vsort: upgrade to onnx runtime 1.23.0 and cudnn 9.12.0

Author: WolframRhodium

.github/workflows/windows-cuda-dependency.yml

@@ -29,7 +29,7 @@ jobs:
 
     steps:
     - name: Download cuDNN inference library
-      run: curl -LJ https://developer.download.nvidia.com/compute/cudnn/redist/cudnn/windows-x86_64/cudnn-windows-x86_64-8.9.7.29_cuda12-archive.zip -o cudnn.zip
+      run: curl -LJ https://developer.download.nvidia.com/compute/cudnn/redist/cudnn/windows-x86_64/cudnn-windows-x86_64-9.12.0.46_cuda12-archive.zip -o cudnn.zip
 
     - name: Extract cuDNN library
       run: unzip cudnn.zip
@@ -38,7 +38,7 @@ jobs:
       run: |
         mkdir -p vsmlrt-cuda
         mv cudnn-windows-*/bin/*.dll vsmlrt-cuda/ -v
-        rm vsmlrt-cuda/cudnn_*_train*.dll -v
+        # rm vsmlrt-cuda/cudnn_*_train*.dll -v
 
     - name: Download TensorRT library
       run: |

.github/workflows/windows-ort.yml

@@ -20,7 +20,7 @@ on:
 
 jobs:
   build-windows:
-    runs-on: windows-2022
+    runs-on: windows-2025
 
     defaults:
       run:
@@ -39,44 +39,12 @@ jobs:
     - name: Setup Ninja
       run: pip install ninja
 
-    - name: Cache protobuf
-      id: cache-protobuf
-      uses: actions/cache@v4
-      with:
-        path: vsort/protobuf/install
-        key: ${{ runner.os }}-vsort-protobuf-v4
-
-    - name: Checkout protobuf
-      uses: actions/checkout@v4
-      if: steps.cache-protobuf.outputs.cache-hit != 'true'
-      with:
-        repository: protocolbuffers/protobuf
-        # follows protobuf in https://github.com/AmusementClub/onnxruntime/blob/master/cmake/external/onnxruntime_external_deps.cmake#L203
-        # if you change this, remember to bump the version of the cache key.
-        ref: v3.21.12
-        fetch-depth: 1
-        path: vsort/protobuf
-
-    - name: Configure protobuf
-      if: steps.cache-protobuf.outputs.cache-hit != 'true'
-      run: cmake -S protobuf -B protobuf\build_rel -G Ninja -LA
-        -D CMAKE_BUILD_TYPE=Release
-        -D protobuf_BUILD_SHARED_LIBS=OFF  -D protobuf_BUILD_TESTS=OFF
-
-    - name: Build protobuf
-      if: steps.cache-protobuf.outputs.cache-hit != 'true'
-      run: cmake --build protobuf\build_rel --verbose
-
-    - name: Install protobuf
-      if: steps.cache-protobuf.outputs.cache-hit != 'true'
-      run: cmake --install protobuf\build_rel --prefix protobuf\install
-
-    - name: Cache onnx
+    - name: Restore cached onnx
       id: cache-onnx
-      uses: actions/cache@v4
+      uses: actions/cache/restore@v4
       with:
         path: vsort/onnx/install
-        key: ${{ runner.os }}-vsort-onnx-v5
+        key: ${{ runner.os }}-vsort-onnx-v6
 
     - name: Checkout onnx
       if: steps.cache-onnx.outputs.cache-hit != 'true'
@@ -85,20 +53,19 @@ jobs:
         repository: onnx/onnx
         # follows onnx in https://github.com/AmusementClub/onnxruntime/tree/master/cmake/external
         # if you change this, remember to bump the version of the cache key.
-        ref: 990217f043af7222348ca8f0301e17fa7b841781
+        ref: v1.19.0
         fetch-depth: 1
         path: vsort/onnx
 
     - name: Configure onnx
       if: steps.cache-onnx.outputs.cache-hit != 'true'
       run: cmake -S onnx -B onnx\build -G Ninja -LA
         -D CMAKE_BUILD_TYPE=Release
-        -D Protobuf_PROTOC_EXECUTABLE=protobuf\install\bin\protoc
-        -D Protobuf_LITE_LIBRARY=protobuf\install\lib
-        -D Protobuf_LIBRARIES=protobuf\install\lib
+        -D CMAKE_PREFIX_PATH=%cd%\protobuf\install\lib\cmake
         -D ONNX_USE_LITE_PROTO=ON -D ONNX_USE_PROTOBUF_SHARED_LIBS=OFF
         -D ONNX_GEN_PB_TYPE_STUBS=OFF -D ONNX_ML=0
         -D ONNX_USE_MSVC_STATIC_RUNTIME=1
+        -D ONNX_BUILD_CUSTOM_PROTOBUF=ON
 
     - name: Build onnx
       if: steps.cache-onnx.outputs.cache-hit != 'true'
@@ -108,6 +75,13 @@ jobs:
       if: steps.cache-onnx.outputs.cache-hit != 'true'
       run: cmake --install onnx\build --prefix onnx\install
 
+    - name: Save onnx
+      if: steps.cache-onnx.outputs.cache-hit != 'true'
+      uses: actions/cache/save@v4
+      with:
+        path: vsort/onnx/install
+        key: ${{ steps.cache-onnx.outputs.cache-primary-key }}
+
     - name: Download VapourSynth headers
       run: |
         curl -s -o vs.zip -L https://github.com/vapoursynth/vapoursynth/archive/refs/tags/R54.zip
@@ -116,33 +90,39 @@ jobs:
 
     - name: Download ONNX Runtime Precompilation
       run: |
-        curl -s -o ortgpu.zip -LJO https://github.com/AmusementClub/onnxruntime/releases/download/orttraining_rc2-8036-geb41d57f21-240425-0428/onnxruntime-gpu-win64.zip
+        curl -s -o ortgpu.zip -LJO https://github.com/AmusementClub/onnxruntime/releases/download/orttraining_rc2-10488-g4754a1d64e-250831-1153/onnxruntime-gpu-win64.zip
         unzip -q ortgpu.zip
 
-    - name: Cache CUDA
+    - name: Restore cached CUDA
       id: cache-cuda
-      uses: actions/cache@v4
+      uses: actions/cache/restore@v4
       with:
         path: C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA
-        key: ${{ runner.os }}-cuda-12.4.1
+        key: ${{ runner.os }}-cuda-12.9.1
 
     - name: Setup CUDA
       if: steps.cache-cuda.outputs.cache-hit != 'true'
       run: |
-        curl -s -o cuda_installer.exe -L https://developer.download.nvidia.com/compute/cuda/12.4.1/network_installers/cuda_12.4.1_windows_network.exe
-        cuda_installer.exe -s nvcc_12.4 cudart_12.4
+        curl -s -o cuda_installer.exe -L https://developer.download.nvidia.com/compute/cuda/12.9.1/network_installers/cuda_12.9.1_windows_network.exe
+        cuda_installer.exe -s nvcc_12.9 cudart_12.9
+
+    - name: Save CUDA
+      if: steps.cache-cuda.outputs.cache-hit != 'true'
+      uses: actions/cache/save@v4
+      with:
+        path: C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA
+        key: ${{ steps.cache-cuda.outputs.cache-primary-key }}
 
     - name: Configure
       run: cmake -S . -B build -G Ninja -LA
         -D CMAKE_BUILD_TYPE=Release
+        -D CMAKE_PREFIX_PATH=onnx\install
         -D CMAKE_MSVC_RUNTIME_LIBRARY=MultiThreaded
         -D VAPOURSYNTH_INCLUDE_DIRECTORY=vapoursynth\include
-        -D protobuf_DIR=protobuf\install\cmake
-        -D ONNX_DIR=onnx\install\lib\cmake\ONNX
         -D ONNX_RUNTIME_API_DIRECTORY=onnxruntime-gpu\include\onnxruntime
         -D ONNX_RUNTIME_LIB_DIRECTORY=onnxruntime-gpu\lib
         -D ENABLE_CUDA=1
-        -D CUDAToolkit_ROOT="C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4"
+        -D CUDAToolkit_ROOT="C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.9"
         -D ENABLE_DML=1
         -D CMAKE_CXX_STANDARD=20
 
@@ -161,7 +141,7 @@ jobs:
     - name: Download DirectML Library
       # follows DirectML in https://github.com/AmusementClub/onnxruntime/blob/master/cmake/external/dml.cmake#L44
       run: |
-        curl -s -o directml.nupkg -LJO https://www.nuget.org/api/v2/package/Microsoft.AI.DirectML/1.14.1
+        curl -s -o directml.nupkg -LJO https://www.nuget.org/api/v2/package/Microsoft.AI.DirectML/1.15.4
         unzip -q directml.nupkg -d dml
         copy dml\bin\x64-win\DirectML.dll artifact\vsort\
 
@@ -171,6 +151,12 @@ jobs:
         name: VSORT-Windows-x64
         path: vsort/artifact
 
+    - name: Describe
+      run: git describe --tags --long
+
+    - name: Dump dependencies
+      run: dumpbin /dependents artifact\vsort.dll
+
     - name: Setup Python portable
       run: |
         curl -s -o python.zip -LJO https://www.python.org/ftp/python/3.9.10/python-3.9.10-embed-amd64.zip
@@ -199,7 +185,7 @@ jobs:
 
     - name: Create script
       shell: bash
-      run: echo "import vapoursynth as vs;from vapoursynth import core;import sys;print(core.ort, file=sys.stderr);print(core.ort.Version(),file=sys.stderr);core.std.BlankClip(format=vs.RGBS).ort.Model(r\"waifu2x\\upconv_7_anime_style_art_rgb\\scale2.0x_model.onnx\", builtin=True).resize.Bicubic(format=vs.YUV420P10, matrix_s='709').set_output()" > test.vpy
+      run: echo "import vapoursynth as vs;from vapoursynth import core;import sys;print(core.ort, file=sys.stderr);print(core.ort.Version(),file=sys.stderr);core.std.BlankClip(format=vs.RGBS).ort.Model(r\"waifu2x\\upconv_7_anime_style_art_rgb\\scale2.0x_model.onnx\", builtin=True, verbosity=4).resize.Bicubic(format=vs.YUV420P10, matrix_s='709').set_output()" > test.vpy
 
     - name: Run vspipe
       shell: bash
@@ -277,12 +263,6 @@ jobs:
         grep -i 'error' x265.log && exit 1
         exit 0
 
-    - name: Describe
-      run: git describe --tags --long
-
-    - name: Dump dependencies
-      run: dumpbin /dependents artifact\vsort.dll
-
     - name: Compress artifact for release
       if: github.event_name == 'workflow_dispatch' && github.event.inputs.tag != ''
       run: |

.github/workflows/windows-release.yml

@@ -121,10 +121,11 @@ jobs:
         pushd models
         for tag in $(echo "${{ github.event.inputs.model-tags }}" | tr ',' ' '); do
           echo "Handling tag $tag"
-          curl -s https://api.github.com/repos/AmusementClub/vs-mlrt/releases/tags/"$tag" > release.json
+          curl -s -H 'Authorization: Bearer ${{ secrets.GITHUB_TOKEN }}' https://api.github.com/repos/AmusementClub/vs-mlrt/releases/tags/"$tag" > release.json
+          cat release.json
           for url in $(cat release.json | jq '.assets | .[] | .url ' | tr -d '"'); do
             echo "Downloading $url"
-            curl -o dl.7z -LJ -H 'Accept: application/octet-stream' "$url"
+            curl -o dl.7z -LJ -H 'Accept: application/octet-stream' -H 'Authorization: Bearer ${{ secrets.GITHUB_TOKEN }}' "$url"
             # later release should overwrite earlier ones
             7za x -y dl.7z
           done

vsort/win32.cpp

@@ -22,7 +22,11 @@ static std::vector<std::wstring> cudaDlls {
     L"cudart64",
     L"cublasLt64", L"cublas64",
     L"cufft64",
-    L"cudnn_ops_infer64", L"cudnn_cnn_infer64", L"cudnn_adv_infer64", L"cudnn64",
+    // follows the dependency graph in 
+    // https://docs.nvidia.com/deeplearning/cudnn/backend/v9.12.0/api/overview.html#backend-api-overview
+    L"cudnn_engines_precompiled64", L"cudnn_heuristic64", L"cudnn_engines_runtime_compiled64",
+    L"cudnn_graph64", L"cudnn_ops64", L"cudnn_cnn64", L"cudnn_adv64",
+    L"cudnn64",
     L"cupti64",
 };