feat: add renderdoc integration

.github/workflows/renderdoc.yml

@@ -0,0 +1,195 @@
+name: Test NCNN RenderDoc Integration
+
+on:
+  push:
+    branches: [ master ]
+  workflow_dispatch:
+
+jobs:
+  test-renderdoc-linux:
+    runs-on: ubuntu-latest
+    strategy:
+      matrix:
+        compiler: [gcc, clang]
+        build_type: [Release, Debug]
+
+    steps:
+    - name: Checkout
+      uses: actions/checkout@v4
+      with:
+        submodules: recursive
+
+    - name: Install dependencies
+      run: |
+        sudo apt-get update
+        sudo apt-get install -y \
+          cmake ninja-build \
+          libvulkan-dev vulkan-utils vulkan-validationlayers \
+          renderdoc \
+          libglfw3-dev libglm-dev \
+          libprotobuf-dev protobuf-compiler \
+          libopencv-dev
+
+    - name: Setup compiler
+      if: matrix.compiler == 'clang'
+      run: |
+        sudo apt-get install -y clang
+        echo "CC=clang" >> $GITHUB_ENV
+        echo "CXX=clang++" >> $GITHUB_ENV
+
+    - name: Configure CMake
+      run: |
+        cmake -B build -G Ninja \
+          -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} \
+          -DNCNN_VULKAN=ON \
+          -DNCNN_SIMPLEVK=OFF \
+          -DNCNN_ENABLE_RENDERDOC_PROFILING=ON \
+          -DNCNN_BUILD_BENCHMARK=ON
+
+    - name: Build
+      run: cmake --build build --parallel
+
+    - name: Build benchncnn
+      run: cmake --build build --target benchncnn --parallel
+
+    - name: Copy benchmark param files
+      run: |
+        echo "Copying benchmark param files..."
+        mkdir -p build/benchmark
+        cp benchmark/*.param build/benchmark/ || echo "No param files found in benchmark directory"
+        ls -la build/benchmark/*.param || echo "No param files copied"
+
+    - name: Test RenderDoc capture with benchncnn
+      run: |
+        export NCNN_RENDERDOC_CAPTURE_PATH="./benchncnn_capture"
+
+        echo "Running benchncnn with RenderDoc headless capture..."
+        renderdoccmd capture ./build/benchmark/benchncnn 10 4 0 0 1
+
+        if [ -f "benchncnn_capture.rdc" ]; then
+          echo "RenderDoc capture file created successfully"
+          ls -la benchncnn_capture.rdc
+        else
+          echo "RenderDoc capture file not found"
+          exit 1
+        fi
+
+    - name: Upload RenderDoc capture
+      uses: actions/upload-artifact@v4
+      if: always()
+      with:
+        name: renderdoc-capture-linux-${{ matrix.compiler }}-${{ matrix.build_type }}
+        path: |
+          benchncnn_capture.rdc
+
+  test-renderdoc-windows:
+    runs-on: windows-latest
+    strategy:
+      matrix:
+        arch: [x64, x86]
+        build_type: [Release, Debug]
+
+    steps:
+    - name: Checkout
+      uses: actions/checkout@v4
+      with:
+        submodules: recursive
+
+    - name: Setup MSVC
+      uses: microsoft/[email protected]
+
+    - name: Install Vulkan SDK
+      run: |
+        $url = "https://sdk.lunarg.com/sdk/download/1.3.268.0/windows/VulkanSDK-1.3.268.0-Installer.exe"
+        $installer = "$env:TEMP\VulkanSDK-Installer.exe"
+        Invoke-WebRequest -Uri $url -OutFile $installer
+        Start-Process -FilePath $installer -ArgumentList "/S" -Wait
+
+    - name: Install RenderDoc
+      run: |
+        $url = "https://renderdoc.org/stable/1.40/RenderDoc_1.40_64.msi"
+        $installer = "$env:TEMP\RenderDoc-Installer.msi"
+        Invoke-WebRequest -Uri $url -OutFile $installer
+        Start-Process -FilePath $installer -ArgumentList "/quiet" -Wait
+
+    - name: Configure CMake
+      run: |
+        cmake -B build -A ${{ matrix.arch }} `
+          -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} `
+          -DNCNN_VULKAN=ON `
+          -DNCNN_SIMPLEVK=OFF `
+          -DNCNN_ENABLE_RENDERDOC_PROFILING=ON `
+          -DNCNN_BUILD_BENCHMARK=ON
+
+    - name: Build
+      run: cmake --build build --config ${{ matrix.build_type }} --parallel
+
+    - name: Build benchncnn
+      run: cmake --build build --config ${{ matrix.build_type }} --target benchncnn --parallel
+
+    - name: Copy benchmark param files (Windows)
+      run: |
+        Write-Host "Copying benchmark param files..."
+        New-Item -ItemType Directory -Force -Path "build\benchmark"
+        if (Test-Path "benchmark\*.param") {
+          Copy-Item "benchmark\*.param" "build\benchmark\" -Force
+          Get-ChildItem "build\benchmark\*.param"
+        } else {
+          Write-Host "No param files found in benchmark directory"
+        }
+
+    - name: Test RenderDoc capture with benchncnn
+      run: |
+        $env:PATH += ";C:\VulkanSDK\1.3.268.0\Bin"
+        $env:PATH += ";C:\Program Files\RenderDoc"
+        $env:NCNN_RENDERDOC_CAPTURE_PATH = "benchncnn_capture.rdc"
+
+        Write-Host "Running benchncnn with RenderDoc capture..."
+        & "C:\Program Files\RenderDoc\renderdoccmd.exe" capture .\build\benchmark\${{ matrix.build_type }}\benchncnn.exe 10 4 0 0 1
+
+        if (Test-Path "benchncnn_capture.rdc") {
+          Write-Host "RenderDoc capture file created successfully"
+          Get-ChildItem benchncnn_capture.rdc
+        } else {
+          Write-Host "RenderDoc capture file not found"
+          exit 1
+        }
+
+    - name: Upload Windows RenderDoc capture
+      uses: actions/upload-artifact@v4
+      if: always()
+      with:
+        name: renderdoc-capture-windows-${{ matrix.arch }}-${{ matrix.build_type }}
+        path: |
+          benchncnn_capture.rdc
+
+
+  integration-test:
+    runs-on: ubuntu-latest
+    needs: [test-renderdoc-linux, test-renderdoc-windows]
+
+    steps:
+    - name: Download all artifacts
+      uses: actions/download-artifact@v4
+      with:
+        pattern: renderdoc-capture-*
+        merge-multiple: true
+
+    - name: Summary with capture analysis
+      run: |
+        echo "## NCNN RenderDoc Cross-Platform Test Results" >> $GITHUB_STEP_SUMMARY
+        echo "Linux benchncnn RenderDoc capture completed" >> $GITHUB_STEP_SUMMARY
+        echo "Windows benchncnn RenderDoc capture completed" >> $GITHUB_STEP_SUMMARY  
+        echo "" >> $GITHUB_STEP_SUMMARY
+
+        echo "### Generated RenderDoc Captures:" >> $GITHUB_STEP_SUMMARY
+        for file in *.rdc; do
+          if [ -f "$file" ]; then
+            size=$(stat -c%s "$file" 2>/dev/null || echo "unknown")
+            echo "- $file (${size} bytes)" >> $GITHUB_STEP_SUMMARY
+          fi
+        done
+
+        echo "" >> $GITHUB_STEP_SUMMARY
+        echo "RenderDoc integration with benchncnn successfully tested!" >> $GITHUB_STEP_SUMMARY
+        echo "Capture files are available in the artifacts for analysis." >> $GITHUB_STEP_SUMMARY

CMakeLists.txt

@@ -89,6 +89,7 @@ option(NCNN_PYTHON "build python api" OFF)
 option(NCNN_INT8 "int8 inference" ON)
 option(NCNN_BF16 "bf16 inference" ON)
 option(NCNN_FORCE_INLINE "force inline some function" ON)
+option(NCNN_ENABLE_RENDERDOC_PROFILING "Enables profiling with the RenderDoc tool." OFF)
 
 if(ANDROID OR IOS OR NCNN_SIMPLESTL)
     option(NCNN_DISABLE_RTTI "disable rtti" ON)
@@ -798,6 +799,9 @@ if(CMAKE_SYSTEM_NAME STREQUAL "Emscripten")
 endif()
 
 if(NCNN_VULKAN)
+    if(NCNN_ENABLE_RENDERDOC_PROFILING)
+        add_definitions(-DNCNN_ENABLE_RENDERDOC_PROFILING=1)
+    endif()
     if(NCNN_SYSTEM_GLSLANG)
         find_package(Threads)
         find_package(SPIRV-Tools QUIET)

benchmark/benchncnn.cpp

@@ -348,6 +348,9 @@ int main(int argc, char** argv)
 
         g_blob_vkallocator = new ncnn::VkBlobAllocator(g_vkdev);
         g_staging_vkallocator = new ncnn::VkStagingAllocator(g_vkdev);
+#if NCNN_ENABLE_RENDERDOC_PROFILING
+        ncnn::start_renderdoc_capture();
+#endif // NCNN_ENABLE_RENDERDOC_PROFILING
     }
 #endif // NCNN_VULKAN
 
@@ -467,6 +470,9 @@ int main(int argc, char** argv)
 #if NCNN_VULKAN
     delete g_blob_vkallocator;
     delete g_staging_vkallocator;
+#if NCNN_ENABLE_RENDERDOC_PROFILING
+    ncnn::end_renderdoc_capture();
+#endif // NCNN_ENABLE_RENDERDOC_PROFILING
 #endif // NCNN_VULKAN
 
     return 0;

docs/developer-guide/glsl-extension.md

@@ -344,6 +344,47 @@ void main()
 
 At runtime, `NCNN_LOGE` will print out the value of `gx`
 
+#### RenderDoc Integration
+
+When validation layer is enabled, ncnn also supports RenderDoc integration for GPU debugging and profiling. To enable RenderDoc support:
+
+* Build with RenderDoc support:
+   ```bash
+   cmake -DNCNN_VULKAN=ON -DNCNN_SIMPLEVK=OFF -DNCNN_ENABLE_RENDERDOC_PROFILING=ON ..
+   ```
+
+* Add capture calls in your program:
+   ```cpp
+   int main() {
+       // Start RenderDoc capture
+       ncnn::start_renderdoc_capture();
+
+       // ... inference operations ...
+
+       // End RenderDoc capture
+       ncnn::end_renderdoc_capture();
+       return 0;
+   }
+   ```
+
+* Specify capture file path (optional):
+   ```bash
+   export NCNN_RENDERDOC_CAPTURE_PATH="/path/to/capture_file"
+   ```
+
+* Capture GPU operations:
+   ```bash
+   renderdoccmd capture your_ncnn_application
+   ```
+
+When RenderDoc integration is enabled, ncnn will:
+- Detect RenderDoc library at runtime
+- Begin capture when GPU operations start
+- End capture when GPU instance is destroyed
+- Save capture files for analysis in RenderDoc
+
+The generated `.rdc` files can be opened in RenderDoc GUI to analyze Vulkan API calls, shader execution, memory usage, and GPU performance metrics for ncnn inference operations.
+
 ### option macros
 
 enable glsl extension only if user enable some options

docs/developer-guide/glsl-extension.zh.md

@@ -345,6 +345,46 @@ void main()
 
 在运行时，`NCNN_LOGE` 将打印出 `gx` 的值
 
+#### RenderDoc 集成
+
+当启用验证层时，ncnn 还支持 RenderDoc 集成，用于 GPU 调试和性能分析。启用 RenderDoc 支持需要：
+
+* 构建时启用 RenderDoc 支持：
+   ```bash
+   cmake -DNCNN_VULKAN=ON -DNCNN_SIMPLEVK=OFF -DNCNN_ENABLE_RENDERDOC_PROFILING=ON ..
+   ```
+
+* 在程序中添加捕获调用：
+   ```cpp
+   int main() {
+       // 开始RenderDoc捕获
+       ncnn::start_renderdoc_capture();
+
+       // ... 推理操作 ...
+
+       // 结束RenderDoc捕获
+       ncnn::end_renderdoc_capture();
+       return 0;
+   }
+   ```
+* 指定捕获文件路径（可选）：
+   ```bash
+   export NCNN_RENDERDOC_CAPTURE_PATH="/path/to/capture_file"
+   ```
+
+* 捕获 GPU 操作：
+   ```bash
+   renderdoccmd capture your_ncnn_application
+   ```
+
+当启用 RenderDoc 集成时，ncnn 将会：
+- 在运行时自动检测 RenderDoc 库
+- 在 GPU 操作开始时开始捕获
+- 在 GPU 实例销毁时结束捕获
+- 保存捕获文件供 RenderDoc 分析
+
+生成的 `.rdc` 文件可以在 RenderDoc GUI 中打开，用于分析 ncnn 推理操作的 Vulkan API 调用、着色器执行、内存使用情况和 GPU 性能指标。
+
 ### 选项宏
 
 仅当用户启用某些选项时才启用 GLSL 扩展