facebookresearch/faiss
1
0
Fork 0
mirror of https://github.com/facebookresearch/faiss.git synced 2026-10-12 02:13:36 +00:00
Code Issues Packages Projects Releases Wiki Activity

Enable dynamic dispatch build on Windows (#5127)

Browse Source 
Summary:
Pull Request resolved: https://github.com/facebookresearch/faiss/pull/5127

Attempt to enable FAISS dynamic dispatch (DD) mode on Windows/MSVC.

Changes:
- CMakeLists.txt: Remove if(NOT WIN32) guard from DD section, add MSVC per-file
  SIMD flags (/arch:AVX2, /arch:AVX512, /bigobj) alongside existing GCC/Clang flags
- build-pull-request.yml: Add windows-x86_64-DD-cmake job that runs immediately
  (no dependency on linux-x86_64-cmake), builds with MSVC and FAISS_OPT_LEVEL=dd,
  runs C++ tests and Python tests

This diff is expected to fail on Windows due to MSVC requiring explicit template
specialization declarations (C++ §17.8.3) which GCC/Clang don't enforce. The CI
failure will surface the exact errors to guide the fix.

UPDATE: in fact, MSVC does not seem to require this. After fixing build scripts and a few classical Windows error, the C++ and python tests pass.

Reviewed By: algoriddle

Differential Revision: D101649751

fbshipit-source-id: 765ef8483d02652ce58625cd22baab8870acf718
This commit is contained in:
Matthijs Douze
2026-04-23 13:31:11 -07:00
committed by meta-codesync[bot]
parent a25bcb1e22
commit 8d1b964c11
12 changed files with 138 additions and 40 deletions
Download Patch File Download Diff File Expand all files Collapse all files
.github/workflows/build-pull-request.yml
+69
View File
@@ -170,6 +170,75 @@ jobs:
fetch-tags: true
- name: Build and Package (conda)
uses: ./.github/actions/build_conda
windows-x86_64-DD-cmake:
name: Windows x86_64 Dynamic Dispatch (cmake)
needs: linux-x86_64-cmake
runs-on: windows-2022
steps:
- name: Checkout
uses: actions/checkout@v4
- name: Setup miniconda
uses: conda-incubator/setup-miniconda@v3
with:
python-version: '3.12'
miniforge-version: latest
channels: conda-forge
- name: Install dependencies
shell: pwsh
run: |
conda install -y -q cmake swig "numpy>=2.0,<3.0" scipy pytest gflags setuptools
conda install -y -q mkl=2024.2.2 mkl-devel=2024.2.2
- name: Build
shell: pwsh
run: |
$pythonExe = (Get-Command python).Source
$pythonDir = Split-Path $pythonExe
# Add Python libs dir so MSVC linker can find python312.lib
$env:LIB = "$pythonDir\libs;$env:LIB"
cmake -B build -T v143 -A x64 -G "Visual Studio 17 2022" `
-DBUILD_TESTING=ON `
-DBUILD_SHARED_LIBS=ON `
-DFAISS_ENABLE_GPU=OFF `
-DFAISS_OPT_LEVEL=dd `
-DFAISS_ENABLE_PYTHON=ON `
-DFAISS_ENABLE_C_API=OFF `
-DBLA_VENDOR=Intel10_64_dyn `
-DCMAKE_BUILD_TYPE=Release `
-DPython_ROOT_DIR="$pythonDir" `
.
# Add faiss.dll and dependency DLLs to PATH for GoogleTest discovery
$env:PATH = "$PWD\build\faiss\Release;$env:CONDA_PREFIX\Library\bin;$env:PATH"
cmake --build build --config Release -j $env:NUMBER_OF_PROCESSORS
- name: C++ tests
shell: pwsh
run: |
# Copy all required DLLs next to the test executable:
# faiss.dll, gtest/gtest_main DLLs, and conda dependency DLLs
Copy-Item build\faiss\Release\*.dll build\tests\Release\
Copy-Item build\bin\Release\*.dll build\tests\Release\
Copy-Item "$env:CONDA_PREFIX\Library\bin\*.dll" build\tests\Release\ -ErrorAction SilentlyContinue
$env:PATH = "$env:CONDA_PREFIX\Library\bin;$env:PATH"
cd build
ctest --output-on-failure -C Release -E "SIMDConfig.modern_hardware_has_simd_support"
- name: Install Python extension
shell: pwsh
run: |
$repoRoot = $PWD.Path
$env:PATH = "$repoRoot\build\faiss\Release;$env:CONDA_PREFIX\Library\bin;$env:PATH"
cd build\faiss\python
python setup.py install
# Copy faiss.dll and MKL DLLs next to the installed _swigfaiss.pyd
$siteDir = python -c "import sysconfig; print(sysconfig.get_paths()['purelib'])"
$siteDir = "$siteDir\faiss"
Copy-Item "$repoRoot\build\faiss\Release\faiss.dll" $siteDir\
Copy-Item "$env:CONDA_PREFIX\Library\bin\mkl_*.dll" $siteDir\ -ErrorAction SilentlyContinue
Copy-Item "$env:CONDA_PREFIX\Library\bin\libiomp5md.dll" $siteDir\ -ErrorAction SilentlyContinue
- name: Python tests
shell: pwsh
run: |
$env:PATH = "$env:CONDA_PREFIX\Library\bin;$env:PATH"
$tests = (Get-ChildItem tests/test_*.py).FullName
pytest $tests
windows-x86_64-conda:
name: Windows x86_64 (conda)
needs: linux-x86_64-cmake
CMakeLists.txt
+3 -1
View File
@@ -115,7 +115,9 @@ endif()
include(CTest)
if(BUILD_TESTING)
add_subdirectory(tests)
add_subdirectory(perf_tests)
if(NOT WIN32)
add_subdirectory(perf_tests)
endif()
if(FAISS_ENABLE_GPU)
if(FAISS_ENABLE_ROCM)
add_subdirectory(faiss/gpu/test)
faiss/CMakeLists.txt
+24 -13
View File
@@ -462,16 +462,16 @@ target_compile_definitions(faiss_sve PRIVATE COMPILE_SIMD_ARM_NEON COMPILE_SIMD_
if(FAISS_OPT_LEVEL STREQUAL "dd")
# Add SIMD source files to main faiss target for DD builds
target_sources(faiss PRIVATE ${FAISS_SIMD_SRC})
if(NOT WIN32)
target_compile_definitions(faiss PRIVATE FAISS_ENABLE_DD)
# Architecture-specific SIMD definitions for Dynamic Dispatch
if(CMAKE_SYSTEM_PROCESSOR MATCHES "(x86_64|amd64|AMD64)")
target_compile_definitions(faiss PRIVATE
COMPILE_SIMD_AVX2 COMPILE_SIMD_AVX512 COMPILE_SIMD_AVX512_SPR)
target_compile_definitions(faiss PRIVATE FAISS_ENABLE_DD)
# Architecture-specific SIMD definitions for Dynamic Dispatch
if(CMAKE_SYSTEM_PROCESSOR MATCHES "(x86_64|amd64|AMD64)")
target_compile_definitions(faiss PRIVATE
COMPILE_SIMD_AVX2 COMPILE_SIMD_AVX512 COMPILE_SIMD_AVX512_SPR)
if(NOT WIN32)
# Baseline flags for common files (prevents auto-vectorization)
target_compile_options(faiss PRIVATE
$<$<COMPILE_LANGUAGE:CXX>:-mpopcnt -msse4 -mno-avx -mno-avx2>)
# Per-file SIMD flags
# Per-file SIMD flags (GCC/Clang)
set_source_files_properties(${FAISS_SIMD_AVX2_SRC}
TARGET_DIRECTORY faiss
PROPERTIES COMPILE_OPTIONS "-mavx2;-mfma;-mf16c;-mpopcnt"
@@ -481,15 +481,26 @@ if(FAISS_OPT_LEVEL STREQUAL "dd")
PROPERTIES COMPILE_OPTIONS
"-mavx512f;-mavx512cd;-mavx512vl;-mavx512dq;-mavx512bw;-mfma;-mf16c;-mpopcnt"
)
elseif(CMAKE_SYSTEM_PROCESSOR MATCHES "(aarch64|arm64|ARM64)")
# ARM NEON is always available on aarch64, no special compiler flags needed
target_compile_definitions(faiss PRIVATE COMPILE_SIMD_ARM_NEON COMPILE_SIMD_ARM_SVE)
# Per-file SVE flags (NEON needs no flags on aarch64)
set_source_files_properties(${FAISS_SIMD_SVE_SRC}
else()
# Per-file SIMD flags (MSVC)
add_compile_options(/bigobj)
set_source_files_properties(${FAISS_SIMD_AVX2_SRC}
TARGET_DIRECTORY faiss
PROPERTIES COMPILE_OPTIONS "-march=armv8.2-a+sve"
PROPERTIES COMPILE_OPTIONS "/arch:AVX2"
)
set_source_files_properties(${FAISS_SIMD_AVX512_SRC}
TARGET_DIRECTORY faiss
PROPERTIES COMPILE_OPTIONS "/arch:AVX512"
)
endif()
elseif(CMAKE_SYSTEM_PROCESSOR MATCHES "(aarch64|arm64|ARM64)")
# ARM NEON is always available on aarch64, no special compiler flags needed
target_compile_definitions(faiss PRIVATE COMPILE_SIMD_ARM_NEON COMPILE_SIMD_ARM_SVE)
# Per-file SVE flags (NEON needs no flags on aarch64)
set_source_files_properties(${FAISS_SIMD_SVE_SRC}
TARGET_DIRECTORY faiss
PROPERTIES COMPILE_OPTIONS "-march=armv8.2-a+sve"
)
endif()
endif()
faiss/python/CMakeLists.txt
+1 -1
View File
@@ -295,7 +295,7 @@ find_package(Python REQUIRED
COMPONENTS Development.Module NumPy
)
add_library(faiss_python_callbacks EXCLUDE_FROM_ALL
add_library(faiss_python_callbacks STATIC EXCLUDE_FROM_ALL
python_callbacks.cpp
)
set_property(TARGET faiss_python_callbacks
faiss/utils/distances_fused/avx512.cpp
+3 -3
View File
@@ -78,7 +78,7 @@ void kernel(
const float* const __restrict xd_0 = x + i * DIM;
// prefetch the next point
_mm_prefetch(xd_0 + DIM * sizeof(float), _MM_HINT_NTA);
_mm_prefetch((char*)(xd_0 + DIM * sizeof(float)), _MM_HINT_NTA);
// load a single point from x
// load -2 * value
@@ -262,10 +262,10 @@ void exhaustive_L2sqr_fused_cmax(
}
}
const size_t nx_p = (nx / NX_POINTS_PER_LOOP) * NX_POINTS_PER_LOOP;
const idx_t nx_p = (nx / NX_POINTS_PER_LOOP) * NX_POINTS_PER_LOOP;
// the main loop.
#pragma omp parallel for schedule(dynamic)
for (size_t i = 0; i < nx_p; i += NX_POINTS_PER_LOOP) {
for (idx_t i = 0; i < nx_p; i += NX_POINTS_PER_LOOP) {
kernel<DIM, NX_POINTS_PER_LOOP, NY_POINTS_PER_LOOP>(
x, y, y_transposed.data(), ny, res, y_norms, i);
}
tests/CMakeLists.txt
+17 -5
View File
@@ -10,9 +10,7 @@ set(FAISS_TEST_SRC
test_ivfpq_indexing.cpp
test_lowlevel_ivf.cpp
test_ivf_index.cpp
test_merge.cpp
test_omp_threads.cpp
test_ondisk_ivf.cpp
test_pairs_decoding.cpp
test_params_override.cpp
test_pq_encoding.cpp
@@ -31,12 +29,10 @@ set(FAISS_TEST_SRC
test_partitioning.cpp
test_fastscan_perf.cpp
test_pqfs_unaligned.cpp
test_disable_pq_sdc_tables.cpp
test_common_ivf_empty_index.cpp
test_callback.cpp
test_utils.cpp
test_hamming.cpp
test_mmap.cpp
test_zerocopy.cpp
test_scalar_quantizer.cpp
test_factory_tools.cpp
@@ -49,6 +45,16 @@ set(FAISS_TEST_SRC
test_fast_scan_distance_to_code.cpp
)
# Tests that use POSIX APIs or have platform-specific issues on Windows
if(NOT WIN32)
list(APPEND FAISS_TEST_SRC
test_merge.cpp
test_ondisk_ivf.cpp
test_disable_pq_sdc_tables.cpp
test_mmap.cpp
)
endif()
if(FAISS_ENABLE_SVS)
list(APPEND FAISS_TEST_SRC test_svs.cpp)
endif()
@@ -138,4 +144,10 @@ target_link_libraries(faiss_test PRIVATE
# Defines `gtest_discover_tests()`.
include(GoogleTest)
gtest_discover_tests(faiss_test)
if(WIN32)
# On Windows, defer test discovery to ctest time (PRE_TEST) instead of
# build time (POST_BUILD) because faiss.dll is not in PATH during build.
gtest_discover_tests(faiss_test DISCOVERY_MODE PRE_TEST)
else()
gtest_discover_tests(faiss_test)
endif()
tests/test_fastscan_perf.cpp
+1
View File
@@ -7,6 +7,7 @@
#include <gtest/gtest.h>
#include <chrono>
#include <cstddef>
#include <cstdint>
#include <memory>
tests/test_mmap.cpp
+1 -1
View File
@@ -46,7 +46,7 @@ std::vector<uint8_t> make_binary_data(
size_t seed) {
std::vector<uint8_t> database(n * d);
std::mt19937 rng(seed);
std::uniform_int_distribution<uint8_t> distrib(0, 255);
std::uniform_int_distribution<int> distrib(0, 255);
for (size_t i = 0; i < n * d; i++) {
database[i] = distrib(rng);
tests/test_pq_code_distance.cpp
+8 -8
View File
@@ -51,7 +51,7 @@ void test(
// rng
std::minstd_rand rng(123);
std::uniform_int_distribution<uint8_t> u(0, 255);
std::uniform_int_distribution<int> u(0, 255);
std::uniform_real_distribution<float> uf(0, 1);
// initialize lookup
@@ -65,10 +65,10 @@ void test(
#pragma omp parallel
{
std::minstd_rand rng0(123);
std::uniform_int_distribution<uint8_t> u1(0, 255);
std::uniform_int_distribution<int> u1(0, 255);
#pragma omp for schedule(guided)
for (size_t i = 0; i < codes.size(); i++) {
for (int64_t i = 0; i < (int64_t)codes.size(); i++) {
codes[i] = u1(rng0);
}
}
@@ -77,7 +77,7 @@ void test(
std::vector<float> resultsRef(n, 0);
for (size_t k = 0; k < 10; k++) {
#pragma omp parallel for schedule(guided)
for (size_t i = 0; i < n; i++) {
for (int64_t i = 0; i < (int64_t)n; i++) {
resultsRef[i] = faiss::PQCodeDistance<
faiss::PQDecoder8,
faiss::SIMDLevel::NONE>::
@@ -93,7 +93,7 @@ void test(
const auto startingTimepoint = std::chrono::steady_clock::now();
for (size_t k = 0; k < 1000; k++) {
#pragma omp parallel for schedule(guided)
for (size_t i = 0; i < n; i++) {
for (int64_t i = 0; i < (int64_t)n; i++) {
resultsScalar1x[i] = faiss::PQCodeDistance<
faiss::PQDecoder8,
faiss::SIMDLevel::NONE>::
@@ -118,7 +118,7 @@ void test(
const auto startingTimepoint = std::chrono::steady_clock::now();
for (size_t k = 0; k < 1000; k++) {
#pragma omp parallel for schedule(guided)
for (size_t i = 0; i < n; i += 4) {
for (int64_t i = 0; i < (int64_t)n; i += 4) {
faiss::PQCodeDistance<
faiss::PQDecoder8,
faiss::SIMDLevel::NONE>::
@@ -151,7 +151,7 @@ void test(
const auto startingTimepoint = std::chrono::steady_clock::now();
for (size_t k = 0; k < 1000; k++) {
#pragma omp parallel for schedule(guided)
for (size_t i = 0; i < n; i++) {
for (int64_t i = 0; i < (int64_t)n; i++) {
resultsDispatched1x[i] = faiss::pq_code_distance_single(
subq, 8, lookup.data(), codes.data() + subq * i);
}
@@ -170,7 +170,7 @@ void test(
const auto startingTimepoint = std::chrono::steady_clock::now();
for (size_t k = 0; k < 1000; k++) {
#pragma omp parallel for schedule(guided)
for (size_t i = 0; i < n; i += 4) {
for (int64_t i = 0; i < (int64_t)n; i += 4) {
faiss::pq_code_distance_four(
subq,
8,
tests/test_pqfs_unaligned.cpp
+1 -3
View File
@@ -59,9 +59,7 @@ TEST(FastScanUnaligned, TestUnalignedCodesAccess) {
#endif
std::random_device rd;
std::mt19937 gen(rd());
std::uniform_int_distribution<uint8_t> dist(
std::numeric_limits<uint8_t>::min(),
std::numeric_limits<uint8_t>::max());
std::uniform_int_distribution<int> dist(0, 255);
std::vector<uint8_t> aligned_buffer =
create_aligned_buffer(code_size, alignment);
tests/test_scalar_quantizer.cpp
+9 -4
View File
@@ -9,6 +9,7 @@
#include <cmath>
#include <memory>
#include <random>
#include <vector>
#include <faiss/IndexFlat.h>
@@ -110,12 +111,14 @@ TEST(TestSQ0bit, CoarseOnlySearch) {
int nq = 10;
int k = 5;
std::mt19937 rng(42);
std::uniform_real_distribution<float> distrib(0.0f, 1.0f);
std::vector<float> xb(nb * d), xq(nq * d);
for (int i = 0; i < nb * d; i++) {
xb[i] = drand48();
xb[i] = distrib(rng);
}
for (int i = 0; i < nq * d; i++) {
xq[i] = drand48();
xq[i] = distrib(rng);
}
faiss::IndexFlatL2 quantizer(d);
@@ -180,12 +183,14 @@ TEST(TestSQ0bit, InnerProduct) {
int nq = 5;
int k = 3;
std::mt19937 rng2(43);
std::uniform_real_distribution<float> distrib2(0.0f, 1.0f);
std::vector<float> xb(nb * d), xq(nq * d);
for (int i = 0; i < nb * d; i++) {
xb[i] = drand48();
xb[i] = distrib2(rng2);
}
for (int i = 0; i < nq * d; i++) {
xq[i] = drand48();
xq[i] = distrib2(rng2);
}
faiss::IndexFlatIP quantizer(d);
tests/test_zerocopy.cpp
+1 -1
View File
@@ -37,7 +37,7 @@ std::vector<uint8_t> make_binary_data(
size_t seed) {
std::vector<uint8_t> database(n * d);
std::mt19937 rng(seed);
std::uniform_int_distribution<uint8_t> distrib(0, 255);
std::uniform_int_distribution<int> distrib(0, 255);
for (size_t i = 0; i < n * d; i++) {
database[i] = distrib(rng);