runtime spir-v compilation with libglslang (#1779)

.github/workflows/android-armv7-gpu.yml

@@ -22,3 +22,27 @@ jobs:
       run: export PATH=`pwd`/1.1.114.0/x86_64/bin:$PATH && mkdir build && cd build && cmake -DCMAKE_TOOLCHAIN_FILE=$ANDROID_HOME/ndk-bundle/build/cmake/android.toolchain.cmake -DANDROID_ABI="armeabi-v7a" -DANDROID_ARM_NEON=ON -DANDROID_PLATFORM=android-24 -DNCNN_VULKAN=ON ..
     - name: build
       run: cmake --build build -j 2
+
+  android-armv7-gpu-online-spirv:
+    runs-on: ubuntu-latest
+    steps:
+    - uses: actions/checkout@v2
+      with:
+        submodules: 'true'
+    - name: cache-vulkansdk
+      id: cache-vulkansdk
+      uses: actions/cache@v1
+      with:
+        path: "1.1.114.0"
+        key: vulkansdk-linux-x86_64-1.1.114.0
+    - name: vulkansdk
+      if: steps.cache-vulkansdk.outputs.cache-hit != 'true'
+      run: |
+        wget https://sdk.lunarg.com/sdk/download/1.1.114.0/linux/vulkansdk-linux-x86_64-1.1.114.0.tar.gz?Human=true -O vulkansdk-linux-x86_64-1.1.114.0.tar.gz
+        tar -xf vulkansdk-linux-x86_64-1.1.114.0.tar.gz
+        rm -rf 1.1.114.0/source 1.1.114.0/samples
+        find 1.1.114.0 -type f | grep -v -E 'vulkan|glslang' | xargs rm
+    - name: configure
+      run: export PATH=`pwd`/1.1.114.0/x86_64/bin:$PATH && mkdir build && cd build && cmake -DCMAKE_TOOLCHAIN_FILE=$ANDROID_HOME/ndk-bundle/build/cmake/android.toolchain.cmake -DANDROID_ABI="armeabi-v7a" -DANDROID_ARM_NEON=ON -DANDROID_PLATFORM=android-24 -DNCNN_VULKAN=ON -DNCNN_VULKAN_ONLINE_SPIRV=ON ..
+    - name: build
+      run: cmake --build build -j 2

.github/workflows/android-armv8-gpu.yml

@@ -22,3 +22,27 @@ jobs:
       run: export PATH=`pwd`/1.1.114.0/x86_64/bin:$PATH && mkdir build && cd build && cmake -DCMAKE_TOOLCHAIN_FILE=$ANDROID_HOME/ndk-bundle/build/cmake/android.toolchain.cmake -DANDROID_ABI="arm64-v8a" -DANDROID_PLATFORM=android-24 -DNCNN_VULKAN=ON ..
     - name: build
       run: cmake --build build -j 2
+
+  android-aarch64-gpu-online-spirv:
+    runs-on: ubuntu-latest
+    steps:
+    - uses: actions/checkout@v2
+      with:
+        submodules: 'true'
+    - name: cache-vulkansdk
+      id: cache-vulkansdk
+      uses: actions/cache@v1
+      with:
+        path: "1.1.114.0"
+        key: vulkansdk-linux-x86_64-1.1.114.0
+    - name: vulkansdk
+      if: steps.cache-vulkansdk.outputs.cache-hit != 'true'
+      run: |
+        wget https://sdk.lunarg.com/sdk/download/1.1.114.0/linux/vulkansdk-linux-x86_64-1.1.114.0.tar.gz?Human=true -O vulkansdk-linux-x86_64-1.1.114.0.tar.gz
+        tar -xf vulkansdk-linux-x86_64-1.1.114.0.tar.gz
+        rm -rf 1.1.114.0/source 1.1.114.0/samples
+        find 1.1.114.0 -type f | grep -v -E 'vulkan|glslang' | xargs rm
+    - name: configure
+      run: export PATH=`pwd`/1.1.114.0/x86_64/bin:$PATH && mkdir build && cd build && cmake -DCMAKE_TOOLCHAIN_FILE=$ANDROID_HOME/ndk-bundle/build/cmake/android.toolchain.cmake -DANDROID_ABI="arm64-v8a" -DANDROID_PLATFORM=android-24 -DNCNN_VULKAN=ON -DNCNN_VULKAN_ONLINE_SPIRV=ON ..
+    - name: build
+      run: cmake --build build -j 2

.github/workflows/android-x64-gpu.yml

@@ -22,3 +22,27 @@ jobs:
       run: export PATH=`pwd`/1.1.114.0/x86_64/bin:$PATH && mkdir build && cd build && cmake -DCMAKE_TOOLCHAIN_FILE=$ANDROID_HOME/ndk-bundle/build/cmake/android.toolchain.cmake -DANDROID_ABI="x86_64" -DANDROID_PLATFORM=android-24 -DNCNN_VULKAN=ON ..
     - name: build
       run: cmake --build build -j 2
+
+  android-x86_64-gpu-online-spirv:
+    runs-on: ubuntu-latest
+    steps:
+    - uses: actions/checkout@v2
+      with:
+        submodules: 'true'
+    - name: cache-vulkansdk
+      id: cache-vulkansdk
+      uses: actions/cache@v1
+      with:
+        path: "1.1.114.0"
+        key: vulkansdk-linux-x86_64-1.1.114.0
+    - name: vulkansdk
+      if: steps.cache-vulkansdk.outputs.cache-hit != 'true'
+      run: |
+        wget https://sdk.lunarg.com/sdk/download/1.1.114.0/linux/vulkansdk-linux-x86_64-1.1.114.0.tar.gz?Human=true -O vulkansdk-linux-x86_64-1.1.114.0.tar.gz
+        tar -xf vulkansdk-linux-x86_64-1.1.114.0.tar.gz
+        rm -rf 1.1.114.0/source 1.1.114.0/samples
+        find 1.1.114.0 -type f | grep -v -E 'vulkan|glslang' | xargs rm
+    - name: configure
+      run: export PATH=`pwd`/1.1.114.0/x86_64/bin:$PATH && mkdir build && cd build && cmake -DCMAKE_TOOLCHAIN_FILE=$ANDROID_HOME/ndk-bundle/build/cmake/android.toolchain.cmake -DANDROID_ABI="x86_64" -DANDROID_PLATFORM=android-24 -DNCNN_VULKAN=ON -DNCNN_VULKAN_ONLINE_SPIRV=ON ..
+    - name: build
+      run: cmake --build build -j 2

.github/workflows/android-x86-gpu.yml

@@ -22,3 +22,27 @@ jobs:
       run: export PATH=`pwd`/1.1.114.0/x86_64/bin:$PATH && mkdir build && cd build && cmake -DCMAKE_TOOLCHAIN_FILE=$ANDROID_HOME/ndk-bundle/build/cmake/android.toolchain.cmake -DANDROID_ABI="x86" -DANDROID_PLATFORM=android-24 -DNCNN_VULKAN=ON ..
     - name: build
       run: cmake --build build -j 2
+
+  android-x86-gpu-online-spirv:
+    runs-on: ubuntu-latest
+    steps:
+    - uses: actions/checkout@v2
+      with:
+        submodules: 'true'
+    - name: cache-vulkansdk
+      id: cache-vulkansdk
+      uses: actions/cache@v1
+      with:
+        path: "1.1.114.0"
+        key: vulkansdk-linux-x86_64-1.1.114.0
+    - name: vulkansdk
+      if: steps.cache-vulkansdk.outputs.cache-hit != 'true'
+      run: |
+        wget https://sdk.lunarg.com/sdk/download/1.1.114.0/linux/vulkansdk-linux-x86_64-1.1.114.0.tar.gz?Human=true -O vulkansdk-linux-x86_64-1.1.114.0.tar.gz
+        tar -xf vulkansdk-linux-x86_64-1.1.114.0.tar.gz
+        rm -rf 1.1.114.0/source 1.1.114.0/samples
+        find 1.1.114.0 -type f | grep -v -E 'vulkan|glslang' | xargs rm
+    - name: configure
+      run: export PATH=`pwd`/1.1.114.0/x86_64/bin:$PATH && mkdir build && cd build && cmake -DCMAKE_TOOLCHAIN_FILE=$ANDROID_HOME/ndk-bundle/build/cmake/android.toolchain.cmake -DANDROID_ABI="x86" -DANDROID_PLATFORM=android-24 -DNCNN_VULKAN=ON -DNCNN_VULKAN_ONLINE_SPIRV=ON ..
+    - name: build
+      run: cmake --build build -j 2

.github/workflows/ios-64bit-gpu.yml

@@ -22,3 +22,27 @@ jobs:
       run: export VULKAN_SDK=`pwd`/vulkansdk-macos-1.1.114.0/macOS && mkdir build && cd build && cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/ios.toolchain.cmake -DIOS_PLATFORM=OS64 -DVulkan_INCLUDE_DIR=`pwd`/../vulkansdk-macos-1.1.114.0/MoltenVK/include -DVulkan_LIBRARY=`pwd`/../vulkansdk-macos-1.1.114.0/MoltenVK/iOS/dynamic/libMoltenVK.dylib -DNCNN_VULKAN=ON ..
     - name: build
       run: cmake --build build -j 2
+
+  ios-iphone-os-gpu-online-spirv:
+    runs-on: macos-latest
+    steps:
+    - uses: actions/checkout@v2
+      with:
+        submodules: 'true'
+    - name: cache-vulkansdk
+      id: cache-vulkansdk
+      uses: actions/cache@v1
+      with:
+        path: "vulkansdk-macos-1.1.114.0"
+        key: vulkansdk-macos-1.1.114.0
+    - name: vulkansdk
+      if: steps.cache-vulkansdk.outputs.cache-hit != 'true'
+      run: |
+        wget https://sdk.lunarg.com/sdk/download/1.1.114.0/mac/vulkansdk-macos-1.1.114.0.tar.gz?Human=true -O vulkansdk-macos-1.1.114.0.tar.gz
+        tar -xf vulkansdk-macos-1.1.114.0.tar.gz
+        rm -rf vulkansdk-macos-1.1.114.0/Applications
+        find vulkansdk-macos-1.1.114.0 -type f | grep -v -E 'vulkan|glslang|MoltenVK' | xargs rm
+    - name: configure
+      run: export VULKAN_SDK=`pwd`/vulkansdk-macos-1.1.114.0/macOS && mkdir build && cd build && cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/ios.toolchain.cmake -DIOS_PLATFORM=OS64 -DVulkan_INCLUDE_DIR=`pwd`/../vulkansdk-macos-1.1.114.0/MoltenVK/include -DVulkan_LIBRARY=`pwd`/../vulkansdk-macos-1.1.114.0/MoltenVK/iOS/dynamic/libMoltenVK.dylib -DNCNN_VULKAN=ON -DNCNN_VULKAN_ONLINE_SPIRV=ON ..
+    - name: build
+      run: cmake --build build -j 2

.github/workflows/linux-x64-gpu-clang.yml

@@ -60,3 +60,65 @@ jobs:
       run: |
         export VK_ICD_FILENAMES="$GITHUB_WORKSPACE/swiftshader-install/vk_swiftshader_icd.json"
         cd build && ctest --output-on-failure -j 2
+
+  linux-clang-gpu-online-spirv:
+    runs-on: ubuntu-latest
+    steps:
+    - uses: actions/checkout@v2
+      with:
+        submodules: 'true'
+    - name: update
+      run: sudo apt-get update
+    - name: protobuf
+      run: sudo apt-get install libprotobuf-dev protobuf-compiler libopencv-dev
+    - name: cache-vulkansdk
+      id: cache-vulkansdk
+      uses: actions/cache@v1
+      with:
+        path: "1.1.114.0"
+        key: vulkansdk-linux-x86_64-1.1.114.0
+    - name: vulkansdk
+      if: steps.cache-vulkansdk.outputs.cache-hit != 'true'
+      run: |
+        wget https://sdk.lunarg.com/sdk/download/1.1.114.0/linux/vulkansdk-linux-x86_64-1.1.114.0.tar.gz?Human=true -O vulkansdk-linux-x86_64-1.1.114.0.tar.gz
+        tar -xf vulkansdk-linux-x86_64-1.1.114.0.tar.gz
+        rm -rf 1.1.114.0/source 1.1.114.0/samples
+        find 1.1.114.0 -type f | grep -v -E 'vulkan|glslang' | xargs rm
+    - name: cache-swiftshader
+      id: cache-swiftshader
+      uses: actions/cache@v1
+      with:
+        path: swiftshader-install
+        key: swiftshader-linux-install-20200508
+    - name: checkout-swiftshader
+      if: steps.cache-swiftshader.outputs.cache-hit != 'true'
+      uses: actions/checkout@v2
+      with:
+        repository: google/swiftshader
+        path: swiftshader
+        ref: 2dd864470e310d173d35fa95ca3a14d216734aab
+    - name: checkout-swiftshader-submodules
+      if: steps.cache-swiftshader.outputs.cache-hit != 'true'
+      run: |
+        cd swiftshader
+        git submodule update --init --recursive
+    - name: swiftshader
+      if: steps.cache-swiftshader.outputs.cache-hit != 'true'
+      run: |
+        cd swiftshader
+        mkdir -p build; cd build
+        cmake -DCMAKE_INSTALL_PREFIX=install -DSWIFTSHADER_BUILD_EGL=FALSE -DSWIFTSHADER_BUILD_GLESv2=FALSE -DSWIFTSHADER_BUILD_GLES_CM=FALSE -DSWIFTSHADER_BUILD_VULKAN=TRUE -DSWIFTSHADER_BUILD_PVR=FALSE -DSWIFTSHADER_BUILD_TESTS=FALSE -DSWIFTSHADER_ENABLE_ASTC=FALSE -DSWIFTSHADER_WARNINGS_AS_ERRORS=FALSE -DREACTOR_BACKEND=Subzero -DREACTOR_DEFAULT_OPT_LEVEL=Default -DCMAKE_BUILD_TYPE=Release ..
+        cmake --build . -j 2
+        mkdir $GITHUB_WORKSPACE/swiftshader-install
+        cp Linux/* $GITHUB_WORKSPACE/swiftshader-install
+    - name: configure
+      env:
+        CC: clang
+        CXX: clang++
+      run: export VULKAN_SDK=`pwd`/1.1.114.0/x86_64 && mkdir build && cd build && cmake -DNCNN_VULKAN=ON -DNCNN_VULKAN_ONLINE_SPIRV=ON ..
+    - name: build
+      run: cmake --build build -j 2
+    - name: test
+      run: |
+        export VK_ICD_FILENAMES="$GITHUB_WORKSPACE/swiftshader-install/vk_swiftshader_icd.json"
+        cd build && ctest --output-on-failure -j 2

.github/workflows/linux-x64-gpu-gcc.yml

@@ -57,3 +57,62 @@ jobs:
       run: |
         export VK_ICD_FILENAMES="$GITHUB_WORKSPACE/swiftshader-install/vk_swiftshader_icd.json"
         cd build && ctest --output-on-failure -j 2
+
+  linux-gcc-gpu-online-spirv:
+    runs-on: ubuntu-latest
+    steps:
+    - uses: actions/checkout@v2
+      with:
+        submodules: 'true'
+    - name: update
+      run: sudo apt-get update
+    - name: protobuf
+      run: sudo apt-get install libprotobuf-dev protobuf-compiler libopencv-dev
+    - name: cache-vulkansdk
+      id: cache-vulkansdk
+      uses: actions/cache@v1
+      with:
+        path: "1.1.114.0"
+        key: vulkansdk-linux-x86_64-1.1.114.0
+    - name: vulkansdk
+      if: steps.cache-vulkansdk.outputs.cache-hit != 'true'
+      run: |
+        wget https://sdk.lunarg.com/sdk/download/1.1.114.0/linux/vulkansdk-linux-x86_64-1.1.114.0.tar.gz?Human=true -O vulkansdk-linux-x86_64-1.1.114.0.tar.gz
+        tar -xf vulkansdk-linux-x86_64-1.1.114.0.tar.gz
+        rm -rf 1.1.114.0/source 1.1.114.0/samples
+        find 1.1.114.0 -type f | grep -v -E 'vulkan|glslang' | xargs rm
+    - name: cache-swiftshader
+      id: cache-swiftshader
+      uses: actions/cache@v1
+      with:
+        path: swiftshader-install
+        key: swiftshader-linux-install-20200508
+    - name: checkout-swiftshader
+      if: steps.cache-swiftshader.outputs.cache-hit != 'true'
+      uses: actions/checkout@v2
+      with:
+        repository: google/swiftshader
+        path: swiftshader
+        ref: 2dd864470e310d173d35fa95ca3a14d216734aab
+    - name: checkout-swiftshader-submodules
+      if: steps.cache-swiftshader.outputs.cache-hit != 'true'
+      run: |
+        cd swiftshader
+        git submodule update --init --recursive
+    - name: swiftshader
+      if: steps.cache-swiftshader.outputs.cache-hit != 'true'
+      run: |
+        cd swiftshader
+        mkdir -p build; cd build
+        cmake -DCMAKE_INSTALL_PREFIX=install -DSWIFTSHADER_BUILD_EGL=FALSE -DSWIFTSHADER_BUILD_GLESv2=FALSE -DSWIFTSHADER_BUILD_GLES_CM=FALSE -DSWIFTSHADER_BUILD_VULKAN=TRUE -DSWIFTSHADER_BUILD_PVR=FALSE -DSWIFTSHADER_BUILD_TESTS=FALSE -DSWIFTSHADER_ENABLE_ASTC=FALSE -DSWIFTSHADER_WARNINGS_AS_ERRORS=FALSE -DREACTOR_BACKEND=Subzero -DREACTOR_DEFAULT_OPT_LEVEL=Default -DCMAKE_BUILD_TYPE=Release ..
+        cmake --build . -j 2
+        mkdir $GITHUB_WORKSPACE/swiftshader-install
+        cp Linux/* $GITHUB_WORKSPACE/swiftshader-install
+    - name: configure
+      run: export VULKAN_SDK=`pwd`/1.1.114.0/x86_64 && mkdir build && cd build && cmake -DNCNN_VULKAN=ON -DNCNN_VULKAN_ONLINE_SPIRV=ON ..
+    - name: build
+      run: cmake --build build -j 2
+    - name: test
+      run: |
+        export VK_ICD_FILENAMES="$GITHUB_WORKSPACE/swiftshader-install/vk_swiftshader_icd.json"
+        cd build && ctest --output-on-failure -j 2

.github/workflows/macos-x64-gpu.yml

@@ -56,3 +56,61 @@ jobs:
         export DYLD_LIBRARY_PATH="vulkansdk-macos-1.1.114.0/macOS/lib":$DYLD_LIBRARY_PATH
         export VK_ICD_FILENAMES="$GITHUB_WORKSPACE/swiftshader-install/vk_swiftshader_icd.json"
         cd build && ctest --output-on-failure -j 2
+
+  macos-clang-gpu-online-spirv:
+    runs-on: macos-latest
+    steps:
+    - uses: actions/checkout@v2
+      with:
+        submodules: 'true'
+    - name: protobuf
+      run: brew install protobuf opencv3
+    - name: cache-vulkansdk
+      id: cache-vulkansdk
+      uses: actions/cache@v1
+      with:
+        path: "vulkansdk-macos-1.1.114.0"
+        key: vulkansdk-macos-1.1.114.0
+    - name: vulkansdk
+      if: steps.cache-vulkansdk.outputs.cache-hit != 'true'
+      run: |
+        wget https://sdk.lunarg.com/sdk/download/1.1.114.0/mac/vulkansdk-macos-1.1.114.0.tar.gz?Human=true -O vulkansdk-macos-1.1.114.0.tar.gz
+        tar -xf vulkansdk-macos-1.1.114.0.tar.gz
+        rm -rf vulkansdk-macos-1.1.114.0/Applications
+        find vulkansdk-macos-1.1.114.0 -type f | grep -v -E 'vulkan|glslang|MoltenVK' | xargs rm
+    - name: cache-swiftshader
+      id: cache-swiftshader
+      uses: actions/cache@v1
+      with:
+        path: swiftshader-install
+        key: swiftshader-macos-install-20200508
+    - name: checkout-swiftshader
+      if: steps.cache-swiftshader.outputs.cache-hit != 'true'
+      uses: actions/checkout@v2
+      with:
+        repository: google/swiftshader
+        path: swiftshader
+        ref: 2dd864470e310d173d35fa95ca3a14d216734aab
+    - name: checkout-swiftshader-submodules
+      if: steps.cache-swiftshader.outputs.cache-hit != 'true'
+      run: |
+        cd swiftshader
+        git submodule update --init --recursive
+    - name: swiftshader
+      if: steps.cache-swiftshader.outputs.cache-hit != 'true'
+      run: |
+        cd swiftshader
+        mkdir -p build; cd build
+        cmake -DCMAKE_INSTALL_PREFIX=install -DSWIFTSHADER_BUILD_EGL=FALSE -DSWIFTSHADER_BUILD_GLESv2=FALSE -DSWIFTSHADER_BUILD_GLES_CM=FALSE -DSWIFTSHADER_BUILD_VULKAN=TRUE -DSWIFTSHADER_BUILD_PVR=FALSE -DSWIFTSHADER_BUILD_TESTS=FALSE -DSWIFTSHADER_ENABLE_ASTC=FALSE -DSWIFTSHADER_WARNINGS_AS_ERRORS=FALSE -DREACTOR_BACKEND=Subzero -DREACTOR_DEFAULT_OPT_LEVEL=Default -DCMAKE_BUILD_TYPE=Release ..
+        cmake --build . -j 2
+        mkdir $GITHUB_WORKSPACE/swiftshader-install
+        cp Darwin/* $GITHUB_WORKSPACE/swiftshader-install
+    - name: configure
+      run: export VULKAN_SDK=`pwd`/vulkansdk-macos-1.1.114.0/macOS && mkdir build && cd build && cmake -DNCNN_VULKAN=ON -DNCNN_VULKAN_ONLINE_SPIRV=ON ..
+    - name: build
+      run: cmake --build build -j 2
+    - name: test
+      run: |
+        export DYLD_LIBRARY_PATH="vulkansdk-macos-1.1.114.0/macOS/lib":$DYLD_LIBRARY_PATH
+        export VK_ICD_FILENAMES="$GITHUB_WORKSPACE/swiftshader-install/vk_swiftshader_icd.json"
+        cd build && ctest --output-on-failure -j 2

.github/workflows/test-coverage.yml

@@ -67,6 +67,74 @@ jobs:
         token: ${{ secrets.CODECOV_TOKEN }}
         file: build/lcov.info
 
+  linux-gcc-gpu-online-spirv:
+    runs-on: ubuntu-latest
+    steps:
+    - uses: actions/checkout@v2
+      with:
+        submodules: 'true'
+    - name: lcov
+      run: sudo apt-get install lcov
+    - name: cache-vulkansdk
+      id: cache-vulkansdk
+      uses: actions/cache@v1
+      with:
+        path: "1.1.114.0"
+        key: vulkansdk-linux-x86_64-1.1.114.0
+    - name: vulkansdk
+      if: steps.cache-vulkansdk.outputs.cache-hit != 'true'
+      run: |
+        wget https://sdk.lunarg.com/sdk/download/1.1.114.0/linux/vulkansdk-linux-x86_64-1.1.114.0.tar.gz?Human=true -O vulkansdk-linux-x86_64-1.1.114.0.tar.gz
+        tar -xf vulkansdk-linux-x86_64-1.1.114.0.tar.gz
+        rm -rf 1.1.114.0/source 1.1.114.0/samples
+        find 1.1.114.0 -type f | grep -v -E 'vulkan|glslang' | xargs rm
+    - name: cache-swiftshader
+      id: cache-swiftshader
+      uses: actions/cache@v1
+      with:
+        path: swiftshader-install
+        key: swiftshader-linux-install-20200508
+    - name: checkout-swiftshader
+      if: steps.cache-swiftshader.outputs.cache-hit != 'true'
+      uses: actions/checkout@v2
+      with:
+        repository: google/swiftshader
+        path: swiftshader
+        ref: 2dd864470e310d173d35fa95ca3a14d216734aab
+    - name: checkout-swiftshader-submodules
+      if: steps.cache-swiftshader.outputs.cache-hit != 'true'
+      run: |
+        cd swiftshader
+        git submodule update --init --recursive
+    - name: swiftshader
+      if: steps.cache-swiftshader.outputs.cache-hit != 'true'
+      run: |
+        cd swiftshader
+        mkdir -p build; cd build
+        cmake -DCMAKE_INSTALL_PREFIX=install -DSWIFTSHADER_BUILD_EGL=FALSE -DSWIFTSHADER_BUILD_GLESv2=FALSE -DSWIFTSHADER_BUILD_GLES_CM=FALSE -DSWIFTSHADER_BUILD_VULKAN=TRUE -DSWIFTSHADER_BUILD_PVR=FALSE -DSWIFTSHADER_BUILD_TESTS=FALSE -DSWIFTSHADER_ENABLE_ASTC=FALSE -DSWIFTSHADER_WARNINGS_AS_ERRORS=FALSE -DREACTOR_BACKEND=Subzero -DREACTOR_DEFAULT_OPT_LEVEL=Default -DCMAKE_BUILD_TYPE=Release ..
+        cmake --build . -j 2
+        mkdir $GITHUB_WORKSPACE/swiftshader-install
+        cp Linux/* $GITHUB_WORKSPACE/swiftshader-install
+    - name: configure
+      run: export VULKAN_SDK=`pwd`/1.1.114.0/x86_64 && mkdir build && cd build && cmake -DCMAKE_BUILD_TYPE=debug -DNCNN_COVERAGE=ON -DNCNN_VULKAN=ON -DNCNN_VULKAN_ONLINE_SPIRV=ON -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF ..
+    - name: build
+      run: cmake --build build -j 2
+    - name: test
+      run: |
+        export VK_ICD_FILENAMES="$GITHUB_WORKSPACE/swiftshader-install/vk_swiftshader_icd.json"
+        cd build && ctest --output-on-failure -j 2
+    - name: lcov-collect
+      run: |
+        cd build
+        lcov -d ./src -c -o lcov.info
+        lcov -r lcov.info '/usr/*' -o lcov.info
+        lcov --list lcov.info
+    - name: codecov
+      uses: codecov/codecov-action@v1
+      with:
+        token: ${{ secrets.CODECOV_TOKEN }}
+        file: build/lcov.info
+
   linux-gcc-avx2:
     runs-on: ubuntu-latest
     steps:

.github/workflows/windows-x64-gpu-vs2019.yml

@@ -71,3 +71,76 @@ jobs:
         Copy-Item -Path '.\VulkanSDK\RunTimeInstaller\x64\vulkan-1.dll' -Destination 'build\tests'
         $env:VK_ICD_FILENAMES="$env:GITHUB_WORKSPACE\swiftshader-install\vk_swiftshader_icd.json"
         cd build; ctest --output-on-failure -j 2
+
+  windows-vs2019-gpu-online-spirv:
+    runs-on: windows-latest
+    steps:
+    - uses: actions/checkout@v2
+      with:
+        submodules: 'true'
+    - name: cache-protobuf
+      id: cache-protobuf
+      uses: actions/cache@v1
+      with:
+        path: "protobuf-install"
+        key: protobuf-windows-install
+    - name: protobuf
+      if: steps.cache-protobuf.outputs.cache-hit != 'true'
+      run: |
+        Invoke-WebRequest -Uri https://github.com/protocolbuffers/protobuf/archive/v3.11.2.zip -OutFile protobuf-3.11.2.zip
+        7z x ./protobuf-3.11.2.zip
+        cd protobuf-3.11.2
+        mkdir build-vs2019; cd build-vs2019; cmake -DCMAKE_INSTALL_PREFIX="$env:GITHUB_WORKSPACE\protobuf-install" -Dprotobuf_BUILD_TESTS=OFF -Dprotobuf_MSVC_STATIC_RUNTIME=OFF ../cmake
+        cmake --build . --config Release -j 2
+        cmake --build . --config Release --target install
+    - name: cache-vulkansdk
+      id: cache-vulkansdk
+      uses: actions/cache@v1
+      with:
+        path: "VulkanSDK"
+        key: VulkanSDK-1.1.114.0-Installer
+    - name: vulkansdk
+      if: steps.cache-vulkansdk.outputs.cache-hit != 'true'
+      run: |
+        Invoke-WebRequest -Uri https://sdk.lunarg.com/sdk/download/1.1.114.0/windows/VulkanSDK-1.1.114.0-Installer.exe?Human=true -OutFile VulkanSDK-1.1.114.0-Installer.exe
+        7z x -aoa ./VulkanSDK-1.1.114.0-Installer.exe -oVulkanSDK
+        Remove-Item .\VulkanSDK\Demos, .\VulkanSDK\glslang, .\VulkanSDK\Samples, .\VulkanSDK\shaderc, .\VulkanSDK\spirv-tools, .\VulkanSDK\Third-Party, .\VulkanSDK\Tools, .\VulkanSDK\Tools32 -Recurse
+    - name: cache-swiftshader
+      id: cache-swiftshader
+      uses: actions/cache@v1
+      with:
+        path: swiftshader-install
+        key: swiftshader-windows-install-20200508
+    - name: checkout-swiftshader
+      if: steps.cache-swiftshader.outputs.cache-hit != 'true'
+      uses: actions/checkout@v2
+      with:
+        repository: google/swiftshader
+        path: swiftshader
+        ref: 2dd864470e310d173d35fa95ca3a14d216734aab
+    - name: checkout-swiftshader-submodules
+      if: steps.cache-swiftshader.outputs.cache-hit != 'true'
+      run: |
+        cd swiftshader
+        git submodule update --init --recursive
+    - name: swiftshader
+      if: steps.cache-swiftshader.outputs.cache-hit != 'true'
+      run: |
+        cd swiftshader
+        mkdir build-vs2019; cd build-vs2019
+        cmake -DCMAKE_INSTALL_PREFIX=install -DSWIFTSHADER_BUILD_EGL=FALSE -DSWIFTSHADER_BUILD_GLESv2=FALSE -DSWIFTSHADER_BUILD_GLES_CM=FALSE -DSWIFTSHADER_BUILD_VULKAN=TRUE -DSWIFTSHADER_BUILD_PVR=FALSE -DSWIFTSHADER_BUILD_TESTS=FALSE -DSWIFTSHADER_ENABLE_ASTC=FALSE -DSWIFTSHADER_WARNINGS_AS_ERRORS=FALSE -DREACTOR_BACKEND=Subzero -DREACTOR_DEFAULT_OPT_LEVEL=Default -DCMAKE_BUILD_TYPE=Release ..
+        cmake --build . --config Release -j 2
+        mkdir "$env:GITHUB_WORKSPACE/swiftshader-install"
+        Copy-Item -Path "Windows\*" -Destination "$env:GITHUB_WORKSPACE\swiftshader-install"
+    - name: configure
+      run: |
+        $env:VULKAN_SDK="$(pwd)/VulkanSDK"
+        mkdir build; cd build
+        cmake -DProtobuf_INCLUDE_DIR="$env:GITHUB_WORKSPACE\protobuf-install\include" -DProtobuf_LIBRARIES="$env:GITHUB_WORKSPACE\protobuf-install\lib\libprotobuf.lib" -DProtobuf_PROTOC_EXECUTABLE="$env:GITHUB_WORKSPACE\protobuf-install\bin\protoc.exe" -DNCNN_VULKAN=ON -DNCNN_VULKAN_ONLINE_SPIRV=ON ..
+    - name: build
+      run: cmake --build build --config Release -j 2
+    - name: test
+      run: |
+        Copy-Item -Path '.\VulkanSDK\RunTimeInstaller\x64\vulkan-1.dll' -Destination 'build\tests'
+        $env:VK_ICD_FILENAMES="$env:GITHUB_WORKSPACE\swiftshader-install\vk_swiftshader_icd.json"
+        cd build; ctest --output-on-failure -j 2

.gitmodules

@@ -0,0 +1,3 @@
+[submodule "glslang"]
+	path = glslang
+	url = https://github.com/KhronosGroup/glslang

CMakeLists.txt

@@ -30,6 +30,7 @@ option(NCNN_PIXEL "convert and resize from/to image pixel" ON)
 option(NCNN_PIXEL_ROTATE "rotate image pixel orientation" ON)
 option(NCNN_CMAKE_VERBOSE "print verbose cmake messages" OFF)
 option(NCNN_VULKAN "vulkan compute support" OFF)
+option(NCNN_VULKAN_ONLINE_SPIRV "online SPIR-V module compilation" OFF)
 option(NCNN_REQUANT "auto merge int8 quant and dequant" OFF)
 option(NCNN_AVX2 "optimize x86 platform with avx2" OFF)
 option(NCNN_DISABLE_PIC "disable position-independent code" OFF)
@@ -64,6 +65,28 @@ if(NCNN_COVERAGE)
     set(CMAKE_EXE_LINKER_FLAGS "${CMAKE_EXE_LINKER_FLAGS} -coverage -lgcov")
 endif()
 
+if(NCNN_VULKAN_ONLINE_SPIRV)
+    if(NOT EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/glslang/CMakeLists.txt")
+        message(WARNING "The submodules were not downloaded! NCNN_VULKAN_ONLINE_SPIRV will be turned off.")
+        message(WARNING "Please update submodules with \"git submodule update --init\" and try again.")
+        set(NCNN_VULKAN_ONLINE_SPIRV OFF)
+    else()
+        # glslang requires c++11
+        set(CMAKE_CXX_STANDARD 11)
+
+        option(BUILD_EXTERNAL "" OFF)
+        option(ENABLE_SPVREMAPPER "" OFF)
+        option(ENABLE_GLSLANG_BINARIES "" OFF)
+        option(ENABLE_HLSL "" OFF)
+        option(ENABLE_RTTI "" OFF)
+        option(ENABLE_EXCEPTIONS "" OFF)
+        option(ENABLE_OPT "" OFF)
+        option(ENABLE_PCH "" OFF)
+        option(ENABLE_CTEST "" OFF)
+        add_subdirectory(glslang)
+    endif()
+endif()
+
 add_subdirectory(src)
 if(NCNN_BUILD_BENCHMARK)
     add_subdirectory(benchmark)

glslang

@@ -0,0 +1 @@
+Subproject commit 59216d5cd87eee78dc979e30645c4c2240a7c351

src/CMakeLists.txt

@@ -3,7 +3,7 @@
 
 configure_file(platform.h.in ${CMAKE_CURRENT_BINARY_DIR}/platform.h)
 
-if(NCNN_VULKAN)
+if(NCNN_VULKAN AND NOT NCNN_VULKAN_ONLINE_SPIRV)
     find_program(GLSLANGVALIDATOR_EXECUTABLE NAMES glslangValidator PATHS $ENV{VULKAN_SDK}/bin NO_CMAKE_FIND_ROOT_PATH)
     message(STATUS "Found glslangValidator: ${GLSLANGVALIDATOR_EXECUTABLE}")
 endif()
@@ -21,6 +21,7 @@ endfunction()
 
 set(ncnn_SRCS
     allocator.cpp
+    benchmark.cpp
     blob.cpp
     command.cpp
     cpu.cpp
@@ -37,59 +38,93 @@ set(ncnn_SRCS
     option.cpp
     paramdict.cpp
     pipeline.cpp
-    benchmark.cpp
 )
 
-ncnn_src_group(ncnn_SRCS "sources")
-
 if(ANDROID)
     list(APPEND ncnn_SRCS mat_pixel_android.cpp)
 endif()
 
+ncnn_src_group(ncnn_SRCS "sources")
+
 include(${CMAKE_CURRENT_SOURCE_DIR}/../cmake/ncnn_generate_shader_spv_header.cmake)
 
 macro(ncnn_add_shader SHADER_SRC)
-    ncnn_generate_shader_spv_header(SHADER_SPV_HEADER SHADER_SPV_HEX_HEADERS ${SHADER_SRC})
-
-    get_filename_component(SHADER_SPV_HEADER_NAME ${SHADER_SPV_HEADER} NAME)
-    string(APPEND layer_shader_spv_data "#include \"${SHADER_SPV_HEADER_NAME}\"\n")
-
-    get_filename_component(SHADER_SRC_NAME_WE ${SHADER_SRC} NAME_WE)
-    string(APPEND layer_shader_registry "{${SHADER_SRC_NAME_WE}_spv_data,sizeof(${SHADER_SRC_NAME_WE}_spv_data)},\n")
-    string(APPEND layer_shader_registry "{${SHADER_SRC_NAME_WE}_fp16p_spv_data,sizeof(${SHADER_SRC_NAME_WE}_fp16p_spv_data)},\n")
-    string(APPEND layer_shader_registry "{${SHADER_SRC_NAME_WE}_fp16pa_spv_data,sizeof(${SHADER_SRC_NAME_WE}_fp16pa_spv_data)},\n")
-    string(APPEND layer_shader_registry "{${SHADER_SRC_NAME_WE}_fp16s_spv_data,sizeof(${SHADER_SRC_NAME_WE}_fp16s_spv_data)},\n")
-    string(APPEND layer_shader_registry "{${SHADER_SRC_NAME_WE}_fp16sa_spv_data,sizeof(${SHADER_SRC_NAME_WE}_fp16sa_spv_data)},\n")
-    string(APPEND layer_shader_registry "{${SHADER_SRC_NAME_WE}_image_spv_data,sizeof(${SHADER_SRC_NAME_WE}_image_spv_data)},\n")
-    string(APPEND layer_shader_registry "{${SHADER_SRC_NAME_WE}_image_fp16p_spv_data,sizeof(${SHADER_SRC_NAME_WE}_image_fp16p_spv_data)},\n")
-    string(APPEND layer_shader_registry "{${SHADER_SRC_NAME_WE}_image_fp16pa_spv_data,sizeof(${SHADER_SRC_NAME_WE}_image_fp16pa_spv_data)},\n")
-    string(APPEND layer_shader_registry "{${SHADER_SRC_NAME_WE}_image_fp16s_spv_data,sizeof(${SHADER_SRC_NAME_WE}_image_fp16s_spv_data)},\n")
-    string(APPEND layer_shader_registry "{${SHADER_SRC_NAME_WE}_image_fp16sa_spv_data,sizeof(${SHADER_SRC_NAME_WE}_image_fp16sa_spv_data)},\n")
-
-    list(APPEND SHADER_SPV_HEX_FILES ${SHADER_SPV_HEADER})
-    list(APPEND SHADER_SPV_HEX_FILES ${SHADER_SPV_HEX_HEADERS})
-
-    # generate layer_shader_type_enum file
-    set(layer_shader_type_enum "${layer_shader_type_enum}${SHADER_SRC_NAME_WE} = ${__LAYER_SHADER_TYPE_ENUM_INDEX},\n")
-    math(EXPR __LAYER_SHADER_TYPE_ENUM_INDEX "${__LAYER_SHADER_TYPE_ENUM_INDEX}+1")
-    set(layer_shader_type_enum "${layer_shader_type_enum}${SHADER_SRC_NAME_WE}_fp16p = ${__LAYER_SHADER_TYPE_ENUM_INDEX},\n")
-    math(EXPR __LAYER_SHADER_TYPE_ENUM_INDEX "${__LAYER_SHADER_TYPE_ENUM_INDEX}+1")
-    set(layer_shader_type_enum "${layer_shader_type_enum}${SHADER_SRC_NAME_WE}_fp16pa = ${__LAYER_SHADER_TYPE_ENUM_INDEX},\n")
-    math(EXPR __LAYER_SHADER_TYPE_ENUM_INDEX "${__LAYER_SHADER_TYPE_ENUM_INDEX}+1")
-    set(layer_shader_type_enum "${layer_shader_type_enum}${SHADER_SRC_NAME_WE}_fp16s = ${__LAYER_SHADER_TYPE_ENUM_INDEX},\n")
-    math(EXPR __LAYER_SHADER_TYPE_ENUM_INDEX "${__LAYER_SHADER_TYPE_ENUM_INDEX}+1")
-    set(layer_shader_type_enum "${layer_shader_type_enum}${SHADER_SRC_NAME_WE}_fp16sa = ${__LAYER_SHADER_TYPE_ENUM_INDEX},\n")
-    math(EXPR __LAYER_SHADER_TYPE_ENUM_INDEX "${__LAYER_SHADER_TYPE_ENUM_INDEX}+1")
-    set(layer_shader_type_enum "${layer_shader_type_enum}${SHADER_SRC_NAME_WE}_image = ${__LAYER_SHADER_TYPE_ENUM_INDEX},\n")
-    math(EXPR __LAYER_SHADER_TYPE_ENUM_INDEX "${__LAYER_SHADER_TYPE_ENUM_INDEX}+1")
-    set(layer_shader_type_enum "${layer_shader_type_enum}${SHADER_SRC_NAME_WE}_image_fp16p = ${__LAYER_SHADER_TYPE_ENUM_INDEX},\n")
-    math(EXPR __LAYER_SHADER_TYPE_ENUM_INDEX "${__LAYER_SHADER_TYPE_ENUM_INDEX}+1")
-    set(layer_shader_type_enum "${layer_shader_type_enum}${SHADER_SRC_NAME_WE}_image_fp16pa = ${__LAYER_SHADER_TYPE_ENUM_INDEX},\n")
-    math(EXPR __LAYER_SHADER_TYPE_ENUM_INDEX "${__LAYER_SHADER_TYPE_ENUM_INDEX}+1")
-    set(layer_shader_type_enum "${layer_shader_type_enum}${SHADER_SRC_NAME_WE}_image_fp16s = ${__LAYER_SHADER_TYPE_ENUM_INDEX},\n")
-    math(EXPR __LAYER_SHADER_TYPE_ENUM_INDEX "${__LAYER_SHADER_TYPE_ENUM_INDEX}+1")
-    set(layer_shader_type_enum "${layer_shader_type_enum}${SHADER_SRC_NAME_WE}_image_fp16sa = ${__LAYER_SHADER_TYPE_ENUM_INDEX},\n")
-    math(EXPR __LAYER_SHADER_TYPE_ENUM_INDEX "${__LAYER_SHADER_TYPE_ENUM_INDEX}+1")
+    if(NCNN_VULKAN_ONLINE_SPIRV)
+
+        file(READ ${SHADER_SRC} comp_data)
+
+        # skip leading comment
+        string(FIND "${comp_data}" "#version" version_start)
+        string(SUBSTRING "${comp_data}" ${version_start} -1 comp_data)
+
+        # remove whitespace
+        string(REGEX REPLACE "\n +" "\n" comp_data "${comp_data}")
+
+        # text to hex
+        file(WRITE ${CMAKE_CURRENT_BINARY_DIR}/text2hex.txt "${comp_data}")
+        file(READ ${CMAKE_CURRENT_BINARY_DIR}/text2hex.txt comp_data_hex HEX)
+        string(REGEX REPLACE "([0-9a-f][0-9a-f])" "0x\\1," comp_data_hex ${comp_data_hex})
+        string(FIND "${comp_data_hex}" "," tail_comma REVERSE)
+        string(SUBSTRING "${comp_data_hex}" 0 ${tail_comma} comp_data_hex)
+
+        get_filename_component(SHADER_SRC_NAME_WE ${SHADER_SRC} NAME_WE)
+        set(SHADER_COMP_HEADER ${CMAKE_CURRENT_BINARY_DIR}/${SHADER_SRC_NAME_WE}.comp.hex.h)
+        file(WRITE ${SHADER_COMP_HEADER} "static const char ${SHADER_SRC_NAME_WE}_comp_data[] = {${comp_data_hex}};\n")
+        set_source_files_properties(${SHADER_COMP_HEADER} PROPERTIES GENERATED TRUE)
+
+        string(APPEND layer_shader_registry "{${SHADER_SRC_NAME_WE}_comp_data,sizeof(${SHADER_SRC_NAME_WE}_comp_data)},\n")
+
+        get_filename_component(SHADER_COMP_HEADER_NAME ${SHADER_COMP_HEADER} NAME)
+        list(APPEND SHADER_SPV_HEX_FILES ${SHADER_COMP_HEADER_NAME})
+
+        string(APPEND layer_shader_spv_data "#include \"${SHADER_COMP_HEADER_NAME}\"\n")
+
+        # generate layer_shader_type_enum file
+        set(layer_shader_type_enum "${layer_shader_type_enum}${SHADER_SRC_NAME_WE} = ${__LAYER_SHADER_TYPE_ENUM_INDEX},\n")
+        math(EXPR __LAYER_SHADER_TYPE_ENUM_INDEX "${__LAYER_SHADER_TYPE_ENUM_INDEX}+1")
+    else()
+        ncnn_generate_shader_spv_header(SHADER_SPV_HEADER SHADER_SPV_HEX_HEADERS ${SHADER_SRC})
+
+        get_filename_component(SHADER_SPV_HEADER_NAME ${SHADER_SPV_HEADER} NAME)
+        string(APPEND layer_shader_spv_data "#include \"${SHADER_SPV_HEADER_NAME}\"\n")
+
+        get_filename_component(SHADER_SRC_NAME_WE ${SHADER_SRC} NAME_WE)
+        string(APPEND layer_shader_registry "{${SHADER_SRC_NAME_WE}_spv_data,sizeof(${SHADER_SRC_NAME_WE}_spv_data)},\n")
+        string(APPEND layer_shader_registry "{${SHADER_SRC_NAME_WE}_fp16p_spv_data,sizeof(${SHADER_SRC_NAME_WE}_fp16p_spv_data)},\n")
+        string(APPEND layer_shader_registry "{${SHADER_SRC_NAME_WE}_fp16pa_spv_data,sizeof(${SHADER_SRC_NAME_WE}_fp16pa_spv_data)},\n")
+        string(APPEND layer_shader_registry "{${SHADER_SRC_NAME_WE}_fp16s_spv_data,sizeof(${SHADER_SRC_NAME_WE}_fp16s_spv_data)},\n")
+        string(APPEND layer_shader_registry "{${SHADER_SRC_NAME_WE}_fp16sa_spv_data,sizeof(${SHADER_SRC_NAME_WE}_fp16sa_spv_data)},\n")
+        string(APPEND layer_shader_registry "{${SHADER_SRC_NAME_WE}_image_spv_data,sizeof(${SHADER_SRC_NAME_WE}_image_spv_data)},\n")
+        string(APPEND layer_shader_registry "{${SHADER_SRC_NAME_WE}_image_fp16p_spv_data,sizeof(${SHADER_SRC_NAME_WE}_image_fp16p_spv_data)},\n")
+        string(APPEND layer_shader_registry "{${SHADER_SRC_NAME_WE}_image_fp16pa_spv_data,sizeof(${SHADER_SRC_NAME_WE}_image_fp16pa_spv_data)},\n")
+        string(APPEND layer_shader_registry "{${SHADER_SRC_NAME_WE}_image_fp16s_spv_data,sizeof(${SHADER_SRC_NAME_WE}_image_fp16s_spv_data)},\n")
+        string(APPEND layer_shader_registry "{${SHADER_SRC_NAME_WE}_image_fp16sa_spv_data,sizeof(${SHADER_SRC_NAME_WE}_image_fp16sa_spv_data)},\n")
+
+        list(APPEND SHADER_SPV_HEX_FILES ${SHADER_SPV_HEADER})
+        list(APPEND SHADER_SPV_HEX_FILES ${SHADER_SPV_HEX_HEADERS})
+
+        # generate layer_shader_type_enum file
+        set(layer_shader_type_enum "${layer_shader_type_enum}${SHADER_SRC_NAME_WE} = ${__LAYER_SHADER_TYPE_ENUM_INDEX},\n")
+        math(EXPR __LAYER_SHADER_TYPE_ENUM_INDEX "${__LAYER_SHADER_TYPE_ENUM_INDEX}+1")
+        set(layer_shader_type_enum "${layer_shader_type_enum}${SHADER_SRC_NAME_WE}_fp16p = ${__LAYER_SHADER_TYPE_ENUM_INDEX},\n")
+        math(EXPR __LAYER_SHADER_TYPE_ENUM_INDEX "${__LAYER_SHADER_TYPE_ENUM_INDEX}+1")
+        set(layer_shader_type_enum "${layer_shader_type_enum}${SHADER_SRC_NAME_WE}_fp16pa = ${__LAYER_SHADER_TYPE_ENUM_INDEX},\n")
+        math(EXPR __LAYER_SHADER_TYPE_ENUM_INDEX "${__LAYER_SHADER_TYPE_ENUM_INDEX}+1")
+        set(layer_shader_type_enum "${layer_shader_type_enum}${SHADER_SRC_NAME_WE}_fp16s = ${__LAYER_SHADER_TYPE_ENUM_INDEX},\n")
+        math(EXPR __LAYER_SHADER_TYPE_ENUM_INDEX "${__LAYER_SHADER_TYPE_ENUM_INDEX}+1")
+        set(layer_shader_type_enum "${layer_shader_type_enum}${SHADER_SRC_NAME_WE}_fp16sa = ${__LAYER_SHADER_TYPE_ENUM_INDEX},\n")
+        math(EXPR __LAYER_SHADER_TYPE_ENUM_INDEX "${__LAYER_SHADER_TYPE_ENUM_INDEX}+1")
+        set(layer_shader_type_enum "${layer_shader_type_enum}${SHADER_SRC_NAME_WE}_image = ${__LAYER_SHADER_TYPE_ENUM_INDEX},\n")
+        math(EXPR __LAYER_SHADER_TYPE_ENUM_INDEX "${__LAYER_SHADER_TYPE_ENUM_INDEX}+1")
+        set(layer_shader_type_enum "${layer_shader_type_enum}${SHADER_SRC_NAME_WE}_image_fp16p = ${__LAYER_SHADER_TYPE_ENUM_INDEX},\n")
+        math(EXPR __LAYER_SHADER_TYPE_ENUM_INDEX "${__LAYER_SHADER_TYPE_ENUM_INDEX}+1")
+        set(layer_shader_type_enum "${layer_shader_type_enum}${SHADER_SRC_NAME_WE}_image_fp16pa = ${__LAYER_SHADER_TYPE_ENUM_INDEX},\n")
+        math(EXPR __LAYER_SHADER_TYPE_ENUM_INDEX "${__LAYER_SHADER_TYPE_ENUM_INDEX}+1")
+        set(layer_shader_type_enum "${layer_shader_type_enum}${SHADER_SRC_NAME_WE}_image_fp16s = ${__LAYER_SHADER_TYPE_ENUM_INDEX},\n")
+        math(EXPR __LAYER_SHADER_TYPE_ENUM_INDEX "${__LAYER_SHADER_TYPE_ENUM_INDEX}+1")
+        set(layer_shader_type_enum "${layer_shader_type_enum}${SHADER_SRC_NAME_WE}_image_fp16sa = ${__LAYER_SHADER_TYPE_ENUM_INDEX},\n")
+        math(EXPR __LAYER_SHADER_TYPE_ENUM_INDEX "${__LAYER_SHADER_TYPE_ENUM_INDEX}+1")
+    endif()
 endmacro()
 
 macro(ncnn_add_layer class)
@@ -303,6 +338,11 @@ endif()
 if(NCNN_VULKAN)
     find_package(Vulkan REQUIRED)
     target_link_libraries(ncnn PUBLIC Vulkan::Vulkan)
+
+    if(NCNN_VULKAN_ONLINE_SPIRV)
+        target_include_directories(ncnn PRIVATE $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/../>)
+        target_link_libraries(ncnn PRIVATE glslang SPIRV)
+    endif()
 endif()
 
 if(ANDROID_NDK)

src/convert_ycbcr.comp

@@ -18,7 +18,7 @@
 #extension GL_EXT_shader_16bit_storage: require
 #endif
 #if NCNN_fp16_arithmetic
-#extension GL_AMD_gpu_shader_half_float: require
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
 layout (constant_id = 0) const int w = 0;

src/gpu.cpp

@@ -23,6 +23,11 @@
 
 #include <algorithm>
 
+#if NCNN_VULKAN_ONLINE_SPIRV
+#include "glslang/glslang/Public/ShaderLang.h"
+#include "glslang/SPIRV/GlslangToSpv.h"
+#endif
+
 #include "mat.h"
 #include "command.h"
 #include "layer_type.h"
@@ -54,11 +59,19 @@ static Mutex g_default_vkdev_lock;
 static VulkanDevice* g_default_vkdev[NCNN_MAX_GPU_COUNT] = {0};
 
 // precompiled spirv
+#if NCNN_VULKAN_ONLINE_SPIRV
+struct layer_shader_registry_entry
+{
+    const char* comp_data;
+    int comp_data_size;
+};
+#else
 struct layer_shader_registry_entry
 {
     const uint32_t* spv_data;
     size_t spv_data_size;
 };
+#endif
 
 #include "layer_shader_spv_data.h"
 
@@ -67,7 +80,9 @@ static const layer_shader_registry_entry layer_shader_registry[] =
 #include "layer_shader_registry.h"
 };
 
+#if !NCNN_VULKAN_ONLINE_SPIRV
 static ShaderInfo layer_shader_infos[sizeof(layer_shader_registry) / sizeof(layer_shader_registry_entry)];
+#endif
 
 static const int layer_shader_registry_entry_count = sizeof(layer_shader_registry) / sizeof(layer_shader_registry_entry);
 
@@ -846,17 +861,25 @@ int create_gpu_instance()
     // the default gpu device
     g_default_gpu_index = find_default_vulkan_device_index();
 
+#if NCNN_VULKAN_ONLINE_SPIRV
+    glslang::InitializeProcess();
+#else
     // resolve shader info
     for (int i=0; i<layer_shader_registry_entry_count; i++)
     {
         resolve_shader_info(layer_shader_registry[i].spv_data, layer_shader_registry[i].spv_data_size, layer_shader_infos[i]);
     }
+#endif
 
     return 0;
 }
 
 void destroy_gpu_instance()
 {
+#if NCNN_VULKAN_ONLINE_SPIRV
+    glslang::FinalizeProcess();
+#endif
+
     for (int i=0; i<NCNN_MAX_GPU_COUNT; i++)
     {
         delete g_default_vkdev[i];
@@ -1051,7 +1074,9 @@ VulkanDevice::VulkanDevice(int device_index) : info(g_gpu_infos[device_index])
 
     init_device_extension();
 
+#if !NCNN_VULKAN_ONLINE_SPIRV
     create_shader_module();
+#endif
 
     compute_queues.resize(info.compute_queue_count);
     blob_allocators.resize(info.compute_queue_count);
@@ -1133,11 +1158,14 @@ VulkanDevice::~VulkanDevice()
     blob_allocators.clear();
     staging_allocators.clear();
 
+#if !NCNN_VULKAN_ONLINE_SPIRV
     destroy_shader_module();
+#endif
 
     vkDestroyDevice(device, 0);
 }
 
+#if !NCNN_VULKAN_ONLINE_SPIRV
 VkShaderModule VulkanDevice::get_shader_module(int shader_type_index) const
 {
     if (shader_type_index < 0 || shader_type_index >= layer_shader_registry_entry_count)
@@ -1162,6 +1190,7 @@ VkShaderModule VulkanDevice::create_shader_module(int shader_type_index, uint32_
 
     return compile_shader_module(spv_data, spv_data_size, local_size_x, local_size_y, local_size_z);
 }
+#endif
 
 VkShaderModule VulkanDevice::compile_shader_module(const uint32_t* spv_data, size_t spv_data_size) const
 {
@@ -1601,6 +1630,7 @@ void VulkanDevice::convert_packing(const VkImageMat& src, VkMat& dst, int dst_el
     uop->forward(src, dst, cmd, opt);
 }
 
+#if !NCNN_VULKAN_ONLINE_SPIRV
 int VulkanDevice::create_shader_module()
 {
     if (info.bug_local_size_spec_const)
@@ -1706,6 +1736,7 @@ void VulkanDevice::destroy_shader_module()
 
     shader_modules.clear();
 }
+#endif
 
 int VulkanDevice::init_device_extension()
 {
@@ -2025,6 +2056,531 @@ VulkanDevice* get_gpu_device(int device_index)
     return g_default_vkdev[device_index];
 }
 
+#if NCNN_VULKAN_ONLINE_SPIRV
+
+const TBuiltInResource default_TBuiltInResource = {
+    /* .MaxLights = */ 32,
+    /* .MaxClipPlanes = */ 6,
+    /* .MaxTextureUnits = */ 32,
+    /* .MaxTextureCoords = */ 32,
+    /* .MaxVertexAttribs = */ 64,
+    /* .MaxVertexUniformComponents = */ 4096,
+    /* .MaxVaryingFloats = */ 64,
+    /* .MaxVertexTextureImageUnits = */ 32,
+    /* .MaxCombinedTextureImageUnits = */ 80,
+    /* .MaxTextureImageUnits = */ 32,
+    /* .MaxFragmentUniformComponents = */ 4096,
+    /* .MaxDrawBuffers = */ 32,
+    /* .MaxVertexUniformVectors = */ 128,
+    /* .MaxVaryingVectors = */ 8,
+    /* .MaxFragmentUniformVectors = */ 16,
+    /* .MaxVertexOutputVectors = */ 16,
+    /* .MaxFragmentInputVectors = */ 15,
+    /* .MinProgramTexelOffset = */ -8,
+    /* .MaxProgramTexelOffset = */ 7,
+    /* .MaxClipDistances = */ 8,
+    /* .MaxComputeWorkGroupCountX = */ 65535,
+    /* .MaxComputeWorkGroupCountY = */ 65535,
+    /* .MaxComputeWorkGroupCountZ = */ 65535,
+    /* .MaxComputeWorkGroupSizeX = */ 1024,
+    /* .MaxComputeWorkGroupSizeY = */ 1024,
+    /* .MaxComputeWorkGroupSizeZ = */ 64,
+    /* .MaxComputeUniformComponents = */ 1024,
+    /* .MaxComputeTextureImageUnits = */ 16,
+    /* .MaxComputeImageUniforms = */ 8,
+    /* .MaxComputeAtomicCounters = */ 8,
+    /* .MaxComputeAtomicCounterBuffers = */ 1,
+    /* .MaxVaryingComponents = */ 60,
+    /* .MaxVertexOutputComponents = */ 64,
+    /* .MaxGeometryInputComponents = */ 64,
+    /* .MaxGeometryOutputComponents = */ 128,
+    /* .MaxFragmentInputComponents = */ 128,
+    /* .MaxImageUnits = */ 8,
+    /* .MaxCombinedImageUnitsAndFragmentOutputs = */ 8,
+    /* .MaxCombinedShaderOutputResources = */ 8,
+    /* .MaxImageSamples = */ 0,
+    /* .MaxVertexImageUniforms = */ 0,
+    /* .MaxTessControlImageUniforms = */ 0,
+    /* .MaxTessEvaluationImageUniforms = */ 0,
+    /* .MaxGeometryImageUniforms = */ 0,
+    /* .MaxFragmentImageUniforms = */ 8,
+    /* .MaxCombinedImageUniforms = */ 8,
+    /* .MaxGeometryTextureImageUnits = */ 16,
+    /* .MaxGeometryOutputVertices = */ 256,
+    /* .MaxGeometryTotalOutputComponents = */ 1024,
+    /* .MaxGeometryUniformComponents = */ 1024,
+    /* .MaxGeometryVaryingComponents = */ 64,
+    /* .MaxTessControlInputComponents = */ 128,
+    /* .MaxTessControlOutputComponents = */ 128,
+    /* .MaxTessControlTextureImageUnits = */ 16,
+    /* .MaxTessControlUniformComponents = */ 1024,
+    /* .MaxTessControlTotalOutputComponents = */ 4096,
+    /* .MaxTessEvaluationInputComponents = */ 128,
+    /* .MaxTessEvaluationOutputComponents = */ 128,
+    /* .MaxTessEvaluationTextureImageUnits = */ 16,
+    /* .MaxTessEvaluationUniformComponents = */ 1024,
+    /* .MaxTessPatchComponents = */ 120,
+    /* .MaxPatchVertices = */ 32,
+    /* .MaxTessGenLevel = */ 64,
+    /* .MaxViewports = */ 16,
+    /* .MaxVertexAtomicCounters = */ 0,
+    /* .MaxTessControlAtomicCounters = */ 0,
+    /* .MaxTessEvaluationAtomicCounters = */ 0,
+    /* .MaxGeometryAtomicCounters = */ 0,
+    /* .MaxFragmentAtomicCounters = */ 8,
+    /* .MaxCombinedAtomicCounters = */ 8,
+    /* .MaxAtomicCounterBindings = */ 1,
+    /* .MaxVertexAtomicCounterBuffers = */ 0,
+    /* .MaxTessControlAtomicCounterBuffers = */ 0,
+    /* .MaxTessEvaluationAtomicCounterBuffers = */ 0,
+    /* .MaxGeometryAtomicCounterBuffers = */ 0,
+    /* .MaxFragmentAtomicCounterBuffers = */ 1,
+    /* .MaxCombinedAtomicCounterBuffers = */ 1,
+    /* .MaxAtomicCounterBufferSize = */ 16384,
+    /* .MaxTransformFeedbackBuffers = */ 4,
+    /* .MaxTransformFeedbackInterleavedComponents = */ 64,
+    /* .MaxCullDistances = */ 8,
+    /* .MaxCombinedClipAndCullDistances = */ 8,
+    /* .MaxSamples = */ 4,
+    /* .maxMeshOutputVerticesNV = */ 256,
+    /* .maxMeshOutputPrimitivesNV = */ 512,
+    /* .maxMeshWorkGroupSizeX_NV = */ 32,
+    /* .maxMeshWorkGroupSizeY_NV = */ 1,
+    /* .maxMeshWorkGroupSizeZ_NV = */ 1,
+    /* .maxTaskWorkGroupSizeX_NV = */ 32,
+    /* .maxTaskWorkGroupSizeY_NV = */ 1,
+    /* .maxTaskWorkGroupSizeZ_NV = */ 1,
+    /* .maxMeshViewCountNV = */ 4,
+    /* .maxDualSourceDrawBuffersEXT = */ 1,
+
+    /* .limits = */ {
+        /* .nonInductiveForLoops = */ 1,
+        /* .whileLoops = */ 1,
+        /* .doWhileLoops = */ 1,
+        /* .generalUniformIndexing = */ 1,
+        /* .generalAttributeMatrixVectorIndexing = */ 1,
+        /* .generalVaryingIndexing = */ 1,
+        /* .generalSamplerIndexing = */ 1,
+        /* .generalVariableIndexing = */ 1,
+        /* .generalConstantMatrixVectorIndexing = */ 1,
+    }
+};
+
+int compile_spirv_module(int shader_type_index, const Option& opt, std::vector<uint32_t>& spirv)
+{
+    if (shader_type_index < 0 || shader_type_index >= layer_shader_registry_entry_count)
+    {
+        NCNN_LOGE("no such shader module %d", shader_type_index);
+        return -1;
+    }
+
+    const char* comp_data = layer_shader_registry[shader_type_index].comp_data;
+    int comp_data_size = layer_shader_registry[shader_type_index].comp_data_size;
+
+    std::vector< std::pair<const char*, const char*> > custom_defines;
+
+    if (opt.use_fp16_storage)
+    {
+        custom_defines.push_back(std::make_pair("sfp", "float16_t"));
+        custom_defines.push_back(std::make_pair("sfpvec2", "f16vec2"));
+        custom_defines.push_back(std::make_pair("sfpvec4", "f16vec4"));
+
+        if (opt.use_fp16_arithmetic)
+        {
+            custom_defines.push_back(std::make_pair("sfpvec8", "f16mat2x4"));
+            custom_defines.push_back(std::make_pair("sfpmat4", "f16mat4"));
+        }
+    }
+    else if (opt.use_fp16_packed)
+    {
+        custom_defines.push_back(std::make_pair("sfp", "float"));
+        custom_defines.push_back(std::make_pair("sfpvec2", "uint"));
+        custom_defines.push_back(std::make_pair("sfpvec4", "uvec2"));
+        custom_defines.push_back(std::make_pair("sfpvec8", "uvec4"));
+    }
+    else
+    {
+        custom_defines.push_back(std::make_pair("sfp", "float"));
+        custom_defines.push_back(std::make_pair("sfpvec2", "vec2"));
+        custom_defines.push_back(std::make_pair("sfpvec4", "vec4"));
+        custom_defines.push_back(std::make_pair("sfpvec8", "mat2x4"));
+        custom_defines.push_back(std::make_pair("sfpmat4", "mat4"));
+    }
+
+    if (opt.use_fp16_arithmetic)
+    {
+        custom_defines.push_back(std::make_pair("afp", "float16_t"));
+        custom_defines.push_back(std::make_pair("afpvec2", "f16vec2"));
+        custom_defines.push_back(std::make_pair("afpvec4", "f16vec4"));
+        custom_defines.push_back(std::make_pair("afpvec8", "f16mat2x4"));
+        custom_defines.push_back(std::make_pair("afpmat4", "f16mat4"));
+    }
+    else
+    {
+        custom_defines.push_back(std::make_pair("afp", "float"));
+        custom_defines.push_back(std::make_pair("afpvec2", "vec2"));
+        custom_defines.push_back(std::make_pair("afpvec4", "vec4"));
+        custom_defines.push_back(std::make_pair("afpvec8", "mat2x4"));
+        custom_defines.push_back(std::make_pair("afpmat4", "mat4"));
+    }
+
+    if (opt.use_fp16_storage && opt.use_fp16_arithmetic)
+    {
+        custom_defines.push_back(std::make_pair("buffer_ld1(buf,i)", "buf[i]"));
+        custom_defines.push_back(std::make_pair("buffer_st1(buf,i,v)", "{buf[i]=v;}"));
+        custom_defines.push_back(std::make_pair("buffer_cp1(buf,i,sbuf,si)", "{buf[i]=sbuf[si];}"));
+        custom_defines.push_back(std::make_pair("buffer_cp1to4(buf,i,sbuf,si4)", "{buf[i]=f16vec4(sbuf[si4.r],sbuf[si4.g],sbuf[si4.b],sbuf[si4.a]);}"));
+        custom_defines.push_back(std::make_pair("buffer_cp1to8(buf,i,sbuf,si4,sii4)", "{buf[i]=f16mat2x4(sbuf[si4.r],sbuf[si4.g],sbuf[si4.b],sbuf[si4.a],sbuf[sii4.r],sbuf[sii4.g],sbuf[sii4.b],sbuf[sii4.a]);}"));
+        custom_defines.push_back(std::make_pair("buffer_ld2(buf,i)", "buf[i]"));
+        custom_defines.push_back(std::make_pair("buffer_st2(buf,i,v)", "{buf[i]=v;}"));
+        custom_defines.push_back(std::make_pair("buffer_cp2(buf,i,sbuf,si)", "{buf[i]=sbuf[si];}"));
+        custom_defines.push_back(std::make_pair("buffer_ld4(buf,i)", "buf[i]"));
+        custom_defines.push_back(std::make_pair("buffer_st4(buf,i,v)", "{buf[i]=v;}"));
+        custom_defines.push_back(std::make_pair("buffer_cp4(buf,i,sbuf,si)", "{buf[i]=sbuf[si];}"));
+        custom_defines.push_back(std::make_pair("buffer_cp4to1(buf,i4,sbuf,si)", "{buf[i4.r]=sbuf[si].r;buf[i4.g]=sbuf[si].g;buf[i4.b]=sbuf[si].b;buf[i4.a]=sbuf[si].a;}"));
+        custom_defines.push_back(std::make_pair("buffer_cp4to8(buf,i,sbuf,si2)", "{buf[i]=f16mat2x4(sbuf[si2.r],sbuf[si2.g]);}"));
+        custom_defines.push_back(std::make_pair("buffer_ld8(buf,i)", "buf[i]"));
+        custom_defines.push_back(std::make_pair("buffer_st8(buf,i,v)", "{buf[i]=v;}"));
+        custom_defines.push_back(std::make_pair("buffer_cp8(buf,i,sbuf,si)", "{buf[i]=sbuf[si];}"));
+        custom_defines.push_back(std::make_pair("buffer_cp8to1(buf,i4,ii4,sbuf,si)", "{f16mat2x4 _v=sbuf[si]; buf[i4.r]=_v[0].r;buf[i4.g]=_v[0].g;buf[i4.b]=_v[0].b;buf[i4.a]=_v[0].a; buf[ii4.r]=_v[1].r;buf[ii4.g]=_v[1].g;buf[ii4.b]=_v[1].b;buf[ii4.a]=_v[1].a;}"));
+        custom_defines.push_back(std::make_pair("buffer_cp8to4(buf,i2,sbuf,si)", "{f16mat2x4 _v=sbuf[si]; buf[i2.r]=_v[0];buf[i2.g]=_v[1];}"));
+        custom_defines.push_back(std::make_pair("sfp2afpmat4(v)", "v"));
+        custom_defines.push_back(std::make_pair("afp2sfpmat4(v)", "v"));
+    }
+    else if (opt.use_fp16_packed && opt.use_fp16_arithmetic)
+    {
+        custom_defines.push_back(std::make_pair("buffer_ld1(buf,i)", "float16_t(buf[i])"));
+        custom_defines.push_back(std::make_pair("buffer_st1(buf,i,v)", "{buf[i]=float(v);}"));
+        custom_defines.push_back(std::make_pair("buffer_cp1(buf,i,sbuf,si)", "{buf[i]=sbuf[si];}"));
+        custom_defines.push_back(std::make_pair("buffer_cp1to4(buf,i,sbuf,si4)", "{buf[i]=uvec2(packHalf2x16(vec2(f16vec2(sbuf[si4.r],sbuf[si4.g]))),packHalf2x16(vec2(f16vec2(sbuf[si4.b],sbuf[si4.a]))));}"));
+        custom_defines.push_back(std::make_pair("buffer_cp1to8(buf,i,sbuf,si4,sii4)", "{buf[i]=uvec4(packHalf2x16(vec2(f16vec2(sbuf[si4.r],sbuf[si4.g]))),packHalf2x16(vec2(f16vec2(sbuf[si4.b],sbuf[si4.a]))),packHalf2x16(vec2(f16vec2(sbuf[sii4.r],sbuf[sii4.g]))),packHalf2x16(vec2(f16vec2(sbuf[sii4.b],sbuf[sii4.a]))));}"));
+        custom_defines.push_back(std::make_pair("buffer_ld2(buf,i)", "f16vec2(unpackHalf2x16(buf[i]))"));
+        custom_defines.push_back(std::make_pair("buffer_st2(buf,i,v)", "{buf[i]=packHalf2x16(vec2(v))}"));
+        custom_defines.push_back(std::make_pair("buffer_cp2(buf,i,sbuf,si)", "{buf[i]=sbuf[si];}"));
+        custom_defines.push_back(std::make_pair("buffer_ld4(buf,i)", "f16vec4(vec4(unpackHalf2x16(buf[i].x),unpackHalf2x16(buf[i].y)))"));
+        custom_defines.push_back(std::make_pair("buffer_st4(buf,i,v)", "{buf[i]=uvec2(packHalf2x16(vec2(v.rg)),packHalf2x16(vec2(v.ba)));}"));
+        custom_defines.push_back(std::make_pair("buffer_cp4(buf,i,sbuf,si)", "{buf[i]=sbuf[si];}"));
+        custom_defines.push_back(std::make_pair("buffer_cp4to1(buf,i4,sbuf,si)", "{uvec2 _v=sbuf[si]; vec2 _v0=unpackHalf2x16(_v.x);vec2 _v1=unpackHalf2x16(_v.y); buf[i4.r]=_v0.r;buf[i4.g]=_v0.g;buf[i4.b]=_v1.r;buf[i4.a]=_v1.g;}"));
+        custom_defines.push_back(std::make_pair("buffer_cp4to8(buf,i,sbuf,si2)", "{buf[i]=uvec4(sbuf[si2.r],sbuf[si2.g]);}"));
+        custom_defines.push_back(std::make_pair("buffer_ld8(buf,i)", "f16mat2x4(f16vec4(vec4(unpackHalf2x16(buf[i].r),unpackHalf2x16(buf[i].g))),f16vec4(vec4(unpackHalf2x16(buf[i].b),unpackHalf2x16(buf[i].a))))"));
+        custom_defines.push_back(std::make_pair("buffer_st8(buf,i,v)", "{buf[i]=uvec4(uvec2(packHalf2x16(vec2(v[0].rg)),packHalf2x16(vec2(v[0].ba))),uvec2(packHalf2x16(vec2(v[1].rg)),packHalf2x16(vec2(v[1].ba))));}"));
+        custom_defines.push_back(std::make_pair("buffer_cp8(buf,i,sbuf,si)", "{buf[i]=sbuf[si];}"));
+        custom_defines.push_back(std::make_pair("buffer_cp8to1(buf,i4,ii4,sbuf,si)", "{uvec4 _v=sbuf[si]; vec2 _v0=unpackHalf2x16(_v.r);vec2 _v1=unpackHalf2x16(_v.g);vec2 _v2=unpackHalf2x16(_v.b);vec2 _v3=unpackHalf2x16(_v.a); buf[i4.r]=_v0.r;buf[i4.g]=_v0.g;buf[i4.b]=_v1.r;buf[i4.a]=_v1.g; buf[ii4.r]=_v2.r;buf[ii4.g]=_v2.g;buf[ii4.b]=_v3.r;buf[ii4.a]=_v3.g;}"));
+        custom_defines.push_back(std::make_pair("buffer_cp8to4(buf,i2,sbuf,si)", "{uvec4 _v=sbuf[si]; buf[i2.r]=_v.rg;buf[i2.g]=_v.ba;}"));
+    }
+    else if (opt.use_fp16_storage)
+    {
+        custom_defines.push_back(std::make_pair("buffer_ld1(buf,i)", "float(buf[i])"));
+        custom_defines.push_back(std::make_pair("buffer_st1(buf,i,v)", "{buf[i]=float16_t(v);}"));
+        custom_defines.push_back(std::make_pair("buffer_cp1(buf,i,sbuf,si)", "{buf[i]=sbuf[si];}"));
+        custom_defines.push_back(std::make_pair("buffer_cp1to4(buf,i,sbuf,si4)", "{buf[i].r=sbuf[si4.r];buf[i].g=sbuf[si4.g];buf[i].b=sbuf[si4.b];buf[i].a=sbuf[si4.a];}"));
+        custom_defines.push_back(std::make_pair("buffer_cp1to8(buf,i,sbuf,si4,sii4)", "{buf[i].abcd.r=sbuf[si4.r];buf[i].abcd.g=sbuf[si4.g];buf[i].abcd.b=sbuf[si4.b];buf[i].abcd.a=sbuf[si4.a];buf[i].efgh.r=sbuf[sii4.r];buf[i].efgh.g=sbuf[sii4.g];buf[i].efgh.b=sbuf[sii4.b];buf[i].efgh.a=sbuf[sii4.a];}"));
+        custom_defines.push_back(std::make_pair("buffer_ld2(buf,i)", "vec2(buf[i])"));
+        custom_defines.push_back(std::make_pair("buffer_st2(buf,i,v)", "{buf[i]=f16vec2(v);}"));
+        custom_defines.push_back(std::make_pair("buffer_cp2(buf,i,sbuf,si)", "{buf[i]=sbuf[si];}"));
+        custom_defines.push_back(std::make_pair("buffer_ld4(buf,i)", "vec4(buf[i])"));
+        custom_defines.push_back(std::make_pair("buffer_st4(buf,i,v)", "{buf[i]=f16vec4(v);}"));
+        custom_defines.push_back(std::make_pair("buffer_cp4(buf,i,sbuf,si)", "{buf[i]=sbuf[si];}"));
+        custom_defines.push_back(std::make_pair("buffer_cp4to1(buf,i4,sbuf,si)", "{buf[i4.r]=sbuf[si].r;buf[i4.g]=sbuf[si].g;buf[i4.b]=sbuf[si].b;buf[i4.a]=sbuf[si].a;}"));
+        custom_defines.push_back(std::make_pair("buffer_cp4to8(buf,i,sbuf,si2)", "{buf[i].abcd=sbuf[si2.r];buf[i].efgh=sbuf[si2.g];}"));
+        custom_defines.push_back(std::make_pair("buffer_ld8(buf,i)", "mat2x4(vec4(buf[i].abcd),vec4(buf[i].efgh))"));
+        custom_defines.push_back(std::make_pair("buffer_st8(buf,i,v)", "{buf[i].abcd=f16vec4(v[0]);buf[i].efgh=f16vec4(v[1]);}"));
+        custom_defines.push_back(std::make_pair("buffer_cp8(buf,i,sbuf,si)", "{buf[i].abcd=sbuf[si].abcd;buf[i].efgh=sbuf[si].efgh;}"));
+        custom_defines.push_back(std::make_pair("buffer_cp8to1(buf,i4,ii4,sbuf,si)", "{buf[i4.r]=sbuf[si].abcd.r;buf[i4.g]=sbuf[si].abcd.g;buf[i4.b]=sbuf[si].abcd.b;buf[i4.a]=sbuf[si].abcd.a; buf[ii4.r]=sbuf[si].efgh.r;buf[ii4.g]=sbuf[si].efgh.g;buf[ii4.b]=sbuf[si].efgh.b;buf[ii4.a]=sbuf[si].efgh.a;}"));
+        custom_defines.push_back(std::make_pair("buffer_cp8to4(buf,i2,sbuf,si)", "{buf[i2.r]=sbuf[si].abcd;buf[i2.g]=sbuf[si].efgh;}"));
+    }
+    else if (opt.use_fp16_packed)
+    {
+        custom_defines.push_back(std::make_pair("buffer_ld1(buf,i)", "buf[i]"));
+        custom_defines.push_back(std::make_pair("buffer_st1(buf,i,v)", "{buf[i]=v;}"));
+        custom_defines.push_back(std::make_pair("buffer_cp1(buf,i,sbuf,si)", "{buf[i]=sbuf[si];}"));
+        custom_defines.push_back(std::make_pair("buffer_cp1to4(buf,i,sbuf,si4)", "{buf[i]=uvec2(packHalf2x16(vec2(sbuf[si4.r],sbuf[si4.g])),packHalf2x16(vec2(sbuf[si4.b],sbuf[si4.a])));}"));
+        custom_defines.push_back(std::make_pair("buffer_cp1to8(buf,i,sbuf,si4,sii4)", "{buf[i]=uvec4(packHalf2x16(vec2(sbuf[si4.r],sbuf[si4.g])),packHalf2x16(vec2(sbuf[si4.b],sbuf[si4.a])),packHalf2x16(vec2(sbuf[sii4.r],sbuf[sii4.g])),packHalf2x16(vec2(sbuf[sii4.b],sbuf[sii4.a])));}"));
+        custom_defines.push_back(std::make_pair("buffer_ld2(buf,i)", "unpackHalf2x16(buf[i])"));
+        custom_defines.push_back(std::make_pair("buffer_st2(buf,i,v)", "{buf[i]=packHalf2x16(v)}"));
+        custom_defines.push_back(std::make_pair("buffer_cp2(buf,i,sbuf,si)", "{buf[i]=sbuf[si];}"));
+        custom_defines.push_back(std::make_pair("buffer_ld4(buf,i)", "vec4(unpackHalf2x16(buf[i].x),unpackHalf2x16(buf[i].y))"));
+        custom_defines.push_back(std::make_pair("buffer_st4(buf,i,v)", "{buf[i]=uvec2(packHalf2x16(v.rg),packHalf2x16(v.ba));}"));
+        custom_defines.push_back(std::make_pair("buffer_cp4(buf,i,sbuf,si)", "{buf[i]=sbuf[si];}"));
+        custom_defines.push_back(std::make_pair("buffer_cp4to1(buf,i4,sbuf,si)", "{uvec2 _v=sbuf[si]; vec2 _v0=unpackHalf2x16(_v.x);vec2 _v1=unpackHalf2x16(_v.y); buf[i4.r]=_v0.r;buf[i4.g]=_v0.g;buf[i4.b]=_v1.r;buf[i4.a]=_v1.g;}"));
+        custom_defines.push_back(std::make_pair("buffer_cp4to8(buf,i,sbuf,si2)", "{buf[i]=uvec4(sbuf[si2.r],sbuf[si2.g]);}"));
+        custom_defines.push_back(std::make_pair("buffer_ld8(buf,i)", "mat2x4(vec4(unpackHalf2x16(buf[i].r),unpackHalf2x16(buf[i].g)),vec4(unpackHalf2x16(buf[i].b),unpackHalf2x16(buf[i].a)))"));
+        custom_defines.push_back(std::make_pair("buffer_st8(buf,i,v)", "{buf[i]=uvec4(uvec2(packHalf2x16(v[0].rg),packHalf2x16(v[0].ba)),uvec2(packHalf2x16(v[1].rg),packHalf2x16(v[1].ba)));}"));
+        custom_defines.push_back(std::make_pair("buffer_cp8(buf,i,sbuf,si)", "{buf[i]=sbuf[si];}"));
+        custom_defines.push_back(std::make_pair("buffer_cp8to1(buf,i4,ii4,sbuf,si)", "{uvec4 _v=sbuf[si]; vec2 _v0=unpackHalf2x16(_v.r);vec2 _v1=unpackHalf2x16(_v.g);vec2 _v2=unpackHalf2x16(_v.b);vec2 _v3=unpackHalf2x16(_v.a); buf[i4.r]=_v0.r;buf[i4.g]=_v0.g;buf[i4.b]=_v1.r;buf[i4.a]=_v1.g; buf[ii4.r]=_v2.r;buf[ii4.g]=_v2.g;buf[ii4.b]=_v3.r;buf[ii4.a]=_v3.g;}"));
+        custom_defines.push_back(std::make_pair("buffer_cp8to4(buf,i2,sbuf,si)", "{uvec4 _v=sbuf[si]; buf[i2.r]=_v.rg;buf[i2.g]=_v.ba;}"));
+    }
+    else
+    {
+        custom_defines.push_back(std::make_pair("buffer_ld1(buf,i)", "buf[i]"));
+        custom_defines.push_back(std::make_pair("buffer_st1(buf,i,v)", "{buf[i]=v;}"));
+        custom_defines.push_back(std::make_pair("buffer_cp1(buf,i,sbuf,si)", "{buf[i]=sbuf[si];}"));
+        custom_defines.push_back(std::make_pair("buffer_cp1to4(buf,i,sbuf,si4)", "{buf[i]=vec4(sbuf[si4.r],sbuf[si4.g],sbuf[si4.b],sbuf[si4.a]);}"));
+        custom_defines.push_back(std::make_pair("buffer_cp1to8(buf,i,sbuf,si4,sii4)", "{buf[i]=mat2x4(sbuf[si4.r],sbuf[si4.g],sbuf[si4.b],sbuf[si4.a],sbuf[sii4.r],sbuf[sii4.g],sbuf[sii4.b],sbuf[sii4.a]);}"));
+        custom_defines.push_back(std::make_pair("buffer_ld2(buf,i)", "buf[i]"));
+        custom_defines.push_back(std::make_pair("buffer_st2(buf,i,v)", "{buf[i]=v;}"));
+        custom_defines.push_back(std::make_pair("buffer_cp2(buf,i,sbuf,si)", "{buf[i]=sbuf[si];}"));
+        custom_defines.push_back(std::make_pair("buffer_ld4(buf,i)", "buf[i]"));
+        custom_defines.push_back(std::make_pair("buffer_st4(buf,i,v)", "{buf[i]=v;}"));
+        custom_defines.push_back(std::make_pair("buffer_cp4(buf,i,sbuf,si)", "{buf[i]=sbuf[si];}"));
+        custom_defines.push_back(std::make_pair("buffer_cp4to1(buf,i4,sbuf,si)", "{vec4 _v=sbuf[si]; buf[i4.r]=_v.r;buf[i4.g]=_v.g;buf[i4.b]=_v.b;buf[i4.a]=_v.a;}"));
+        custom_defines.push_back(std::make_pair("buffer_cp4to8(buf,i,sbuf,si2)", "{buf[i]=mat2x4(sbuf[si2.r],sbuf[si2.g]);}"));
+        custom_defines.push_back(std::make_pair("buffer_ld8(buf,i)", "buf[i]"));
+        custom_defines.push_back(std::make_pair("buffer_st8(buf,i,v)", "{buf[i]=v;}"));
+        custom_defines.push_back(std::make_pair("buffer_cp8(buf,i,sbuf,si)", "{buf[i]=sbuf[si];}"));
+        custom_defines.push_back(std::make_pair("buffer_cp8to1(buf,i4,ii4,sbuf,si)", "{mat2x4 _v=sbuf[si]; buf[i4.r]=_v[0].r;buf[i4.g]=_v[0].g;buf[i4.b]=_v[0].b;buf[i4.a]=_v[0].a; buf[ii4.r]=_v[1].r;buf[ii4.g]=_v[1].g;buf[ii4.b]=_v[1].b;buf[ii4.a]=_v[1].a;}"));
+        custom_defines.push_back(std::make_pair("buffer_cp8to4(buf,i2,sbuf,si)", "{mat2x4 _v=sbuf[si]; buf[i2.r]=_v[0];buf[i2.g]=_v[1];}"));
+        custom_defines.push_back(std::make_pair("sfp2afpmat4(v)", "v"));
+        custom_defines.push_back(std::make_pair("afp2sfpmat4(v)", "v"));
+    }
+
+    if (opt.use_image_storage)
+    {
+        if (opt.use_fp16_storage)
+        {
+            custom_defines.push_back(std::make_pair("imfmtc1", "r16f"));
+            custom_defines.push_back(std::make_pair("imfmtc4", "rgba16f"));
+            custom_defines.push_back(std::make_pair("unfp", "mediump"));
+        }
+        else if (opt.use_fp16_packed)
+        {
+            custom_defines.push_back(std::make_pair("imfmtc1", "r32f"));
+            custom_defines.push_back(std::make_pair("imfmtc4", "rgba16f"));
+            custom_defines.push_back(std::make_pair("unfp", "mediump"));
+        }
+        else
+        {
+            custom_defines.push_back(std::make_pair("imfmtc1", "r32f"));
+            custom_defines.push_back(std::make_pair("imfmtc4", "rgba32f"));
+            custom_defines.push_back(std::make_pair("unfp", "highp"));
+        }
+
+        if (opt.use_fp16_storage && opt.use_fp16_arithmetic)
+        {
+            custom_defines.push_back(std::make_pair("image1d_ld1(tex,p)", "float16_t(texelFetch(tex,p,0).r)"));
+            custom_defines.push_back(std::make_pair("image2d_ld1(tex,p)", "float16_t(texelFetch(tex,p,0).r)"));
+            custom_defines.push_back(std::make_pair("image3d_ld1(tex,p)", "float16_t(texelFetch(tex,p,0).r)"));
+            custom_defines.push_back(std::make_pair("image1d_st1(img,p,v)", "{f16vec4 _v;_v.r=float16_t(v);imageStore(img,p,_v);}"));
+            custom_defines.push_back(std::make_pair("image2d_st1(img,p,v)", "{f16vec4 _v;_v.r=float16_t(v);imageStore(img,p,_v);}"));
+            custom_defines.push_back(std::make_pair("image3d_st1(img,p,v)", "{f16vec4 _v;_v.r=float16_t(v);imageStore(img,p,_v);}"));
+            custom_defines.push_back(std::make_pair("image1d_cp1(img,p,tex,sp)", "{imageStore(img,p,texelFetch(tex,sp,0));}"));
+            custom_defines.push_back(std::make_pair("image2d_cp1(img,p,tex,sp)", "{imageStore(img,p,texelFetch(tex,sp,0));}"));
+            custom_defines.push_back(std::make_pair("image3d_cp1(img,p,tex,sp)", "{imageStore(img,p,texelFetch(tex,sp,0));}"));
+            custom_defines.push_back(std::make_pair("image1d_ld4(tex,p)", "f16vec4(texelFetch(tex,p,0))"));
+            custom_defines.push_back(std::make_pair("image2d_ld4(tex,p)", "f16vec4(texelFetch(tex,p,0))"));
+            custom_defines.push_back(std::make_pair("image3d_ld4(tex,p)", "f16vec4(texelFetch(tex,p,0))"));
+            custom_defines.push_back(std::make_pair("image1d_st4(img,p,v)", "{imageStore(img,p,vec4(v));}"));
+            custom_defines.push_back(std::make_pair("image2d_st4(img,p,v)", "{imageStore(img,p,vec4(v));}"));
+            custom_defines.push_back(std::make_pair("image3d_st4(img,p,v)", "{imageStore(img,p,vec4(v));}"));
+            custom_defines.push_back(std::make_pair("image1d_cp4(img,p,tex,sp)", "{imageStore(img,p,texelFetch(tex,sp,0));}"));
+            custom_defines.push_back(std::make_pair("image2d_cp4(img,p,tex,sp)", "{imageStore(img,p,texelFetch(tex,sp,0));}"));
+            custom_defines.push_back(std::make_pair("image3d_cp4(img,p,tex,sp)", "{imageStore(img,p,texelFetch(tex,sp,0));}"));
+            custom_defines.push_back(std::make_pair("image1d_ld8(tex,p)", "f16mat2x4(texelFetch(tex,(p)*2,0),texelFetch(tex,(p)*2+1,0))"));
+            custom_defines.push_back(std::make_pair("image2d_ld8(tex,p)", "f16mat2x4(texelFetch(tex,ivec2(p.x*2,p.y),0),texelFetch(tex,ivec2(p.x*2+1,p.y),0))"));
+            custom_defines.push_back(std::make_pair("image3d_ld8(tex,p)", "f16mat2x4(texelFetch(tex,ivec3(p.x*2,p.y,p.z),0),texelFetch(tex,ivec3(p.x*2+1,p.y,p.z),0))"));
+            custom_defines.push_back(std::make_pair("image1d_st8(img,p,v)", "{imageStore(img,(p)*2,vec4(v[0]));imageStore(img,(p)*2+1,vec4(v[1]));}"));
+            custom_defines.push_back(std::make_pair("image2d_st8(img,p,v)", "{imageStore(img,ivec2(p.x*2,p.y),vec4(v[0]));imageStore(img,ivec2(p.x*2+1,p.y),vec4(v[1]));}"));
+            custom_defines.push_back(std::make_pair("image3d_st8(img,p,v)", "{imageStore(img,ivec3(p.x*2,p.y,p.z),vec4(v[0]));imageStore(img,ivec3(p.x*2+1,p.y,p.z),vec4(v[1]));}"));
+            custom_defines.push_back(std::make_pair("image1d_cp8(img,p,tex,sp)", "{imageStore(img,(p)*2,texelFetch(tex,sp*2,0));imageStore(img,(p)*2+1,texelFetch(tex,sp*2+1,0));}"));
+            custom_defines.push_back(std::make_pair("image2d_cp8(img,p,tex,sp)", "{imageStore(img,ivec2(p.x*2,p.y),texelFetch(tex,ivec2(sp.x*2,sp.y),0));imageStore(img,ivec2(p.x*2+1,p.y),texelFetch(tex,ivec2(sp.x*2+1,sp.y),0));}"));
+            custom_defines.push_back(std::make_pair("image3d_cp8(img,p,tex,sp)", "{imageStore(img,ivec3(p.x*2,p.y,p.z),texelFetch(tex,ivec3(sp.x*2,sp.y,sp.z),0));imageStore(img,ivec3(p.x*2+1,p.y,p.z),texelFetch(tex,ivec3(sp.x*2+1,sp.y,sp.z),0));}"));
+        }
+        else if (opt.use_fp16_packed && opt.use_fp16_arithmetic)
+        {
+            custom_defines.push_back(std::make_pair("image1d_ld1(tex,p)", "float16_t(texelFetch(tex,p,0).r)"));
+            custom_defines.push_back(std::make_pair("image2d_ld1(tex,p)", "float16_t(texelFetch(tex,p,0).r)"));
+            custom_defines.push_back(std::make_pair("image3d_ld1(tex,p)", "float16_t(texelFetch(tex,p,0).r)"));
+            custom_defines.push_back(std::make_pair("image1d_st1(img,p,v)", "{vec4 _v;_v.r=v;imageStore(img,p,_v);}"));
+            custom_defines.push_back(std::make_pair("image2d_st1(img,p,v)", "{vec4 _v;_v.r=v;imageStore(img,p,_v);}"));
+            custom_defines.push_back(std::make_pair("image3d_st1(img,p,v)", "{vec4 _v;_v.r=v;imageStore(img,p,_v);}"));
+            custom_defines.push_back(std::make_pair("image1d_cp1(img,p,tex,sp)", "{imageStore(img,p,texelFetch(tex,sp,0));}"));
+            custom_defines.push_back(std::make_pair("image2d_cp1(img,p,tex,sp)", "{imageStore(img,p,texelFetch(tex,sp,0));}"));
+            custom_defines.push_back(std::make_pair("image3d_cp1(img,p,tex,sp)", "{imageStore(img,p,texelFetch(tex,sp,0));}"));
+            custom_defines.push_back(std::make_pair("image1d_ld4(tex,p)", "f16vec4(texelFetch(tex,p,0))"));
+            custom_defines.push_back(std::make_pair("image2d_ld4(tex,p)", "f16vec4(texelFetch(tex,p,0))"));
+            custom_defines.push_back(std::make_pair("image3d_ld4(tex,p)", "f16vec4(texelFetch(tex,p,0))"));
+            custom_defines.push_back(std::make_pair("image1d_st4(img,p,v)", "{imageStore(img,p,v);}"));
+            custom_defines.push_back(std::make_pair("image2d_st4(img,p,v)", "{imageStore(img,p,v);}"));
+            custom_defines.push_back(std::make_pair("image3d_st4(img,p,v)", "{imageStore(img,p,v);}"));
+            custom_defines.push_back(std::make_pair("image1d_cp4(img,p,tex,sp)", "{imageStore(img,p,texelFetch(tex,sp,0));}"));
+            custom_defines.push_back(std::make_pair("image2d_cp4(img,p,tex,sp)", "{imageStore(img,p,texelFetch(tex,sp,0));}"));
+            custom_defines.push_back(std::make_pair("image3d_cp4(img,p,tex,sp)", "{imageStore(img,p,texelFetch(tex,sp,0));}"));
+            custom_defines.push_back(std::make_pair("image1d_ld8(tex,p)", "f16mat2x4(texelFetch(tex,(p)*2,0),texelFetch(tex,(p)*2+1,0))"));
+            custom_defines.push_back(std::make_pair("image2d_ld8(tex,p)", "f16mat2x4(texelFetch(tex,ivec2(p.x*2,p.y),0),texelFetch(tex,ivec2(p.x*2+1,p.y),0))"));
+            custom_defines.push_back(std::make_pair("image3d_ld8(tex,p)", "f16mat2x4(texelFetch(tex,ivec3(p.x*2,p.y,p.z),0),texelFetch(tex,ivec3(p.x*2+1,p.y,p.z),0))"));
+            custom_defines.push_back(std::make_pair("image1d_st8(img,p,v)", "{imageStore(img,(p)*2,v[0]);imageStore(img,(p)*2+1,v[1]);}"));
+            custom_defines.push_back(std::make_pair("image2d_st8(img,p,v)", "{imageStore(img,ivec2(p.x*2,p.y),v[0]);imageStore(img,ivec2(p.x*2+1,p.y),v[1]);}"));
+            custom_defines.push_back(std::make_pair("image3d_st8(img,p,v)", "{imageStore(img,ivec3(p.x*2,p.y,p.z),v[0]);imageStore(img,ivec3(p.x*2+1,p.y,p.z),v[1]);}"));
+            custom_defines.push_back(std::make_pair("image1d_cp8(img,p,tex,sp)", "{imageStore(img,(p)*2,texelFetch(tex,sp*2,0));imageStore(img,(p)*2+1,texelFetch(tex,sp*2+1,0));}"));
+            custom_defines.push_back(std::make_pair("image2d_cp8(img,p,tex,sp)", "{imageStore(img,ivec2(p.x*2,p.y),texelFetch(tex,ivec2(sp.x*2,sp.y),0));imageStore(img,ivec2(p.x*2+1,p.y),texelFetch(tex,ivec2(sp.x*2+1,sp.y),0));}"));
+            custom_defines.push_back(std::make_pair("image3d_cp8(img,p,tex,sp)", "{imageStore(img,ivec3(p.x*2,p.y,p.z),texelFetch(tex,ivec3(sp.x*2,sp.y,sp.z),0));imageStore(img,ivec3(p.x*2+1,p.y,p.z),texelFetch(tex,ivec3(sp.x*2+1,sp.y,sp.z),0));}"));
+        }
+        else if (opt.use_fp16_storage)
+        {
+            custom_defines.push_back(std::make_pair("image1d_ld1(tex,p)", "texelFetch(tex,p,0).r"));
+            custom_defines.push_back(std::make_pair("image2d_ld1(tex,p)", "texelFetch(tex,p,0).r"));
+            custom_defines.push_back(std::make_pair("image3d_ld1(tex,p)", "texelFetch(tex,p,0).r"));
+            custom_defines.push_back(std::make_pair("image1d_st1(img,p,v)", "{vec4 _v;_v.r=v;imageStore(img,p,_v);}"));
+            custom_defines.push_back(std::make_pair("image2d_st1(img,p,v)", "{vec4 _v;_v.r=v;imageStore(img,p,_v);}"));
+            custom_defines.push_back(std::make_pair("image3d_st1(img,p,v)", "{vec4 _v;_v.r=v;imageStore(img,p,_v);}"));
+            custom_defines.push_back(std::make_pair("image1d_cp1(img,p,tex,sp)", "{imageStore(img,p,texelFetch(tex,sp,0));}"));
+            custom_defines.push_back(std::make_pair("image2d_cp1(img,p,tex,sp)", "{imageStore(img,p,texelFetch(tex,sp,0));}"));
+            custom_defines.push_back(std::make_pair("image3d_cp1(img,p,tex,sp)", "{imageStore(img,p,texelFetch(tex,sp,0));}"));
+            custom_defines.push_back(std::make_pair("image1d_ld4(tex,p)", "texelFetch(tex,p,0)"));
+            custom_defines.push_back(std::make_pair("image2d_ld4(tex,p)", "texelFetch(tex,p,0)"));
+            custom_defines.push_back(std::make_pair("image3d_ld4(tex,p)", "texelFetch(tex,p,0)"));
+            custom_defines.push_back(std::make_pair("image1d_st4(img,p,v)", "{imageStore(img,p,v);}"));
+            custom_defines.push_back(std::make_pair("image2d_st4(img,p,v)", "{imageStore(img,p,v);}"));
+            custom_defines.push_back(std::make_pair("image3d_st4(img,p,v)", "{imageStore(img,p,v);}"));
+            custom_defines.push_back(std::make_pair("image1d_cp4(img,p,tex,sp)", "{imageStore(img,p,texelFetch(tex,sp,0));}"));
+            custom_defines.push_back(std::make_pair("image2d_cp4(img,p,tex,sp)", "{imageStore(img,p,texelFetch(tex,sp,0));}"));
+            custom_defines.push_back(std::make_pair("image3d_cp4(img,p,tex,sp)", "{imageStore(img,p,texelFetch(tex,sp,0));}"));
+            custom_defines.push_back(std::make_pair("image1d_ld8(tex,p)", "mat2x4(texelFetch(tex,(p)*2,0),texelFetch(tex,(p)*2+1,0))"));
+            custom_defines.push_back(std::make_pair("image2d_ld8(tex,p)", "mat2x4(texelFetch(tex,ivec2(p.x*2,p.y),0),texelFetch(tex,ivec2(p.x*2+1,p.y),0))"));
+            custom_defines.push_back(std::make_pair("image3d_ld8(tex,p)", "mat2x4(texelFetch(tex,ivec3(p.x*2,p.y,p.z),0),texelFetch(tex,ivec3(p.x*2+1,p.y,p.z),0))"));
+            custom_defines.push_back(std::make_pair("image1d_st8(img,p,v)", "{imageStore(img,(p)*2,v[0]);imageStore(img,(p)*2+1,v[1]);}"));
+            custom_defines.push_back(std::make_pair("image2d_st8(img,p,v)", "{imageStore(img,ivec2(p.x*2,p.y),v[0]);imageStore(img,ivec2(p.x*2+1,p.y),v[1]);}"));
+            custom_defines.push_back(std::make_pair("image3d_st8(img,p,v)", "{imageStore(img,ivec3(p.x*2,p.y,p.z),v[0]);imageStore(img,ivec3(p.x*2+1,p.y,p.z),v[1]);}"));
+            custom_defines.push_back(std::make_pair("image1d_cp8(img,p,tex,sp)", "{imageStore(img,(p)*2,texelFetch(tex,sp*2,0));imageStore(img,(p)*2+1,texelFetch(tex,sp*2+1,0));}"));
+            custom_defines.push_back(std::make_pair("image2d_cp8(img,p,tex,sp)", "{imageStore(img,ivec2(p.x*2,p.y),texelFetch(tex,ivec2(sp.x*2,sp.y),0));imageStore(img,ivec2(p.x*2+1,p.y),texelFetch(tex,ivec2(sp.x*2+1,sp.y),0));}"));
+            custom_defines.push_back(std::make_pair("image3d_cp8(img,p,tex,sp)", "{imageStore(img,ivec3(p.x*2,p.y,p.z),texelFetch(tex,ivec3(sp.x*2,sp.y,sp.z),0));imageStore(img,ivec3(p.x*2+1,p.y,p.z),texelFetch(tex,ivec3(sp.x*2+1,sp.y,sp.z),0));}"));
+        }
+        else if (opt.use_fp16_packed)
+        {
+            custom_defines.push_back(std::make_pair("image1d_ld1(tex,p)", "texelFetch(tex,p,0).r"));
+            custom_defines.push_back(std::make_pair("image2d_ld1(tex,p)", "texelFetch(tex,p,0).r"));
+            custom_defines.push_back(std::make_pair("image3d_ld1(tex,p)", "texelFetch(tex,p,0).r"));
+            custom_defines.push_back(std::make_pair("image1d_st1(img,p,v)", "{vec4 _v;_v.r=v;imageStore(img,p,_v);}"));
+            custom_defines.push_back(std::make_pair("image2d_st1(img,p,v)", "{vec4 _v;_v.r=v;imageStore(img,p,_v);}"));
+            custom_defines.push_back(std::make_pair("image3d_st1(img,p,v)", "{vec4 _v;_v.r=v;imageStore(img,p,_v);}"));
+            custom_defines.push_back(std::make_pair("image1d_cp1(img,p,tex,sp)", "{imageStore(img,p,texelFetch(tex,sp,0));}"));
+            custom_defines.push_back(std::make_pair("image2d_cp1(img,p,tex,sp)", "{imageStore(img,p,texelFetch(tex,sp,0));}"));
+            custom_defines.push_back(std::make_pair("image3d_cp1(img,p,tex,sp)", "{imageStore(img,p,texelFetch(tex,sp,0));}"));
+            custom_defines.push_back(std::make_pair("image1d_ld4(tex,p)", "texelFetch(tex,p,0)"));
+            custom_defines.push_back(std::make_pair("image2d_ld4(tex,p)", "texelFetch(tex,p,0)"));
+            custom_defines.push_back(std::make_pair("image3d_ld4(tex,p)", "texelFetch(tex,p,0)"));
+            custom_defines.push_back(std::make_pair("image1d_st4(img,p,v)", "{imageStore(img,p,v);}"));
+            custom_defines.push_back(std::make_pair("image2d_st4(img,p,v)", "{imageStore(img,p,v);}"));
+            custom_defines.push_back(std::make_pair("image3d_st4(img,p,v)", "{imageStore(img,p,v);}"));
+            custom_defines.push_back(std::make_pair("image1d_cp4(img,p,tex,sp)", "{imageStore(img,p,texelFetch(tex,sp,0));}"));
+            custom_defines.push_back(std::make_pair("image2d_cp4(img,p,tex,sp)", "{imageStore(img,p,texelFetch(tex,sp,0));}"));
+            custom_defines.push_back(std::make_pair("image3d_cp4(img,p,tex,sp)", "{imageStore(img,p,texelFetch(tex,sp,0));}"));
+            custom_defines.push_back(std::make_pair("image1d_ld8(tex,p)", "mat2x4(texelFetch(tex,(p)*2,0),texelFetch(tex,(p)*2+1,0))"));
+            custom_defines.push_back(std::make_pair("image2d_ld8(tex,p)", "mat2x4(texelFetch(tex,ivec2(p.x*2,p.y),0),texelFetch(tex,ivec2(p.x*2+1,p.y),0))"));
+            custom_defines.push_back(std::make_pair("image3d_ld8(tex,p)", "mat2x4(texelFetch(tex,ivec3(p.x*2,p.y,p.z),0),texelFetch(tex,ivec3(p.x*2+1,p.y,p.z),0))"));
+            custom_defines.push_back(std::make_pair("image1d_st8(img,p,v)", "{imageStore(img,(p)*2,v[0]);imageStore(img,(p)*2+1,v[1]);}"));
+            custom_defines.push_back(std::make_pair("image2d_st8(img,p,v)", "{imageStore(img,ivec2(p.x*2,p.y),v[0]);imageStore(img,ivec2(p.x*2+1,p.y),v[1]);}"));
+            custom_defines.push_back(std::make_pair("image3d_st8(img,p,v)", "{imageStore(img,ivec3(p.x*2,p.y,p.z),v[0]);imageStore(img,ivec3(p.x*2+1,p.y,p.z),v[1]);}"));
+            custom_defines.push_back(std::make_pair("image1d_cp8(img,p,tex,sp)", "{imageStore(img,(p)*2,texelFetch(tex,sp*2,0));imageStore(img,(p)*2+1,texelFetch(tex,sp*2+1,0));}"));
+            custom_defines.push_back(std::make_pair("image2d_cp8(img,p,tex,sp)", "{imageStore(img,ivec2(p.x*2,p.y),texelFetch(tex,ivec2(sp.x*2,sp.y),0));imageStore(img,ivec2(p.x*2+1,p.y),texelFetch(tex,ivec2(sp.x*2+1,sp.y),0));}"));
+            custom_defines.push_back(std::make_pair("image3d_cp8(img,p,tex,sp)", "{imageStore(img,ivec3(p.x*2,p.y,p.z),texelFetch(tex,ivec3(sp.x*2,sp.y,sp.z),0));imageStore(img,ivec3(p.x*2+1,p.y,p.z),texelFetch(tex,ivec3(sp.x*2+1,sp.y,sp.z),0));}"));
+        }
+        else
+        {
+            custom_defines.push_back(std::make_pair("image1d_ld1(tex,p)", "texelFetch(tex,p,0).r"));
+            custom_defines.push_back(std::make_pair("image2d_ld1(tex,p)", "texelFetch(tex,p,0).r"));
+            custom_defines.push_back(std::make_pair("image3d_ld1(tex,p)", "texelFetch(tex,p,0).r"));
+            custom_defines.push_back(std::make_pair("image1d_st1(img,p,v)", "{vec4 _v;_v.r=v;imageStore(img,p,_v);}"));
+            custom_defines.push_back(std::make_pair("image2d_st1(img,p,v)", "{vec4 _v;_v.r=v;imageStore(img,p,_v);}"));
+            custom_defines.push_back(std::make_pair("image3d_st1(img,p,v)", "{vec4 _v;_v.r=v;imageStore(img,p,_v);}"));
+            custom_defines.push_back(std::make_pair("image1d_cp1(img,p,tex,sp)", "{imageStore(img,p,texelFetch(tex,sp,0));}"));
+            custom_defines.push_back(std::make_pair("image2d_cp1(img,p,tex,sp)", "{imageStore(img,p,texelFetch(tex,sp,0));}"));
+            custom_defines.push_back(std::make_pair("image3d_cp1(img,p,tex,sp)", "{imageStore(img,p,texelFetch(tex,sp,0));}"));
+            custom_defines.push_back(std::make_pair("image1d_ld4(tex,p)", "texelFetch(tex,p,0)"));
+            custom_defines.push_back(std::make_pair("image2d_ld4(tex,p)", "texelFetch(tex,p,0)"));
+            custom_defines.push_back(std::make_pair("image3d_ld4(tex,p)", "texelFetch(tex,p,0)"));
+            custom_defines.push_back(std::make_pair("image1d_st4(img,p,v)", "{imageStore(img,p,v);}"));
+            custom_defines.push_back(std::make_pair("image2d_st4(img,p,v)", "{imageStore(img,p,v);}"));
+            custom_defines.push_back(std::make_pair("image3d_st4(img,p,v)", "{imageStore(img,p,v);}"));
+            custom_defines.push_back(std::make_pair("image1d_cp4(img,p,tex,sp)", "{imageStore(img,p,texelFetch(tex,sp,0));}"));
+            custom_defines.push_back(std::make_pair("image2d_cp4(img,p,tex,sp)", "{imageStore(img,p,texelFetch(tex,sp,0));}"));
+            custom_defines.push_back(std::make_pair("image3d_cp4(img,p,tex,sp)", "{imageStore(img,p,texelFetch(tex,sp,0));}"));
+            custom_defines.push_back(std::make_pair("image1d_ld8(tex,p)", "mat2x4(texelFetch(tex,(p)*2,0),texelFetch(tex,(p)*2+1,0))"));
+            custom_defines.push_back(std::make_pair("image2d_ld8(tex,p)", "mat2x4(texelFetch(tex,ivec2(p.x*2,p.y),0),texelFetch(tex,ivec2(p.x*2+1,p.y),0))"));
+            custom_defines.push_back(std::make_pair("image3d_ld8(tex,p)", "mat2x4(texelFetch(tex,ivec3(p.x*2,p.y,p.z),0),texelFetch(tex,ivec3(p.x*2+1,p.y,p.z),0))"));
+            custom_defines.push_back(std::make_pair("image1d_st8(img,p,v)", "{imageStore(img,(p)*2,v[0]);imageStore(img,(p)*2+1,v[1]);}"));
+            custom_defines.push_back(std::make_pair("image2d_st8(img,p,v)", "{imageStore(img,ivec2(p.x*2,p.y),v[0]);imageStore(img,ivec2(p.x*2+1,p.y),v[1]);}"));
+            custom_defines.push_back(std::make_pair("image3d_st8(img,p,v)", "{imageStore(img,ivec3(p.x*2,p.y,p.z),v[0]);imageStore(img,ivec3(p.x*2+1,p.y,p.z),v[1]);}"));
+            custom_defines.push_back(std::make_pair("image1d_cp8(img,p,tex,sp)", "{imageStore(img,(p)*2,texelFetch(tex,sp*2,0));imageStore(img,(p)*2+1,texelFetch(tex,sp*2+1,0));}"));
+            custom_defines.push_back(std::make_pair("image2d_cp8(img,p,tex,sp)", "{imageStore(img,ivec2(p.x*2,p.y),texelFetch(tex,ivec2(sp.x*2,sp.y),0));imageStore(img,ivec2(p.x*2+1,p.y),texelFetch(tex,ivec2(sp.x*2+1,sp.y),0));}"));
+            custom_defines.push_back(std::make_pair("image3d_cp8(img,p,tex,sp)", "{imageStore(img,ivec3(p.x*2,p.y,p.z),texelFetch(tex,ivec3(sp.x*2,sp.y,sp.z),0));imageStore(img,ivec3(p.x*2+1,p.y,p.z),texelFetch(tex,ivec3(sp.x*2+1,sp.y,sp.z),0));}"));
+        }
+    }
+
+    custom_defines.push_back(std::make_pair("psc(x)", "(x==0?p.x:x)"));
+
+    if (opt.use_fp16_packed)
+    {
+        custom_defines.push_back(std::make_pair("NCNN_fp16_packed", "1"));
+    }
+    if (opt.use_fp16_storage)
+    {
+        custom_defines.push_back(std::make_pair("NCNN_fp16_storage", "1"));
+    }
+    if (opt.use_fp16_arithmetic)
+    {
+        custom_defines.push_back(std::make_pair("NCNN_fp16_arithmetic", "1"));
+    }
+
+    if (opt.use_image_storage)
+    {
+        custom_defines.push_back(std::make_pair("NCNN_image_shader", "1"));
+    }
+
+    std::string preamble;
+    std::vector<std::string> processes;
+
+    processes.resize(custom_defines.size());
+    for (size_t i = 0; i < custom_defines.size(); i++)
+    {
+        const char* key = custom_defines[i].first;
+        const char* def = custom_defines[i].second;
+
+        preamble += std::string("#define ") + key + " " + def + "\n";
+        processes[i] = std::string("define-macro ") + key + "=" + def;
+    }
+
+    bool compile_success = true;
+
+    {
+        glslang::TShader s(EShLangCompute);
+
+        s.setStringsWithLengths(&comp_data, &comp_data_size, 1);
+
+        s.setPreamble(preamble.c_str());
+        s.addProcesses(processes);
+        s.setEntryPoint("main");
+        s.setSourceEntryPoint("main");
+
+        s.setEnvInput(glslang::EShSourceGlsl, EShLangCompute, glslang::EShClientVulkan, 1);
+        s.setEnvClient(glslang::EShClientVulkan, glslang::EShTargetVulkan_1_0);
+        s.setEnvTarget(glslang::EshTargetSpv, glslang::EShTargetSpv_1_0);
+
+        TBuiltInResource resources = default_TBuiltInResource;
+
+        // although vulkan 1.1 accept glsl directly
+        // ncnn resolve_shader_info() only works with the intermediate spirv code
+        bool pr = s.parse(&resources, 100, false, EShMsgDefault);
+        if (!pr)
+        {
+            NCNN_LOGE("compile spir-v module failed");
+            NCNN_LOGE("%s", s.getInfoLog());
+            NCNN_LOGE("%s", s.getInfoDebugLog());
+
+            compile_success = false;
+        }
+        else
+        {
+            glslang::TIntermediate* ir = s.getIntermediate();
+            glslang::GlslangToSpv(*ir, spirv);
+        }
+    }
+
+    return compile_success ? 0 : -1;
+}
+#endif
+
+#if !NCNN_VULKAN_ONLINE_SPIRV
 const ShaderInfo& get_shader_info(int shader_type_index)
 {
     if (shader_type_index < 0 || shader_type_index >= layer_shader_registry_entry_count)
@@ -2035,6 +2591,7 @@ const ShaderInfo& get_shader_info(int shader_type_index)
 
     return layer_shader_infos[shader_type_index];
 }
+#endif
 
 int resolve_shader_info(const uint32_t* spv_data, size_t spv_data_size, ShaderInfo& shader_info)
 {

src/gpu.h

@@ -173,10 +173,12 @@ public:
 
     VkDevice vkdevice() const { return device; }
 
+#if !NCNN_VULKAN_ONLINE_SPIRV
     VkShaderModule get_shader_module(int shader_type_index) const;
 
     // with fixed workgroup size
     VkShaderModule create_shader_module(int shader_type_index, uint32_t local_size_x, uint32_t local_size_y, uint32_t local_size_z) const;
+#endif
 
     VkShaderModule compile_shader_module(const uint32_t* spv_data, size_t spv_data_size) const;
 
@@ -252,9 +254,11 @@ public:
 #endif // __ANDROID_API__ >= 26
 
 protected:
+#if !NCNN_VULKAN_ONLINE_SPIRV
     // shader management
     int create_shader_module();
     void destroy_shader_module();
+#endif
 
     // device extension
     int init_device_extension();
@@ -269,7 +273,9 @@ protected:
 
 private:
     VkDevice device;
+#if !NCNN_VULKAN_ONLINE_SPIRV
     std::vector<VkShaderModule> shader_modules;
+#endif
 
     // hardware queue
     mutable std::vector<VkQueue> compute_queues;
@@ -304,6 +310,11 @@ private:
 
 VulkanDevice* get_gpu_device(int device_index = get_default_gpu_index());
 
+#if NCNN_VULKAN_ONLINE_SPIRV
+// online spirv compilation
+int compile_spirv_module(int shader_type_index, const Option& opt, std::vector<uint32_t>& spirv);
+#endif
+
 // info from spirv
 class ShaderInfo
 {
@@ -319,7 +330,9 @@ public:
     int binding_types[16];// 16 is large enough I think ...
 };
 
+#if !NCNN_VULKAN_ONLINE_SPIRV
 const ShaderInfo& get_shader_info(int shader_type_index);
+#endif
 int resolve_shader_info(const uint32_t* spv_data, size_t spv_data_size, ShaderInfo& shader_info);
 
 } // namespace ncnn

src/pipeline.cpp

@@ -79,6 +79,47 @@ int Pipeline::create(const uint32_t* spv_data, size_t spv_data_size, const std::
 
 int Pipeline::create(int shader_type_index, const Option& opt, const std::vector<vk_specialization_type>& specializations)
 {
+#if NCNN_VULKAN_ONLINE_SPIRV
+    std::vector<uint32_t> spirv;
+    int retc = compile_spirv_module(shader_type_index, opt, spirv);
+    if (retc != 0)
+    {
+        NCNN_LOGE("compile_spirv_module failed %d", retc);
+        return -1;
+    }
+
+    const uint32_t* spv_data = spirv.data();
+    size_t spv_data_size = spirv.size() * 4;
+
+    ShaderInfo si;
+    int ret = resolve_shader_info(spv_data, spv_data_size, si);
+    if (ret != 0)
+    {
+        NCNN_LOGE("resolve_shader_info failed %d", ret);
+        return -1;
+    }
+
+    // -3 for local_size_xyz
+    int specialization_count_expected = si.specialization_count - 3;
+    if ((int)specializations.size() != specialization_count_expected)
+    {
+        NCNN_LOGE("pipeline specialization count mismatch, expect %d but got %d", specialization_count_expected, (int)specializations.size());
+        return -1;
+    }
+
+    if (vkdev->info.bug_local_size_spec_const)
+    {
+        local_shader_module = vkdev->compile_shader_module(spv_data, spv_data_size, local_size_x, local_size_y, local_size_z);
+    }
+    else
+    {
+        local_shader_module = vkdev->compile_shader_module(spv_data, spv_data_size);
+    }
+
+//     NCNN_LOGE("local_shader_module %p created", local_shader_module);
+
+    return create(local_shader_module, si, specializations);
+#else
     // ncnn_add_shader cmake macro
     // 0 = fp32
     // 1 = fp16p
@@ -148,6 +189,7 @@ int Pipeline::create(int shader_type_index, const Option& opt, const std::vector
     VkShaderModule shader_module = vkdev->get_shader_module(shader_type_index);
 
     return create(shader_module, si, specializations);
+#endif
 }
 
 int Pipeline::create(VkShaderModule shader_module, const ShaderInfo& _shader_info, const std::vector<vk_specialization_type>& specializations)
@@ -590,6 +632,44 @@ int ImportAndroidHardwareBufferPipeline::create(VkAndroidHardwareBufferImageAllo
 
     int shader_type_index = LayerShaderType::convert_ycbcr;
 
+#if NCNN_VULKAN_ONLINE_SPIRV
+    std::vector<uint32_t> spirv;
+    int retc = compile_spirv_module(shader_type_index, opt, spirv);
+    if (retc != 0)
+    {
+        NCNN_LOGE("compile_spirv_module failed %d", retc);
+        return -1;
+    }
+
+    const uint32_t* spv_data = spirv.data();
+    size_t spv_data_size = spirv.size() * 4;
+
+    ShaderInfo si;
+    int ret = resolve_shader_info(spv_data, spv_data_size, si);
+    if (ret != 0)
+    {
+        NCNN_LOGE("resolve_shader_info failed %d", ret);
+        return -1;
+    }
+
+    // -3 for local_size_xyz
+    int specialization_count_expected = si.specialization_count - 3;
+    if ((int)specializations.size() != specialization_count_expected)
+    {
+        NCNN_LOGE("pipeline specialization count mismatch, expect %d but got %d", specialization_count_expected, (int)specializations.size());
+        return -1;
+    }
+
+    VkShaderModule shader_module;
+    if (vkdev->info.bug_local_size_spec_const)
+    {
+        shader_module = vkdev->compile_shader_module(spv_data, spv_data_size, local_size_x, local_size_y, local_size_z);
+    }
+    else
+    {
+        shader_module = vkdev->compile_shader_module(spv_data, spv_data_size);
+    }
+#else
     // ncnn_add_shader cmake macro
     // 0 = fp32
     // 1 = fp16p
@@ -640,7 +720,7 @@ int ImportAndroidHardwareBufferPipeline::create(VkAndroidHardwareBufferImageAllo
     }
 
     VkShaderModule shader_module = vkdev->get_shader_module(shader_type_index);
-
+#endif
     create_pipeline(shader_module, specializations);
 
     if (vkdev->info.support_VK_KHR_descriptor_update_template)

src/platform.h.in

@@ -23,6 +23,7 @@
 #cmakedefine01 NCNN_PIXEL
 #cmakedefine01 NCNN_PIXEL_ROTATE
 #cmakedefine01 NCNN_VULKAN
+#cmakedefine01 NCNN_VULKAN_ONLINE_SPIRV
 #cmakedefine01 NCNN_REQUANT
 #cmakedefine01 NCNN_AVX2