Skip to content

windows.md

Latest commit

 

History

History
89 lines (66 loc) · 3.82 KB

windows.md

File metadata and controls

89 lines (66 loc) · 3.82 KB
Raw

Installing on Windows

Use PowerShell as an administrator:

powershell -ExecutionPolicy Bypass -File scripts\setup\windows.ps1

The script uses winget (falling back to Chocolatey) to install:

  • Visual Studio 2022 Build Tools (with the Desktop C++ workload)
  • meson, ninja, nasm, python (3.11+ for core use; 3.14.5+ required for ai/), llvm
  • Optional: CUDA 12.6, Intel oneAPI Base Toolkit

Environment

After install, open a x64 Native Tools Command Prompt for VS 2022 so that cl.exe and the Windows SDK are on PATH. From there:

cd core
meson setup ..\build --buildtype=release
ninja -C ..\build

CUDA

Download the installer from the NVIDIA CUDA Toolkit page. Re-run meson with -Denable_cuda=true after install.

oneAPI / SYCL

Install the Intel oneAPI Base Toolkit, then initialize in the shell via:

"C:\Program Files (x86)\Intel\oneAPI\setvars.bat"

Re-run meson with -Denable_sycl=true.

Intel QSV (optional, for h264_qsv / hevc_qsv / av1_qsv)

The fork's three QSV codec adapters in tools/vmaf-tune/ require an FFmpeg built with the Intel oneVPL dispatcher (libvpl) — Intel archived Media SDK / libmfx in May 2023.

On Windows, the oneVPL runtime is bundled with Intel's standard graphics drivers — install the latest driver for your iGPU / Arc card from the Intel Driver & Support Assistant or grab the Arc & Iris Xe Graphics driver package directly. No additional SDK install is needed for runtime use.

To build FFmpeg yourself with QSV support, install the oneVPL development headers and pass --enable-libvpl (FFmpeg n6.0+ — the "oneVPL support for QSV" line in the FFmpeg 6.0 changelog) or the legacy --enable-libmfx for older FFmpeg builds. Pre-built FFmpeg binaries from BtbN's gyan.dev / FFmpeg Builds ship with --enable-libvpl enabled.

Intel Quick Sync hardware capability matrix

Verified 2026-05-08 against Wikipedia: Intel Quick Sync Video — Hardware decoding and encoding.

CPU / GPU generation H.264 enc/dec HEVC 8-bit enc/dec HEVC 10-bit enc/dec AV1 decode AV1 encode
Skylake / Kaby Lake / Coffee Lake (Gen 9) yes yes decode only no no
Ice Lake (Gen 11) yes yes yes no no
Tiger Lake / Alder Lake / Raptor Lake (Xe LP) yes yes yes yes no
Arc Alchemist (Xe HPG, A-series, 2022) yes yes yes yes yes
Arc Battlemage (Xe2, B-series) yes yes yes yes yes

av1_qsv therefore requires Arc Alchemist or newer; hevc_qsv 10-bit requires Ice Lake or newer.

Notes

  • Windows CI is Linux's canary for MSVC quirks (forbidden VLAs, different 64-bit typedefs, narrowing-conversion strictness). If your change builds on Linux but fails on Windows, look for one of those first.