build: refactor CUDA macros, adjust CUDA_TARGET_ARCH by CUDA version

Shuffle some options and macros among cmake files. Ensure that most
"find dependencies" logic is in externalpackages.cmake, and move more
options and logic related to "how do we compile to ptx and link
against cuda/optix" into cuda_macros.

The CUDA target architecture had been sm_60 for a long time, but CUDA
Toolkit 13.0+ doesn't target that old an architecture. So make the
architecture dependent on CUDA vrsion -- 13+ gets sm_75 as its default
now (still overrideable by setting build option
CUDA_TARGET_ARCH_NUMBER, but if CUDA <= 12.x is found, it gets a
default target of sm_60 in order to allow a wider range of target
cards.

Some other miscellaneous cleanup here and there.

Signed-off-by: Larry Gritz <lg@larrygritz.com>

Author: lgritz

CMakeLists.txt

@@ -127,9 +127,6 @@ option (OSL_BUILD_PLUGINS "Bool OSL plugins, for example OIIO plugin" ON)
 option (OSL_BUILD_SHADERS "Build shaders" ON)
 option (OSL_USE_OPTIX "Include OptiX support" OFF)
 message(STATUS "OSL_USE_OPTIX: ${OSL_USE_OPTIX}")
-set (OPTIX_EXTRA_LIBS CACHE STRING "Extra lib targets needed for OptiX")
-set (CUDA_EXTRA_LIBS CACHE STRING "Extra lib targets needed for CUDA")
-set (CUDA_TARGET_ARCH "sm_60" CACHE STRING "CUDA GPU architecture (e.g. sm_50)")
 set (OSL_SHADER_INSTALL_DIR "${CMAKE_INSTALL_FULL_DATADIR}/${PROJECT_NAME}/shaders"
      CACHE STRING "Directory where shaders will be installed")
 set (OSL_PTX_INSTALL_DIR "${CMAKE_INSTALL_FULL_DATADIR}/${PROJECT_NAME}/ptx"
@@ -210,8 +207,8 @@ include (pythonutils)
 # Dependency finding utilities and all dependency-related options
 include (externalpackages)
 include (flexbison)
-include (cuda_macros)
 include (llvm_macros)
+include (cuda_macros)
 
 # Include all our testing apparatus and utils, but not if it's a subproject
 if (${PROJECT_NAME}_BUILD_TESTS AND PROJECT_IS_TOP_LEVEL)

src/cmake/cuda_macros.cmake

@@ -2,9 +2,23 @@
 # SPDX-License-Identifier: BSD-3-Clause
 # https://github.com/AcademySoftwareFoundation/OpenShadingLanguage
 
-set ($OSL_EXTRA_NVCC_ARGS "" CACHE STRING "Custom args passed to nvcc when compiling CUDA code")
+# NOTE: externalpackages.cmake is included before this file
+
+if (CUDA_VERSION VERSION_GREATER 13.0)
+    # CUDA 13+ requires sm_75 at least, so this is the minimum
+    set (CUDA_TARGET_ARCH_NUMBER "75" CACHE STRING "CUDA GPU architecture numeric code (e.g. 60)")
+else ()
+    # CUDA <= 12.x supports old GPUs, so conservatively target sm_60 to give
+    # the widest possible range of GPUs that OSL shaders will run on.
+    set (CUDA_TARGET_ARCH_NUMBER "60" CACHE STRING "CUDA GPU architecture numeric code (e.g. 60)")
+endif ()
+set (CUDA_TARGET_ARCH "sm_${CUDA_TARGET_ARCH_NUMBER}" CACHE STRING "CUDA GPU architecture (e.g. 60)")
+set (OSL_EXTRA_NVCC_ARGS "" CACHE STRING "Custom args passed to nvcc when compiling CUDA code")
 set (CUDA_OPT_FLAG_NVCC "-O3" CACHE STRING "The optimization level to use when compiling CUDA/C++ files with nvcc")
 set (CUDA_OPT_FLAG_CLANG "-O3" CACHE STRING "The optimization level to use when compiling CUDA/C++ files with clang")
+set (CUDA_EXTRA_LIBS CACHE STRING "Extra lib targets needed for CUDA")
+set (OPTIX_EXTRA_LIBS CACHE STRING "Extra lib targets needed for OptiX")
+set (CUDA_LIB_FLAGS "--cuda-path=${CUDA_TOOLKIT_ROOT_DIR}")
 
 # FTZ is enabled by default since it offers the best performance,
 # but it may be disabled to debug numerical issues, to better match
@@ -14,6 +28,15 @@ if (CUDA_NO_FTZ)
     add_compile_definitions (DOSL_CUDA_NO_FTZ=1)
 endif ()
 
+
+function (osl_optix_target TARGET)
+    target_include_directories (${TARGET} BEFORE PRIVATE ${CUDA_INCLUDES} ${OPTIX_INCLUDES})
+    ## XXX: Should -DPTX_PATH point to (or include) CMAKE_CURRENT_BINARY_DIR so tests can run before installation ?
+    target_compile_definitions (${TARGET} PRIVATE PTX_PATH="${OSL_PTX_INSTALL_DIR}")
+    target_link_libraries (${TARGET} PRIVATE ${CUDA_LIBRARIES} ${CUDA_EXTRA_LIBS} ${OPTIX_LIBRARIES} ${OPTIX_EXTRA_LIBS})
+endfunction()
+
+
 # Compile a CUDA file to PTX using NVCC
 function ( NVCC_COMPILE cuda_src extra_headers ptx_generated extra_nvcc_args extra_libs)
     get_filename_component ( cuda_src_we ${cuda_src} NAME_WE )
@@ -167,7 +190,7 @@ function ( MAKE_CUDA_BITCODE src suffix generated_bc extra_clang_args )
             ${ALL_OpenImageIO_INCLUDES}
             ${ALL_IMATH_INCLUDES}
             ${ALL_OPENEXR_INCLUDES}
-            ${LLVM_COMPILE_FLAGS} ${CUDA_LIB_FLAGS} ${CLANG_MSVC_FIX} ${CUDA_TEXREF_FIX}
+    ${LLVM_COMPILE_FLAGS} ${CUDA_LIB_FLAGS} ${CLANG_MSVC_FIX} ${CUDA_TEXREF_FIX}
             -D__CUDACC__ -DOSL_COMPILING_TO_BITCODE=1 -DNDEBUG -DOIIO_NO_SSE -D__CUDADEVRT_INTERNAL__
             --language=cuda --cuda-device-only --cuda-gpu-arch=${CUDA_TARGET_ARCH}
             -Wno-deprecated-register -Wno-format-security

src/cmake/externalpackages.cmake

@@ -118,7 +118,7 @@ if (OSL_USE_OPTIX)
             message (STATUS "CUDA_TOOLKIT_ROOT_DIR = ${CUDA_TOOLKIT_ROOT_DIR}")
         endif ()
 
-        checked_find_package (CUDA REQUIRED
+        checked_find_package(CUDA REQUIRED
                              VERSION_MIN 9.0
                              RECOMMEND_MIN 11.0
                              RECOMMEND_MIN_REASON
@@ -142,17 +142,16 @@ if (OSL_USE_OPTIX)
                 "first shipped in LLVM 22.1.2. Either upgrade LLVM or downgrade CUDA.${ColorReset}")
         endif ()
 
-        set (CUDA_LIB_FLAGS "--cuda-path=${CUDA_TOOLKIT_ROOT_DIR}")
-
-        # If the user wants, try to use static libs here to putting static lib
-        # suffixes earlier in the suffix list. Don't forget to restore after
-        # so that this only applies to these library searches right here.
-        set (save_lib_path ${CMAKE_FIND_LIBRARY_SUFFIXES})
         if (CUDA_PREFER_STATIC_LIBS)
+            # If the user wants, try to use static libs here by putting static
+            # lib suffixes earlier in the suffix list.
+            set (save_lib_path ${CMAKE_FIND_LIBRARY_SUFFIXES})
             set (CMAKE_FIND_LIBRARY_SUFFIXES .lib .a ${CMAKE_FIND_LIBRARY_SUFFIXES})
             find_library(cudart_lib REQUIRED
                          NAMES cudart_static cudart
                          PATHS "${CUDA_TOOLKIT_ROOT_DIR}/lib64" "${CUDA_TOOLKIT_ROOT_DIR}/x64" "${CUDA_TOOLKIT_ROOT_DIR}/lib/x64")
+            set (CMAKE_FIND_LIBRARY_SUFFIXES ${save_lib_path})
+            unset (save_lib_path)
         else ()
             find_library(cudart_lib REQUIRED
                          NAMES cudart
@@ -161,8 +160,6 @@ if (OSL_USE_OPTIX)
         # Is it really a good idea to completely reset CUDA_LIBRARIES here?
         set(CUDA_LIBRARIES ${cudart_lib})
         set(CUDA_EXTRA_LIBS ${CUDA_EXTRA_LIBS} dl rt)
-        set (CMAKE_FIND_LIBRARY_SUFFIXES ${save_lib_path})
-        unset (save_lib_path)
     endif()
 
     # OptiX setup
@@ -174,17 +171,6 @@ if (OSL_USE_OPTIX)
             message (FATAL_ERROR "Enabling OptiX requires USE_LLVM_BITCODE=1 and USE_FAST_MATH=1")
         endif ()
     endif ()
-
-    function (osl_optix_target TARGET)
-        target_include_directories (${TARGET} BEFORE PRIVATE ${CUDA_INCLUDES} ${OPTIX_INCLUDES})
-        ## XXX: Should -DPTX_PATH point to (or include) CMAKE_CURRENT_BINARY_DIR so tests can run before installation ?
-        target_compile_definitions (${TARGET} PRIVATE PTX_PATH="${OSL_PTX_INSTALL_DIR}")
-        target_link_libraries (${TARGET} PRIVATE ${CUDA_LIBRARIES} ${CUDA_EXTRA_LIBS} ${OPTIX_LIBRARIES} ${OPTIX_EXTRA_LIBS})
-    endfunction()
-else ()
-    message(STATUS "CUDA/OptiX support disabled")
-    function (osl_optix_target TARGET)
-    endfunction()
 endif ()
 
 

src/testrender/CMakeLists.txt

@@ -85,6 +85,8 @@ target_link_libraries (testrender
         pugixml::pugixml
         Threads::Threads)
 
-osl_optix_target (testrender)
+if (OSL_USE_OPTIX)
+    osl_optix_target (testrender)
+endif ()
 
 install ( TARGETS testrender RUNTIME DESTINATION ${CMAKE_INSTALL_BINDIR} )

src/testshade/CMakeLists.txt

@@ -88,8 +88,9 @@ endif ()
 
 install (TARGETS testshade RUNTIME DESTINATION ${CMAKE_INSTALL_BINDIR} )
 
-osl_optix_target(testshade)
-
+if (OSL_USE_OPTIX)
+    osl_optix_target(testshade)
+endif ()
 
 # As explained in PR #39, problems were reported with calling OSL from a
 # Houdini plugin.  So this section sets up a version of testshade as a DSO, to

testsuite/example-cuda/CMakeLists.txt

@@ -2,7 +2,7 @@
 # SPDX-License-Identifier: BSD-3-Clause
 # https://github.com/AcademySoftwareFoundation/OpenShadingLanguage
 
-cmake_minimum_required(VERSION 3.15)
+cmake_minimum_required(VERSION 3.19)
 project(examplecuda LANGUAGES CXX)
 
 if (NOT CMAKE_BUILD_TYPE)
@@ -20,9 +20,9 @@ include(checked_find_package)
 
 find_package(OSL REQUIRED)
 find_package(CUDA REQUIRED)
-checked_find_package(LLVM 7.0 REQUIRED)
+checked_find_package(LLVM 14.0 REQUIRED)
 checked_find_package(Imath 3.1 REQUIRED)
-checked_find_package(OpenImageIO 2.4 REQUIRED)
+checked_find_package(OpenImageIO 3.0 REQUIRED)
 checked_find_package(OptiX REQUIRED)
 
 
@@ -35,8 +35,7 @@ set (CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin)
 find_library(CUDA_nvrtc_LIBRARY nvrtc HINTS ${CUDA_TOOLKIT_ROOT_DIR} PATH_SUFFIXES lib lib64)
 find_library(CUDA_cuda_LIBRARY cuda HINTS ${CUDA_TOOLKIT_ROOT_DIR} PATH_SUFFIXES lib/stubs lib64/stubs)
 
-# TODO: move to sm_60?
-set(CUDA_TARGET_ARCH sm_35)
+set(CUDA_TARGET_ARCH sm_75)
 
 set (CMAKE_CXX_STANDARD 17 CACHE STRING
      "C++ standard to build with (17, 20, etc.)")