Update CMake configuration for CUDA support and simplify pybind11 int…

…egration

CMakeLists.txt

@@ -1,7 +1,7 @@
 cmake_minimum_required(VERSION 3.19)
 project(SCE)
 set(CMAKE_CXX_STANDARD 14 CACHE STRING "C++ version selection") 
-cmake_policy(SET CMP0104 NEW) # Can't get CUDA_ARCHITECTURES to work with NEW
+cmake_policy(SET CMP0104 OLD) # Can't get CUDA_ARCHITECTURES to work with NEW
 
 # Variable definitions
 set(TARGET_NAME SCE)
@@ -13,8 +13,6 @@ ENDIF()
 
 set(CMAKE_CXX_FLAGS_DEBUG "${CMAKE_CXX_FLAGS_DEBUG} -O0")
 set(CMAKE_C_FLAGS_DEBUG "${CMAKE_C_FLAGS_DEBUG} -O0")
-
-
 # Set these to profile
 # set(CMAKE_CXX_FLAGS_RELEASE "-O2 -g")
 # set(CMAKE_CUDA_FLAGS_RELEASE "-O2 -DUSE_CUDA_PROFILER")
@@ -37,8 +35,7 @@ endif()
 # Add libraries
 find_package(ZLIB REQUIRED)
 find_package(Boost REQUIRED)
-find_package(Python COMPONENTS Interpreter Development REQUIRED)
-find_package(pybind11 CONFIG REQUIRED)
+find_package(pybind11 REQUIRED)
 include_directories(${Boost_INCLUDE_DIRS})
 
 # Check for CUDA and compiles GPU library
@@ -62,36 +59,40 @@ if(CMAKE_CUDA_COMPILER)
     set(CMAKE_CUDA_FLAGS "${CUDA_OPTS}")
 
     add_compile_definitions(GPU_AVAILABLE)
-    # Create Python extension module with CUDA support
-    pybind11_add_module(${TARGET_NAME}_CUDA src/wtsne_gpu.cu)
-
-    # Set CUDA properties (optional, but recommended)
-    set_target_properties(${TARGET_NAME}_CUDA PROPERTIES
-        CUDA_SEPARABLE_COMPILATION ON  # Allows linking multiple CUDA files
-        POSITION_INDEPENDENT_CODE ON    # Required for shared libraries
-    )
+    add_library("${TARGET_NAME}_CUDA" OBJECT src/wtsne_gpu.cu)
+    target_include_directories("${TARGET_NAME}_CUDA" PRIVATE "${pybind11_INCLUDE_DIRS}")
+    set_property(TARGET "${TARGET_NAME}_CUDA"
+                 PROPERTY POSITION_INDEPENDENT_CODE ON
+                 CUDA_SEPARABLE_COMPILATION OFF
+                 CUDA_RESOLVE_DEVICE_SYMBOLS ON   # try and ensure device link with nvcc
+                 CUDA_VISIBILITY_PRESET "hidden")
 else()
     message(STATUS "CUDA not found, compiling CPU code only")
 endif()
 
 # Compile CPU library
-pybind11_add_module("${TARGET_NAME}" src/wtsne_cpu.cpp src/python_bindings.cpp)
-
+add_library("${TARGET_NAME}" MODULE src/wtsne_cpu.cpp src/python_bindings.cpp)
+set_target_properties("${TARGET_NAME}" PROPERTIES
+    CXX_VISIBILITY_PRESET "hidden"
+    PREFIX "${PYTHON_MODULE_PREFIX}"
+    SUFFIX "${PYTHON_MODULE_EXTENSION}"
+)
 if(UNIX AND (NOT APPLE OR NOT CMAKE_COMPILER_IS_GNUCC))
     set_target_properties("${TARGET_NAME}" PROPERTIES
                           INTERPROCEDURAL_OPTIMIZATION ON)
 endif()
 
 if(CMAKE_CUDA_COMPILER)
     target_link_libraries("${TARGET_NAME}" PRIVATE "${TARGET_NAME}_CUDA")
+    set_property(TARGET "${TARGET_NAME}" PROPERTY CUDA_RESOLVE_DEVICE_SYMBOLS ON)
 endif()
 
-if(DEFINED ENV{CONDA_PREFIX} AND (NOT APPLE OR CMAKE_COMPILER_IS_GNUCC OR "$ENV{SCE_BUILD_ENV}" STREQUAL "conda_forge"))
+target_link_libraries("${TARGET_NAME}" PRIVATE pybind11::module)
+if(DEFINED ENV{CONDA_PREFIX} AND (NOT APPLE OR CMAKE_COMPILER_IS_GNUCC OR ENV{SCE_BUILD_ENV} EQUAL "conda_forge"))
   target_link_libraries("${TARGET_NAME}" PRIVATE gomp z)
 else()
   target_link_libraries("${TARGET_NAME}" PRIVATE ZLIB::ZLIB)
   if(OpenMP_CXX_FOUND)
       target_link_libraries("${TARGET_NAME}" PRIVATE OpenMP::OpenMP_CXX)
   endif()
 endif()
-

environment.yml

@@ -16,7 +16,7 @@ dependencies:
   - hdbscan
   - h5py
   - pp-sketchlib >=1.7.0
-  - pybind11 >= 2.12.0
+  - pybind11
   - zlib
   - cmake >= 3.19
   - libgomp