Enable dynamic dispatch build on Windows (#5127)

Summary:

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.

Differential Revision: D101649751

Author: mdouze

.github/workflows/build-pull-request.yml

@@ -170,6 +170,54 @@ 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)
+    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
+          conda install -y -q "pytorch<2.5" -c pytorch
+      - name: Build
+        shell: pwsh
+        run: |
+          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=ON `
+            -DPYTHON_EXECUTABLE="$env:CONDA\python.exe" `
+            -DBLA_VENDOR=Intel10_64_dyn `
+            -DCMAKE_BUILD_TYPE=Release `
+            .
+          cmake --build build --config Release -j $env:NUMBER_OF_PROCESSORS
+      - name: C++ tests
+        shell: pwsh
+        run: |
+          cd build
+          ctest --output-on-failure -C Release
+      - name: Install Python extension
+        shell: pwsh
+        working-directory: build/faiss/python
+        run: |
+          & "$env:CONDA\python.exe" setup.py install
+      - name: Python tests
+        shell: pwsh
+        run: |
+          & "$env:CONDA\python.exe" -m pytest tests/test_*.py
+          & "$env:CONDA\python.exe" -m pytest tests/torch_*.py
   windows-x86_64-conda:
     name: Windows x86_64 (conda)
     needs: linux-x86_64-cmake

faiss/CMakeLists.txt

@@ -452,16 +452,16 @@ 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(${FAISS_SIMD_AVX2_SRC}
         TARGET_DIRECTORY faiss
         PROPERTIES COMPILE_OPTIONS "-mavx2;-mfma;-mf16c;-mpopcnt"
@@ -471,15 +471,26 @@ if(FAISS_OPT_LEVEL STREQUAL "dd")
         PROPERTIES COMPILE_OPTIONS
           "-mavx512f;-mavx512cd;-mavx512vl;-mavx512dq;-mavx512bw;-mfma;-mf16c;-mpopcnt"
       )
-    elseif(CMAKE_SYSTEM_PROCESSOR MATCHES "(aarch64|arm64|ARM64)")
-      # ARM NEON is always available on aarch64, no special compiler flags needed
-      target_compile_definitions(faiss PRIVATE COMPILE_SIMD_ARM_NEON COMPILE_SIMD_ARM_SVE)
-      # Per-file SVE flags (NEON needs no flags on aarch64)
-      set_source_files_properties(${FAISS_SIMD_SVE_SRC}
+    else()
+      # Per-file SIMD flags (MSVC)
+      add_compile_options(/bigobj)
+      set_source_files_properties(${FAISS_SIMD_AVX2_SRC}
+        TARGET_DIRECTORY faiss
+        PROPERTIES COMPILE_OPTIONS "/arch:AVX2"
+      )
+      set_source_files_properties(${FAISS_SIMD_AVX512_SRC}
         TARGET_DIRECTORY faiss
-        PROPERTIES COMPILE_OPTIONS "-march=armv8.2-a+sve"
+        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/utils/distances_fused/avx512.cpp

@@ -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);
     }