Merge branch 'development' into feature/glare_frontend

CMakeLists.txt

@@ -110,14 +110,6 @@ find_package(TBB REQUIRED)
 # ZLib is used in the generator
 find_package(ZLIB REQUIRED)
 
-# Locate stuff for windows
-# On Linux this is properly propagated
-# So it needs a fixing in AnyDSL
-if(WIN32)
-	find_package(CUDAToolkit)
-	find_package(OpenCL)
-endif()
-
 # Locate AnyDSL runtime (general purpose)
 find_package(AnyDSL_runtime REQUIRED)
 

cmake/Install.cmake

@@ -53,7 +53,6 @@ if(IG_INSTALL_RUNTIME_DEPENDENCIES)
     else()
         set(exclude_regexes 
             "libX.*" "libxkb.*" "libwayland.*" "libvorbis.*" "libexpat.*" "libFLAC.*" "libpulse.*" "libasound.*" # Skip media related libraries
-            "libcuda.*" "libOpenCL.*" "libnvvm.*" "libnvrtc.*" # Ignore CUDA related stuff, as the system should be providing it!
             )
     endif()
 
@@ -62,6 +61,8 @@ if(IG_INSTALL_RUNTIME_DEPENDENCIES)
         list(APPEND search_dirs "${Python_RUNTIME_LIBRARY_DIRS}")
         list(APPEND search_dirs "${Python_RUNTIME_SABI_LIBRARY_DIRS}")
     endif()
+
+    find_package(CUDAToolkit)
     if(CUDAToolkit_FOUND)
         list(APPEND search_dirs "${CUDAToolkit_BIN_DIR}")
     endif()

src/device/CMakeLists.txt

@@ -9,27 +9,39 @@ list(REMOVE_DUPLICATES _DISABLED_DEVICES)
 list(TRANSFORM _DISABLED_DEVICES TOUPPER)
 
 ###########################################################
-option(IG_ASSUME_NVIDIA "Assume CUDA libraries are available" OFF)
-mark_as_advanced(IG_ASSUME_NVIDIA)
-if(IG_ASSUME_NVIDIA OR AnyDSL_runtime_HAS_CUDA_SUPPORT)
-    set(_HAS_CUDA ON)
-else()
-    find_library(CUDA_LIBRARY cudart HINTS ${CUDA_DIR} PATHS /usr/local/cuda
-    ENV CUDA_PATH
-    ENV CUDA_LIB_PATH
-    PATH_SUFFIXES "lib/Win32" "lib" "libWin32")
-    if(EXISTS "${CUDA_LIBRARY}")
-        set(_HAS_CUDA ON)
+find_package(CUDAToolkit)
+
+if(CUDAToolkit_FOUND)
+    if(CUDAToolkit_LIBRARY_ROOT)
+        set(_CUDA_ROOT ${CUDAToolkit_LIBRARY_ROOT})
+    else()
+        set(_CUDA_ROOT "${CUDAToolkit_LIBRARY_DIR}/..")
+    endif()
+
+    find_file(_LIBDEVICE_LIB
+        NAMES libdevice.10.bc
+        HINTS ${_CUDA_ROOT}
+        PATH_SUFFIXES nvvm/libdevice)
+
+    if(_LIBDEVICE_LIB)
+        message(STATUS "Redistributing ${_LIBDEVICE_LIB}")
+        install(FILES ${_LIBDEVICE_LIB} ${_IG_RUNTIME_SET} DESTINATION bin COMPONENT runtime)
+    else()
+        message(WARNING "Could not find 'libdevice.10.bc'. Redistribution might fail")
     endif()
+
+    set(_HAS_CUDA ON)
 endif()
 
 ###########################################################
 option(IG_ASSUME_AMD "Assume AMD rocm libraries are available" OFF)
 mark_as_advanced(IG_ASSUME_AMD)
+
 if(IG_ASSUME_AMD OR AnyDSL_runtime_HAS_HSA_SUPPORT)
     set(_HAS_HSA ON)
 else()
     find_library(HSA_LIBRARY hsa-runtime64 HINTS ${HSA_DIR} PATHS /opt/rocm PATH_SUFFIXES lib)
+
     if(EXISTS "${HSA_LIBRARY}")
         set(_HAS_HSA ON)
     endif()
@@ -39,42 +51,47 @@ endif()
 include(CMakeDependentOption)
 option(IG_WITH_DEVICE_CPU "Enable CPU device" ON)
 option(IG_WITH_DEVICE_GPU "Enable GPU device" ON)
+
 # cmake_dependent_option(IG_WITH_DEVICE_CPU_SIMD    "Enable CPU vectorized device"  ON "IG_WITH_DEVICE_CPU" OFF)
-cmake_dependent_option(IG_WITH_DEVICE_GPU_NVIDIA  "Enable NVidia GPU device"      ON "IG_WITH_DEVICE_GPU;_HAS_CUDA" OFF)
-cmake_dependent_option(IG_WITH_DEVICE_GPU_AMD_HSA "Enable AMD GPU device via HSA" ON "IG_WITH_DEVICE_GPU;_HAS_HSA"  OFF)
+cmake_dependent_option(IG_WITH_DEVICE_GPU_NVIDIA "Enable NVidia GPU device" ON "IG_WITH_DEVICE_GPU;_HAS_CUDA" OFF)
+cmake_dependent_option(IG_WITH_DEVICE_GPU_AMD_HSA "Enable AMD GPU device via HSA" ON "IG_WITH_DEVICE_GPU;_HAS_HSA" OFF)
 
 # ##########################################################
 set(VARIANTS)
-if (IG_WITH_DEVICE_CPU)
+
+if(IG_WITH_DEVICE_CPU)
     list(APPEND VARIANTS cpu)
 endif()
-if (IG_WITH_DEVICE_GPU_NVIDIA)
+
+if(IG_WITH_DEVICE_GPU_NVIDIA)
     list(APPEND VARIANTS cuda)
 endif()
-if (IG_WITH_DEVICE_GPU_AMD_HSA)
+
+if(IG_WITH_DEVICE_GPU_AMD_HSA)
     list(APPEND VARIANTS hsa)
 endif()
 
-# ##########################################################
-set(_targets )
+###########################################################
+set(_targets)
+
 foreach(var ${VARIANTS})
     string(TOUPPER "${var}" VAR_UPPER)
     set(_target_name ig_device_${var})
     message(STATUS "Adding device for ${var}")
 
-    set(AnyDSLRuntimeDevice_LIBRARIES )
-    set(AnyDSLRuntimeDevice_INCLUDE_DIRS )
+    set(AnyDSLRuntimeDevice_LIBRARIES)
+    set(AnyDSLRuntimeDevice_INCLUDE_DIRS)
 
     find_package(AnyDSLRuntimeDevice COMPONENTS ${var})
-    
-    if (NOT AnyDSLRuntimeDevice_${var}_FOUND)
+
+    if(NOT AnyDSLRuntimeDevice_${var}_FOUND)
         message(WARNING "No specific AnyDSL runtime for device ${var} found. Using default one")
         set(AnyDSLRuntimeDevice_LIBRARIES ${AnyDSL_runtime_LIBRARIES} ${AnyDSL_runtime_ARTIC_JIT_LIBRARIES})
         set(AnyDSLRuntimeDevice_INCLUDE_DIRS ${AnyDSL_runtime_INCLUDE_DIRS})
     endif()
 
     add_library(${_target_name} MODULE ${SRC})
-    target_link_libraries(${_target_name} PUBLIC ig_lib_runtime PRIVATE ${AnyDSLRuntimeDevice_LIBRARIES} TBB::tbb TBB::tbbmalloc )
+    target_link_libraries(${_target_name} PUBLIC ig_lib_runtime PRIVATE ${AnyDSLRuntimeDevice_LIBRARIES} TBB::tbb TBB::tbbmalloc)
     target_include_directories(${_target_name} SYSTEM PRIVATE ${AnyDSLRuntimeDevice_INCLUDE_DIRS})
     target_compile_definitions(${_target_name} PRIVATE $<BUILD_INTERFACE:"IG_BUILD_DEVICE_${VAR_UPPER}">)
     set_target_properties(${_target_name} PROPERTIES PREFIX "")
@@ -89,6 +106,7 @@ foreach(var ${VARIANTS})
     # however devices are dynamically defined and ment to be rebuild anyway.
     if(NOT IG_BUILD_DEVICE_PARALLEL)
         list(LENGTH _targets _length)
+
         if(_length GREATER 0)
             list(GET _targets -1 _prev)
             add_dependencies(${_target_name} ${_prev})

src/runtime/RuntimeInfo.cpp

@@ -76,6 +76,21 @@ Path RuntimeInfo::igcPath()
         return {};
 }
 
+Path RuntimeInfo::libdevicePath()
+{
+    // Should reside in the same directory!
+    const Path exePath = executablePath();
+    if (exePath.empty())
+        return {};
+
+    const Path libPath = exePath.parent_path() / "libdevice.10.bc";
+
+    if (std::filesystem::exists(libPath))
+        return libPath;
+    else
+        return {};
+}
+
 Path RuntimeInfo::modulePath(void* func)
 {
     if (func == nullptr)

src/runtime/RuntimeInfo.h

@@ -7,6 +7,7 @@ class IG_LIB RuntimeInfo {
 public:
     [[nodiscard]] static Path executablePath();                 // Path to the current executable
     [[nodiscard]] static Path igcPath();                        // Path to the ignis jit compiler 'igc'
+    [[nodiscard]] static Path libdevicePath();                  // Path to 'libdevice.10.bc' needed by the cuda device
     [[nodiscard]] static Path modulePath(void* func = nullptr); // Path to the module/dll for the given function
     [[nodiscard]] static Path cacheDirectory();
     [[nodiscard]] static size_t sizeOfDirectory(const Path& dir);

src/runtime/device/DeviceManager.cpp

@@ -4,12 +4,16 @@
 #include "RuntimeInfo.h"
 #include "StringUtils.h"
 
+#include <cstdlib>
+
 #include <unordered_set>
 
 #define _IG_DEVICE_ENV_PATH_NAME "IG_DEVICE_PATH"
 #define _IG_DEVICE_ENV_SKIP_SYSTEM_PATH "IG_DEVICE_SKIP_SYSTEM_PATH"
 #define _IG_DEVICE_LIB_PREFIX "ig_device_"
 
+#define ANYDSL_CUDA_LIBDEVICE_PATH_ENV "ANYDSL_CUDA_LIBDEVICE_PATH"
+
 namespace IG {
 
 using GetInterfaceFunction = const IDeviceInterface* (*)();
@@ -64,6 +68,16 @@ bool DeviceManager::init(const Path& dir, bool ignoreEnv, bool force)
     if (!force && !mAvailableDevices.empty())
         return true;
 
+    const auto libdevicePath = RuntimeInfo::libdevicePath();
+    if (!libdevicePath.empty()) {
+        IG_LOG(L_DEBUG) << "Setting libdevice path " << libdevicePath << " to environment variable " << ANYDSL_CUDA_LIBDEVICE_PATH_ENV << std::endl;
+#ifdef IG_OS_WINDOWS
+        _putenv_s(ANYDSL_CUDA_LIBDEVICE_PATH_ENV, libdevicePath.generic_string().c_str());
+#else
+        setenv(ANYDSL_CUDA_LIBDEVICE_PATH_ENV, libdevicePath.generic_string().c_str(), 1);
+#endif
+    }
+
     path_set paths;
 
     bool skipSystem = false; // Skip the system search path