x4r (xess (xmx) 4 radeon)

want to run dlss on radeon? check out d4r!

Experimental Intel XeSS XMX kernel translation for AMD GPUs under Wine and vkd3d-proton. x4r translates Intel CM/GenX LLVM IR into wave-based HLSL, with an int8 cooperative-matrix path for RDNA3. An igdext shim supplies the translated kernels to the D3D12 application.

This is literally just another PoC, it's not really practical (yet atleast).

currently doesnt work as intended, wip. if you do compile expect ghosting and iffy game support

Status

In Ghost of Tsushima through OptiScaler (856×480 input, 2560×1440 output, XeSS network model KPSS) XeSS takes about 2.05 ms of GPU time per frame on the RX 7700 XT, measured with GPU timestamps around every XeSS dispatch (stock dot4 kernels: about 4 ms). With steady GPU clocks the offline harness gives 2.13 ms at that setup and 2.22 ms at 1280×720 → 2560×1440; the kernels' speed does not depend on image content. Performance measurements exclude rendering, presentation and pipeline creation.

Quality: on a static 3D fixture the WMMA, RDNA4-set and dot4 outputs match byte-for-byte. On moving-camera fixtures (x4r/tools/quality-motion.py, with true motion vectors and a moving object) they match the stock kernels to about 60 dB and score 28.5–31 dB sRGB against a supersampled reference, within about 1 dB of XeSS's own DP4a path; the moving object is where XeSS's XMX network is weakest, and a mirror-image test (quality-mirror.py) shows no direction-dependent error in the translation. Severe in-game ghosting has been reported with both the WMMA and stock DXIL profiles, including through native XeSS and FSR inputs. The moving-object fixture also shows stronger trails than DP4a; the cause is still unresolved. The XMX path has not been shown to look better than DP4a: that needs real Intel XMX output as a reference.

XeSS selects its Intel extension/CM kernel path. x4r translates those kernels to AMD instructions; matrix work runs on RDNA3 WMMA units. This is the XMX-path network running through translation, with no Intel XMX hardware present.

Install (release zip)

Drag and drop, OptiScaler style:

1. Extract the release zip into the folder holding the game's main .exe (it brings OptiScaler as dxgi.dll, XeSS 2.0.2.68 as libxess.dll and a x4r/ folder). Use GE-Proton 11; no launch options.
2. Start the game and pick FSR 3.1 (or DLSS) in its graphics settings. OptiScaler replaces it with XeSS. The first launch sets x4r up by itself (shim into the Proton prefix, Intel vendor id in the game's dxvk.conf).
3. After playing a few seconds, run bash x4r/x4r-check.sh. It prints PASS only if the translated XMX kernels were dispatched, how many run on WMMA, and the XeSS GPU time.

The kernels are translated from XeSS 2.0.2.68. x4r does not fall back quietly: if a kernel cannot be built, XeSS fails and the check prints FAIL. RDNA3 (gfx11) and RDNA4 (gfx12) each get their own WMMA kernel set, picked by GPU; other AMD GPUs get the stock dot4 kernels. The RDNA4 set is untested on RDNA4 hardware (it matches the RDNA3 output byte-for-byte on an RX 7700 XT). Details: packaging/README.txt.

Build the tools

The development environment uses Linux, LLVM 23.1.1, a C++20-capable compiler (the translator builds as C++17; the oracle uses C++20 barriers), Vulkan headers/loader, Python 3 with NumPy and Pillow, and systemd user scopes. The Windows quality harness additionally requires MinGW-w64. Shader builds use the Linux DXC distribution; tested shaders are compiled at -O0.

./x4r/tools/setup-workspace.sh
./x4r/tools/build-tools.sh

The source repository does not contain Intel SDK binaries, captured kernels, generated DXIL/SPIR-V, a Wine prefix, or prebuilt third-party tools. See development setup for the expected local dependencies and third-party notices for their separate licences.

On a prepared runtime with baseline DXIL, build the selected fast shader profiles:

./x4r/tools/build-fast.sh

License

x4r's own code is licensed under GPL-3.0-only; see LICENSE. Third-party components retain their original licences. This licence does not grant rights to Intel's SDK, captured model/kernel material, or generated shader artifacts. Review the applicable SDK terms before distributing derived kernels or binaries; no such material is included in this source tree.