From e254b70e5027a19b97e1942d230a8bdf4999cac6 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Matthias=20M=C3=B6ller?= <m_moeller@live.de>
Date: Fri, 6 Sep 2024 19:37:57 +0200
Subject: [PATCH 1/5] replaces cmake CUDA finder with FindCUDAToolkit

---
 .github/workflows/build_cpp.yml  |   1 +
 dlib/CMakeLists.txt              | 190 +++----------------------------
 dlib/cmake_utils/findCUDNN.cmake |  76 +++++++++++++
 3 files changed, 95 insertions(+), 172 deletions(-)
 create mode 100644 dlib/cmake_utils/findCUDNN.cmake

diff --git a/.github/workflows/build_cpp.yml b/.github/workflows/build_cpp.yml
index 7a309065b7..e7ebcd097f 100644
--- a/.github/workflows/build_cpp.yml
+++ b/.github/workflows/build_cpp.yml
@@ -4,6 +4,7 @@ on:
   push:
     branches:
     - master
+    - cuda-cmake
     paths:
     - ".github/workflows/build_cpp.yml"
     - "**.cpp"
diff --git a/dlib/CMakeLists.txt b/dlib/CMakeLists.txt
index 44045ef3c2..488a66967e 100644
--- a/dlib/CMakeLists.txt
+++ b/dlib/CMakeLists.txt
@@ -4,7 +4,7 @@
 #
 
 
-cmake_minimum_required(VERSION 3.8.0)
+cmake_minimum_required(VERSION 3.17.0)
 
 set(CMAKE_DISABLE_SOURCE_CHANGES ON)
 set(CMAKE_DISABLE_IN_SOURCE_BUILD ON)
@@ -647,160 +647,16 @@ if (NOT TARGET dlib)
          endif()
       endif()
 
+      find_package(OpenMP)
+      set(openmp_libraries ${OpenMP_CXX_FLAGS}) 
 
-      if (DLIB_USE_CUDA)
-         find_package(CUDA 7.5)
-
-         if (CUDA_VERSION VERSION_GREATER 9.1 AND CMAKE_VERSION VERSION_LESS 3.12.2)
-            # This bit of weirdness is to work around a bug in cmake 
-            list(REMOVE_ITEM CUDA_CUBLAS_LIBRARIES "CUDA_cublas_device_LIBRARY-NOTFOUND")
-         endif()
-
-
-         if (CUDA_FOUND AND MSVC AND NOT CUDA_CUBLAS_LIBRARIES AND "${CMAKE_SIZEOF_VOID_P}" EQUAL "4")
-            message(WARNING "You have CUDA installed, but we can't use it unless you put visual studio in 64bit mode.")
-            set(CUDA_FOUND 0)
-         endif()
-
-         if (NOT CUDA_CUBLAS_LIBRARIES)
-            message(STATUS "Found CUDA, but CMake was unable to find the cuBLAS libraries that should be part of every basic CUDA "
-               "install. Your CUDA install is somehow broken or incomplete. Since cuBLAS is required for dlib to use CUDA we won't use CUDA.")
-            set(CUDA_FOUND 0)
-         endif()
-
-         if (CUDA_FOUND)
-
-            # There is some bug in cmake that causes it to mess up the
-            # -std=c++11 option if you let it propagate it to nvcc in some
-            # cases.  So instead we disable this and manually include
-            # things from CMAKE_CXX_FLAGS in the CUDA_NVCC_FLAGS list below.
-            if (APPLE)
-               set(CUDA_PROPAGATE_HOST_FLAGS OFF)
-               # Grab all the -D flags from CMAKE_CXX_FLAGS so we can pass them
-               # to nvcc.
-               string(REGEX MATCHALL "-D[^ ]*" FLAGS_FOR_NVCC "${CMAKE_CXX_FLAGS}")
-
-               # Check if we are being built as part of a pybind11 module. 
-               if (COMMAND pybind11_add_module)
-                  # Don't export unnecessary symbols.
-                  list(APPEND FLAGS_FOR_NVCC "-Xcompiler=-fvisibility=hidden")
-               endif()
-            endif()
-
-            set(CUDA_HOST_COMPILATION_CPP ON)
-            string(REPLACE "," ";" DLIB_CUDA_COMPUTE_CAPABILITIES ${DLIB_USE_CUDA_COMPUTE_CAPABILITIES})
-            foreach(CAP ${DLIB_CUDA_COMPUTE_CAPABILITIES})
-                list(APPEND CUDA_NVCC_FLAGS "-gencode arch=compute_${CAP},code=[sm_${CAP},compute_${CAP}]")
-            endforeach()
-            # Note that we add __STRICT_ANSI__ to avoid freaking out nvcc with gcc specific
-            # magic in the standard C++ header files (since nvcc uses gcc headers on linux).
-            list(APPEND CUDA_NVCC_FLAGS "-D__STRICT_ANSI__;-D_MWAITXINTRIN_H_INCLUDED;-D_FORCE_INLINES;${FLAGS_FOR_NVCC}")
-            list(APPEND CUDA_NVCC_FLAGS ${active_preprocessor_switches})
-            if (NOT DLIB_IN_PROJECT_BUILD)
-               LIST(APPEND CUDA_NVCC_FLAGS -DDLIB__CMAKE_GENERATED_A_CONFIG_H_FILE)
-            endif()
-            if (NOT MSVC)
-               list(APPEND CUDA_NVCC_FLAGS "-std=c++14")
-            endif()
-            if (CMAKE_POSITION_INDEPENDENT_CODE)
-               # sometimes this setting isn't propagated to NVCC, which then causes the
-               # compile to fail.  So make sure it's propagated.
-               if (NOT MSVC) # Visual studio doesn't have -fPIC so don't do it in that case.
-                  list(APPEND CUDA_NVCC_FLAGS "-Xcompiler -fPIC")
-               endif()
-            endif()
-
-            include(cmake_utils/test_for_cudnn/find_cudnn.txt)
-
-            if (cudnn AND cudnn_include AND NOT DEFINED cuda_test_compile_worked AND NOT DEFINED cudnn_test_compile_worked)
-               # make sure cuda is really working by doing a test compile
-               message(STATUS "Building a CUDA test project to see if your compiler is compatible with CUDA...")
 
-               set(CUDA_TEST_CMAKE_FLAGS 
-                  "-DCMAKE_PREFIX_PATH=${CMAKE_PREFIX_PATH}"
-                  "-DCMAKE_INCLUDE_PATH=${CMAKE_INCLUDE_PATH}"
-                  "-DCMAKE_LIBRARY_PATH=${CMAKE_LIBRARY_PATH}")
-
-               if (NOT MSVC) # see https://github.com/davisking/dlib/issues/363
-                  list(APPEND CUDA_TEST_CMAKE_FLAGS "-DCUDA_HOST_COMPILER=${CUDA_HOST_COMPILER}")
-               endif()
-
-               try_compile(cuda_test_compile_worked 
-                  ${PROJECT_BINARY_DIR}/cuda_test_build 
-                  ${PROJECT_SOURCE_DIR}/cmake_utils/test_for_cuda cuda_test
-                  CMAKE_FLAGS ${CUDA_TEST_CMAKE_FLAGS}
-                  OUTPUT_VARIABLE try_compile_output_message
-                  )
-               if (NOT cuda_test_compile_worked)
-                  string(REPLACE "\n" "\n   ***   " try_compile_output_message "${try_compile_output_message}")
-                  message(STATUS "*****************************************************************************************************************")
-                  message(STATUS "*** CUDA was found but your compiler failed to compile a simple CUDA program so dlib isn't going to use CUDA. ")
-                  message(STATUS "*** The output of the failed CUDA test compile is shown below: ")
-                  message(STATUS "*** ")
-                  message(STATUS "***   ${try_compile_output_message}")
-                  message(STATUS "*****************************************************************************************************************")
-               else()
-                  message(STATUS "Building a cuDNN test project to check if you have the right version of cuDNN installed...")
-                  try_compile(cudnn_test_compile_worked 
-                     ${PROJECT_BINARY_DIR}/cudnn_test_build 
-                     ${PROJECT_SOURCE_DIR}/cmake_utils/test_for_cudnn cudnn_test
-                     CMAKE_FLAGS ${CUDA_TEST_CMAKE_FLAGS}
-                     OUTPUT_VARIABLE try_compile_output_message
-                     )
-                  if (NOT cudnn_test_compile_worked)
-                     string(REPLACE "\n" "\n   ***   " try_compile_output_message "${try_compile_output_message}")
-                     message(STATUS "*****************************************************************************************************")
-                     message(STATUS "*** Found cuDNN, but we failed to compile the dlib/cmake_utils/test_for_cudnn project. ")
-                     message(STATUS "*** You either have an unsupported version of cuDNN or something is wrong with your cudDNN install.")
-                     message(STATUS "*** Since a functional cuDNN is not found DLIB WILL NOT USE CUDA. ")
-                     message(STATUS "*** The output of the failed test_for_cudnn build is: ")
-                     message(STATUS "*** ")
-                     message(STATUS "***   ${try_compile_output_message}")
-                     message(STATUS "*****************************************************************************************************")
-                  endif()
-               endif()
-            endif()
-
-            # Find where cuSOLVER is since the FindCUDA cmake package doesn't
-            # bother to look for it in older versions of cmake.
-            if (NOT CUDA_cusolver_LIBRARY)
-               get_filename_component(cuda_blas_path "${CUDA_CUBLAS_LIBRARIES}" DIRECTORY)
-               find_library(CUDA_cusolver_LIBRARY cusolver HINTS ${cuda_blas_path})
-               # CUDA 10.1 doesn't install symbolic links to libcusolver.so in
-               # the usual place.  This is probably a bug in the cuda
-               # installer.  In any case, If we haven't found cusolver yet go
-               # look in the cuda install folder for it.  New versions of cmake
-               # do this correctly, but older versions need help.
-               if (NOT CUDA_cusolver_LIBRARY)
-                  find_library(CUDA_cusolver_LIBRARY cusolver HINTS 
-                     /usr/local/cuda/lib64/
-                     )
-               endif()
-               mark_as_advanced(CUDA_cusolver_LIBRARY)
-            endif()
-            # Also find OpenMP since cuSOLVER needs it.  Importantly, we only
-            # look for one to link to if our use of BLAS, specifically the
-            # Intel MKL, hasn't already decided what to use.  This is because
-            # it makes the MKL bug out if you link to another openmp lib other
-            # than Intel's when you use the MKL. I'm also not really sure when
-            # explicit linking to openmp became unnecessary, but for
-            # sufficiently older versions of cuda it was needed.  Then in
-            # versions of cmake newer than 3.11 linking to openmp started to
-            # mess up the switches passed to nvcc, so you can't just leave
-            # these "try to link to openmp" statements here going forward.  Fun
-            # times.
-            if (CUDA_VERSION VERSION_LESS "9.1" AND NOT openmp_libraries AND NOT MSVC AND NOT XCODE AND NOT APPLE)
-               find_package(OpenMP)
-               if (OPENMP_FOUND)
-                  set(openmp_libraries ${OpenMP_CXX_FLAGS}) 
-               else()
-                  message(STATUS "*** Didn't find OpenMP, which is required to use CUDA. ***")
-                  set(CUDA_FOUND 0)
-               endif()
-            endif()
-         endif()
+      if (DLIB_USE_CUDA)
+         find_package(CUDAToolkit 9.1)
+         set(CMAKE_MODULE_PATH ${CMAKE_MODULE_PATH} ${CMAKE_CURRENT_SOURCE_DIR}/cmake_utils)
+         find_package(CUDNN)
 
-         if (CUDA_FOUND AND cudnn AND cuda_test_compile_worked AND cudnn_test_compile_worked AND cudnn_include)
+         if (CUDAToolkit_FOUND AND CUDNN_FOUND)
             set(source_files ${source_files} 
                cuda/cuda_dlib.cu 
                cuda/cudnn_dlibapi.cpp
@@ -810,19 +666,17 @@ if (NOT TARGET dlib)
                cuda/cuda_data_ptr.cpp
                cuda/gpu_data.cpp
                )
-            list (APPEND dlib_needed_private_libraries ${CUDA_CUBLAS_LIBRARIES})
-            list (APPEND dlib_needed_private_libraries ${cudnn})
-            list (APPEND dlib_needed_private_libraries ${CUDA_curand_LIBRARY})
-            list (APPEND dlib_needed_private_libraries ${CUDA_cusolver_LIBRARY})
-            list (APPEND dlib_needed_private_libraries ${CUDA_CUDART_LIBRARY})
-            if(openmp_libraries)
-               list (APPEND dlib_needed_private_libraries ${openmp_libraries})
-            endif()
-
-            include_directories(${cudnn_include})
+            list (APPEND dlib_needed_private_libraries CUDA::cublas)
+            list (APPEND dlib_needed_private_libraries ${CUDNN_LIBRARY_PATH})
+            list (APPEND dlib_needed_private_libraries CUDA::curand)
+            list (APPEND dlib_needed_private_libraries CUDA::cusolver)
+            list (APPEND dlib_needed_private_libraries CUDA::cudart)
+            list (APPEND dlib_needed_private_libraries ${openmp_libraries})
+
+            include_directories(${CUDAToolkit_INCLUDE_DIRS} ${CUDNN_INCLUDE_PATH})
             message(STATUS "Enabling CUDA support for dlib.  DLIB WILL USE CUDA, compute capabilities: ${DLIB_CUDA_COMPUTE_CAPABILITIES}")
          else()
-            set(DLIB_USE_CUDA OFF CACHE STRING ${DLIB_USE_BLAS_STR} FORCE )
+            set(DLIB_USE_CUDA OFF CACHE BOOL ${DLIB_USE_BLAS_STR} FORCE )
             toggle_preprocessor_switch(DLIB_USE_CUDA)
             if (NOT CUDA_FOUND)
                message(STATUS "DID NOT FIND CUDA")
@@ -875,15 +729,7 @@ if (NOT TARGET dlib)
          endif()
       endif()
 
-      # Tell CMake to build dlib via add_library()/cuda_add_library()
-      if (DLIB_USE_CUDA)
-         # The old cuda_add_library() command doesn't support CMake's newer dependency
-         # stuff, so we have to set the include path manually still, which we do here.
-         include_directories(${dlib_needed_public_includes})
-         cuda_add_library(dlib ${source_files} )
-      else()
-         add_library(dlib ${source_files} )
-      endif()
+      add_library(dlib ${source_files} )
 
    endif ()  ##### end of if NOT DLIB_ISO_CPP_ONLY ##########################################################
 
diff --git a/dlib/cmake_utils/findCUDNN.cmake b/dlib/cmake_utils/findCUDNN.cmake
new file mode 100644
index 0000000000..0c900fcd4e
--- /dev/null
+++ b/dlib/cmake_utils/findCUDNN.cmake
@@ -0,0 +1,76 @@
+# Find the CUDNN libraries
+#
+# The following variables are optionally searched for defaults
+#  CUDNN_ROOT: Base directory where CUDNN is found
+#  CUDNN_INCLUDE_DIR: Directory where CUDNN header is searched for
+#  CUDNN_LIBRARY: Directory where CUDNN library is searched for
+#  CUDNN_STATIC: Are we looking for a static library? (default: no)
+#
+# The following are set after configuration is done:
+#  CUDNN_FOUND
+#  CUDNN_INCLUDE_PATH
+#  CUDNN_LIBRARY_PATH
+#
+
+include(FindPackageHandleStandardArgs)
+
+set(CUDNN_ROOT $ENV{CUDNN_ROOT_DIR} CACHE PATH "Folder containing NVIDIA cuDNN")
+if (DEFINED $ENV{CUDNN_ROOT_DIR})
+  message(WARNING "CUDNN_ROOT_DIR is deprecated. Please set CUDNN_ROOT instead.")
+endif()
+list(APPEND CUDNN_ROOT $ENV{CUDNN_ROOT_DIR} ${CUDA_TOOLKIT_ROOT_DIR})
+
+# Compatible layer for CMake <3.12. CUDNN_ROOT will be accounted in for searching paths and libraries for CMake >=3.12.
+list(APPEND CMAKE_PREFIX_PATH ${CUDNN_ROOT})
+
+set(CUDNN_INCLUDE_DIR $ENV{CUDNN_INCLUDE_DIR} CACHE PATH "Folder containing NVIDIA cuDNN header files")
+
+find_path(CUDNN_INCLUDE_PATH cudnn.h
+  HINTS ${CUDNN_INCLUDE_DIR}
+  PATH_SUFFIXES cuda/include cuda include)
+
+option(CUDNN_STATIC "Look for static CUDNN" OFF)
+if (CUDNN_STATIC)
+  set(CUDNN_LIBNAME "libcudnn_static.a")
+else()
+  set(CUDNN_LIBNAME "cudnn")
+endif()
+
+set(CUDNN_LIBRARY $ENV{CUDNN_LIBRARY} CACHE PATH "Path to the cudnn library file (e.g., libcudnn.so)")
+if (CUDNN_LIBRARY MATCHES ".*cudnn_static.a" AND NOT CUDNN_STATIC)
+  message(WARNING "CUDNN_LIBRARY points to a static library (${CUDNN_LIBRARY}) but CUDNN_STATIC is OFF.")
+endif()
+
+find_library(CUDNN_LIBRARY_PATH ${CUDNN_LIBNAME}
+  PATHS ${CUDNN_LIBRARY}
+  PATH_SUFFIXES lib lib64 cuda/lib cuda/lib64 lib/x64)
+
+find_package_handle_standard_args(CUDNN DEFAULT_MSG CUDNN_LIBRARY_PATH CUDNN_INCLUDE_PATH)
+
+if(CUDNN_FOUND)
+  # Get cuDNN version
+  if(EXISTS ${CUDNN_INCLUDE_PATH}/cudnn_version.h)
+    file(READ ${CUDNN_INCLUDE_PATH}/cudnn_version.h CUDNN_HEADER_CONTENTS)
+  else()
+    file(READ ${CUDNN_INCLUDE_PATH}/cudnn.h CUDNN_HEADER_CONTENTS)
+  endif()
+  string(REGEX MATCH "define CUDNN_MAJOR * +([0-9]+)"
+               CUDNN_VERSION_MAJOR "${CUDNN_HEADER_CONTENTS}")
+  string(REGEX REPLACE "define CUDNN_MAJOR * +([0-9]+)" "\\1"
+               CUDNN_VERSION_MAJOR "${CUDNN_VERSION_MAJOR}")
+  string(REGEX MATCH "define CUDNN_MINOR * +([0-9]+)"
+               CUDNN_VERSION_MINOR "${CUDNN_HEADER_CONTENTS}")
+  string(REGEX REPLACE "define CUDNN_MINOR * +([0-9]+)" "\\1"
+               CUDNN_VERSION_MINOR "${CUDNN_VERSION_MINOR}")
+  string(REGEX MATCH "define CUDNN_PATCHLEVEL * +([0-9]+)"
+               CUDNN_VERSION_PATCH "${CUDNN_HEADER_CONTENTS}")
+  string(REGEX REPLACE "define CUDNN_PATCHLEVEL * +([0-9]+)" "\\1"
+               CUDNN_VERSION_PATCH "${CUDNN_VERSION_PATCH}")
+  # Assemble cuDNN version
+  if(NOT CUDNN_VERSION_MAJOR)
+    set(CUDNN_VERSION "?")
+  else()
+    set(CUDNN_VERSION
+        "${CUDNN_VERSION_MAJOR}.${CUDNN_VERSION_MINOR}.${CUDNN_VERSION_PATCH}")
+  endif()
+endif()
\ No newline at end of file

From 5f94ad5d53a45e8e3e6301f04c1c5f95d9f55c84 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Matthias=20M=C3=B6ller?= <m_moeller@live.de>
Date: Fri, 6 Sep 2024 20:17:21 +0200
Subject: [PATCH 2/5] increase cmake version to 3.17

---
 .github/workflows/build_cpp.yml                   | 15 +++++++--------
 CMakeLists.txt                                    |  2 +-
 dlib/CMakeLists.txt                               |  7 +++----
 dlib/cmake                                        |  2 +-
 ...k_if_avx_instructions_executable_on_host.cmake |  2 +-
 dlib/cmake_utils/check_if_neon_available.cmake    |  2 +-
 ..._if_sse4_instructions_executable_on_host.cmake |  2 +-
 dlib/cmake_utils/find_ffmpeg.cmake                |  2 +-
 dlib/cmake_utils/find_libjpeg.cmake               |  2 +-
 dlib/cmake_utils/find_libpng.cmake                |  2 +-
 .../set_compiler_specific_options.cmake           |  2 +-
 ...tell_visual_studio_to_use_static_runtime.cmake |  2 +-
 dlib/cmake_utils/test_for_avx/CMakeLists.txt      |  2 +-
 dlib/cmake_utils/test_for_cuda/CMakeLists.txt     |  2 +-
 dlib/cmake_utils/test_for_cudnn/CMakeLists.txt    |  2 +-
 dlib/cmake_utils/test_for_libjpeg/CMakeLists.txt  |  2 +-
 dlib/cmake_utils/test_for_libjxl/CMakeLists.txt   |  2 +-
 dlib/cmake_utils/test_for_libpng/CMakeLists.txt   |  2 +-
 dlib/cmake_utils/test_for_libwebp/CMakeLists.txt  |  2 +-
 dlib/cmake_utils/test_for_neon/CMakeLists.txt     |  2 +-
 dlib/cmake_utils/test_for_sse4/CMakeLists.txt     |  2 +-
 dlib/external/cblas/CMakeLists.txt                |  2 +-
 dlib/external/pybind11/CMakeLists.txt             |  6 +++---
 dlib/java/CMakeLists.txt                          |  2 +-
 dlib/java/cmake_swig_jni                          |  4 ++--
 dlib/matlab/CMakeLists.txt                        |  2 +-
 dlib/matlab/cmake_mex_wrapper                     |  2 +-
 dlib/test/CMakeLists.txt                          |  8 +++-----
 dlib/test/blas_bindings/CMakeLists.txt            |  2 +-
 dlib/test/tools/CMakeLists.txt                    |  2 +-
 docs/docs/release_notes.xml                       |  2 +-
 examples/CMakeLists.txt                           |  2 +-
 tools/archive/CMakeLists.txt                      |  2 +-
 tools/convert_dlib_nets_to_caffe/CMakeLists.txt   |  2 +-
 tools/htmlify/CMakeLists.txt                      |  2 +-
 tools/imglab/CMakeLists.txt                       |  2 +-
 tools/python/CMakeLists.txt                       |  2 +-
 37 files changed, 50 insertions(+), 54 deletions(-)

diff --git a/.github/workflows/build_cpp.yml b/.github/workflows/build_cpp.yml
index e7ebcd097f..5889d9d68e 100644
--- a/.github/workflows/build_cpp.yml
+++ b/.github/workflows/build_cpp.yml
@@ -4,7 +4,6 @@ on:
   push:
     branches:
     - master
-    - cuda-cmake
     paths:
     - ".github/workflows/build_cpp.yml"
     - "**.cpp"
@@ -31,7 +30,7 @@ defaults:
     working-directory: dlib/test
 
 jobs:
-  ubuntu-latest-gcc-default-cmake-3-8-ffmpeg5:
+  ubuntu-latest-gcc-default-cmake-3-17-ffmpeg5:
     runs-on: 'ubuntu-latest'
     steps:
     - uses: actions/checkout@v2
@@ -41,18 +40,18 @@ jobs:
         sudo apt update
         sudo apt install libwebp-dev make yasm
         
-    - name: Cache cmake 3.8.0
+    - name: Cache cmake 3.17.0
       uses: actions/cache@v3
       id: cache-cmake-download
       with:
         # cache this folder:
-        path: ~/cmake-3.8.0-Linux-x86_64
-        key: cmake-3.8.0_try3
+        path: ~/cmake-3.17.0-Linux-x86_64
+        key: cmake-3.17.0_try3
         
     - run: |
         # Get the minimum version of cmake dlib supports
-        wget https://cmake.org/files/v3.8/cmake-3.8.0-Linux-x86_64.tar.gz
-        tar -xf cmake-3.8.0-Linux-x86_64.tar.gz -C ~
+        wget https://cmake.org/files/v3.17/cmake-3.17.0-Linux-x86_64.tar.gz
+        tar -xf cmake-3.17.0-Linux-x86_64.tar.gz -C ~
       if: steps.cache-cmake-download.outputs.cache-hit != 'true'
 
     - name: Cache FFmpeg 5
@@ -77,7 +76,7 @@ jobs:
       run: |
         mkdir build
         cd build
-        ~/cmake-3.8.0-Linux-x86_64/bin/cmake -DCMAKE_PREFIX_PATH=/home/runner/ffmpeg-n5.1.3_installation .. 
+        ~/cmake-3.17.0-Linux-x86_64/bin/cmake -DCMAKE_PREFIX_PATH=/home/runner/ffmpeg-n5.1.3_installation .. 
     
     - name: Build just tests
       run: |
diff --git a/CMakeLists.txt b/CMakeLists.txt
index da92454aab..ca00843f18 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -1,4 +1,4 @@
-cmake_minimum_required(VERSION 3.8.0)
+cmake_minimum_required(VERSION 3.17.0)
 
 project(dlib_project)
 
diff --git a/dlib/CMakeLists.txt b/dlib/CMakeLists.txt
index 488a66967e..e7bfd29c74 100644
--- a/dlib/CMakeLists.txt
+++ b/dlib/CMakeLists.txt
@@ -647,16 +647,15 @@ if (NOT TARGET dlib)
          endif()
       endif()
 
-      find_package(OpenMP)
-      set(openmp_libraries ${OpenMP_CXX_FLAGS}) 
-
 
       if (DLIB_USE_CUDA)
          find_package(CUDAToolkit 9.1)
          set(CMAKE_MODULE_PATH ${CMAKE_MODULE_PATH} ${CMAKE_CURRENT_SOURCE_DIR}/cmake_utils)
          find_package(CUDNN)
+         find_package(OpenMP)
+         set(openmp_libraries ${OpenMP_CXX_FLAGS}) 
 
-         if (CUDAToolkit_FOUND AND CUDNN_FOUND)
+         if (CUDAToolkit_FOUND AND CUDNN_FOUND AND OPENMP_FOUND)
             set(source_files ${source_files} 
                cuda/cuda_dlib.cu 
                cuda/cudnn_dlibapi.cpp
diff --git a/dlib/cmake b/dlib/cmake
index 224ba3a491..4a9d7b3985 100644
--- a/dlib/cmake
+++ b/dlib/cmake
@@ -1,5 +1,5 @@
 
-cmake_minimum_required(VERSION 3.8.0)
+cmake_minimum_required(VERSION 3.17.0)
 
 add_subdirectory(${CMAKE_CURRENT_LIST_DIR} dlib_build)
 
diff --git a/dlib/cmake_utils/check_if_avx_instructions_executable_on_host.cmake b/dlib/cmake_utils/check_if_avx_instructions_executable_on_host.cmake
index 4f2cfef933..717812678d 100644
--- a/dlib/cmake_utils/check_if_avx_instructions_executable_on_host.cmake
+++ b/dlib/cmake_utils/check_if_avx_instructions_executable_on_host.cmake
@@ -1,6 +1,6 @@
 # This script checks if your compiler and host processor can generate and then run programs with AVX instructions.
 
-cmake_minimum_required(VERSION 3.8.0)
+cmake_minimum_required(VERSION 3.17.0)
 
 # Don't rerun this script if its already been executed.
 if (DEFINED AVX_IS_AVAILABLE_ON_HOST)
diff --git a/dlib/cmake_utils/check_if_neon_available.cmake b/dlib/cmake_utils/check_if_neon_available.cmake
index 895c810b74..6ed6725cac 100644
--- a/dlib/cmake_utils/check_if_neon_available.cmake
+++ b/dlib/cmake_utils/check_if_neon_available.cmake
@@ -1,6 +1,6 @@
 # This script checks if __ARM_NEON__ is defined for your compiler
 
-cmake_minimum_required(VERSION 3.8.0)
+cmake_minimum_required(VERSION 3.17.0)
 
 # Don't rerun this script if its already been executed.
 if (DEFINED ARM_NEON_IS_AVAILABLE)
diff --git a/dlib/cmake_utils/check_if_sse4_instructions_executable_on_host.cmake b/dlib/cmake_utils/check_if_sse4_instructions_executable_on_host.cmake
index c475609970..7fbe182aed 100644
--- a/dlib/cmake_utils/check_if_sse4_instructions_executable_on_host.cmake
+++ b/dlib/cmake_utils/check_if_sse4_instructions_executable_on_host.cmake
@@ -1,6 +1,6 @@
 # This script checks if your compiler and host processor can generate and then run programs with SSE4 instructions.
 
-cmake_minimum_required(VERSION 3.8.0)
+cmake_minimum_required(VERSION 3.17.0)
 
 # Don't rerun this script if its already been executed.
 if (DEFINED SSE4_IS_AVAILABLE_ON_HOST)
diff --git a/dlib/cmake_utils/find_ffmpeg.cmake b/dlib/cmake_utils/find_ffmpeg.cmake
index b05955387b..4179013147 100644
--- a/dlib/cmake_utils/find_ffmpeg.cmake
+++ b/dlib/cmake_utils/find_ffmpeg.cmake
@@ -1,4 +1,4 @@
-cmake_minimum_required(VERSION 3.8.0)
+cmake_minimum_required(VERSION 3.17.0)
 
 message(STATUS "Searching for FFMPEG/LIBAV")
 find_package(PkgConfig)
diff --git a/dlib/cmake_utils/find_libjpeg.cmake b/dlib/cmake_utils/find_libjpeg.cmake
index 028217b075..1ae3b07ed7 100644
--- a/dlib/cmake_utils/find_libjpeg.cmake
+++ b/dlib/cmake_utils/find_libjpeg.cmake
@@ -1,7 +1,7 @@
 #This script just runs CMake's built in JPEG finding tool.  But it also checks that the
 #copy of libjpeg that cmake finds actually builds and links.
 
-cmake_minimum_required(VERSION 3.8.0)
+cmake_minimum_required(VERSION 3.17.0)
 
 if (BUILDING_PYTHON_IN_MSVC)
    # Never use any system copy of libjpeg when building python in visual studio
diff --git a/dlib/cmake_utils/find_libpng.cmake b/dlib/cmake_utils/find_libpng.cmake
index 6760739226..87e593c9a6 100644
--- a/dlib/cmake_utils/find_libpng.cmake
+++ b/dlib/cmake_utils/find_libpng.cmake
@@ -1,7 +1,7 @@
 #This script just runs CMake's built in PNG finding tool.  But it also checks that the
 #copy of libpng that cmake finds actually builds and links.
 
-cmake_minimum_required(VERSION 3.8.0)
+cmake_minimum_required(VERSION 3.17.0)
 
 if (BUILDING_PYTHON_IN_MSVC)
    # Never use any system copy of libpng when building python in visual studio
diff --git a/dlib/cmake_utils/set_compiler_specific_options.cmake b/dlib/cmake_utils/set_compiler_specific_options.cmake
index 8093ca6d30..06220effc4 100644
--- a/dlib/cmake_utils/set_compiler_specific_options.cmake
+++ b/dlib/cmake_utils/set_compiler_specific_options.cmake
@@ -1,5 +1,5 @@
 
-cmake_minimum_required(VERSION 3.8.0)
+cmake_minimum_required(VERSION 3.17.0)
 
 
 # Check if we are being built as part of a pybind11 module. 
diff --git a/dlib/cmake_utils/tell_visual_studio_to_use_static_runtime.cmake b/dlib/cmake_utils/tell_visual_studio_to_use_static_runtime.cmake
index 80122d8018..0d83f932bf 100644
--- a/dlib/cmake_utils/tell_visual_studio_to_use_static_runtime.cmake
+++ b/dlib/cmake_utils/tell_visual_studio_to_use_static_runtime.cmake
@@ -2,7 +2,7 @@
 # Including this cmake script into your cmake project will cause visual studio
 # to build your project against the static C runtime.
 
-cmake_minimum_required(VERSION 3.8.0)
+cmake_minimum_required(VERSION 3.17.0)
 
 if (MSVC OR "${CMAKE_CXX_COMPILER_ID}" STREQUAL "MSVC")
    option (DLIB_FORCE_MSVC_STATIC_RUNTIME "use static runtime" ON)
diff --git a/dlib/cmake_utils/test_for_avx/CMakeLists.txt b/dlib/cmake_utils/test_for_avx/CMakeLists.txt
index 10890153cf..a19ca62bf9 100644
--- a/dlib/cmake_utils/test_for_avx/CMakeLists.txt
+++ b/dlib/cmake_utils/test_for_avx/CMakeLists.txt
@@ -1,5 +1,5 @@
 
-cmake_minimum_required(VERSION 3.8.0)
+cmake_minimum_required(VERSION 3.17.0)
 project(avx_test)
 
 set(USE_AVX_INSTRUCTIONS ON CACHE BOOL "Use AVX instructions")
diff --git a/dlib/cmake_utils/test_for_cuda/CMakeLists.txt b/dlib/cmake_utils/test_for_cuda/CMakeLists.txt
index 8b6cd7e7b0..7392501f15 100644
--- a/dlib/cmake_utils/test_for_cuda/CMakeLists.txt
+++ b/dlib/cmake_utils/test_for_cuda/CMakeLists.txt
@@ -1,5 +1,5 @@
 
-cmake_minimum_required(VERSION 3.8.0)
+cmake_minimum_required(VERSION 3.17.0)
 project(cuda_test)
 
 include_directories(../../cuda)
diff --git a/dlib/cmake_utils/test_for_cudnn/CMakeLists.txt b/dlib/cmake_utils/test_for_cudnn/CMakeLists.txt
index 0ca0f6b6a2..4a0315a327 100644
--- a/dlib/cmake_utils/test_for_cudnn/CMakeLists.txt
+++ b/dlib/cmake_utils/test_for_cudnn/CMakeLists.txt
@@ -1,5 +1,5 @@
 
-cmake_minimum_required(VERSION 3.8.0)
+cmake_minimum_required(VERSION 3.17.0)
 project(cudnn_test)
 
 # Override the FindCUDA.cmake setting to avoid duplication of host flags if using a toolchain:
diff --git a/dlib/cmake_utils/test_for_libjpeg/CMakeLists.txt b/dlib/cmake_utils/test_for_libjpeg/CMakeLists.txt
index a3c6910dbe..db740bca50 100644
--- a/dlib/cmake_utils/test_for_libjpeg/CMakeLists.txt
+++ b/dlib/cmake_utils/test_for_libjpeg/CMakeLists.txt
@@ -1,5 +1,5 @@
 
-cmake_minimum_required(VERSION 3.8.0)
+cmake_minimum_required(VERSION 3.17.0)
 project(test_if_libjpeg_is_broken)
 
 find_package(JPEG)
diff --git a/dlib/cmake_utils/test_for_libjxl/CMakeLists.txt b/dlib/cmake_utils/test_for_libjxl/CMakeLists.txt
index c91b5d8ff6..b89e46fcbb 100644
--- a/dlib/cmake_utils/test_for_libjxl/CMakeLists.txt
+++ b/dlib/cmake_utils/test_for_libjxl/CMakeLists.txt
@@ -1,5 +1,5 @@
 
-cmake_minimum_required(VERSION 3.8.0)
+cmake_minimum_required(VERSION 3.17.0)
 project(test_if_libjxl_is_broken)
 
 include_directories(${JXL_INCLUDE_DIR})
diff --git a/dlib/cmake_utils/test_for_libpng/CMakeLists.txt b/dlib/cmake_utils/test_for_libpng/CMakeLists.txt
index 0be9206132..94ce87826d 100644
--- a/dlib/cmake_utils/test_for_libpng/CMakeLists.txt
+++ b/dlib/cmake_utils/test_for_libpng/CMakeLists.txt
@@ -1,5 +1,5 @@
 
-cmake_minimum_required(VERSION 3.8.0)
+cmake_minimum_required(VERSION 3.17.0)
 project(test_if_libpng_is_broken)
 
 find_package(PNG)
diff --git a/dlib/cmake_utils/test_for_libwebp/CMakeLists.txt b/dlib/cmake_utils/test_for_libwebp/CMakeLists.txt
index 9ea3ed10ea..83c4502638 100644
--- a/dlib/cmake_utils/test_for_libwebp/CMakeLists.txt
+++ b/dlib/cmake_utils/test_for_libwebp/CMakeLists.txt
@@ -1,5 +1,5 @@
 
-cmake_minimum_required(VERSION 3.8.0)
+cmake_minimum_required(VERSION 3.17.0)
 project(test_if_libwebp_is_broken)
 
 include_directories(${WEBP_INCLUDE_DIR})
diff --git a/dlib/cmake_utils/test_for_neon/CMakeLists.txt b/dlib/cmake_utils/test_for_neon/CMakeLists.txt
index b9c6e51e8a..c27d339723 100644
--- a/dlib/cmake_utils/test_for_neon/CMakeLists.txt
+++ b/dlib/cmake_utils/test_for_neon/CMakeLists.txt
@@ -1,5 +1,5 @@
 
-cmake_minimum_required(VERSION 3.8.0)
+cmake_minimum_required(VERSION 3.17.0)
 project(neon_test)
 
 add_library(neon_test STATIC neon_test.cpp )
diff --git a/dlib/cmake_utils/test_for_sse4/CMakeLists.txt b/dlib/cmake_utils/test_for_sse4/CMakeLists.txt
index 4cd4e95cef..e01aeafd4d 100644
--- a/dlib/cmake_utils/test_for_sse4/CMakeLists.txt
+++ b/dlib/cmake_utils/test_for_sse4/CMakeLists.txt
@@ -1,5 +1,5 @@
 
-cmake_minimum_required(VERSION 3.8.0)
+cmake_minimum_required(VERSION 3.17.0)
 project(sse4_test)
 
 set(USE_SSE4_INSTRUCTIONS ON CACHE BOOL "Use SSE4 instructions")
diff --git a/dlib/external/cblas/CMakeLists.txt b/dlib/external/cblas/CMakeLists.txt
index c5379dd9fa..78ade59f3c 100644
--- a/dlib/external/cblas/CMakeLists.txt
+++ b/dlib/external/cblas/CMakeLists.txt
@@ -4,7 +4,7 @@
 #
 
 
-cmake_minimum_required(VERSION 3.8.0)
+cmake_minimum_required(VERSION 3.17.0)
 project(cblas)
 
 
diff --git a/dlib/external/pybind11/CMakeLists.txt b/dlib/external/pybind11/CMakeLists.txt
index 7db1bf668f..9156a623df 100644
--- a/dlib/external/pybind11/CMakeLists.txt
+++ b/dlib/external/pybind11/CMakeLists.txt
@@ -77,9 +77,9 @@ if(CMAKE_SOURCE_DIR STREQUAL PROJECT_SOURCE_DIR)
   if(OSX AND CMAKE_VERSION VERSION_LESS 3.7)
     # Bug in macOS CMake < 3.7 is unable to download catch
     message(WARNING "CMAKE 3.7+ needed on macOS to download catch, and newer HIGHLY recommended")
-  elseif(WINDOWS AND CMAKE_VERSION VERSION_LESS 3.8)
-    # Only tested with 3.8+ in CI.
-    message(WARNING "CMAKE 3.8+ tested on Windows, previous versions untested")
+  elseif(WINDOWS AND CMAKE_VERSION VERSION_LESS 3.17)
+    # Only tested with 3.8´17+ in CI.
+    message(WARNING "CMAKE 3.17+ tested on Windows, previous versions untested")
   endif()
 
   message(STATUS "CMake ${CMAKE_VERSION}")
diff --git a/dlib/java/CMakeLists.txt b/dlib/java/CMakeLists.txt
index b31bfad0d4..c87258ccc8 100644
--- a/dlib/java/CMakeLists.txt
+++ b/dlib/java/CMakeLists.txt
@@ -1,5 +1,5 @@
 
-cmake_minimum_required(VERSION 3.8.0)
+cmake_minimum_required(VERSION 3.17.0)
 project (myproject)
 set(java_package_name net.dlib)
 set(source_files
diff --git a/dlib/java/cmake_swig_jni b/dlib/java/cmake_swig_jni
index 48f960f41f..d4127aa918 100644
--- a/dlib/java/cmake_swig_jni
+++ b/dlib/java/cmake_swig_jni
@@ -2,7 +2,7 @@
 # it by defining some CMake variables and then include(cmake_swig_jni).  You
 # would make a CMakeLists.txt file that looks like the following:
 #
-#   cmake_minimum_required (VERSION 3.8.0)
+#   cmake_minimum_required (VERSION 3.17.0)
 #   project (example)
 #   set(java_package_name "org.mycompany")
 #   set(source_files
@@ -40,7 +40,7 @@
 ################################################################################
 ################################################################################
 
-cmake_minimum_required (VERSION 3.8.0)
+cmake_minimum_required (VERSION 3.17.0)
 
 # This block of code tries to figure out what the JAVA_HOME environment
 # variable should be by looking at the folder that contains the java
diff --git a/dlib/matlab/CMakeLists.txt b/dlib/matlab/CMakeLists.txt
index 2dee98c60c..b9a5a12289 100644
--- a/dlib/matlab/CMakeLists.txt
+++ b/dlib/matlab/CMakeLists.txt
@@ -1,5 +1,5 @@
 
-cmake_minimum_required(VERSION 3.8.0)
+cmake_minimum_required(VERSION 3.17.0)
 
 PROJECT(mex_functions)
 
diff --git a/dlib/matlab/cmake_mex_wrapper b/dlib/matlab/cmake_mex_wrapper
index 6cbc27afc6..0621b7dc6a 100644
--- a/dlib/matlab/cmake_mex_wrapper
+++ b/dlib/matlab/cmake_mex_wrapper
@@ -3,7 +3,7 @@
 # that additional library dependencies can be added like this: add_mex_function(name lib1 dlib libetc).
 # That is, just add more libraries after the name and they will be build into the mex file.
 
-cmake_minimum_required(VERSION 3.8.0)
+cmake_minimum_required(VERSION 3.17.0)
 
 set(BUILDING_MATLAB_MEX_FILE true)
 set(CMAKE_POSITION_INDEPENDENT_CODE True)
diff --git a/dlib/test/CMakeLists.txt b/dlib/test/CMakeLists.txt
index 4154fa0a3b..bb06619b3d 100644
--- a/dlib/test/CMakeLists.txt
+++ b/dlib/test/CMakeLists.txt
@@ -3,7 +3,7 @@
 # information about it at http://www.cmake.org
 #
 
-cmake_minimum_required(VERSION 3.8.0)
+cmake_minimum_required(VERSION 3.17.0)
 
 # create a variable called target_name and set it to the string "dtest"
 set (target_name dtest)
@@ -213,10 +213,8 @@ else() # basically Clang
    # This is for the comment in face_detection_ex.cpp that says "faces/*.jpg"
    add_compile_options(-Wno-comment)
 
-   if(${CMAKE_VERSION} VERSION_GREATER "3.8.0") 
-      # strip debug symbols to make the binary smaller
-      target_link_options(${target_name} PRIVATE $<$<CONFIG:RELEASE>:-s>)
-   endif()
+   # strip debug symbols to make the binary smaller
+   target_link_options(${target_name} PRIVATE $<$<CONFIG:RELEASE>:-s>)
 endif()
 
 
diff --git a/dlib/test/blas_bindings/CMakeLists.txt b/dlib/test/blas_bindings/CMakeLists.txt
index f9c0b05a12..883d237bd4 100644
--- a/dlib/test/blas_bindings/CMakeLists.txt
+++ b/dlib/test/blas_bindings/CMakeLists.txt
@@ -3,7 +3,7 @@
 # information about it at http://www.cmake.org
 #
 
-cmake_minimum_required(VERSION 3.8.0)
+cmake_minimum_required(VERSION 3.17.0)
 
 # This variable contains a list of all the tests we are building
 # into the regression test suite.
diff --git a/dlib/test/tools/CMakeLists.txt b/dlib/test/tools/CMakeLists.txt
index 80371d6d84..f3db1a0cb2 100644
--- a/dlib/test/tools/CMakeLists.txt
+++ b/dlib/test/tools/CMakeLists.txt
@@ -1,4 +1,4 @@
-cmake_minimum_required(VERSION 3.8.0)
+cmake_minimum_required(VERSION 3.17.0)
 
 add_subdirectory(../../../tools/imglab imglab_build)
 add_subdirectory(../../../tools/htmlify htmlify_build)
diff --git a/docs/docs/release_notes.xml b/docs/docs/release_notes.xml
index 97cb945486..38b3304e35 100644
--- a/docs/docs/release_notes.xml
+++ b/docs/docs/release_notes.xml
@@ -58,7 +58,7 @@ New Features and Improvements:
    - FFMPEG wrappers: dlib::ffmpeg::decoder and dlib::ffmpeg::demuxer (PR #2707)
 
 Non-Backwards Compatible Changes:
-   - Dlib now requires C++14 and at least CMake version 3.8.0.
+   - Dlib now requires C++14 and at least CMake version 3.17.0.
 
 Bug fixes:
    - Various adjustments to avoid errors on unusual platforms of system configurations.
diff --git a/examples/CMakeLists.txt b/examples/CMakeLists.txt
index 61af199179..e3357ab44b 100644
--- a/examples/CMakeLists.txt
+++ b/examples/CMakeLists.txt
@@ -32,7 +32,7 @@
 #
 
 
-cmake_minimum_required(VERSION 3.8.0)
+cmake_minimum_required(VERSION 3.17.0)
 # Every project needs a name.  We call this the "examples" project.
 project(examples)
 
diff --git a/tools/archive/CMakeLists.txt b/tools/archive/CMakeLists.txt
index 796e469852..1753969e5e 100644
--- a/tools/archive/CMakeLists.txt
+++ b/tools/archive/CMakeLists.txt
@@ -4,7 +4,7 @@
 #
 
 
-cmake_minimum_required(VERSION 3.8.0)
+cmake_minimum_required(VERSION 3.17.0)
 PROJECT(archive)
 
 
diff --git a/tools/convert_dlib_nets_to_caffe/CMakeLists.txt b/tools/convert_dlib_nets_to_caffe/CMakeLists.txt
index 2325366fe0..e043786042 100644
--- a/tools/convert_dlib_nets_to_caffe/CMakeLists.txt
+++ b/tools/convert_dlib_nets_to_caffe/CMakeLists.txt
@@ -3,7 +3,7 @@
 # information about it at http://www.cmake.org
 #
 
-cmake_minimum_required(VERSION 3.8.0)
+cmake_minimum_required(VERSION 3.17.0)
 
 set (target_name dtoc)
 
diff --git a/tools/htmlify/CMakeLists.txt b/tools/htmlify/CMakeLists.txt
index 35f327f275..8d8644fc18 100644
--- a/tools/htmlify/CMakeLists.txt
+++ b/tools/htmlify/CMakeLists.txt
@@ -3,7 +3,7 @@
 # information about it at http://www.cmake.org
 #
 
-cmake_minimum_required(VERSION 3.8.0)
+cmake_minimum_required(VERSION 3.17.0)
 
 # create a variable called target_name and set it to the string "htmlify"
 set (target_name htmlify)
diff --git a/tools/imglab/CMakeLists.txt b/tools/imglab/CMakeLists.txt
index 5fb1331d15..316a879cd1 100644
--- a/tools/imglab/CMakeLists.txt
+++ b/tools/imglab/CMakeLists.txt
@@ -3,7 +3,7 @@
 # information about it at http://www.cmake.org
 #
 
-cmake_minimum_required(VERSION 3.8.0)
+cmake_minimum_required(VERSION 3.17.0)
 
 # create a variable called target_name and set it to the string "imglab"
 set (target_name imglab)
diff --git a/tools/python/CMakeLists.txt b/tools/python/CMakeLists.txt
index 8309d5fb93..97759bd26b 100644
--- a/tools/python/CMakeLists.txt
+++ b/tools/python/CMakeLists.txt
@@ -1,5 +1,5 @@
 
-CMAKE_MINIMUM_REQUIRED(VERSION 3.8.0)
+CMAKE_MINIMUM_REQUIRED(VERSION 3.17.0)
 
 if (WIN32 AND NOT "${CMAKE_GENERATOR}" MATCHES "Visual Studio") 
    message(FATAL_ERROR "\n"

From 9da5c52a0ec5c366143eaf701c35a5ccd89d7048 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Matthias=20M=C3=B6ller?= <m_moeller@live.de>
Date: Sat, 7 Sep 2024 22:23:14 +0200
Subject: [PATCH 3/5] re-introduce cuda test compile

---
 dlib/CMakeLists.txt                           | 157 ++++++++++++++++--
 dlib/cmake_utils/findCUDNN.cmake              |  76 ---------
 dlib/cmake_utils/test_for_cuda/CMakeLists.txt |   5 +-
 .../cmake_utils/test_for_cudnn/CMakeLists.txt |   5 +-
 4 files changed, 153 insertions(+), 90 deletions(-)
 delete mode 100644 dlib/cmake_utils/findCUDNN.cmake

diff --git a/dlib/CMakeLists.txt b/dlib/CMakeLists.txt
index e7bfd29c74..491b78f57e 100644
--- a/dlib/CMakeLists.txt
+++ b/dlib/CMakeLists.txt
@@ -649,13 +649,144 @@ if (NOT TARGET dlib)
 
 
       if (DLIB_USE_CUDA)
-         find_package(CUDAToolkit 9.1)
-         set(CMAKE_MODULE_PATH ${CMAKE_MODULE_PATH} ${CMAKE_CURRENT_SOURCE_DIR}/cmake_utils)
-         find_package(CUDNN)
-         find_package(OpenMP)
-         set(openmp_libraries ${OpenMP_CXX_FLAGS}) 
+         find_package(CUDAToolkit QUIET)
 
-         if (CUDAToolkit_FOUND AND CUDNN_FOUND AND OPENMP_FOUND)
+         if (CUDA_VERSION VERSION_GREATER 9.1 AND CMAKE_VERSION VERSION_LESS 3.12.2)
+            # This bit of weirdness is to work around a bug in cmake 
+            list(REMOVE_ITEM CUDA_CUBLAS_LIBRARIES "CUDA_cublas_device_LIBRARY-NOTFOUND")
+         endif()
+
+
+         if (CUDAToolkit_FOUND AND MSVC AND NOT TARGET CUDA::cublas AND "${CMAKE_SIZEOF_VOID_P}" EQUAL "4")
+            message(WARNING "You have CUDA installed, but we can't use it unless you put visual studio in 64bit mode.")
+            set(CUDA_FOUND 0)
+         endif()
+
+         if (CUDAToolkit_FOUND)
+
+            # There is some bug in cmake that causes it to mess up the
+            # -std=c++11 option if you let it propagate it to nvcc in some
+            # cases.  So instead we disable this and manually include
+            # things from CMAKE_CXX_FLAGS in the CUDA_NVCC_FLAGS list below.
+            if (APPLE)
+               set(CUDA_PROPAGATE_HOST_FLAGS OFF)
+               # Grab all the -D flags from CMAKE_CXX_FLAGS so we can pass them
+               # to nvcc.
+               string(REGEX MATCHALL "-D[^ ]*" FLAGS_FOR_NVCC "${CMAKE_CXX_FLAGS}")
+
+               # Check if we are being built as part of a pybind11 module. 
+               if (COMMAND pybind11_add_module)
+                  # Don't export unnecessary symbols.
+                  list(APPEND FLAGS_FOR_NVCC "-Xcompiler=-fvisibility=hidden")
+               endif()
+            endif()
+
+            set(CUDA_HOST_COMPILATION_CPP ON)
+            string(REPLACE "," ";" DLIB_CUDA_COMPUTE_CAPABILITIES ${DLIB_USE_CUDA_COMPUTE_CAPABILITIES})
+            foreach(CAP ${DLIB_CUDA_COMPUTE_CAPABILITIES})
+                list(APPEND CUDA_NVCC_FLAGS "-gencode arch=compute_${CAP},code=[sm_${CAP},compute_${CAP}]")
+            endforeach()
+            # Note that we add __STRICT_ANSI__ to avoid freaking out nvcc with gcc specific
+            # magic in the standard C++ header files (since nvcc uses gcc headers on linux).
+            list(APPEND CUDA_NVCC_FLAGS "-D__STRICT_ANSI__;-D_MWAITXINTRIN_H_INCLUDED;-D_FORCE_INLINES;${FLAGS_FOR_NVCC}")
+            list(APPEND CUDA_NVCC_FLAGS ${active_preprocessor_switches})
+            if (NOT DLIB_IN_PROJECT_BUILD)
+               LIST(APPEND CUDA_NVCC_FLAGS -DDLIB__CMAKE_GENERATED_A_CONFIG_H_FILE)
+            endif()
+            if (NOT MSVC)
+               list(APPEND CUDA_NVCC_FLAGS "-std=c++14")
+            endif()
+            if (CMAKE_POSITION_INDEPENDENT_CODE)
+               # sometimes this setting isn't propagated to NVCC, which then causes the
+               # compile to fail.  So make sure it's propagated.
+               if (NOT MSVC) # Visual studio doesn't have -fPIC so don't do it in that case.
+                  list(APPEND CUDA_NVCC_FLAGS "-Xcompiler -fPIC")
+               endif()
+            endif()
+
+            include(cmake_utils/test_for_cudnn/find_cudnn.txt)
+
+            include(CheckLanguage)
+            check_language(CUDA)
+            set(cuda_compiler_found OFF)
+            if (CMAKE_CUDA_COMPILER)
+               set(cuda_compiler_found ON)
+            else()
+               message(STATUS " *** Cannot find CUDA compiler. If you are on windows using Visual Studio, make sure to install 'Visual Studio Integration' in the cuda installer.")
+            endif()
+
+            if (cudnn AND cudnn_include AND cuda_compiler_found AND NOT DEFINED cuda_test_compile_worked AND NOT DEFINED cudnn_test_compile_worked)
+               # make sure cuda is really working by doing a test compile
+               enable_language(CUDA)
+               message(STATUS "Building a CUDA test project to see if your compiler is compatible with CUDA...")
+
+               set(CUDA_TEST_CMAKE_FLAGS 
+                  "-DCMAKE_PREFIX_PATH=${CMAKE_PREFIX_PATH}"
+                  "-DCMAKE_INCLUDE_PATH=${CMAKE_INCLUDE_PATH}"
+                  "-DCMAKE_LIBRARY_PATH=${CMAKE_LIBRARY_PATH}")
+
+               if (NOT MSVC) # see https://github.com/davisking/dlib/issues/363
+                  list(APPEND CUDA_TEST_CMAKE_FLAGS "-DCUDA_HOST_COMPILER=${CUDA_HOST_COMPILER}")
+               endif()
+
+               try_compile(cuda_test_compile_worked 
+                  ${PROJECT_BINARY_DIR}/cuda_test_build 
+                  ${PROJECT_SOURCE_DIR}/cmake_utils/test_for_cuda cuda_test
+                  CMAKE_FLAGS ${CUDA_TEST_CMAKE_FLAGS}
+                  OUTPUT_VARIABLE try_compile_output_message
+                  )
+               if (NOT cuda_test_compile_worked)
+                  string(REPLACE "\n" "\n   ***   " try_compile_output_message "${try_compile_output_message}")
+                  message(STATUS "*****************************************************************************************************************")
+                  message(STATUS "*** CUDA was found but your compiler failed to compile a simple CUDA program so dlib isn't going to use CUDA. ")
+                  message(STATUS "*** The output of the failed CUDA test compile is shown below: ")
+                  message(STATUS "*** ")
+                  message(STATUS "***   ${try_compile_output_message}")
+                  message(STATUS "*****************************************************************************************************************")
+               else()
+                  message(STATUS "Building a cuDNN test project to check if you have the right version of cuDNN installed...")
+                  try_compile(cudnn_test_compile_worked 
+                     ${PROJECT_BINARY_DIR}/cudnn_test_build 
+                     ${PROJECT_SOURCE_DIR}/cmake_utils/test_for_cudnn cudnn_test
+                     CMAKE_FLAGS ${CUDA_TEST_CMAKE_FLAGS}
+                     OUTPUT_VARIABLE try_compile_output_message
+                     )
+                  if (NOT cudnn_test_compile_worked)
+                     string(REPLACE "\n" "\n   ***   " try_compile_output_message "${try_compile_output_message}")
+                     message(STATUS "*****************************************************************************************************")
+                     message(STATUS "*** Found cuDNN, but we failed to compile the dlib/cmake_utils/test_for_cudnn project. ")
+                     message(STATUS "*** You either have an unsupported version of cuDNN or something is wrong with your cudDNN install.")
+                     message(STATUS "*** Since a functional cuDNN is not found DLIB WILL NOT USE CUDA. ")
+                     message(STATUS "*** The output of the failed test_for_cudnn build is: ")
+                     message(STATUS "*** ")
+                     message(STATUS "***   ${try_compile_output_message}")
+                     message(STATUS "*****************************************************************************************************")
+                  endif()
+               endif()
+            endif()
+            # Also find OpenMP since cuSOLVER needs it.  Importantly, we only
+            # look for one to link to if our use of BLAS, specifically the
+            # Intel MKL, hasn't already decided what to use.  This is because
+            # it makes the MKL bug out if you link to another openmp lib other
+            # than Intel's when you use the MKL. I'm also not really sure when
+            # explicit linking to openmp became unnecessary, but for
+            # sufficiently older versions of cuda it was needed.  Then in
+            # versions of cmake newer than 3.11 linking to openmp started to
+            # mess up the switches passed to nvcc, so you can't just leave
+            # these "try to link to openmp" statements here going forward.  Fun
+            # times.
+            if (CUDA_VERSION VERSION_LESS "9.1" AND NOT openmp_libraries AND NOT MSVC AND NOT XCODE AND NOT APPLE)
+               find_package(OpenMP)
+               if (OPENMP_FOUND)
+                  set(openmp_libraries ${OpenMP_CXX_FLAGS}) 
+               else()
+                  message(STATUS "*** Didn't find OpenMP, which is required to use CUDA. ***")
+                  set(CUDA_FOUND 0)
+               endif()
+            endif()
+         endif()
+
+         if (CUDAToolkit_FOUND AND cudnn AND cuda_test_compile_worked AND cudnn_test_compile_worked AND cudnn_include)
             set(source_files ${source_files} 
                cuda/cuda_dlib.cu 
                cuda/cudnn_dlibapi.cpp
@@ -665,19 +796,22 @@ if (NOT TARGET dlib)
                cuda/cuda_data_ptr.cpp
                cuda/gpu_data.cpp
                )
+            list (APPEND dlib_needed_private_libraries CUDA::toolkit)
             list (APPEND dlib_needed_private_libraries CUDA::cublas)
-            list (APPEND dlib_needed_private_libraries ${CUDNN_LIBRARY_PATH})
+            list (APPEND dlib_needed_private_libraries ${cudnn})
             list (APPEND dlib_needed_private_libraries CUDA::curand)
             list (APPEND dlib_needed_private_libraries CUDA::cusolver)
             list (APPEND dlib_needed_private_libraries CUDA::cudart)
-            list (APPEND dlib_needed_private_libraries ${openmp_libraries})
+            if(openmp_libraries)
+               list (APPEND dlib_needed_private_libraries ${openmp_libraries})
+            endif()
 
-            include_directories(${CUDAToolkit_INCLUDE_DIRS} ${CUDNN_INCLUDE_PATH})
+            include_directories(${cudnn_include})
             message(STATUS "Enabling CUDA support for dlib.  DLIB WILL USE CUDA, compute capabilities: ${DLIB_CUDA_COMPUTE_CAPABILITIES}")
          else()
             set(DLIB_USE_CUDA OFF CACHE BOOL ${DLIB_USE_BLAS_STR} FORCE )
             toggle_preprocessor_switch(DLIB_USE_CUDA)
-            if (NOT CUDA_FOUND)
+            if (NOT CUDAToolkit_FOUND)
                message(STATUS "DID NOT FIND CUDA")
             endif()
             message(STATUS "Disabling CUDA support for dlib.  DLIB WILL NOT USE CUDA")
@@ -762,6 +896,9 @@ if (NOT TARGET dlib)
       target_compile_options(dlib PRIVATE "-DDLIB_CHECK_FOR_VERSION_MISMATCH=${DLIB_CHECK_FOR_VERSION_MISMATCH}")
    endif()
 
+   if (DLIB_USE_CUDA)
+      set_target_properties(dlib PROPERTIES CUDA_ARCHITECTURES ${DLIB_CUDA_COMPUTE_CAPABILITIES})
+   endif()
 
    # Allow the unit tests to ask us to compile the all/source.cpp file just to make sure it compiles.
    if (DLIB_TEST_COMPILE_ALL_SOURCE_CPP)
diff --git a/dlib/cmake_utils/findCUDNN.cmake b/dlib/cmake_utils/findCUDNN.cmake
deleted file mode 100644
index 0c900fcd4e..0000000000
--- a/dlib/cmake_utils/findCUDNN.cmake
+++ /dev/null
@@ -1,76 +0,0 @@
-# Find the CUDNN libraries
-#
-# The following variables are optionally searched for defaults
-#  CUDNN_ROOT: Base directory where CUDNN is found
-#  CUDNN_INCLUDE_DIR: Directory where CUDNN header is searched for
-#  CUDNN_LIBRARY: Directory where CUDNN library is searched for
-#  CUDNN_STATIC: Are we looking for a static library? (default: no)
-#
-# The following are set after configuration is done:
-#  CUDNN_FOUND
-#  CUDNN_INCLUDE_PATH
-#  CUDNN_LIBRARY_PATH
-#
-
-include(FindPackageHandleStandardArgs)
-
-set(CUDNN_ROOT $ENV{CUDNN_ROOT_DIR} CACHE PATH "Folder containing NVIDIA cuDNN")
-if (DEFINED $ENV{CUDNN_ROOT_DIR})
-  message(WARNING "CUDNN_ROOT_DIR is deprecated. Please set CUDNN_ROOT instead.")
-endif()
-list(APPEND CUDNN_ROOT $ENV{CUDNN_ROOT_DIR} ${CUDA_TOOLKIT_ROOT_DIR})
-
-# Compatible layer for CMake <3.12. CUDNN_ROOT will be accounted in for searching paths and libraries for CMake >=3.12.
-list(APPEND CMAKE_PREFIX_PATH ${CUDNN_ROOT})
-
-set(CUDNN_INCLUDE_DIR $ENV{CUDNN_INCLUDE_DIR} CACHE PATH "Folder containing NVIDIA cuDNN header files")
-
-find_path(CUDNN_INCLUDE_PATH cudnn.h
-  HINTS ${CUDNN_INCLUDE_DIR}
-  PATH_SUFFIXES cuda/include cuda include)
-
-option(CUDNN_STATIC "Look for static CUDNN" OFF)
-if (CUDNN_STATIC)
-  set(CUDNN_LIBNAME "libcudnn_static.a")
-else()
-  set(CUDNN_LIBNAME "cudnn")
-endif()
-
-set(CUDNN_LIBRARY $ENV{CUDNN_LIBRARY} CACHE PATH "Path to the cudnn library file (e.g., libcudnn.so)")
-if (CUDNN_LIBRARY MATCHES ".*cudnn_static.a" AND NOT CUDNN_STATIC)
-  message(WARNING "CUDNN_LIBRARY points to a static library (${CUDNN_LIBRARY}) but CUDNN_STATIC is OFF.")
-endif()
-
-find_library(CUDNN_LIBRARY_PATH ${CUDNN_LIBNAME}
-  PATHS ${CUDNN_LIBRARY}
-  PATH_SUFFIXES lib lib64 cuda/lib cuda/lib64 lib/x64)
-
-find_package_handle_standard_args(CUDNN DEFAULT_MSG CUDNN_LIBRARY_PATH CUDNN_INCLUDE_PATH)
-
-if(CUDNN_FOUND)
-  # Get cuDNN version
-  if(EXISTS ${CUDNN_INCLUDE_PATH}/cudnn_version.h)
-    file(READ ${CUDNN_INCLUDE_PATH}/cudnn_version.h CUDNN_HEADER_CONTENTS)
-  else()
-    file(READ ${CUDNN_INCLUDE_PATH}/cudnn.h CUDNN_HEADER_CONTENTS)
-  endif()
-  string(REGEX MATCH "define CUDNN_MAJOR * +([0-9]+)"
-               CUDNN_VERSION_MAJOR "${CUDNN_HEADER_CONTENTS}")
-  string(REGEX REPLACE "define CUDNN_MAJOR * +([0-9]+)" "\\1"
-               CUDNN_VERSION_MAJOR "${CUDNN_VERSION_MAJOR}")
-  string(REGEX MATCH "define CUDNN_MINOR * +([0-9]+)"
-               CUDNN_VERSION_MINOR "${CUDNN_HEADER_CONTENTS}")
-  string(REGEX REPLACE "define CUDNN_MINOR * +([0-9]+)" "\\1"
-               CUDNN_VERSION_MINOR "${CUDNN_VERSION_MINOR}")
-  string(REGEX MATCH "define CUDNN_PATCHLEVEL * +([0-9]+)"
-               CUDNN_VERSION_PATCH "${CUDNN_HEADER_CONTENTS}")
-  string(REGEX REPLACE "define CUDNN_PATCHLEVEL * +([0-9]+)" "\\1"
-               CUDNN_VERSION_PATCH "${CUDNN_VERSION_PATCH}")
-  # Assemble cuDNN version
-  if(NOT CUDNN_VERSION_MAJOR)
-    set(CUDNN_VERSION "?")
-  else()
-    set(CUDNN_VERSION
-        "${CUDNN_VERSION_MAJOR}.${CUDNN_VERSION_MINOR}.${CUDNN_VERSION_PATCH}")
-  endif()
-endif()
\ No newline at end of file
diff --git a/dlib/cmake_utils/test_for_cuda/CMakeLists.txt b/dlib/cmake_utils/test_for_cuda/CMakeLists.txt
index 7392501f15..f0a0fd0b00 100644
--- a/dlib/cmake_utils/test_for_cuda/CMakeLists.txt
+++ b/dlib/cmake_utils/test_for_cuda/CMakeLists.txt
@@ -7,8 +7,9 @@ add_definitions(-DDLIB_USE_CUDA)
 
 # Override the FindCUDA.cmake setting to avoid duplication of host flags if using a toolchain:
 option(CUDA_PROPAGATE_HOST_FLAGS "Propage C/CXX_FLAGS and friends to the host compiler via -Xcompile" OFF)
-find_package(CUDA 7.5 REQUIRED)
+find_package(CUDAToolkit 7.5 REQUIRED)
 set(CUDA_HOST_COMPILATION_CPP ON)
 list(APPEND CUDA_NVCC_FLAGS "-arch=sm_50;-std=c++14;-D__STRICT_ANSI__;-D_MWAITXINTRIN_H_INCLUDED;-D_FORCE_INLINES")
 
-cuda_add_library(cuda_test STATIC cuda_test.cu )
+enable_language(CUDA)
+add_library(cuda_test STATIC cuda_test.cu )
diff --git a/dlib/cmake_utils/test_for_cudnn/CMakeLists.txt b/dlib/cmake_utils/test_for_cudnn/CMakeLists.txt
index 4a0315a327..9bfed91430 100644
--- a/dlib/cmake_utils/test_for_cudnn/CMakeLists.txt
+++ b/dlib/cmake_utils/test_for_cudnn/CMakeLists.txt
@@ -4,7 +4,7 @@ project(cudnn_test)
 
 # Override the FindCUDA.cmake setting to avoid duplication of host flags if using a toolchain:
 option(CUDA_PROPAGATE_HOST_FLAGS "Propage C/CXX_FLAGS and friends to the host compiler via -Xcompile" OFF)
-find_package(CUDA 7.5 REQUIRED)
+find_package(CUDAToolkit 7.5 REQUIRED)
 set(CUDA_HOST_COMPILATION_CPP ON)
 list(APPEND CUDA_NVCC_FLAGS "-arch=sm_50;-std=c++14;-D__STRICT_ANSI__")
 add_definitions(-DDLIB_USE_CUDA)
@@ -13,6 +13,7 @@ include(find_cudnn.txt)
 
 if (cudnn_include AND cudnn)
    include_directories(${cudnn_include})
-   cuda_add_library(cudnn_test STATIC ../../cuda/cudnn_dlibapi.cpp ${cudnn} )
+   enable_language(CUDA)
+   add_library(cudnn_test STATIC ../../cuda/cudnn_dlibapi.cpp ${cudnn} )
    target_compile_features(cudnn_test PUBLIC cxx_std_14)
 endif()

From d6780734f2260d320a0fe01b97e7f08f02e7df74 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Matthias=20M=C3=B6ller?= <m_moeller@live.de>
Date: Sun, 8 Sep 2024 20:13:09 +0200
Subject: [PATCH 4/5] links cudnn tests against cuda runtime

---
 dlib/cmake_utils/test_for_cudnn/CMakeLists.txt | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/dlib/cmake_utils/test_for_cudnn/CMakeLists.txt b/dlib/cmake_utils/test_for_cudnn/CMakeLists.txt
index 9bfed91430..c1a12abafe 100644
--- a/dlib/cmake_utils/test_for_cudnn/CMakeLists.txt
+++ b/dlib/cmake_utils/test_for_cudnn/CMakeLists.txt
@@ -16,4 +16,6 @@ if (cudnn_include AND cudnn)
    enable_language(CUDA)
    add_library(cudnn_test STATIC ../../cuda/cudnn_dlibapi.cpp ${cudnn} )
    target_compile_features(cudnn_test PUBLIC cxx_std_14)
+   target_link_libraries(cudnn_test PRIVATE CUDA::toolkit)
+   target_link_libraries(cudnn_test PRIVATE CUDA::cudart)
 endif()

From 426c93d36ae112d5c2d27573256d3633c0b21315 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Matthias=20M=C3=B6ller?= <m_moeller@live.de>
Date: Sat, 14 Sep 2024 22:46:06 +0200
Subject: [PATCH 5/5] add cudatoolkit root for test compilation

---
 dlib/CMakeLists.txt | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/dlib/CMakeLists.txt b/dlib/CMakeLists.txt
index 491b78f57e..e4da4a8bf5 100644
--- a/dlib/CMakeLists.txt
+++ b/dlib/CMakeLists.txt
@@ -723,7 +723,8 @@ if (NOT TARGET dlib)
                set(CUDA_TEST_CMAKE_FLAGS 
                   "-DCMAKE_PREFIX_PATH=${CMAKE_PREFIX_PATH}"
                   "-DCMAKE_INCLUDE_PATH=${CMAKE_INCLUDE_PATH}"
-                  "-DCMAKE_LIBRARY_PATH=${CMAKE_LIBRARY_PATH}")
+                  "-DCMAKE_LIBRARY_PATH=${CMAKE_LIBRARY_PATH}"
+                  "-DCUDAToolkit_ROOT=${CUDAToolkit_ROOT}")
 
                if (NOT MSVC) # see https://github.com/davisking/dlib/issues/363
                   list(APPEND CUDA_TEST_CMAKE_FLAGS "-DCUDA_HOST_COMPILER=${CUDA_HOST_COMPILER}")