Windows support

.gitignore

@@ -40,3 +40,7 @@ visualization/animation.gif
 
 # FAISS third_party dependencies
 !build/.gitkeep
+
+# WINDOWS
+.vs/
+

CMakeLists.txt

@@ -1,5 +1,7 @@
 cmake_minimum_required (VERSION 3.20)
-project(tsnecuda)
+project(tsnecuda LANGUAGES CXX)
+set(CMAKE_WINDOWS_EXPORT_ALL_SYMBOLS 1) # https://stackoverflow.com/questions/37128555/getting-cmake-to-build-shared-library-for-msvc
+
 set (CMAKE_PROJECT_VERSION 3)
 set (CMAKE_PROJECT_VERSION_MAJOR 3)
 set (CMAKE_PROJECT_VERSION_MINOR 0)
@@ -37,11 +39,11 @@ set(OPT_FLAGS
     -msse4
     -fPIC
     -m64
-    -Wno-sign-compare
+    #-Wno-sign-compare
     -g
     -O3
     -Wall
-    -Wextra
+    #-Wextra
     -DFINTEGER=int
 )
 string(REPLACE ";" " " CXX_FLAGS_STR "${CMAKE_CXX_FLAGS} ${OPT_FLAGS}")
@@ -120,18 +122,25 @@ else()
 endif()
 
 
+#set(CUDA_OPTS
+#    -O3
+#    -g
+#    -Xptxas '-dlcm=cg'
+#    -Xcompiler '-O3'
+#    -Xcompiler '-fPIC'
+#    -Xcompiler '-fopenmp'
+#    -Xcompiler '-msse4'
+#    -Xcompiler '-m64'
+#    -Xcompiler '-mpopcnt'
+#    -Xcompiler '-g'
+#    -Xlinker 'muldefs'
+#)
 set(CUDA_OPTS
     -O3
     -g
-    -Xptxas '-dlcm=cg'
-    -Xcompiler '-O3'
-    -Xcompiler '-fPIC'
-    -Xcompiler '-fopenmp'
-    -Xcompiler '-msse4'
-    -Xcompiler '-m64'
-    -Xcompiler '-mpopcnt'
-    -Xcompiler '-g'
-    -Xlinker 'muldefs'
+    -Xptxas -dlcm=cg
+    -Xcompiler /DYNAMICBASE,/openmp,/arch:AVX,/Zi,/Ox #-mpopcnt,
+    -Xlinker muldefs
 )
 string (REPLACE ";" " " NVCC_FLAGS_STR "${CUDA_ARCH} ${CUDA_OPTS}")
 set(CMAKE_CUDA_FLAGS "${NVCC_FLAGS_STR}")
@@ -145,9 +154,13 @@ find_package(OpenMP REQUIRED)
 
 # GTEST Configuration
 #-------------------------------------------------------------------------------
-find_package(gflags REQUIRED)
+find_package(GFlags REQUIRED)
+include_directories(${GFLAGS_INCLUDE_DIR})
+
 # find_package(GLog REQUIRED)
 find_package(GTest REQUIRED)
+include_directories(${GTEST_INCLUDE_DIR})
+
 #-------------------------------------------------------------------------------
 
 # ZMQ Configuration
@@ -223,17 +236,17 @@ set(PYTHON_SOURCES
 add_library(tsnecuda SHARED ${SOURCES})
 # set_property(TARGET tsnecuda PROPERTY POSITION_INDEPENDENT_CODE TRUE)
 target_link_libraries(tsnecuda PRIVATE
-    gflags
-    gtest
+    ${GFLAGS_LIBRARIES}
+    ${GTEST_LIBRARIES}
     CUDA::cudart
     CUDA::cublas
     # CUDA::cublasLt
     CUDA::cufft
     CUDA::cufftw
     CUDA::cusparse
     OpenMP::OpenMP_CXX
-    pthread
-    -Wl,--allow-multiple-definition
+    # pthread
+    # -Wl,--allow-multiple-definition
     ${ZMQ_LIBRARIES}
     ${FAISS_LIBRARIES}
 )
@@ -244,7 +257,7 @@ find_package(MKL)
 if(MKL_FOUND)
   target_link_libraries(tsnecuda PRIVATE ${MKL_LIBRARIES})
 else()
-  find_package(BLAS REQUIRED)
+  find_package(OpenBLAS REQUIRED)
   target_link_libraries(tsnecuda PRIVATE ${BLAS_LIBRARIES})
 
   find_package(LAPACK REQUIRED)
@@ -255,8 +268,8 @@ endif()
 #-------------------------------------------------------------------------------
 # Main executable
 #-------------------------------------------------------------------------------
-add_executable(tsne src/exe/main.cu)
-target_link_libraries(tsne tsnecuda)
+#add_executable(tsne src/exe/main.cu)
+#target_link_libraries(tsne tsnecuda)
 
 #-------------------------------------------------------------------------------
 # Python library copy and build

cmake/Modules/FindFAISS.cmake

@@ -11,12 +11,14 @@
 
 include(FindPackageHandleStandardArgs)
 
-set(FAISS_ROOT_DIR "" CACHE PATH "Folder contains FAISS")
+if(NOT DEFINED FAISS_ROOT_DIR)
+    set(FAISS_ROOT_DIR "" CACHE PATH "Folder contains FAISS")
+endif()
 
 # We are testing only a couple of files in the include directories
-find_path(FAISS_INCLUDE_DIR faiss/Index.h PATHS ${FAISS_ROOT_DIR})
-find_path(FAISS_GPU_INCLUDE_DIR faiss/gpu/GpuIndex.h PATHS ${FAISS_ROOT_DIR})
-find_library(FAISS_LIBRARY faiss)
+find_path(FAISS_INCLUDE_DIR faiss/Index.h PATHS "${FAISS_ROOT_DIR}/include")
+find_path(FAISS_GPU_INCLUDE_DIR faiss/gpu/GpuIndex.h PATHS "${FAISS_ROOT_DIR}/include")
+find_library(FAISS_LIBRARY faiss PATHS "${FAISS_ROOT_DIR}/lib")
 
 find_package_handle_standard_args(FAISS DEFAULT_MSG FAISS_INCLUDE_DIR FAISS_GPU_INCLUDE_DIR FAISS_LIBRARY)
 

cmake/Modules/FindGFlags.cmake

@@ -14,29 +14,8 @@ include(FindPackageHandleStandardArgs)
 set(GFLAGS_ROOT_DIR "" CACHE PATH "Folder contains Gflags")
 
 # We are testing only a couple of files in the include directories
-if(WIN32)
-    find_path(GFLAGS_INCLUDE_DIR gflags/gflags.h
-        PATHS ${GFLAGS_ROOT_DIR}/src/windows)
-else()
-    find_path(GFLAGS_INCLUDE_DIR gflags/gflags.h
-        PATHS ${GFLAGS_ROOT_DIR})
-endif()
-
-if(MSVC)
-    find_library(GFLAGS_LIBRARY_RELEASE
-        NAMES libgflags
-        PATHS ${GFLAGS_ROOT_DIR}
-        PATH_SUFFIXES Release)
-
-    find_library(GFLAGS_LIBRARY_DEBUG
-        NAMES libgflags-debug
-        PATHS ${GFLAGS_ROOT_DIR}
-        PATH_SUFFIXES Debug)
-
-    set(GFLAGS_LIBRARY optimized ${GFLAGS_LIBRARY_RELEASE} debug ${GFLAGS_LIBRARY_DEBUG})
-else()
-    find_library(GFLAGS_LIBRARY gflags)
-endif()
+find_path(GFLAGS_INCLUDE_DIR gflags/gflags.h PATHS "${GFLAGS_ROOT_DIR}/include")
+find_library(GFLAGS_LIBRARY gflags PATHS "${GFLAGS_ROOT_DIR}/lib")
 
 find_package_handle_standard_args(GFlags DEFAULT_MSG GFLAGS_INCLUDE_DIR GFLAGS_LIBRARY)
 

cmake/Modules/FindGTest.cmake

@@ -0,0 +1,29 @@
+# - Try to find GTEST
+#
+# The following variables are optionally searched for defaults
+#  GTEST_ROOT_DIR:            Base directory where all GTEST components are found
+#
+# The following are set after configuration is done:
+#  GTEST_FOUND
+#  GTEST_INCLUDE_DIRS
+#  GTEST_LIBRARIES
+#  GTEST_LIBRARYRARY_DIRS
+
+include(FindPackageHandleStandardArgs)
+
+set(GTEST_ROOT_DIR "" CACHE PATH "Folder contains GTest")
+
+# We are testing only a couple of files in the include directories
+find_path(GTEST_INCLUDE_DIR gtest/gtest.h PATHS "${GTEST_ROOT_DIR}/include")
+find_library(GTEST_LIBRARY gtest PATHS "${GTEST_ROOT_DIR}/lib")
+
+find_package_handle_standard_args(GTest DEFAULT_MSG GTEST_INCLUDE_DIR GTEST_LIBRARY)
+
+
+if(GTEST_FOUND)
+    set(GTEST_INCLUDE_DIRS ${GTEST_INCLUDE_DIR})
+    set(GTEST_LIBRARIES ${GTEST_LIBRARY})
+    message(STATUS "Found gtest  (include: ${GTEST_INCLUDE_DIR}, library: ${GTEST_LIBRARY})")
+    mark_as_advanced(GTEST_LIBRARY_DEBUG GTEST_LIBRARY_RELEASE
+                     GTEST_LIBRARY GTEST_INCLUDE_DIR GTEST_ROOT_DIR)
+endif()

cmake/Modules/FindMKL.cmake

@@ -58,10 +58,10 @@ find_path(MKL_ROOT_DIR
     /opt/intel/mkl/
     /opt/intel/cmkl/
     /Library/Frameworks/Intel_MKL.framework/Versions/Current/lib/universal
-    "Program Files (x86)/Intel/ComposerXE-2011/mkl"
+    "C:/Program Files (x86)/Intel/oneAPI/mkl/latest"
 )
 
-#MESSAGE("-- MKL_ROOT_DIR : ${MKL_ROOT_DIR}") # for debug
+MESSAGE("-- MKL_ROOT_DIR : ${MKL_ROOT_DIR}") # for debug
 
 find_path(MKL_INCLUDE_DIR
   mkl_cblas.h

cmake/Modules/FindOpenBLAS.cmake

@@ -11,6 +11,7 @@ SET(Open_BLAS_INCLUDE_SEARCH_PATHS
   /opt/OpenBLAS/include
   $ENV{OpenBLAS_HOME}
   $ENV{OpenBLAS_HOME}/include
+  ./blas/include
 )
 
 SET(Open_BLAS_LIB_SEARCH_PATHS
@@ -28,10 +29,11 @@ SET(Open_BLAS_LIB_SEARCH_PATHS
         $ENV{OpenBLAS}/lib
         $ENV{OpenBLAS_HOME}
         $ENV{OpenBLAS_HOME}/lib
+        ./blas/lib
  )
 
 FIND_PATH(OpenBLAS_INCLUDE_DIR NAMES cblas.h PATHS ${Open_BLAS_INCLUDE_SEARCH_PATHS})
-FIND_LIBRARY(OpenBLAS_LIB NAMES openblas PATHS ${Open_BLAS_LIB_SEARCH_PATHS})
+FIND_LIBRARY(OpenBLAS_LIB NAMES libopenblas PATHS ${Open_BLAS_LIB_SEARCH_PATHS})
 
 SET(OpenBLAS_FOUND ON)
 

src/fit_tsne.cu

@@ -97,7 +97,7 @@ void tsnecuda::RunTsne(tsnecuda::Options &opt,
     // TODO: Investigate what it takes to use unified memory + Async fetch and execution
     float *knn_squared_distances = new float[num_points * num_neighbors];
     memset(knn_squared_distances, 0, num_points * num_neighbors * sizeof(float));
-    long *knn_indices = new long[num_points * num_neighbors];
+    int64_t *knn_indices = new int64_t[num_points * num_neighbors];
 
     // Set cache configs
     // cudaFuncSetCacheConfig(tsnecuda::IntegrationKernel, cudaFuncCachePreferL1);
@@ -116,7 +116,7 @@ void tsnecuda::RunTsne(tsnecuda::Options &opt,
     // TODO: Add suport for arbitrary metrics on GPU (Introduced by recent FAISS computation)
     // TODO: Expose Multi-GPU computation (+ Add streaming memory support for GPU optimization)
     tsnecuda::util::KNearestNeighbors(gpu_opt, opt, knn_indices, knn_squared_distances, high_dim_points, high_dim, num_points, num_neighbors);
-    thrust::device_vector<long> knn_indices_long_device(knn_indices, knn_indices + num_points * num_neighbors);
+    thrust::device_vector<int64_t> knn_indices_long_device(knn_indices, knn_indices + num_points * num_neighbors);
     thrust::device_vector<int> pij_indices_device(num_points * num_neighbors);
     tsnecuda::util::PostprocessNeighborIndices(gpu_opt, pij_indices_device, knn_indices_long_device,
                                                num_points, num_neighbors);
@@ -243,7 +243,7 @@ void tsnecuda::RunTsne(tsnecuda::Options &opt,
     }
 
     // FIT-TNSE Parameters
-    int n_interpolation_points = 3;
+    const int n_interpolation_points = 3;
     // float intervals_per_integer = 1;
     int min_num_intervals = 125;
     int N = num_points;

src/include/common.h

@@ -44,7 +44,7 @@
 #include <string.h>
 #include <math.h>
 #include <time.h>
-#include <sys/time.h>
+//#include <sys/time.h>
 
 // C++ Library includes
 #include <random>
@@ -58,3 +58,10 @@
 #include <cmath>
 
 #endif
+
+#if defined(_MSC_VER)
+#  pragma warning(push)
+#  pragma warning(disable: 4127) // warning C4127: Conditional expression is constant
+#include <BaseTsd.h>
+typedef SSIZE_T ssize_t;
+#endif

src/include/util/distance_utils.h

@@ -97,7 +97,7 @@ void KNearestNeighbors(tsnecuda::GpuOptions &gpu_opt, tsnecuda::Options &base_op
 
 __global__ void PostprocessNeighborIndicesKernel(
     volatile int *__restrict__ indices,
-    const long *__restrict__ long_indices,
+    const int64_t *__restrict__ long_indices,
     const int num_points,
     const int num_neighbors);
 

src/python/setup.py

@@ -10,7 +10,7 @@
     author='Chan, David M., Huang, Forrest., Rao, Roshan.',
     author_email='[email protected]',
     packages=['tsnecuda', 'tsnecuda.test'],
-    package_data={'tsnecuda': ['libtsnecuda.so']},
+    package_data={'tsnecuda': ['libtsnecuda.so', 'tsnecuda.dll']},
     scripts=[],
     url='https://github.com/CannyLab/tsne-cuda',
     license='LICENSE.txt',

src/python/tsnecuda/TSNE.py

@@ -136,8 +136,7 @@ def __init__(self,
 
         # Build the hooks for the BH T-SNE library
         self._path = os.path.dirname(__file__)
-        self._lib = N.ctypeslib.load_library(
-            'libtsnecuda', self._path)  # Load the ctypes library
+        self._lib = ctypes.CDLL('tsnecuda.dll', winmode=0)
 
         # Hook the BH T-SNE function
         self._lib.pymodule_tsne.restype = None

src/util/distance_utils.cu

@@ -131,7 +131,7 @@ void tsnecuda::util::KNearestNeighbors(tsnecuda::GpuOptions &gpu_opt,
 // TODO: Maybe think about getting FAISS to return integers (long-term todo)
 __global__ void tsnecuda::util::PostprocessNeighborIndicesKernel(
     volatile int *__restrict__ indices,
-    const long *__restrict__ long_indices,
+    const int64_t *__restrict__ long_indices,
     const int num_points,
     const int num_neighbors)
 {

windows.txt

@@ -0,0 +1,19 @@
+The CLI commands I did to compile tsne-cuda. Mind the paths, also MKL/OpenBLAS search paths are still hardcoded into CMAKE, might need to change them:
+
+conda install conda-forge::faiss-gpu
+conda install conda-forge::gflags
+conda install conda-forge::gtest
+
+cmake -G "Visual Studio 17 2022" -A x64  -DCMAKE_CONFIGURATION_TYPES:STRING="RelWithDebInfo" -DGFLAGS_ROOT_DIR="C:\Users\Michael\miniconda3\Library" -DFAISS_ROOT_DIR="C:/Users/Michael/miniconda3/Library" -DGTEST_ROOT_DIR="C:\Users\Michael\miniconda3\Library" -B  "C:\Users\Michael\CProjects\tsne-cuda\build" -T cuda=11.8  "C:\Users\Michael\CProjects\tsne-cuda"
+
+"C:\Program Files\Microsoft Visual Studio\2022\Community\MSBuild\Current\Bin\MSBuild.exe" "C:\Users\Michael\CProjects\tsne-cuda\build\tsnecuda.sln"
+
+pip install "C:\Users\Michael\CProjects\tsne-cuda\build\RelWithDebInfo\python"
+
+" if it failed to, copy tsnecuda.dll into "C:\Users\Michael\miniconda3\envs\transformers\Library\bin\tsnecuda.dll manually""
+
+----Questionable things I did not find windows analogue for and commented out----
+pthread
+sys/time.h
+cxx compiler flag -mpopcnt
+-Wl,--allow-multiple-definition unrecognized options