Xbox2Windows researches static recompilation from Xbox Series X|S (Scarlett) to x64 Windows. It translates selected x64 machine code into C++ and selected gfx1013 shaders into desktop DXIL. The public tests build their own inputs and run without a game installation.
Startup status. Source-built Windows programs run from entry to process exit. The main executable was absent from the inspected package. Execution of the title's CPU code, full-game startup and frame rate remain unverified.
Results · Coverage chart · Community and contributions · Build on Windows · Choose an experiment · Run the GPU check
| What runs | What was compared |
|---|---|
| x64 programs from PE entry | A two-record archive reader runs with blast 1.3 decompression under Windows, the PE mapper and the recompiled build. The three agree on output bytes, callback order, completed records, allocations, releases and exit status, including truncation and a refused output window. Reproduce the archive comparison |
| Pixel shaders | At c849f7a, 851 of 12,101 PS records compiled, built D3D12 pipelines and matched controlled output on RTX 4060 Laptop and WARP. A two-mip draw verifies explicit LOD 0. A separate probe follows a constant-buffer read shared by 529 blocked records. Swapping buffers changes all 1,024 probe pixels. Color readback · Depth and bindings · Read probe and latest counts |
| Compute shaders | Of 2,689 package records, 12 ran under the tested UAV bindings. Each adapter passed 14 full-buffer comparisons, including dispatches with different group counts. Dispatch and readback |
| Mesh shaders | A specialized entry fetched indexed vertices from a GPU buffer. Two position sets covered 1,024 and 276 pixels on a 32×32 target, matching the CPU reference on hardware and WARP. Mesh draw and specialization limits |
| Windows runtime | Bink frame-buffer allocations reach the Windows heap through the shared provider. The FidelityFX experiment follows scratch writes from two passes through the RCAS output on an 8×4 tile. Allocation, failure and retry · Upscaler pixel comparison |
The chart counts instruction forms exercised by passing source-built fixtures. The experiments above measure program outputs and package shader records.
Snapshot at a46dac3 · Latest successful Windows runs on main · Find forms without a passing fixture
To see a newer measurement, download coverage-dashboard from the run's Artifacts section.
- Claim a task. Search for the behavior or test name in open issues and open PRs. On the issue, say what you will test, on which platform, and which shared interface or other PR your work depends on. For a first patch, start with good first issues.
- Submit a change. Keep the PR focused on one behavior. Link the task it advances and any prerequisite PR. Include the exact command, expected and observed results, and a negative control for a new fixture. Patch and evidence guide
- Review a PR. For CPU translation or PE loading, reproduce the comparison with the original fixture. For shaders or GPU compatibility, check the DXIL and relevant WARP or hardware results. Both reviewers cover shared CPU/GPU interfaces. Record your verdict and the commit you tested. The Maintainer merges after review threads are resolved and the required Windows check passes. Review responsibilities
- Contribute a hardware result. Run the standalone GPU check and post the report in issue #1. You can also claim a batch there without owning a GPU. Verify the original logs, add new adapters as rows in the hardware table, and link driver retests in the existing row.
New contributor guide · Build and debugging Q&A · Report a bug or propose research · Community conduct · Private security reports
On x64 Windows, install Python 3.11 or newer and Visual Studio's Desktop development with C++ workload, including x64 MSVC, CMake, Ninja, and Windows SDK 10.0.26100 with DXC. Windows CI uses Python 3.12. Open PowerShell at the repository root and run the following commands.
.\tools\dev.ps1
python -m venv out/pe-env
.\out\pe-env\Scripts\python.exe -m pip install -r requirements-dev.txt
cmake --preset dev
cmake --build --preset dev
ctest --preset dev --no-tests=error -R 'desktop_root_signature|era_heap_allocation|offscreen_warp'The smoke tests exercise a D3D12 root signature, heap allocation and software rendering. For just those checks, you can omit the two Python setup commands. With the Python environment installed, run a focused experiment or run the full suite with ctest --preset dev --no-tests=error.
To build and test at /O2 as Windows CI also does, use the release preset in place of dev. Its build directory is out/build/release.
Contributing from macOS or Linux? See the non-Windows host guide for the platform-independent Python and LLVM 15 checks.
- DXC validation.
shader_metadata_probeskips when the validator lacks module-only validation. DXC 1.8.2502 is the first verified release that passes. To use a separate DXC installation, configure withcmake --preset dev -DDXC_EXECUTABLE=C:/path/to/dxc/bin/x64/dxc.exe. Results by release · Original report · DXC download - Windows 10. The in-box WARP runtime lacks the features used by the BVH and DXR checks. Use the app-local Agility SDK and WARP packages to run those checks. Windows 10 setup
- Independent build. A Windows 11 contributor reproduced the three smoke tests from a clean checkout at
71bb9f4and again after an incremental rebuild at2465837. Toolchain, full-suite results and capability skips
After the full Windows build, run one comparison from the repository root.
| Experiment | Focused test |
|---|---|
| Native, mapped and recompiled archive reader | ctest --preset dev --no-tests=error -R '^cpu_native_archive_blast$' -V |
| Indexed mesh draw on WARP | ctest --preset dev --no-tests=error -R '^mesh_indexed_draw_warp$' -V |
| Compute shader buffer writes on WARP | ctest --preset dev --no-tests=error -R '^compute_store_aot_warp$' -V |
| Area | A contribution you can verify | Focused CTest command |
|---|---|---|
| Static CPU recompilation | Compare leaf, chained-range and cross-function translations with their source-built originals and negative controls, including Bink frame plane requests and the Bink group descriptor builders. | ctest --preset dev --no-tests=error -R '^cpu_' -V |
| Script states | Reproduce a parent/child lookup, an ambiguous selection or a damaged sibling link with a synthetic tree. | ctest --preset dev --no-tests=error -R '^script_(state_literals|path_links)$' -V |
| Resource metadata | Check a relocated resource root against its block directory and type schema. | ctest --preset dev --no-tests=error -R '^resource_relocated_root$' -V |
| Texture recovery | Check a packed mip tail, a first refill byte, the fields of a generated BC7 block, or a resident mip drawn on the software adapter. Look up a texture key across archives and inspect a matching descriptor. | ctest --preset dev --no-tests=error -R '^(bcpack_prefix_and_mip_tail|bc7_block_reference|bc7_draw_warp|resource_range_controls)$' -V |
| Audio metadata | Inventory stream counts, codec IDs and sample rates in a bounded AWC sound-effect bank. | ctest --preset dev --no-tests=error -R '^audio_awc_metadata$' -V |
| Compute shader recompilation | Translate fixed and indexed raw UAV stores, including an optional runtime constant buffer, and compare the complete buffer on D3D12. Bindings and commands | ctest --preset dev --no-tests=error -R '^compute_store_aot_' |
| Static shader recompilation | Translate pixel shader fixtures to desktop DXIL and check the output, including all 32 bits of an unsigned conversion. Select a local record by name and, for duplicate names, by decoded DXBC SHA-256. Supported instructions and bindings · Float-to-uint readback and entry inputs | ctest --preset dev -R '^pixel_aot_' -V |
| Shader effect metadata | Inventory the candidate programs for a named SGD5 effect using independently located archive ranges. | ctest --preset dev --no-tests=error -R "^sgd5_effect_candidates$" -V |
| Pipeline readiness | Join explicit pipeline references to the native shader manifest and identify blockers for PSO creation. Inputs and WARP control | ctest --preset dev -R '^pso_' -V |
| D3D12 and ray tracing | Compare the CPU reference, custom BVH shader and DXR on another adapter. Record the driver and native wave width. | ctest --preset dev -R '^bvh_(differential|wave64)_hardware$' -V |
| Upscaler memory accounting | Compare post-query counter writes on success and failure. | ctest --preset dev --no-tests=error -R 'working_set_delta$' |
| Windows runtime | Add a source-built allocation case with a measurable Windows result. | ctest --preset dev -R 'era_heap|desktop_memory_query' |
| PE tooling | Reproduce an import-resolution edge case with a DLL or EXE built from source. | ctest --preset dev --no-tests=error -R 'static_exe_intake|conditional_import_paths|bundled_export_forwarders' |
| Build and documentation | Try the first build from a clean clone. Identify the first missing prerequisite or unclear instruction. | Use the three-test command above. |
Full experiment notes and reproduction commands
Download the Windows x64 GPU check and SHA-256, verify and extract the ZIP, then run this in PowerShell from the extracted folder. The package contains its own shaders and needs no compiler, console or game files.
.\Run-GpuCheck.ps1With a package built from current source, use .\Run-GpuCheck.ps1 -Adapter Radeon to select a hardware adapter by part of its name. For source builds, set XBOX2WINDOWS_ADAPTER before running CTest.
Review gpu-report-hardware.md and paste its full contents as a comment on the hardware coverage thread (#1). Do not open a new GitHub issue. Reports with a failure or capability skip are welcome. Package instructions
Results are grouped by adapter. Open a report for the driver, build and any capability skips.
Show adapter results and original reports
| Adapter or backend | Default native lanes | Forced wave64 | BVH differential | Evidence |
|---|---|---|---|---|
| NVIDIA GeForce RTX 2060 | 32 | Skipped | 408 queries passed | Maintainer baseline |
| NVIDIA GeForce RTX 3070 (driver 32.0.16.1088) | 32 | Skipped, lane range 32–32 | 408 queries passed | RTX 3070 report · 32.0.16.2101 report · Independent 32.0.16.2101 run, source build at b860dd5 |
| NVIDIA GeForce RTX 4060 Laptop GPU | 32 | Skipped | 408 queries passed | Second maintainer workstation · Independent Windows run, driver 32.0.16.2124 · 32.0.16.1742 report |
| NVIDIA GeForce RTX 5070 | 32 on every run | Skipped on every run, lane range 32–32 | 408 queries passed per run | 32.0.16.1692, debug layer unavailable · 32.0.16.2135 retest, debug layer enabled · 32.0.16.2124 report · 32.0.16.1686 report · Independent 32.0.16.2135 run |
| AMD Radeon RX 6600 (driver 32.0.21045.16010) | 32 | Passed, 64 lanes | 408 queries passed | RX 6600 report · 32.0.11037.4004 retest · 32.0.21045.29003 report |
| AMD Radeon RX 6750 XT (driver 32.0.21045.29003) | 32 | Passed, 64 lanes | 408 queries passed | RX 6750 XT report |
| AMD Radeon RX 6800 XT (driver 32.0.21043.5001) | 32 | Passed, 64 lanes | 408 queries passed | RX 6800 XT report |
| AMD Radeon RX 9070 (driver 32.0.22042.14002) | 64 | Passed, 64 lanes | 408 queries passed | RX 9070 report · 32.0.32029.2010 report |
| AMD Radeon RX 9070 XT (driver 32.0.32015.16001) | 64 | Passed, 64 lanes | 408 queries passed | RX 9070 XT report · 32.0.32029.2010 retest, D3D12 warning 989 · 32.0.32015.16001 report · 32.0.32015.7029 report · 32.0.32029.2010 report · 32.0.32029.2010 run 3 · 32.0.32029.2010 run 4 · 32.0.32029.2010 run 5 · 32.0.32015.12012 report · 32.0.32029.2010 source build at bc24dfd |
| NVIDIA GeForce RTX 4060 Ti (driver 32.0.16.1742) | 32 | Skipped, lane range 32–32 | 408 queries passed | Hardware report · 32.0.16.2124 report |
| AMD Radeon RX 7900 GRE (driver 32.0.32015.7029) | 32 | Passed, 64 lanes | 408 queries passed | Hardware report |
| NVIDIA GeForce RTX 4070 Ti SUPER (driver 32.0.16.1088) | 32 | Skipped, lane range 32–32 | 408 queries passed | Hardware report · 32.0.15.7270 report, Windows 10 |
| NVIDIA GeForce RTX 5080 (driver 32.0.16.2124) | 32 | Skipped, lane range 32–32 | 408 queries passed | Hardware report · 32.0.16.1088 report · 32.0.16.1664 report · Independent run, same driver · 32.0.16.1692 report · 32.0.16.1714 report |
| AMD Radeon RX 9060 XT (driver 32.0.32029.2010) | 64 | Passed, 64 lanes | 408 queries passed | Hardware report · Independent run, same driver · Third run, same driver · 32.0.31036.15 report · Fourth run, same driver |
| NVIDIA GeForce RTX 4060 (driver 32.0.16.1088) | 32 | Skipped, lane range 32–32 | 408 queries passed | Hardware report · 32.0.16.2124 report · Independent 32.0.16.2124 run by FZNull |
| NVIDIA GeForce RTX 4050 Laptop GPU (driver 32.0.16.1088) | 32 | Skipped, lane range 32–32 | 408 queries passed | Hardware report |
| NVIDIA GeForce RTX 2060 SUPER (driver 32.0.16.2124) | 32 | Skipped, lane range 32–32 | 408 queries passed | Windows hardware report |
| NVIDIA GeForce RTX 3060 (driver 32.0.16.2124) | 32 | Skipped, lane range 32–32 | 408 queries passed | Windows hardware report · 32.0.16.2101 retest, device 2503 · 32.0.15.9636 GPU passthrough retest · 32.0.16.1714 report |
| Intel Arc B570 (driver 32.0.101.9175) | 32 | Skipped, lane range 16–32 | 408 queries passed | Windows hardware report |
| AMD Radeon RX 6600 LE | 32 | Skipped, Shader Model 6.5 | 408 queries passed | 32.0.21045.29003 report |
| NVIDIA GeForce RTX 3080 Ti | 32 | Skipped, lane range 32–32 | 408 queries passed | 32.0.16.1742 report · 32.0.16.1088 report |
| NVIDIA GeForce RTX 4070 SUPER | 32 | Skipped, lane range 32–32 | 408 queries passed | 32.0.16.2135 report · 32.0.16.2114 report · 32.0.16.1742 source build at ccfa703 · 32.0.16.1088 report |
| NVIDIA GeForce GTX 1080 Ti | 32 | Skipped, lane range 32–32 | Skipped, DXR tier 1.0 | 32.0.15.8266 report |
| AMD Radeon RX 6700 | 32 | Passed, 64 lanes | 408 queries passed | 32.0.21045.16010 report |
| NVIDIA GeForce RTX 5060 | 32 | Skipped, lane range 32–32 | 408 queries passed | 32.0.16.1742 report · 32.0.16.1656 report · 32.0.16.2114 report |
| NVIDIA GeForce RTX 4090 | 32 | Skipped, lane range 32–32 | 408 queries passed | 32.0.16.2135 report |
| NVIDIA GeForce RTX 3080 Ti Laptop GPU | 32 | Skipped, lane range 32–32 | 408 queries passed | 32.0.16.2135 report |
| AMD Radeon RX 6700 XT | 32 | Passed, 64 lanes | 408 queries passed | 32.0.21045.16010 report · 32.0.21045.29003 report, debug layer enabled |
| NVIDIA GeForce RTX 4090 Laptop GPU | 32 | Skipped, lane range 32–32 | 408 queries passed | 32.0.16.1656 report |
| NVIDIA GeForce RTX 5070 Ti | 32 | Skipped, lane range 32–32 | 408 queries passed | 32.0.16.2124 report · Independent 32.0.16.2124 run · 32.0.16.1664 report · 32.0.16.1742 report |
| NVIDIA GeForce RTX 5050 | 32 | Skipped, lane range 32–32 | 408 queries passed | 32.0.16.2101 report |
| AMD Radeon RX 7800 XT | 32 on both drivers | Skipped, Shader Model 6.5 on 32.0.32015.16001; passed, 64 lanes on 32.0.32029.2010 | 408 queries passed per run | 32.0.32015.16001 report · 32.0.32029.2010 report, debug layer enabled · Independent 32.0.32029.2010 run |
| NVIDIA GeForce GTX 1070 | 32 | Skipped, lane range 32–32 | Skipped, DXR tier 1.0 | 32.0.15.8180 report |
| AMD Radeon RX 6650 XT | 32 | Passed, 64 lanes | 408 queries passed | 32.0.21016.1010 report · 32.0.21045.16010 report |
| NVIDIA GeForce RTX 3080 | 32 | Skipped, lane range 32–32 | 408 queries passed | 32.0.16.1088 report · 32.0.16.2124 report · Independent 32.0.16.2124 run · Independent 32.0.16.1088 run, source build at bc24dfd |
| NVIDIA GeForce RTX 3060 Ti | 32 on every run | Skipped on every run, lane range 32–32 | 408 queries passed per run | 32.0.16.2101 report · 32.0.16.1742 report · Independent 32.0.16.2101 run · 32.0.16.1088 report |
| Intel Arc B580 | 32 | Skipped, lane range 16–32 | 408 queries passed per run | 32.0.101.9175 report · Independent run, same driver |
| AMD Radeon RX 5700 XT | 32 | Passed, 64 lanes | Skipped, DXR unsupported | 32.0.21045.29003 report |
| NVIDIA GeForce RTX 3050 | 32 | Skipped, lane range 32–32 | 408 queries passed | 32.0.16.2114 report · 32.0.16.2135 report |
| AMD Radeon RX 7600 | 32 | Passed, 64 lanes | 408 queries passed | 32.0.31021.5001 report · 32.0.31035.1003 report · 32.0.32015.7029 report · Independent 32.0.32015.7029 run |
| Radeon RX 570 Series | 64 | Skipped, Shader Model 6.5 | Skipped, DXR unsupported | 31.0.21029.6002 report |
| AMD Radeon RX 6900 XT | 32 | Passed, 64 lanes | 408 queries passed | 32.0.21043.19003 report |
| NVIDIA GeForce GTX 1660 Ti | 32 | Skipped, lane range 32–32 | Skipped, DXR tier 1.0 | 32.0.16.2124 report |
| NVIDIA GeForce RTX 4070 | 32 | Skipped, lane range 32–32 | 408 queries passed | 32.0.16.1714 report · 32.0.16.2124 report · Independent 32.0.16.2124 run, source build at b7f7e3f |
| NVIDIA GeForce RTX 2080 | 32 | Skipped, lane range 32–32 | 408 queries passed | 32.0.16.1088 report |
| NVIDIA GeForce RTX 4080 | 32 | Skipped, lane range 32–32 | 408 queries passed | 32.0.16.2124 report |
| NVIDIA GeForce RTX 5090 | 32 | Skipped, lane range 32–32 | 408 queries passed per run | 32.0.16.2124 report · Full source-built hardware suite at ee39536 · Independent run, same driver |
| AMD Radeon(TM) Graphics (Ryzen 5 7600 and 7600X, device 164e) | 32 | Passed, 64 lanes | 408 queries passed | 32.0.21045.29003 report, package built from #116 at c659916 · 32.0.13066.12 source build at bc24dfd, Ryzen 5 7600X |
| NVIDIA GeForce GTX 1070 Ti | 32 | Skipped, lane range 32–32 | Skipped, DXR tier 1.0 | 32.0.15.8266 report |
| NVIDIA GeForce RTX 5070 Ti Laptop GPU | 32 | Skipped, lane range 32–32 | 408 queries passed | 32.0.16.2135 report |
| NVIDIA GeForce RTX 3050 Ti Laptop GPU | 32 | Skipped, lane range 32–32 | 408 queries passed | 32.0.16.2101 report |
| AMD Radeon RX 6600 XT (driver 32.0.21045.29003) | 32 | Skipped, Shader Model 6.5 on build 19045; passed, 64 lanes on build 26200 | 408 queries passed per run | Windows build 19045 report · Windows build 26200 report |
| NVIDIA GeForce RTX 5060 Ti | 32 | Skipped, lane range 32–32 | 408 queries passed per run | 32.0.16.2101 report · 32.0.16.1714 report · 32.0.16.2124 report · Independent 32.0.16.2124 run by abilliontokens · 32.0.16.1088 report |
| Radeon RX 590 Series | 64 | Passed, 64 lanes | Skipped, DXR unsupported | 31.0.21923.11000 report |
| NVIDIA GeForce RTX 4070 Ti | 32 | Skipped, lane range 32–32 | 408 queries passed per run | 32.0.16.1088 report · Independent run, same driver |
| AMD Radeon(TM) Graphics (device 164c) | 64 | Passed, 64 lanes | Skipped, DXR unsupported | 31.0.21925.1001 report |
| NVIDIA GeForce GTX 1660 SUPER | 32 | Skipped, lane range 32–32 | Skipped, DXR tier 1.0 | 32.0.16.2101 report |
| NVIDIA GeForce RTX 2070 SUPER | 32 | Skipped, lane range 32–32, Shader Model 6.5 | 408 queries passed | 32.0.16.1088 report |
| NVIDIA GeForce RTX 5090 Laptop GPU | 32 | Skipped, lane range 32–32 | 408 queries passed | 32.0.16.2124 report |
| NVIDIA GeForce RTX 3090 | 32 | Skipped, lane range 32–32 | 408 queries passed | 32.0.16.2124 report · 32.0.16.1062 report, debug layer enabled |
| NVIDIA RTX 3500 Ada Generation Laptop GPU | 32 | Skipped, lane range 32–32 | 408 queries passed | 32.0.16.2124 report |
| AMD Radeon Graphics (device 15bf) | 32 | Passed, 64 lanes | 408 queries passed | 32.0.32015.14001 report |
| AMD Radeon RX 5600 XT | 32 | Passed, 64 lanes | Skipped, DXR unsupported | 32.0.21045.16010 report |
| AMD Radeon RX 7900 XTX | 32 | Passed, 64 lanes | 408 queries passed | 32.0.32029.2010 report, debug layer enabled · Independent 32.0.32029.2010 run, source build at bc24dfd |
| Intel Arc B390 GPU (Panther Lake) | 32 | Skipped, lane range 16–32 | 408 queries passed | 32.0.101.8509 report |
| NVIDIA GeForce RTX 4080 SUPER | 32 | Skipped, lane range 32–32 | 408 queries passed | 32.0.16.2101 report |
| NVIDIA GeForce GTX 1650 | 32 | Skipped, lane range 32–32 | Skipped, DXR unsupported | 32.0.16.2124 report |
| NVIDIA GeForce GTX 1060 6GB | 32 | Skipped, lane range 32–32, Shader Model 6.5 | Skipped, DXR tier 1.0 | 32.0.15.6094 report, Windows 10 build 19045, source build at 6549244 |
| NVIDIA GeForce RTX 3050 Laptop GPU | 32 | Skipped, lane range 32–32 | 408 queries passed | 32.0.16.2124 report · 32.0.16.2114 report, debug layer enabled |
| NVIDIA GeForce RTX 3050 6GB Laptop GPU | 32 | Skipped, lane range 32–32 | 408 queries passed | 32.0.16.1062 report |
| Microsoft WARP (software) | 4 | Skipped | 408 queries passed | Maintainer baseline · 10.0.26100.9711 report · Independent 10.0.26100.9711 run · 10.0.26100.9711 source build at bc24dfd · 10.0.26100.9818 report |
Under Linux translation layers (Wine or Proton with vkd3d-proton), Mesa RADV on an AMD Ryzen 5 integrated GPU and an RX 9070 XT reported by @barbiscool passed all forced-wave64 groups with zero mismatches; RADV defaults to 32-lane waves on the 9070 XT where Windows drivers default to 64. NVIDIA translation-layer runs on an RTX 4070, RTX 5060 Ti and RTX 3060 reported a 32-lane maximum, skipped forced wave64 and passed 408 BVH comparisons.
An earlier batch at commit def386a covered seven GPU models in eight community reports. Each Radeon completed all 38 forced-wave64 groups with zero mismatches. The RX 9070 XT, RTX 3070 and RTX 5070 driver retest reported zero D3D12 debug warnings and errors in their passing checks.
Maintainers link independent results to the original report and record the adapter, driver, commit and outcome. Names or handles are added on request. The hardware-report and independent-build milestones are complete. The remaining community milestone is a fix verified by the contributor who reported the problem.
The installation metadata inspected for this case study lists a Scarlett executable. The main executable was absent from the inspected files. The public regression suite uses source-built fixtures and independently built shaders. Intake methods and evidence
Windows Authenticode verification on the local copy reported HashMismatch for 12 DLLs and NotSigned for four. Three independently compared whole-file SHA-256 values matched the community reports. The cause of the digest mismatch and the original signing path remain undetermined. Methods and fingerprints
Original project code and documentation use the MIT License. Third-party games, assets, SDKs and trademarks belong to their respective owners. Xbox2Windows is an independent community project, unaffiliated with Rockstar Games, Take-Two, Microsoft or AMD.