Merge pull request #156 from alicevision/dev/cmake-lang-cuda

Update to Continuous integration and CMake/CUDA

Author: fabiencastan

.github/workflows/continuous-integration.yml

@@ -21,11 +21,11 @@ jobs:
     runs-on: ubuntu-latest
     strategy:
       matrix:
-        container: ["alicevision/popsift-deps:cuda10.2-ubuntu18.04", "alicevision/popsift-deps:cuda11.8.0-ubuntu20.04", "alicevision/popsift-deps:cuda12.1.0-ubuntu22.04"]
+        container: ["alicevision/popsift-deps:cuda11.8.0-ubuntu20.04", "alicevision/popsift-deps:cuda12.1.0-ubuntu22.04"]
         build_tpe: ["Release", "Debug"]
         exclude:
         # excludes debug on this one as it has a segmentation fault during the compilation (!)
-        - container: "alicevision/popsift-deps:cuda12.1.0-ubuntu22.04"
+        - container: "alicevision/popsift-deps:cuda11.8.0-ubuntu20.04"
           build_tpe: "Debug"
 
     container:

.travis.yml

@@ -1,4 +1,4 @@
-dist: xenial
+dist: jammy
 
 language: cpp
 compiler: gcc
@@ -13,10 +13,8 @@ addons:
 
 env:
   matrix:
-    - CUDA_VERSION_MAJOR="8" CUDA_VERSION_MINOR="0" CUDA_PKG_LONGVERSION="${CUDA_VERSION_MAJOR}.${CUDA_VERSION_MINOR}.61-1" CUDA_PKG_VERSION="${CUDA_VERSION_MAJOR}-${CUDA_VERSION_MINOR}"
-    - CUDA_VERSION_MAJOR="9" CUDA_VERSION_MINOR="2" CUDA_PKG_LONGVERSION="${CUDA_VERSION_MAJOR}.${CUDA_VERSION_MINOR}.148-1" CUDA_PKG_VERSION="${CUDA_VERSION_MAJOR}-${CUDA_VERSION_MINOR}"
-    - CUDA_VERSION_MAJOR="10" CUDA_VERSION_MINOR="2" CUDA_PKG_LONGVERSION="${CUDA_VERSION_MAJOR}.${CUDA_VERSION_MINOR}.89-1" CUDA_PKG_VERSION="${CUDA_VERSION_MAJOR}-${CUDA_VERSION_MINOR}"
-    - CUDA_VERSION_MAJOR="11" CUDA_VERSION_MINOR="0" CUDA_PKG_LONGVERSION="${CUDA_VERSION_MAJOR}.${CUDA_VERSION_MINOR}.2-1" CUDA_PKG_VERSION="${CUDA_VERSION_MAJOR}-${CUDA_VERSION_MINOR}"
+    - CUDA_VERSION_MAJOR="11" CUDA_VERSION_MINOR="8" CUDA_PKG_LONGVERSION="${CUDA_VERSION_MAJOR}.${CUDA_VERSION_MINOR}.0-1" CUDA_PKG_VERSION="${CUDA_VERSION_MAJOR}-${CUDA_VERSION_MINOR}"
+    - CUDA_VERSION_MAJOR="12" CUDA_VERSION_MINOR="5" CUDA_PKG_LONGVERSION="${CUDA_VERSION_MAJOR}.${CUDA_VERSION_MINOR}.1-1" CUDA_PKG_VERSION="${CUDA_VERSION_MAJOR}-${CUDA_VERSION_MINOR}"
 
 
   global:
@@ -34,7 +32,7 @@ env:
     - POPSIFT_APP_INSTALL_RELEASE=${POPSIFT_APP_BUILD_RELEASE}/install
     - POPSIFT_APP_INSTALL_DEBUG=${POPSIFT_APP_BUILD_DEBUG}/install
     # CMAKE
-    - CMAKE_URL="https://cmake.org/files/v3.13/cmake-3.13.5-Linux-x86_64.tar.gz"
+    - CMAKE_URL="https://github.com/Kitware/CMake/releases/download/v3.30.1/cmake-3.30.1-linux-x86_64.tar.gz"
     - CMAKE_ROOT=${TRAVIS_BUILD_DIR}/cmake
     - CMAKE_SOURCE=${CMAKE_ROOT}/source
     - CMAKE_INSTALL=${CMAKE_ROOT}/install
@@ -57,30 +55,12 @@ before_install:
     fi
 
 install:
-  - UBUNTU_VERSION=ubuntu1604
-  - >
-    if [ ${CUDA_VERSION_MAJOR} -lt 11 ]; then
-        CUDA_REPO_PKG=cuda-repo-${UBUNTU_VERSION}_${CUDA_PKG_LONGVERSION}_amd64.deb
-        wget http://developer.download.nvidia.com/compute/cuda/repos/${UBUNTU_VERSION}/x86_64/$CUDA_REPO_PKG
-        travis_retry sudo apt-key adv --fetch-keys http://developer.download.nvidia.com/compute/cuda/repos/${UBUNTU_VERSION}/x86_64/7fa2af80.pub
-        sudo dpkg -i $CUDA_REPO_PKG
-        rm ${CUDA_REPO_PKG}
-        travis_retry sudo apt-get -y update
-        # cuda > 10.0 changed cublas naming
-        if [ ${CUDA_VERSION_MAJOR} -lt 10 ]; then
-              CUBLAS_PKG=cuda-cublas-dev-$CUDA_PKG_VERSION
-        else
-              CUBLAS_PKG=libcublas-dev
-        fi
-        travis_retry sudo apt-get install -y --no-install-recommends --allow-unauthenticated cuda-core-$CUDA_PKG_VERSION  cuda-cudart-dev-$CUDA_PKG_VERSION  ${CUBLAS_PKG} cuda-curand-dev-$CUDA_PKG_VERSION
-        sudo ln -s /usr/local/cuda-${CUDA_VERSION_MAJOR}.${CUDA_VERSION_MINOR} /usr/local/cuda
-    else
-        wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu1604/x86_64/cuda-ubuntu1604.pin
-        travis_retry sudo mv cuda-ubuntu1604.pin /etc/apt/preferences.d/cuda-repository-pin-600
-        travis_retry sudo apt-key adv --fetch-keys http://developer.download.nvidia.com/compute/cuda/repos/ubuntu1604/x86_64/7fa2af80.pub
-        travis_retry sudo add-apt-repository "deb http://developer.download.nvidia.com/compute/cuda/repos/ubuntu1604/x86_64/ /"
-        sudo apt-get update && sudo apt-get -y install cuda
-    fi
+  - UBUNTU_VERSION=ubuntu2204
+  - wget https://developer.download.nvidia.com/compute/cuda/repos/${UBUNTU_VERSION}/x86_64/cuda-${UBUNTU_VERSION}.pin
+  - travis_retry sudo mv cuda-${UBUNTU_VERSION}.pin /etc/apt/preferences.d/cuda-repository-pin-600
+  - travis_retry sudo apt-key adv --fetch-keys http://developer.download.nvidia.com/compute/cuda/repos/${UBUNTU_VERSION}/x86_64/3bf863cc.pub
+  - travis_retry sudo add-apt-repository "deb http://developer.download.nvidia.com/compute/cuda/repos/${UBUNTU_VERSION}/x86_64/ /"
+  - sudo apt-get update && sudo apt-get -y install cuda
 #  - CUDA_REPO_PKG=cuda-repo-${UBUNTU_VERSION}_${CUDA_PKG_LONGVERSION}_amd64.deb
 #  - wget http://developer.download.nvidia.com/compute/cuda/repos/${UBUNTU_VERSION}/x86_64/$CUDA_REPO_PKG
 
@@ -131,3 +111,4 @@ cache:
   apt: true
   directories:
     - ${CMAKE_INSTALL}
+

CHANGES.md

@@ -17,6 +17,10 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [1.0.0] - YYYY-MM-DD
 
+## 2024
+
+- CMake: CUDA as first-order language, different CC selection
+
 ### Added
 - Improved checks for CUDA textures [PR](https://github.com/alicevision/popsift/pull/89)
 - CMake: Improved support for all Cuda CC [PR](https://github.com/alicevision/popsift/pull/75)

CMakeLists.txt

@@ -1,7 +1,16 @@
-# CMake below 3.4 does not work with CUDA separable compilation at all
-cmake_minimum_required(VERSION 3.12)
+# First-order language CUDA requires at least CMake 3.18
+cmake_minimum_required(VERSION 3.24)
 
-project(PopSift VERSION 1.0.0 LANGUAGES CXX)
+project(PopSift VERSION 1.0.0 LANGUAGES CXX CUDA)
+
+# Policy to support CUDA as a first-order language for CMake.
+# Since CMake 3.18. See https://cmake.org/cmake/help/latest/policy/CMP0104.html
+cmake_policy(SET CMP0104 NEW)
+
+set(CMAKE_CUDA_ARCHITECTURES "all-major"
+    CACHE
+    STRING "Which CUDA CCs to support: native, all, all-major or an explicit list delimited by semicolons"
+    FORCE)
 
 # Set build path as a folder named as the platform (linux, windows, darwin...) plus the processor type
 set(CMAKE_RUNTIME_OUTPUT_DIRECTORY "${PROJECT_BINARY_DIR}/${CMAKE_SYSTEM_NAME}-${CMAKE_SYSTEM_PROCESSOR}")
@@ -16,7 +25,6 @@ option(PopSift_USE_GRID_FILTER "Switch off grid filtering to massively reduce co
 option(PopSift_USE_NORMF "The __normf function computes Euclidean distance on large arrays. Fast but stability is uncertain." OFF)
 option(PopSift_NVCC_WARNINGS "Switch on several additional warning for CUDA nvcc" OFF)
 option(PopSift_USE_TEST_CMD "Add testing step for functional verification" OFF)
-option(PopSift_NO_DEPRECATED_CUDA_SM_WARNINGS "Suppress warnings about soon to be deprecated cuda SM" ON)
 option(BUILD_SHARED_LIBS "Build shared libraries" ON)
 
 if(PopSift_USE_POSITION_INDEPENDENT_CODE AND NOT MSVC)
@@ -55,17 +63,15 @@ include(GNUInstallDirs)
 
 if(BUILD_SHARED_LIBS)
   message(STATUS "BUILD_SHARED_LIBS ON")
-  # Need to declare CUDA_USE_STATIC_CUDA_RUNTIME as an option to ensure that it is not overwritten in FindCUDA.
-  option(CUDA_USE_STATIC_CUDA_RUNTIME "Use the static version of the CUDA runtime library if available" OFF)
-  set(CUDA_USE_STATIC_CUDA_RUNTIME OFF)
-  # Workaround to force deactivation of cuda static runtime for cmake < 3.10
-  set(CUDA_cudart_static_LIBRARY 0)
+
   # Auto-build dll exports on Windows
   set(CMAKE_WINDOWS_EXPORT_ALL_SYMBOLS ON)
+
+  set(CMAKE_CUDA_RUNTIME_LIBRARY Shared)
 else()
   message(STATUS "BUILD_SHARED_LIBS OFF")
-  option(CUDA_USE_STATIC_CUDA_RUNTIME "Use the static version of the CUDA runtime library if available" ON)
-  set(CUDA_USE_STATIC_CUDA_RUNTIME ON)
+
+  set(CMAKE_CUDA_RUNTIME_LIBRARY Static)
 endif()
 
 # Require threads because of std::thread.
@@ -74,26 +80,15 @@ find_package(Threads REQUIRED)
 ###################
 #  CUDA
 ###################
-find_package(CUDA 7.0 REQUIRED)
+include(CheckLanguage)
+check_language(CUDA)
 
-if(NOT CUDA_FOUND)
-  message(FATAL_ERROR "Could not find CUDA >= 7.0")
-endif()
+# Use this if necessary: "cmake -DCUDAToolkit_ROOT=/some/path"
+# target_link_libraries(binary_linking_to_cudart PRIVATE CUDA::cudart)
+find_package(CUDAToolkit)
 
-message(STATUS "CUDA Version is ${CUDA_VERSION}")
-
-include(ChooseCudaCC)
-if(NOT DEFINED PopSift_CUDA_CC_LIST)
-  chooseCudaCC(PopSift_CUDA_CC_LIST_BASIC
-               PopSift_CUDA_GENCODE_FLAGS
-               MIN_CC 30
-               MIN_CUDA_VERSION 7.0)
-  set(PopSift_CUDA_CC_LIST ${PopSift_CUDA_CC_LIST_BASIC} CACHE STRING "CUDA CC versions to compile")
-else()
-  getFlagsForCudaCCList(PopSift_CUDA_CC_LIST
-                        PopSift_CUDA_GENCODE_FLAGS)
-endif()
-list(APPEND CUDA_NVCC_FLAGS "${PopSift_CUDA_GENCODE_FLAGS}")
+message(STATUS "CUDA Version is ${CUDAToolkit_VERSION}")
+set(CUDA_VERSION ${CUDAToolkit_VERSION})
 
 if(PopSift_USE_NVTX_PROFILING)
   message(STATUS "PROFILING CPU CODE: NVTX is in use")
@@ -104,46 +99,8 @@ if(PopSift_ERRCHK_AFTER_KERNEL)
   list(APPEND CUDA_NVCC_FLAGS "-DERRCHK_AFTER_KERNEL")
 endif()
 
-set(CUDA_SEPARABLE_COMPILATION ON)
-
-if(PopSift_NO_DEPRECATED_CUDA_SM_WARNINGS)
-  list(APPEND CUDA_NVCC_FLAGS "-Wno-deprecated-gpu-targets")
-endif()
-
-if(UNIX AND NOT APPLE)
-  list(APPEND CUDA_NVCC_FLAGS         "-Xcompiler;-rdynamic")
-  # set(CUDA_NVCC_FLAGS       "${CUDA_NVCC_FLAGS};-Xptxas;-v")
-  # set(CUDA_NVCC_FLAGS       "${CUDA_NVCC_FLAGS};-Xptxas;-warn-double-usage")
-  list(APPEND CUDA_NVCC_FLAGS_DEBUG   "--keep")
-  list(APPEND CUDA_NVCC_FLAGS_DEBUG   "--source-in-ptx")
-endif()
-
-# The following if should not be necessary, but apparently there is a bug in FindCUDA.cmake that
-# generate an empty string in the nvcc command line causing the compilation to fail.
-# see https://gitlab.kitware.com/cmake/cmake/issues/16411
-if(CMAKE_BUILD_TYPE STREQUAL "Debug")
-  message(STATUS "Building in debug mode")
-  list(APPEND CUDA_NVCC_FLAGS_DEBUG   "-G")
-endif()
-list(APPEND CUDA_NVCC_FLAGS_RELEASE "-O3")
-
-if(PopSift_USE_POSITION_INDEPENDENT_CODE AND NOT MSVC)
-  list(APPEND CUDA_NVCC_FLAGS "-Xcompiler;-fPIC")
-endif()
-
-# this is to ensure that on MSVC the flags for the linker are properly propagate even to the intermediate
-# linking step. This seems not the case e.g. on vcpkg using ninja build.
-if(MSVC)
-  if(BUILD_SHARED_LIBS)
-    set(PopSift_MVSC_LINKER "/MD")
-  else()
-    set(PopSift_MVSC_LINKER "/MT")
-  endif()
-  if(CMAKE_BUILD_TYPE STREQUAL "Debug")
-      set(PopSift_MVSC_LINKER "${PopSift_MVSC_LINKER}d")
-  endif()
-  list(APPEND CUDA_NVCC_FLAGS   -Xcompiler ${PopSift_MVSC_LINKER})
-endif()
+# This may not be required any more.
+set(CMAKE_CUDA_SEPARABLE_COMPILATION ON)
 
 # default stream per-thread implies that each host thread has one non-synchronizing 0-stream
 # currently, the code requires legacy mode
@@ -159,7 +116,7 @@ if(CUDA_VERSION VERSION_GREATER_EQUAL "7.5")
   endif()
 endif()
 
-set(PopSift_CXX_STANDARD 14) # Thrust/CUB requires C++14 starting with CUDA SDK 11
+set(PopSift_CXX_STANDARD 17) # Thrust/CUB requires C++14 starting with CUDA SDK 11
 if(CUDA_VERSION_MAJOR LESS_EQUAL 8)
   set(PopSift_CXX_STANDARD 11)
 endif()
@@ -181,9 +138,9 @@ else()
 endif()
 
 if(CUDA_VERSION VERSION_GREATER_EQUAL "9.0")
-  set(HAVE_SHFL_DOWN_SYNC   1)
+  set(PopSift_HAVE_SHFL_DOWN_SYNC   1)
 else()
-  set(HAVE_SHFL_DOWN_SYNC   0)
+  set(PopSift_HAVE_SHFL_DOWN_SYNC   0)
 endif()
 
 if(NOT PopSift_USE_GRID_FILTER)
@@ -193,12 +150,8 @@ else()
   set(DISABLE_GRID_FILTER   0)
 endif()
 
-# library required for CUDA dynamic parallelism, forgotten by CMake 3.4
-cuda_find_library_local_first(CUDA_CUDADEVRT_LIBRARY cudadevrt "\"cudadevrt\" library")
-
 if(PopSift_USE_NVTX_PROFILING)
   # library required for NVTX profiling of the CPU
-  cuda_find_library_local_first(CUDA_NVTX_LIBRARY nvToolsExt "NVTX library")
   set(PopSift_USE_NVTX 1)
 else()
   set(PopSift_USE_NVTX 0)
@@ -245,9 +198,14 @@ message(STATUS "Use CUDA NVTX for profiling: " ${PopSift_USE_NVTX_PROFILING})
 message(STATUS "Synchronize and check CUDA error after every kernel: " ${PopSift_ERRCHK_AFTER_KERNEL})
 message(STATUS "Grid filtering: " ${PopSift_USE_GRID_FILTER})
 message(STATUS "Additional warning for CUDA nvcc: " ${PopSift_NVCC_WARNINGS})
-message(STATUS "Compiling for CUDA CCs: ${PopSift_CUDA_CC_LIST}")
 message(STATUS "Install path: " ${CMAKE_INSTALL_PREFIX})
 message(STATUS "Testing step: " ${PopSift_USE_TEST_CMD})
+
+message(STATUS "CMAKE_CUDA_COMPILER = ${CMAKE_CUDA_COMPILER}")
+message(STATUS "CMAKE_CUDA_COMPILER_ID = ${CMAKE_CUDA_COMPILER_ID}")
+message(STATUS "CMAKE_CUDA_COMPILER_VERSION = ${CMAKE_CUDA_COMPILER_VERSION}")
+message(STATUS "CMAKE_CUDA_ARCHITECTURES = ${CMAKE_CUDA_ARCHITECTURES}")
+
 if(PopSift_USE_TEST_CMD)
   message(STATUS "Path for test input: " ${PopSift_TESTFILE_PATH})
 endif()

README.md

@@ -25,7 +25,7 @@ PopSift depends on:
 
 Optionally, for the provided applications:
 
-* Boost >= 1.55 (required components {atomic, chrono, date-time, system, thread}-dev)
+* Boost >= 1.71 (required components {atomic, chrono, date-time, system, thread}-dev)
 
 * DevIL (libdevil-dev) can be used to load a broader range of image formats, otherwise only pgm is supported.
 

appveyor.yml

@@ -1,6 +1,9 @@
+#
+# Build system for the PopSift library, including its demo programs.
+#
 version: '1.0.{build}'
 
-image: Visual Studio 2015
+image: Visual Studio 2022
 
 platform:
   - x64
@@ -14,6 +17,9 @@ configuration:
 #    - DBUILD_SHARED_LIBS: 0
 #    - DBUILD_SHARED_LIBS: 1
 
+#
+# Check the separate file cudaInstallAppveyor for the installation of CUDA
+#
 install:
     - cmd: >-
           call cudaInstallAppveyor.cmd
@@ -23,10 +29,22 @@ install:
           --triplet %PLATFORM%-windows
           # devil
 
+#
+# When updating to a new version of visual studio, change the generation string after
+# -G and find the suitable toolkit version that is listed after -T (v143 in this case).
+# The CUDA Toolkit and the VS version must match. The matches are found in the CUDA
+# documentation.
+# The platform in this case is x64. Apparently, you need in after -T for VS and after -A
+# for CUDA.
+# You can only have one -T parameter, but you can separate several options with a comma.
+#
+# PopSift_USE_GRID_FILTER is off in this build because the installation of CUDA Thrust
+# in cudaInstallAppveyor is not happening yet.
+#
 before_build:
     - md build
     - cd build
-    - cmake -G "Visual Studio 14 2015" -A x64 -T v140,host=x64 -DBUILD_SHARED_LIBS=%DBUILD_SHARED_LIBS% -DPopSift_BUILD_DOCS:BOOL=OFF -DPopSift_USE_POSITION_INDEPENDENT_CODE:BOOL=%DBUILD_SHARED_LIBS% -DPopSift_BUILD_EXAMPLES:BOOL=ON -DCMAKE_BUILD_TYPE=%configuration% -DCMAKE_TOOLCHAIN_FILE=c:/tools/vcpkg/scripts/buildsystems/vcpkg.cmake ..
+    - cmake -G "Visual Studio 17 2022" -A x64 -T v143,host=x64,cuda="%ProgramFiles%\NVIDIA GPU Computing Toolkit\CUDA\v12.5" -DBUILD_SHARED_LIBS:BOOL=ON -DPopSift_USE_NVTX_PROFILING:BOOL=OFF -DPopSift_USE_GRID_FILTER:BOOL=OFF -DPopSift_BUILD_DOCS:BOOL=OFF -DPopSift_USE_POSITION_INDEPENDENT_CODE:BOOL=ON -DPopSift_BUILD_EXAMPLES:BOOL=ON -DCMAKE_BUILD_TYPE=%configuration% -DCMAKE_TOOLCHAIN_FILE=c:/tools/vcpkg/scripts/buildsystems/vcpkg.cmake ..
     - ls -l
 
 build: