CMake LTO and CUDA arch (#9677)

CMakeLists.txt

@@ -2,15 +2,14 @@ cmake_minimum_required(VERSION 3.18 FATAL_ERROR)
 project(xgboost LANGUAGES CXX C VERSION 2.1.0)
 include(cmake/Utils.cmake)
 list(APPEND CMAKE_MODULE_PATH "${xgboost_SOURCE_DIR}/cmake/modules")
-cmake_policy(SET CMP0022 NEW)
-cmake_policy(SET CMP0079 NEW)
-cmake_policy(SET CMP0076 NEW)
-set(CMAKE_POLICY_DEFAULT_CMP0063 NEW)
-cmake_policy(SET CMP0063 NEW)
 
-if((${CMAKE_VERSION} VERSION_GREATER 3.13) OR (${CMAKE_VERSION} VERSION_EQUAL 3.13))
+# These policies are already set from 3.18 but we still need to set the policy
-  cmake_policy(SET CMP0077 NEW)
+# default variables here for lower minimum versions in the submodules
-endif()
+set(CMAKE_POLICY_DEFAULT_CMP0063 NEW)
+set(CMAKE_POLICY_DEFAULT_CMP0069 NEW)
+set(CMAKE_POLICY_DEFAULT_CMP0076 NEW)
+set(CMAKE_POLICY_DEFAULT_CMP0077 NEW)
+set(CMAKE_POLICY_DEFAULT_CMP0079 NEW)
 
 message(STATUS "CMake version ${CMAKE_VERSION}")
 
@@ -41,6 +40,8 @@ write_version()
 set_default_configuration_release()
 
 #-- Options
+include(CMakeDependentOption)
+
 ## User options
 option(BUILD_C_DOC "Build documentation for C APIs using Doxygen." OFF)
 option(USE_OPENMP "Build with OpenMP support." ON)
@@ -69,8 +70,24 @@ option(USE_CUDA  "Build with GPU acceleration" OFF)
 option(USE_PER_THREAD_DEFAULT_STREAM "Build with per-thread default stream" ON)
 option(USE_NCCL  "Build with NCCL to enable distributed GPU support." OFF)
 option(BUILD_WITH_SHARED_NCCL "Build with shared NCCL library." OFF)
+if(USE_CUDA)
+  if(NOT DEFINED CMAKE_CUDA_ARCHITECTURES AND NOT DEFINED ENV{CUDAARCHS})
     set(GPU_COMPUTE_VER "" CACHE STRING
       "Semicolon separated list of compute versions to be built against, e.g. '35;61'")
+  else()
+    # Clear any cached values from previous runs
+    unset(GPU_COMPUTE_VER)
+    unset(GPU_COMPUTE_VER CACHE)
+  endif()
+endif()
+# CUDA device LTO was introduced in CMake v3.25 and requires host LTO to also be enabled but can still
+# be explicitly disabled allowing for LTO on host only, host and device, or neither, but device-only LTO
+# is not a supproted configuration
+cmake_dependent_option(USE_CUDA_LTO
+  "Enable link-time optimization for CUDA device code"
+  "${CMAKE_INTERPROCEDURAL_OPTIMIZATION}"
+  "CMAKE_VERSION VERSION_GREATER_EQUAL 3.25;USE_CUDA;CMAKE_INTERPROCEDURAL_OPTIMIZATION"
+  OFF)
 ## Sanitizers
 option(USE_SANITIZER "Use santizer flags" OFF)
 option(SANITIZER_PATH "Path to sanitizes.")
@@ -168,15 +185,24 @@ endif()
 if(USE_CUDA)
   set(USE_OPENMP ON CACHE BOOL "CUDA requires OpenMP" FORCE)
   # `export CXX=' is ignored by CMake CUDA.
-  set(CMAKE_CUDA_HOST_COMPILER ${CMAKE_CXX_COMPILER})
+  if(NOT DEFINED CMAKE_CUDA_HOST_COMPILER AND NOT DEFINED ENV{CUDAHOSTCXX})
+    set(CMAKE_CUDA_HOST_COMPILER ${CMAKE_CXX_COMPILER} CACHE FILEPATH
+      "The compiler executable to use when compiling host code for CUDA or HIP language files.")
+    mark_as_advanced(CMAKE_CUDA_HOST_COMPILER)
     message(STATUS "Configured CUDA host compiler: ${CMAKE_CUDA_HOST_COMPILER}")
+  endif()
+
+  if(NOT DEFINED CMAKE_CUDA_RUNTIME_LIBRARY)
+    set(CMAKE_CUDA_RUNTIME_LIBRARY Static)
+  endif()
 
   enable_language(CUDA)
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_LESS 11.0)
     message(FATAL_ERROR "CUDA version must be at least 11.0!")
   endif()
-  set(GEN_CODE "")
+  if(DEFINED GPU_COMPUTE_VER)
-  format_gencode_flags("${GPU_COMPUTE_VER}" GEN_CODE)
+    compute_cmake_cuda_archs("${GPU_COMPUTE_VER}")
+  endif()
   add_subdirectory(${PROJECT_SOURCE_DIR}/gputreeshap)
 
   find_package(CUDAToolkit REQUIRED)

cmake/Utils.cmake

@@ -82,46 +82,35 @@ function(set_default_configuration_release)
     endif()
 endfunction()
 
-# Generate nvcc compiler flags given a list of architectures
+# Generate CMAKE_CUDA_ARCHITECTURES form a list of architectures
 # Also generates PTX for the most recent architecture for forwards compatibility
-function(format_gencode_flags flags out)
+function(compute_cmake_cuda_archs archs)
   if(CMAKE_CUDA_COMPILER_VERSION MATCHES "^([0-9]+\\.[0-9]+)")
     set(CUDA_VERSION "${CMAKE_MATCH_1}")
   endif()
-  # Set up architecture flags
+  list(SORT archs)
-  if(NOT flags)
+  unset(CMAKE_CUDA_ARCHITECTURES CACHE)
+  set(CMAKE_CUDA_ARCHITECTURES ${archs})
+
+  # Set up defaults based on CUDA varsion
+  if(NOT CMAKE_CUDA_ARCHITECTURES)
     if(CUDA_VERSION VERSION_GREATER_EQUAL "11.8")
-      set(flags "50;60;70;80;90")
+      set(CMAKE_CUDA_ARCHITECTURES 50 60 70 80 90)
     elseif(CUDA_VERSION VERSION_GREATER_EQUAL "11.0")
-      set(flags "50;60;70;80")
+      set(CMAKE_CUDA_ARCHITECTURES 50 60 70 80)
     elseif(CUDA_VERSION VERSION_GREATER_EQUAL "10.0")
-      set(flags "35;50;60;70")
+      set(CMAKE_CUDA_ARCHITECTURES 35 50 60 70)
     elseif(CUDA_VERSION VERSION_GREATER_EQUAL "9.0")
-      set(flags "35;50;60;70")
+      set(CMAKE_CUDA_ARCHITECTURES 35 50 60 70)
     else()
-      set(flags "35;50;60")
+      set(CMAKE_CUDA_ARCHITECTURES 35 50 60)
     endif()
   endif()
 
-  if(CMAKE_VERSION VERSION_GREATER_EQUAL "3.18")
+  list(TRANSFORM CMAKE_CUDA_ARCHITECTURES APPEND "-real")
-    cmake_policy(SET CMP0104 NEW)
+  list(TRANSFORM CMAKE_CUDA_ARCHITECTURES REPLACE "([0-9]+)-real" "\\0;\\1-virtual" AT -1)
-    list(GET flags -1 latest_arch)
-    list(TRANSFORM flags APPEND "-real")
-    list(APPEND flags ${latest_arch})
-    set(CMAKE_CUDA_ARCHITECTURES ${flags})
   set(CMAKE_CUDA_ARCHITECTURES "${CMAKE_CUDA_ARCHITECTURES}" PARENT_SCOPE)
   message(STATUS "CMAKE_CUDA_ARCHITECTURES: ${CMAKE_CUDA_ARCHITECTURES}")
-  else()
-    # Generate SASS
-    foreach(ver ${flags})
-      set(${out} "${${out}}--generate-code=arch=compute_${ver},code=sm_${ver};")
-    endforeach()
-    # Generate PTX for last architecture
-    list(GET flags -1 ver)
-    set(${out} "${${out}}--generate-code=arch=compute_${ver},code=compute_${ver};")
-    set(${out} "${${out}}" PARENT_SCOPE)
-    message(STATUS "CUDA GEN_CODE: ${GEN_CODE}")
-  endif()
 endfunction()
 
 # Set CUDA related flags to target.  Must be used after code `format_gencode_flags`.
@@ -129,7 +118,6 @@ function(xgboost_set_cuda_flags target)
   target_compile_options(${target} PRIVATE
     $<$<COMPILE_LANGUAGE:CUDA>:--expt-extended-lambda>
     $<$<COMPILE_LANGUAGE:CUDA>:--expt-relaxed-constexpr>
-    $<$<COMPILE_LANGUAGE:CUDA>:${GEN_CODE}>
     $<$<COMPILE_LANGUAGE:CUDA>:-Xcompiler=${OpenMP_CXX_FLAGS}>
     $<$<COMPILE_LANGUAGE:CUDA>:-Xfatbin=-compress-all>)
 
@@ -138,10 +126,6 @@ function(xgboost_set_cuda_flags target)
             $<$<COMPILE_LANGUAGE:CUDA>:--default-stream per-thread>)
   endif()
 
-  if(CMAKE_VERSION VERSION_GREATER_EQUAL "3.18")
-    set_property(TARGET ${target} PROPERTY CUDA_ARCHITECTURES ${CMAKE_CUDA_ARCHITECTURES})
-  endif()
-
   if(FORCE_COLORED_OUTPUT)
     if(FORCE_COLORED_OUTPUT AND (CMAKE_GENERATOR STREQUAL "Ninja") AND
         ((CMAKE_CXX_COMPILER_ID STREQUAL "GNU") OR
@@ -176,9 +160,15 @@ function(xgboost_set_cuda_flags target)
 
   set_target_properties(${target} PROPERTIES
     CUDA_STANDARD 17
-    CUDA_STANDARD_REQUIRED ON
+    CUDA_STANDARD_REQUIRED ON)
-    CUDA_SEPARABLE_COMPILATION OFF
+  if(USE_CUDA_LTO)
-    CUDA_RUNTIME_LIBRARY Static)
+    set_target_properties(${target} PROPERTIES
+      INTERPROCEDURAL_OPTIMIZATION ON
+      CUDA_SEPARABLE_COMPILATION ON)
+  else()
+    set_target_properties(${target} PROPERTIES
+      CUDA_SEPARABLE_COMPILATION OFF)
+  endif()
 endfunction()
 
 macro(xgboost_link_nccl target)

tests/ci_build/prune_libnccl.sh

@@ -19,10 +19,8 @@ cmake_policy(SET CMP0104 NEW)
 set(CMAKE_CUDA_HOST_COMPILER \${CMAKE_CXX_COMPILER})
 enable_language(CUDA)
 include(../cmake/Utils.cmake)
-set(GEN_CODE "")
+compute_cmake_cuda_archs("")
-format_gencode_flags("" GEN_CODE)
 add_library(test OBJECT test.cu)
-set_property(TARGET test PROPERTY CUDA_ARCHITECTURES \${CMAKE_CUDA_ARCHITECTURES})
 set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
 EOF