Repository navigation
Description
Why
Under Proton, WrongGPU meets a different D3D12 stack: vkd3d-proton with dxil-spirv on RADV, DXVK's DXGI, and Wine's d3dcompiler_47. #7 (Wine's shader compiler) and the AGS matrix-instruction translation (dxil-spirv PR #362) only show up there.
A small runner makes those checks repeatable on any Linux machine with an RX 9000 card and a Proton install. That includes your retest request in #7.
What I have (clean/linux/, Python with NumPy plus MinGW-w64; on my fork as test/linux-wine-runner, two commits on main)
clean/linux/run_wine_tests.py --proton <Proton folder> [--vkd3d <self-built vkd3d-proton folder>] [--runtime <wronggpu_runtime.dll>] sets up a Wine prefix the way Proton sets up a game's: vkd3d-proton, DXVK's dxgi and Wine's vkd3d libraries. It then runs four tests, each reporting PASS, FAIL or SKIP:
interop: the existingclean/runtime/interop/tests/run-tests.sh, 31 tests, compiled against Wine's ownd3dcompiler_47;rr-leak: the RR test host'stagloopjob from RR runtime: free released input copies while no RR instance exists #11, with the 256 MiB limit of the Windows end-to-end test;ags-layout:wg_ags.hlsli's WMMA and conversion helpers on fixed register values, against a CPU reference of the gfx12 register layout (that's what dxil-spirv's translation has to preserve);launcher: the launcher's four unit-test programs, aWRONGGPU_TESTINGMinGW build run under Wine.
run-tests.sh needs two small changes: -fms-extensions for clang-based MinGW (the __try in descriptor_publish.h), and a way to pick Proton's wine over a system wine64.
Results on current main (a8500cb), RX 9070 XT, RADV, synthetic data only, with a vkd3d-proton built from vkd3d-proton 5b1089c8 plus #362's three code commits:
- All four tests pass under GE-Proton 11-11: interop 31/31, rr-leak +37 MiB, ags-layout 7/7, launcher 71/225/186/18 checks.
- interop, ags-layout and rr-leak also pass under Proton Experimental 11.0-20261105c, Proton-CachyOS 11.0-20260703, DW-Proton 11.0-15 and Proton 11.0.
- Model M fails under Proton: the interop spin shader's globallycoherent is rejected by Wine's d3dcompiler_47 #7 retest: interop passes 31/31 on main under all five. Before Interop: declare the spin pass's buffer without globallycoherent #12 it was 29/31, with "Unknown modifier globallycoherent".
- Model M driver probe: distinct values per lane and element #14 under Proton: your
native_probe_test(MinGW build, run under Wine) passes on the RX 9070 XT with that vkd3d-proton under all five Protons. WithWG_TEST_DRIVER_PROBE_FAIL=1it reports "probe mismatch at float 7". With Proton's stock vkd3d-proton the probe kernel fails atCreateComputePipelineState, as expected without #362.test_wmma_probe.cppalso passes with Linux clang.
Each test fails on what it targets:
- rr-leak: +2636 MiB with a runtime from before RR runtime: free released input copies while no RR instance exists #11;
- ags-layout: fails on stock vkd3d-proton, and rejects lane-swapped and half-swapped results;
- launcher: fails when a check is broken on purpose.
Nothing in it runs on Windows or in CI.
If you'd like it in the repository, I can open a PR from that branch.
One more Linux note: tools/ci/run_python_tests.py fails 18 emulator and vector-table modules on Linux, the same on unchanged main. Most come from compiling the HIP kernels with Linux's clang, which lacks the AMD GPU builtins (__builtin_amdgcn_readfirstlane, the "+s" asm constraint). One test also writes a file called nul. That may be useful for #4.
Metadata
Metadata
Assignees
Labels
Projects
- StatusShow more project fieldsTodo
Activity