raft C++ code only. Also builds the C-wrapper library
@@ -28,7 +28,8 @@ HELP="$0 [<target> ...] [<flag> ...] [--cmake-args=\"<args>\"] [--cache-tool=<to
    raft-dask        - build the raft-dask Python package. this also requires pylibraft.
    docs             - build the documentation
    tests            - build the tests
-   bench            - build the benchmarks
+   bench-prims      - build micro-benchmarks for primitives
+   bench-ann        - build end-to-end ann benchmarks
    template         - build the example RAFT application template
 
  and <flag> is:
@@ -39,7 +40,8 @@ HELP="$0 [<target> ...] [<flag> ...] [--cmake-args=\"<args>\"] [--cache-tool=<to
    --compile-lib               - compile shared libraries for all components
                                  can be useful for a pure header-only install
    --limit-tests               - semicolon-separated list of test executables to compile (e.g. NEIGHBORS_TEST;CLUSTER_TEST)
-   --limit-bench               - semicolon-separated list of benchmark executables to compute (e.g. NEIGHBORS_BENCH;CLUSTER_BENCH)
+   --limit-bench-prims         - semicolon-separated list of prims benchmark executables to compute (e.g. NEIGHBORS_PRIMS_BENCH;CLUSTER_PRIMS_BENCH)
+   --limit-bench-ann           - semicolon-separated list of ann benchmark executables to compute (e.g. HNSWLIB_ANN_BENCH;RAFT_IVF_PQ_ANN_BENCH)
    --allgpuarch                - build for all supported GPU architectures
    --no-nvtx                   - disable nvtx (profiling markers), but allow enabling it in downstream projects
    --show_depr_warn            - show cmake deprecation warnings
@@ -63,7 +65,8 @@ VERBOSE_FLAG=""
 BUILD_ALL_GPU_ARCH=0
 BUILD_TESTS=OFF
 BUILD_TYPE=Release
-BUILD_BENCH=OFF
+BUILD_PRIMS_BENCH=OFF
+BUILD_ANN_BENCH=OFF
 COMPILE_LIBRARY=OFF
 INSTALL_TARGET=install
 
@@ -151,15 +154,30 @@ function limitTests {
 
 function limitBench {
     # Check for option to limit the set of test binaries to build
-    if [[ -n $(echo $ARGS | { grep -E "\-\-limit\-bench" || true; } ) ]]; then
+    if [[ -n $(echo $ARGS | { grep -E "\-\-limit\-bench-prims" || true; } ) ]]; then
         # There are possible weird edge cases that may cause this regex filter to output nothing and fail silently
         # the true pipe will catch any weird edge cases that may happen and will cause the program to fall back
         # on the invalid option error
-        LIMIT_BENCH_TARGETS=$(echo $ARGS | sed -e 's/.*--limit-bench=//' -e 's/ .*//')
-        if [[ -n ${LIMIT_BENCH_TARGETS} ]]; then
+        LIMIT_PRIMS_BENCH_TARGETS=$(echo $ARGS | sed -e 's/.*--limit-bench-prims=//' -e 's/ .*//')
+        if [[ -n ${LIMIT_PRIMS_BENCH_TARGETS} ]]; then
+            # Remove the full LIMIT_PRIMS_BENCH_TARGETS argument from list of args so that it passes validArgs function
+            ARGS=${ARGS//--limit-bench-prims=$LIMIT_PRIMS_BENCH_TARGETS/}
+            PRIMS_BENCH_TARGETS=${LIMIT_PRIMS_BENCH_TARGETS}
+        fi
+    fi
+}
+
+function limitAnnBench {
+    # Check for option to limit the set of test binaries to build
+    if [[ -n $(echo $ARGS | { grep -E "\-\-limit\-bench-ann" || true; } ) ]]; then
+        # There are possible weird edge cases that may cause this regex filter to output nothing and fail silently
+        # the true pipe will catch any weird edge cases that may happen and will cause the program to fall back
+        # on the invalid option error
+        LIMIT_ANN_BENCH_TARGETS=$(echo $ARGS | sed -e 's/.*--limit-bench-ann=//' -e 's/ .*//')
+        if [[ -n ${LIMIT_ANN_BENCH_TARGETS} ]]; then
             # Remove the full LIMIT_TEST_TARGETS argument from list of args so that it passes validArgs function
-            ARGS=${ARGS//--limit-bench=$LIMIT_BENCH_TARGETS/}
-            BENCH_TARGETS=${LIMIT_BENCH_TARGETS}
+            ARGS=${ARGS//--limit-bench-ann=$LIMIT_ANN_BENCH_TARGETS/}
+            ANN_BENCH_TARGETS=${LIMIT_ANN_BENCH_TARGETS}
         fi
     fi
 }
@@ -175,6 +193,7 @@ if (( ${NUMARGS} != 0 )); then
     cacheTool
     limitTests
     limitBench
+    limitAnnBench
     for a in ${ARGS}; do
         if ! (echo " ${VALIDARGS} " | grep -q " ${a} "); then
             echo "Invalid option: ${a}"
@@ -281,18 +300,23 @@ if hasArg tests || (( ${NUMARGS} == 0 )); then
     fi
 fi
 
-if hasArg bench || (( ${NUMARGS} == 0 )); then
-    BUILD_BENCH=ON
-    CMAKE_TARGET="${CMAKE_TARGET};${BENCH_TARGETS}"
+if hasArg bench-prims || (( ${NUMARGS} == 0 )); then
+    BUILD_PRIMS_BENCH=ON
+    CMAKE_TARGET="${CMAKE_TARGET};${PRIMS_BENCH_TARGETS}"
 
     # Force compile library when needed benchmark targets are specified
-    if [[ $CMAKE_TARGET == *"CLUSTER_BENCH"* || \
-          $CMAKE_TARGET == *"MATRIX_BENCH"* || \
-          $CMAKE_TARGET == *"NEIGHBORS_BENCH"* ]]; then
+    if [[ $CMAKE_TARGET == *"CLUSTER_PRIMS_BENCH"* || \
+          $CMAKE_TARGET == *"MATRIX_PRIMS_BENCH"* || \
+          $CMAKE_TARGET == *"NEIGHBORS_PRIMS_BENCH"* ]]; then
       echo "-- Enabling compiled lib for benchmarks"
       COMPILE_LIBRARY=ON
     fi
+fi
 
+if hasArg bench-ann || (( ${NUMARGS} == 0 )); then
+    BUILD_ANN_BENCH=ON
+    CMAKE_TARGET="${CMAKE_TARGET};${ANN_BENCH_TARGETS}"
+    COMPILE_LIBRARY=ON
 fi
 
 if hasArg --no-nvtx; then
@@ -305,8 +329,6 @@ if hasArg clean; then
     CLEAN=1
 fi
 
-
-
 if [[ ${CMAKE_TARGET} == "" ]]; then
     CMAKE_TARGET="all"
 fi
@@ -340,7 +362,7 @@ fi
 
 ################################################################################
 # Configure for building all C++ targets
-if (( ${NUMARGS} == 0 )) || hasArg libraft || hasArg docs || hasArg tests || hasArg bench; then
+if (( ${NUMARGS} == 0 )) || hasArg libraft || hasArg docs || hasArg tests || hasArg bench-prims || hasArg bench-ann; then
     if (( ${BUILD_ALL_GPU_ARCH} == 0 )); then
         RAFT_CMAKE_CUDA_ARCHITECTURES="NATIVE"
         echo "Building for the architecture of the GPU in the system..."
@@ -359,7 +381,8 @@ if (( ${NUMARGS} == 0 )) || hasArg libraft || hasArg docs || hasArg tests || has
           -DRAFT_NVTX=${NVTX} \
           -DDISABLE_DEPRECATION_WARNINGS=${DISABLE_DEPRECATION_WARNINGS} \
           -DBUILD_TESTS=${BUILD_TESTS} \
-          -DBUILD_BENCH=${BUILD_BENCH} \
+          -DBUILD_PRIMS_BENCH=${BUILD_PRIMS_BENCH} \
+          -DBUILD_ANN_BENCH=${BUILD_ANN_BENCH} \
           -DCMAKE_MESSAGE_LOG_LEVEL=${CMAKE_LOG_LEVEL} \
           ${CACHE_ARGS} \
           ${EXTRA_CMAKE_ARGS}
@@ -380,7 +403,6 @@ if (( ${NUMARGS} == 0 )) || hasArg pylibraft; then
     if [[ "${EXTRA_CMAKE_ARGS}" != *"DFIND_RAFT_CPP"* ]]; then
         EXTRA_CMAKE_ARGS="${EXTRA_CMAKE_ARGS} -DFIND_RAFT_CPP=ON"
     fi
-
     cd ${REPODIR}/python/pylibraft
     python setup.py build_ext --inplace -- -DCMAKE_PREFIX_PATH="${RAFT_DASK_BUILD_DIR};${INSTALL_PREFIX}" -DCMAKE_LIBRARY_PATH=${LIBRAFT_BUILD_DIR} ${EXTRA_CMAKE_ARGS} -- -j${PARALLEL_LEVEL:-1}
     if [[ ${INSTALL_TARGET} != "" ]]; then

diff --git a/ci/checks/copyright.py b/ci/checks/copyright.py
@@ -192,7 +192,8 @@ def checkCopyright_main():
                            action="append",
                            required=False,
                            default=["python/cuml/_thirdparty/",
-                                    "cpp/include/raft/thirdparty/"],
+                                    "cpp/include/raft/thirdparty/",
+                                    "cpp/cmake/modules/FindAVX.cmake"],
                            help=("Exclude the paths specified (regexp). "
                                  "Can be specified multiple times."))
 

diff --git a/conda/environments/bench_ann_cuda-118_arch-x86_64.yaml b/conda/environments/bench_ann_cuda-118_arch-x86_64.yaml
@@ -0,0 +1,37 @@
+# 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:
+- c-compiler
+- clang-tools=11.1.0
+- clang=11.1.0
+- cmake>=3.23.1,!=3.25.0
+- cuda-profiler-api=11.8.86
+- cudatoolkit=11.8
+- cxx-compiler
+- cython>=0.29,<0.30
+- faiss-proc=*=cuda
+- gcc_linux-64=11.*
+- glog>=0.6.0
+- h5py>=3.8.0
+- hnswlib=0.7.0
+- 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
+- libfaiss>=1.7.1
+- nccl>=2.9.9
+- ninja
+- nlohmann_json>=3.11.2
+- scikit-build>=0.13.1
+- sysroot_linux-64==2.17
+name: bench_ann_cuda-118_arch-x86_64
diff --git a/conda/recipes/libraft/build_libraft_nn_bench.sh b/conda/recipes/libraft/build_libraft_nn_bench.sh
@@ -0,0 +1,5 @@
+#!/usr/bin/env bash
+# Copyright (c) 2023, NVIDIA CORPORATION.
+
+./build.sh tests bench-ann --allgpuarch --no-nvtx
+cmake --install cpp/build --component ann_bench
diff --git a/conda/recipes/libraft/build_libraft_tests.sh b/conda/recipes/libraft/build_libraft_tests.sh
@@ -1,5 +1,5 @@
 #!/usr/bin/env bash
 # Copyright (c) 2022-2023, NVIDIA CORPORATION.
 
-./build.sh tests bench --allgpuarch --no-nvtx
+./build.sh tests bench-prims --allgpuarch --no-nvtx
 cmake --install cpp/build --component testing
diff --git a/conda/recipes/libraft/conda_build_config.yaml b/conda/recipes/libraft/conda_build_config.yaml
@@ -19,6 +19,18 @@ nccl_version:
 gtest_version:
   - "=1.10.0"
 
+glog_version:
+  - ">=0.6.0"
+
+faiss_version:
+  - ">=1.7.1"
+
+h5py_version:
+  - ">=3.8.0"
+
+nlohmann_json_version:
+  - ">=3.11.2"
+
 # The CTK libraries below are missing from the conda-forge::cudatoolkit
 # package. The "*_host_*" version specifiers correspond to `11.8` packages and the
 # "*_run_*" version specifiers correspond to `11.x` packages.

diff --git a/conda/recipes/libraft/meta.yaml b/conda/recipes/libraft/meta.yaml
@@ -186,3 +186,47 @@ outputs:
       home: https://rapids.ai/
       license: Apache-2.0
       summary: libraft template
+  - name: libraft-ann-bench
+    version: {{ version }}
+    script: build_libraft_nn_bench.sh
+    build:
+      script_env: *script_env
+      number: {{ GIT_DESCRIBE_NUMBER }}
+      string: cuda{{ cuda_major }}_{{ date_string }}_{{ GIT_DESCRIBE_HASH }}_{{ GIT_DESCRIBE_NUMBER }}
+      ignore_run_exports_from:
+        - {{ compiler('cuda') }}
+    requirements:
+      build:
+        - {{ compiler('c') }}
+        - {{ compiler('cuda') }} {{ cuda_version }}
+        - {{ compiler('cxx') }}
+        - cmake {{ cmake_version }}
+        - ninja
+        - sysroot_{{ target_platform }} {{ sysroot_version }}
+      host:
+        - {{ pin_subpackage('libraft', exact=True) }}
+        - {{ pin_subpackage('libraft-headers', exact=True) }}
+        - cuda-profiler-api {{ cuda_profiler_api_host_version }}
+        - libcublas {{ libcublas_host_version }}
+        - libcublas-dev {{ libcublas_host_version }}
+        - libcurand {{ libcurand_host_version }}
+        - libcurand-dev {{ libcurand_host_version }}
+        - libcusolver {{ libcusolver_host_version }}
+        - libcusolver-dev {{ libcusolver_host_version }}
+        - libcusparse {{ libcusparse_host_version }}
+        - libcusparse-dev {{ libcusparse_host_version }}
+        - glog {{ glog_version }}
+        - nlohmann_json {{ nlohmann_json_version }}
+        - libfaiss>=1.7.1
+        - faiss-proc=*=cuda
+      run:
+        - {{ pin_subpackage('libraft', exact=True) }}
+        - {{ pin_subpackage('libraft-headers', exact=True) }}
+        - glog {{ glog_version }}
+        - faiss-proc=*=cuda
+        - libfaiss {{ faiss_version }}
+        - h5py {{ h5py_version }}
+    about:
+      home: https://rapids.ai/
+      license: Apache-2.0
+      summary: libraft ann bench
diff --git a/cpp/CMakeLists.txt b/cpp/CMakeLists.txt
@@ -46,7 +46,8 @@ set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
 
 option(BUILD_SHARED_LIBS "Build raft shared libraries" ON)
 option(BUILD_TESTS "Build raft unit-tests" ON)
-option(BUILD_BENCH "Build raft C++ benchmark tests" OFF)
+option(BUILD_PRIMS_BENCH "Build raft C++ benchmark tests" OFF)
+option(BUILD_ANN_BENCH "Build raft ann benchmarks" OFF)
 option(CUDA_ENABLE_KERNELINFO "Enable kernel resource usage info" OFF)
 option(CUDA_ENABLE_LINEINFO
        "Enable the -lineinfo option for nvcc (useful for cuda-memcheck / profiler)" OFF
@@ -58,14 +59,20 @@ option(DISABLE_OPENMP "Disable OpenMP" OFF)
 option(RAFT_NVTX "Enable nvtx markers" OFF)
 
 set(RAFT_COMPILE_LIBRARY_DEFAULT OFF)
-if(BUILD_TESTS OR BUILD_BENCH)
+if(BUILD_TESTS
+   OR BUILD_PRIMS_BENCH
+   OR BUILD_ANN_BENCH
+)
   set(RAFT_COMPILE_LIBRARY_DEFAULT ON)
 endif()
 option(RAFT_COMPILE_LIBRARY "Enable building raft shared library instantiations"
        ${RAFT_COMPILE_LIBRARY_DEFAULT}
 )
 
-if(BUILD_TESTS OR BUILD_BENCH)
+if(BUILD_TESTS
+   OR BUILD_PRIMS_BENCH
+   OR BUILD_ANN_BENCH
+)
   # Needed because GoogleBenchmark changes the state of FindThreads.cmake, causing subsequent runs
   # to have different values for the `Threads::Threads` target. Setting this flag ensures
   # `Threads::Threads` is the same value in first run and subsequent runs.
@@ -78,7 +85,7 @@ include(CMakeDependentOption)
 
 message(VERBOSE "RAFT: Building optional components: ${raft_FIND_COMPONENTS}")
 message(VERBOSE "RAFT: Build RAFT unit-tests: ${BUILD_TESTS}")
-message(VERBOSE "RAFT: Building raft C++ benchmarks: ${BUILD_BENCH}")
+message(VERBOSE "RAFT: Building raft C++ benchmarks: ${BUILD_PRIMS_BENCH}")
 message(VERBOSE "RAFT: Enable detection of conda environment for dependencies: ${DETECT_CONDA_ENV}")
 message(VERBOSE "RAFT: Disable depreaction warnings " ${DISABLE_DEPRECATION_WARNINGS})
 message(VERBOSE "RAFT: Disable OpenMP: ${DISABLE_OPENMP}")
@@ -159,7 +166,7 @@ if(BUILD_TESTS)
   include(cmake/thirdparty/get_gtest.cmake)
 endif()
 
-if(BUILD_BENCH)
+if(BUILD_PRIMS_BENCH)
   include(${rapids-cmake-dir}/cpm/gbench.cmake)
   rapids_cpm_gbench()
 endif()
@@ -647,7 +654,7 @@ raft_export(
 # ##################################################################################################
 # * shared test/bench headers ------------------------------------------------
 
-if(BUILD_TESTS OR BUILD_BENCH)
+if(BUILD_TESTS OR BUILD_PRIMS_BENCH)
   include(internal/CMakeLists.txt)
 endif()
 
@@ -661,6 +668,13 @@ endif()
 # ##################################################################################################
 # * build benchmark executable -----------------------------------------------
 
-if(BUILD_BENCH)
-  include(bench/CMakeLists.txt)
+if(BUILD_PRIMS_BENCH)
+  include(bench/prims/CMakeLists.txt)
+endif()
+
+# ##################################################################################################
+# * build ann benchmark executable -----------------------------------------------
+
+if(BUILD_ANN_BENCH)
+  include(bench/ann/CMakeLists.txt)
 endif()