Skip to content

build.md

Latest commit

 

History

History
240 lines (165 loc) · 8.11 KB

build.md

File metadata and controls

240 lines (165 loc) · 8.11 KB
Raw

Build Guide

This guide explains the current build system in plain terms first, then adds the technical details needed for local development.

Which Builder Should I Use?

Tool Best for What it does
BUILDER.bat Most people building from source. Opens an interactive choice for DML or CUDA, prepares dependencies, configures CMake, and builds.
BUILDER.ps1 Same as above, but easier to script. Lets you pass -Backend DML or -Backend CUDA directly.
build_no-options.bat Fast local rebuilds after the full build already worked once. Rebuilds ai from an existing build tree. No downloads, updates, NuGet restore, OpenCV setup, or dependency prompts.
tools/build_dml.ps1 Direct DML backend build. Restores DML/ONNX dependencies and prepares the DML build tree.
tools/build_cuda.ps1 Direct CUDA/TensorRT backend build. Resolves CUDA/TensorRT/OpenCV CUDA dependencies and prepares the CUDA build tree.

If you are unsure, run BUILDER.bat.

Requirements

Required for Both Builds

  • Windows 10 or Windows 11.
  • Visual Studio 2022 or Build Tools with the Desktop development with C++ workload.
  • Windows SDK with C++/WinRT headers.
  • CMake.
  • PowerShell.

The builder can download or restore some project dependencies when you allow it. It cannot install Visual Studio for you.

DML Build Requirements

  • A Windows GPU/driver that supports DirectML.
  • NuGet packages restored by the build scripts:
    • Microsoft.ML.OnnxRuntime.DirectML
    • Microsoft.AI.DirectML
  • OpenCV DML layout under:
sunone_aimbot_2\modules\opencv\build\dml

The DML builder can prepare the OpenCV layout when downloads are enabled.

CUDA Build Requirements

  • NVIDIA GPU.
  • CUDA Toolkit.
  • TensorRT 10 Windows binary SDK.
  • OpenCV built with CUDA under:
sunone_aimbot_2\modules\opencv\build\cuda\install

The CUDA builder can guide dependency downloads and build OpenCV CUDA when downloads are enabled. OpenCV CUDA builds are large and can take a while.

Full Interactive Build

From the repository root:

.\BUILDER.bat

Choose:

  • DML for the DirectML build.
  • CUDA for the CUDA + TensorRT build.

The batch file keeps the console open after a double-click run so you can read errors.

Full Scripted Build

DML:

powershell -NoProfile -ExecutionPolicy Bypass -File .\BUILDER.ps1 -Backend DML

CUDA:

powershell -NoProfile -ExecutionPolicy Bypass -File .\BUILDER.ps1 -Backend CUDA

Useful options:

powershell -NoProfile -ExecutionPolicy Bypass -File .\BUILDER.ps1 -Backend CUDA -OpenCvAlreadyBuilt true
powershell -NoProfile -ExecutionPolicy Bypass -File .\BUILDER.ps1 -Backend CUDA -DownloadOrUpdateNeeded false
powershell -NoProfile -ExecutionPolicy Bypass -File .\BUILDER.ps1 -Backend DML -DryRun

BUILDER.ps1 forwards backend-specific arguments to tools/build_dml.ps1 or tools/build_cuda.ps1.

No-Options Rebuild

Use this after the full builder has already prepared the selected backend at least once:

.\build_no-options.bat

or:

powershell -NoProfile -ExecutionPolicy Bypass -File .\build_no-options.ps1 -Backend DML
powershell -NoProfile -ExecutionPolicy Bypass -File .\build_no-options.ps1 -Backend CUDA

This script intentionally does only this:

cmake --build build\<backend> --config Release --target ai --parallel

It is useful when you changed C++ or UI code and want a quick rebuild. It is not a dependency setup tool. If the build tree is missing, stale, or configured for the wrong dependency paths, run BUILDER.bat again.

Direct Backend Scripts

DML:

powershell -NoProfile -ExecutionPolicy Bypass -File .\tools\build_dml.ps1

CUDA:

powershell -NoProfile -ExecutionPolicy Bypass -File .\tools\build_cuda.ps1

Examples:

powershell -NoProfile -ExecutionPolicy Bypass -File .\tools\build_cuda.ps1 -CudaArchBin 8.6
powershell -NoProfile -ExecutionPolicy Bypass -File .\tools\build_cuda.ps1 -CudaArchBin all
powershell -NoProfile -ExecutionPolicy Bypass -File .\tools\build_cuda.ps1 -SkipOpenCvBuild
powershell -NoProfile -ExecutionPolicy Bypass -File .\tools\build_dml.ps1 -UseLatestPackages

-CudaArchBin all expands to:

7.5;8.0;8.6;8.7;8.8;8.9;9.0;10.0;10.3;11.0;12.0;12.1

Output Folders

Default outputs:

build\dml\Release\ai.exe
build\cuda\Release\ai.exe

The output folder also receives runtime DLLs and selected assets that the executable needs at runtime.

Runtime Files Copied by CMake

When AIMBOT_COPY_RUNTIME_DLLS is enabled, CMake copies available runtime DLLs next to ai.exe, including:

  • OpenCV runtime DLL.
  • ghub_mouse.dll, when present.
  • chroma_lighting.dll, when present.
  • ONNX Runtime DLLs.
  • DirectML DLL for DML support.
  • TensorRT and CUDA provider DLLs for CUDA builds.
  • cuDNN DLL if a compatible cuDNN layout is found.

The Razer control DLL is expected at:

sunone_aimbot_2\modules\razer-controls\x64\Release\chroma_lighting.dll

You can override its CMake cache path with AIMBOT_RAZER_CONTROL_DLL.

Important CMake Options

Option Meaning
AIMBOT_USE_CUDA ON for CUDA + TensorRT, OFF for DML.
AIMBOT_COPY_RUNTIME_DLLS Copies runtime DLLs beside ai.exe.
AIMBOT_RAZER_CONTROL_DLL Source path for chroma_lighting.dll.
AIMBOT_OPENCV_DML_ROOT DML OpenCV layout root.
AIMBOT_OPENCV_CUDA_ROOT CUDA OpenCV layout root.
AIMBOT_TENSORRT_ROOT TensorRT SDK root.
AIMBOT_CUDNN_ROOT Optional cuDNN root.

CUDA Build Notes

The CUDA build uses TensorRT for the main TRT backend and still links ONNX Runtime/DirectML libraries because DML-related runtime paths and tooling remain available in the shared project.

For best performance:

  • Use a TensorRT .engine model with the CUDA build.
  • Use Circle FOV for normal circular aim limiting.
  • Turn off preview/debug windows when measuring capture performance, because preview requires CPU-readable pixels.

The current CUDA preprocess path can intentionally run CPU preprocessing and then copy the tensor to CUDA input. That is useful for diagnosis and compatibility, but pure GPU paths are faster when available.

DML Build Notes

The DML build expects ONNX models. A TensorRT .engine model will not run through the DML backend.

If DML launches but does not detect targets:

  • Check that the selected AI model is .onnx.
  • Lower confidence_threshold temporarily.
  • Confirm the configured class IDs match the model.
  • Confirm backend = DML.
  • Check dml_device_id if the system has more than one GPU.

Validation

Run the regression checks after build-system or integration changes:

powershell -NoProfile -ExecutionPolicy Bypass -File .\tools\regression_checks.ps1

This checks important source contracts such as:

  • Razer and Teensy runtime wiring.
  • no-options builder behavior.
  • Circle FOV behavior.
  • frame-age latency compensation while preserving detection confidences.
  • config and UI integration contracts.

Troubleshooting

Problem What to check
Build tree not found from build_no-options Run BUILDER.bat first for that backend.
CUDA dependency missing Install CUDA Toolkit and TensorRT Windows binary SDK, then rerun the CUDA builder.
OpenCV CUDA compile errors Check CUDA, Visual Studio compiler, OpenCV, and contrib compatibility. If you only need CPU preprocess for diagnosis, keep legacy GPU-heavy capture features off and rebuild the main app after the dependency tree is healthy.
DML build cannot find OpenCV Run the DML builder with downloads enabled or set OpenCV CMake paths manually.
Missing chroma_lighting.dll Build or copy the Razer controls DLL to sunone_aimbot_2\modules\razer-controls\x64\Release\chroma_lighting.dll or next to ai.exe.
Training bootstrap fails Install training requirements or place the expected .onnx files in sunone_aimbot_2\modules\training\models.
App runs but Razer or Teensy control does nothing The selected control method does not fall back. Check the selected input_method, the Razer DLL, or the Teensy RawHID endpoint.