Add iomp on windows

.github/actions/deploy-windows/action.yml

@@ -29,9 +29,9 @@ runs:
         rm "C:/msys64/mingw32/bin/clang-cl.exe" "C:/msys64/mingw64/bin/clang-cl.exe" "C:/msys64/mingw32/bin/cmake.exe" "C:/msys64/mingw64/bin/cmake.exe"
         rm "C:/Strawberry/c/lib/libz.a" "C:/Strawberry/c/lib/libzlib.a" "C:/Strawberry/c/lib/libzdll.a" "C:/Strawberry/c/bin/cmake.exe"
 
-        curl -LO https://github.com/Kitware/CMake/releases/download/v3.26.4/cmake-3.26.4-win64-x64.msi
-        msiexec /i cmake-3.26.4-win64-x64.msi /qn
-        rm cmake-3.26.4-win64-x64.msi
+        rem curl -LO https://github.com/Kitware/CMake/releases/download/v3.30.0-rc3/cmake-3.30.0-rc3-win64-x64.msi
+        rem msiexec /i cmake-3.30.0-rc3-win64-x64.msi /qn
+        rem rm cmake-3.30.0-rc3-win64-x64.msi
         
         curl -LO https://archive.apache.org/dist/maven/maven-3/3.6.3/binaries/apache-maven-3.6.3-bin.tar.gz
         bash -c "tar -xzf apache-maven-3.6.3-bin.tar.gz -C 'C:/Program Files/'"
@@ -159,10 +159,15 @@ runs:
           move TensorRT-10.0.1.6 "%ProgramFiles%\NVIDIA GPU Computing Toolkit\TensorRT"
         )
 
-        if "%CI_DEPLOY_MODULE%"=="mkl" (
+        if not "%CI_DEPLOY_NEED_MKL%"=="" (
           echo Installing MKL
-          curl -LO https://registrationcenter-download.intel.com/akdlm/IRC_NAS/b5a4ff98-2c98-4b28-87f7-32082cac359e/w_onemkl_p_2024.0.0.49672_offline.exe
-          w_onemkl_p_2024.0.0.49672_offline.exe -s -a -s --eula accept
+          if "%CI_DEPLOY_NEED_MKL%"=="2022" (
+            curl -LO https://registrationcenter-download.intel.com/akdlm/IRC_NAS/18899/w_onemkl_p_2022.2.0.9563_offline.exe
+            w_onemkl_p_2022.2.0.9563_offline.exe -s -a -s --eula accept
+          ) else (
+            curl -LO https://registrationcenter-download.intel.com/akdlm/IRC_NAS/b5a4ff98-2c98-4b28-87f7-32082cac359e/w_onemkl_p_2024.0.0.49672_offline.exe
+            w_onemkl_p_2024.0.0.49672_offline.exe -s -a -s --eula accept
+          )
         )
 
         if "%CI_DEPLOY_PLATFORM%"=="windows-x86" if "%CI_DEPLOY_MODULE%"=="flycapture" (

.github/workflows/mkl.yml

@@ -16,6 +16,7 @@ env:
   CI_DEPLOY_USERNAME: ${{ secrets.CI_DEPLOY_USERNAME }}
   CI_DEPLOY_PASSWORD: ${{ secrets.CI_DEPLOY_PASSWORD }}
   STAGING_REPOSITORY: ${{ secrets.STAGING_REPOSITORY }}
+  CI_DEPLOY_NEED_MKL: 1
 jobs:
 #  linux-x86:
 #    runs-on: ubuntu-20.04

pytorch/cppbuild.sh

@@ -61,8 +61,13 @@ if [[ $PLATFORM == windows* ]]; then
         cmake --build . --config Release
         cmake --install . --config Release --prefix ../dist
         cd ../..
+
+        #nuget install  intelopenmp.devel.win -Version 2024.1.0.964
+        #export CMAKE_INCLUDE_PATH=${INSTALL_PATH}/intelopenmp.devel.win.2024.1.0.964/build/native/include
+        #export CMAKE_LIBRARY_PATH="${INSTALL_PATH}/intelopenmp.devel.win.2024.1.0.964/build/native/win-x64;${INSTALL_PATH}/intelopenmp.redist.win.2024.1.0.964/runtimes/win-x64/native"
+
     fi
-    export libuv_ROOT=`pwd`/libuv/dist
+    export libuv_ROOT=${INSTALL_PATH}/libuv/dist
 fi
 
 if [[ ! -d pytorch ]]; then
@@ -205,9 +210,23 @@ sedinplace 's/char(\(.*\))/\1/g' torch/csrc/jit/serialization/pickler.h
 # some windows header defines a macro named "interface"
 sedinplace 's/const std::string& interface)/const std::string\& interface_name)/g' torch/csrc/distributed/c10d/ProcessGroupGloo.hpp
 
+if [[ $PLATFORM == windows* ]]; then
+    # Remove pytorch adaptations of FindOpenMP.cmake that, without iomp and with
+    # new versions of VS 2019 including -openmp:experimental and libomp, causes
+    # final binary to be linked to both libomp and vcomp and produce incorrect results.
+    # Wait for eventual upstream fix, or for cmake 2.30 that allows to choose between -openmp and -openmp:experimental
+    # and see if choosing experimental works.
+    rm cmake/Modules/FindOpenMP.cmake
+    sedinplace 's/include(${CMAKE_CURRENT_LIST_DIR}\/Modules\/FindOpenMP.cmake)/find_package(OpenMP)/g' cmake/Dependencies.cmake
+fi
+
 #USE_FBGEMM=0 USE_KINETO=0 USE_GLOO=0 USE_MKLDNN=0 \
 "$PYTHON_BIN_PATH" setup.py build
 
+echo "CMAKECACHE"
+cat $BUILD_DIR/CMakeCache.txt
+cat $BUILD_DIR/CMakeFiles/CMakeConfigureLog.yaml
+
 rm -Rf ../lib
 if [[ ! -e torch/include/gloo ]]; then
     ln -sf ../../third_party/gloo/gloo torch/include
@@ -216,9 +235,9 @@ ln -sf pytorch/torch/include ../include
 ln -sf pytorch/torch/lib ../lib
 ln -sf pytorch/torch/bin ../bin
 
-# fix library with correct rpath on Mac
 case $PLATFORM in
     macosx-*)
+        # fix library with correct rpath
         cp /usr/local/lib/libomp.dylib ../lib/libiomp5.dylib
         chmod +w ../lib/libiomp5.dylib
         install_name_tool -id @rpath/libiomp5.dylib ../lib/libiomp5.dylib

pytorch/pom.xml

@@ -72,9 +72,15 @@
         <dependencies>
           <dependency>
             <groupId>org.bytedeco</groupId>
-            <artifactId>openblas-platform</artifactId>
+            <artifactId>openblas</artifactId>
             <version>0.3.27-${project.parent.version}</version>
           </dependency>
+          <dependency>
+            <groupId>org.bytedeco</groupId>
+            <artifactId>openblas</artifactId>
+            <version>0.3.27-${project.parent.version}</version>
+            <classifier>${javacpp.platform}</classifier>
+          </dependency>
           <dependency>
             <groupId>org.bytedeco</groupId>
             <artifactId>cuda-platform</artifactId>

pytorch/src/main/java/org/bytedeco/pytorch/presets/torch.java

@@ -86,7 +86,7 @@
             },
             exclude = {"openblas_config.h", "cblas.h", "lapacke_config.h", "lapacke_mangling.h", "lapack.h", "lapacke.h", "lapacke_utils.h"},
             link = {"c10", "torch_cpu", "torch"},
-            preload = {"[email protected]", "iomp5", "omp", "[email protected]", "asmjit", "fbgemm", "uv"}
+            preload = {"[email protected]", "iomp5", "omp", "[email protected]", "asmjit", "fbgemm", "uv", "iomp5md"}
         ),
         @Platform(
             value = {"linux", "macosx", "windows"},