mirkobrombin/raven

Installation

This minimal build runs console programs and simple windows. Installers and most desktop programs also need the common controls, the shell dialogs and the installer services, which Raven builds from pinned ReactOS sources, for both 64-bit and 32-bit programs. Without them a program stops at its first call into them, such as InitCommonControls:

DXVK and vkd3d-proton are only needed for Direct3D programs and games.

On an X11 session, --desktop=x11 takes the place of --desktop=wayland. The X11 desktop is experimental; -DRAVENDESKTOPWAYLAND=OFF builds Raven without Wayland at all.

RAVENPREFIX is the Windows installation the program sees, with its own C: drive and registry; give each application its own. Raven takes a snapshot of the prefix before every installer, and build/native/raven --restore-prefix brings it back.

git clone https://github.com/mirkobrombin/raven.git
cd raven
cmake -S . -B build -DCMAKE_BUILD_TYPE=RelWithDebInfo
cmake --build build -j2

Requirements

The tested setup also uses libpng 1.6.59, a CPU build of Skia with the Liberation fonts, DXVK and vkd3d-proton; the scripts/provision-.py tools download the pinned versions.

• Linux on x8664
• MinGW-w64 for x8664 and i686, to build the Windows test programs
• For desktop applications: Wayland, xkbcommon, pixman, FreeType and Vulkan
• Optionally, for the experimental X11 desktop: libxcb with its xkb, xfixes and xinput extensions, and xkbcommon-x11

What works today

Raven's own system libraries stay in charge even when an installer writes its own kernel32 or user32, while the redistributables an installer brings are the ones that load.

• 32 and 64-bit programs — Both, side by side in the same prefix
• Installers — NSIS, Inno Setup and Windows Installer (MSI) packages, with custom actions, rollback and elevation
• Visual C++ 2005 to 2022 — x86 and x64, alone or all together, reinstalled, older over newer
• .NET Framework 4.8 — Installs from Microsoft's offline installer; console, WinForms and WPF programs run
• .NET Framework 3.5 — Enabled from its package as Windows does; CLR 2.0 programs run
• DirectX June 2010, XNA 4.0, OpenAL — Install from their official redistributables
• Direct3D 8 and 9 — Through DXVK, including 32-bit programs
• Direct3D 11 — Through DXVK: every shader stage, compute, MSAA, deferred contexts
• Direct3D 12 — Through vkd3d-proton, including ray tracing
• Side-by-side assemblies — Manifests, WinSxS and activation contexts

Tested applications

Each application is installed with its own installer into a fresh prefix and then started. Every stage that works adds 25%: install, start, main window, and a real workflow where one is automated.

KeePass needs the .NET Framework 4.8, which a fresh prefix does not have yet. The Notepad++ workflow saves a file, closes and reopens it, appends to it, uses the clipboard and switches theme. scripts/benchmark-app-matrix.py runs the whole table again.

• Notepad++ 8.7.1 — yes — yes — yes — editing — 100%
• 7-Zip 24.09 (EXE installer) — yes — yes — yes — not automated yet — 75%
• 7-Zip 24.09 (MSI) — yes — yes — yes — not automated yet — 75%
• Audacity 3.6.4 — yes — yes — yes — not automated yet — 75%
• KeePass 2.57 — yes — yes — no — not automated yet — 50%
• VLC 3.0.21 — yes — no — no — not automated yet — 25%
• ClassiCube 1.3.8 — portable — yes — yes — singleplayer — 75%
• Minesweeper — portable — yes — yes — gameplay — 75%

Benchmarks

These replace the earlier ClassiCube series; its scene configuration was not verified against the retained samples. Raven uses its default execution mode.

The workloads: a program that exits at once, a SQLite query over eight million rows, a Direct3D 11 triangle checked pixel by pixel, 2,000 Direct3D 11 frames at 256 by 256, the time to the main window of Notepad++ and of the ClassiCube launcher, and ClassiCube 1.3.8 itself (Direct3D 11 build, world seed 1, 640 by 480, view distance 128, frames counted from the fifth to the fifteenth second after the world loads).

Every runtime starts from a copy of a prefix it initialised beforehand. Each workload runs three times and the charts show the median; ClassiCube runs four rounds alternating the runtimes, so a change in machine load weighs on both. The Soda Direct3D 11 triangle did not finish within its 600-second limit. Software Vulkan keeps the numbers reproducible anywhere; a hardware GPU raises them for every runtime.

The earlier chart data, with CPU time and memory, remains available; its ClassiCube entry is superseded by the table above. The scripts scripts/benchmark-wine.py, benchmark-apps.py, benchmark-frames.py and benchmark-classicube.py run these workloads. The ClassiCube runner groups trials by runtime; use separate one-trial runs to alternate runtimes.

• Raven — 71.5 — 72, 73, 54, 71
• Wine 11.0 — 63.0 — 65, 68, 61, 61
• Soda 11.0-32 exp. — 63.0 — 64, 64, 61, 62
• CPU — Intel Core i7-13700H (14 cores, 20 threads, up to 5.0 GHz); 8 logical CPUs available to the benchmark environment
• Memory — 31 GiB
• Kernel — Linux 7.2.11
• Userland — Ubuntu 26.04.1 LTS
• Graphics — Mesa 26.0.8 lavapipe (software Vulkan), two rendering threads
• Display — labwc 0.9.3, headless, one isolated instance per run