Enable experimental native Windows RDNA3 and RDNA4 targets

Share explicit target metadata across build/runtime/package guards, preserve native gfx11 WMMA with a Windows LDS lane exchange, select HIP devices by D3D12 LUID, and reject wrong-target/vkd3d installations before changing game files. Review all issue #10 logs and document compile-only coverage plus RX9070XT regressions.

Co-authored-by: realdody <[email protected]>

CMakeLists.txt

@@ -15,15 +15,36 @@ endif()
 find_program(D4R_HIP_CLANG NAMES clang++.exe PATHS "${D4R_HIP_ROOT}/bin" "${D4R_HIP_ROOT}/lib/llvm/bin" NO_DEFAULT_PATH REQUIRED)
 # Architecture parameterization adapted from realdody/d4r ae530b0.
 # Keep explicit subtargets: no gfx11 substitution or untested future gfx12 ISA.
+file(READ "${CMAKE_CURRENT_SOURCE_DIR}/tools/windows/gpu-targets.json" d4r_target_json)
+set_property(DIRECTORY APPEND PROPERTY CMAKE_CONFIGURE_DEPENDS tools/windows/gpu-targets.json)
+string(JSON d4r_target_count LENGTH "${d4r_target_json}" targets)
+math(EXPR d4r_target_last "${d4r_target_count} - 1")
+set(d4r_supported_targets)
+set(d4r_target_cpp_entries "")
+foreach(index RANGE ${d4r_target_last})
+  foreach(field architecture elfMachine isaMajor family)
+    string(JSON target_${field} GET "${d4r_target_json}" targets ${index} ${field})
+  endforeach()
+  list(APPEND d4r_supported_targets "${target_architecture}")
+  string(APPEND d4r_target_cpp_entries "  GpuTarget{\"${target_architecture}\", ${target_elfMachine}, ${target_isaMajor}, \"${target_family}\"},\n")
+endforeach()
+configure_file(tools/windows/gpu_targets.generated.h.in gpu_targets.generated.h @ONLY)
+include_directories("${CMAKE_CURRENT_BINARY_DIR}")
 set(D4R_GPU_ARCH "gfx1201" CACHE STRING "Native Windows package GPU target")
-set_property(CACHE D4R_GPU_ARCH PROPERTY STRINGS gfx1201 gfx1200)
-set(d4r_supported_targets gfx1200 gfx1201)
+set_property(CACHE D4R_GPU_ARCH PROPERTY STRINGS ${d4r_supported_targets})
 if(NOT D4R_GPU_ARCH IN_LIST d4r_supported_targets)
-  message(FATAL_ERROR "D4R_GPU_ARCH must be gfx1200 or gfx1201")
+  message(FATAL_ERROR "D4R_GPU_ARCH must be one of: ${d4r_supported_targets}")
 endif()
 add_compile_definitions(D4R_TARGET_ARCH="${D4R_GPU_ARCH}")
 configure_file(tools/windows/gpu-target.json.in gpu-target.json @ONLY)
 install(FILES "${CMAKE_CURRENT_BINARY_DIR}/gpu-target.json" DESTINATION .)
+install(FILES tools/windows/gpu-targets.json DESTINATION .)
+set(d4r_device_flags)
+if(D4R_GPU_ARCH MATCHES "^gfx11")
+  # Windows-only workaround informed by realdody/d4r 1897b4a. Keep the real
+  # gfx11 layout; do not route RDNA3 through gfx12 fragments.
+  list(APPEND d4r_device_flags -DD4R_WINDOWS_GFX11_LANE_EXCHANGE)
+endif()
 set(probe_object "${CMAKE_CURRENT_BINARY_DIR}/probe_${D4R_GPU_ARCH}.hsaco")
 add_custom_command(OUTPUT "${probe_object}"
   COMMAND "${D4R_HIP_CLANG}" -x hip --offload-arch=${D4R_GPU_ARCH}
@@ -44,10 +65,10 @@ set(wmma_object "${CMAKE_CURRENT_BINARY_DIR}/wmma_${D4R_GPU_ARCH}.hsaco")
 add_custom_command(OUTPUT "${wmma_object}"
   COMMAND "${D4R_HIP_CLANG}" -x hip --offload-arch=${D4R_GPU_ARCH}
     --offload-device-only --no-gpu-bundle-output -mno-wavefrontsize64
-    -nogpuinc -nogpulib -O2 "${CMAKE_CURRENT_SOURCE_DIR}/tools/windows/wmma_gfx1201.hip"
+    -nogpuinc -nogpulib -O2 ${d4r_device_flags} "${CMAKE_CURRENT_SOURCE_DIR}/tools/windows/wmma_gfx1201.hip"
     -o "${wmma_object}"
   DEPENDS tools/windows/wmma_gfx1201.hip kernels/common/wmma_backend.h
-    kernels/common/wmma_layout.h VERBATIM)
+    kernels/common/wmma_layout.h kernels/common/wave32_exchange.h VERBATIM)
 add_custom_target(d4r_wmma_code_object ALL DEPENDS "${wmma_object}")
 # K/M objects stay in experimental/, outside the ZLUDA native override
 # directory, until each layer and then the complete network pass replay.
@@ -78,9 +99,9 @@ foreach(family k m)
     add_custom_command(OUTPUT "${native_object}"
       COMMAND "${D4R_HIP_CLANG}" -x hip -std=c++17 --offload-arch=${D4R_GPU_ARCH}
         --offload-device-only --no-gpu-bundle-output -mno-wavefrontsize64
-        -nogpuinc -nogpulib -O3 -DD4R_DEVICE_ONLY_MINIMAL ${native_flags}
+        -nogpuinc -nogpulib -O3 -DD4R_DEVICE_ONLY_MINIMAL ${d4r_device_flags} ${native_flags}
         "${source}" -o "${native_object}"
-      DEPENDS "${source}" ${native_headers} kernels/common/wmma_layout.h
+      DEPENDS "${source}" ${native_headers} kernels/common/wmma_layout.h kernels/common/wave32_exchange.h
         kernels/common/hip_device_minimal.h VERBATIM)
     list(APPEND native_objects "${native_object}")
     install(FILES "${native_object}" DESTINATION "experimental/${family}")
@@ -89,7 +110,7 @@ endforeach()
 add_custom_target(d4r_native_compile ALL DEPENDS ${native_objects})
 # Retain the old manual build target for existing scripts.
 add_custom_target(d4r_native_gfx1201_compile DEPENDS d4r_native_compile)
-foreach(probe hip_gfx1201_probe gfx12_wmma_probe k_module_probe native_replay_probe cuda_driver_probe cuda_mma_probe cuda_image_probe d3d12_command_probe d3d12_pixel_probe d3d12_hip_interop_probe hip_stream_lifecycle_probe ngx_cuda_init_probe d3d12_dlss_probe)
+foreach(probe gpu_inventory hip_gfx1201_probe gfx12_wmma_probe k_module_probe native_replay_probe cuda_driver_probe cuda_mma_probe cuda_image_probe d3d12_command_probe d3d12_pixel_probe d3d12_hip_interop_probe hip_stream_lifecycle_probe ngx_cuda_init_probe d3d12_dlss_probe)
   add_executable(d4r_${probe} tools/windows/${probe}.cpp)
   target_compile_definitions(d4r_${probe} PRIVATE WIN32_LEAN_AND_MEAN NOMINMAX __HIP_PLATFORM_AMD__)
   target_include_directories(d4r_${probe} PRIVATE "${D4R_HIP_ROOT}/include")
@@ -200,6 +221,7 @@ install(FILES "${pixel_object}" DESTINATION bin)
 install(FILES "${wmma_object}" DESTINATION bin)
 install(FILES scripts/windows/test-windows-rdna4.ps1 DESTINATION .)
 install(FILES scripts/windows/gpu-target.ps1 DESTINATION .)
+install(FILES scripts/windows/gpu-preflight.ps1 DESTINATION .)
 install(FILES scripts/windows/test-gpu-target-contract.ps1 DESTINATION .)
 install(FILES scripts/windows/test-k-replay.ps1 DESTINATION .)
 install(FILES scripts/windows/test-k-captures.ps1 DESTINATION .)

docs/windows-game.md

@@ -1,21 +1,29 @@
-# Native Windows RDNA4 game development package
+# Native Windows RDNA3 / RDNA4 game development package
 
 Validated hardware: Windows 11 x64 / Radeon RX 9070 XT / gfx1201. Native K and
 M transformer kernels pass standalone D3D12 + OptiScaler checks. Silent Hill 2
 K now renders real menu frames and M renders a saved gameplay level with
 GPU NaN/Inf checks. Both K and M pass 3840x2160 output on this machine.
-The user reports 49-51 FPS for K in the measured scene after the command-hook
-fix and locally built output-store optimization. This is a local
+Performance measurements and their controls are recorded in
+[windows-performance.md](windows-performance.md). This is a
 development package; no NVIDIA proprietary DLL is included.
 
-Builds also accept gfx1200 (for example RX 9060/9060 XT). This target is
-**compile-tested only**; no physical gfx1200 validation has been performed.
+Experimental builds also accept gfx1100..gfx1103 (RDNA3), gfx1150..gfx1154
+(RDNA3.5) and gfx1200 (for example RX 9060/9060 XT). These targets are
+**compile-tested only**; no physical validation here has been performed.
 RX 9070 and other gfx1201 cards also remain untested. Use a package built for
 the GPU's actual HIP architecture: its `gpu-target.json`, host shim and ELF
 code objects must agree. Installation/runtime reject mixed targets. See
 [windows-gpu-support.md](windows-gpu-support.md) for separate-target build
 commands, the fork review and the first hardware diagnostic gate.
 
+Before changing game files, the runner inventories HIP devices without
+executing kernels. It rejects another architecture's package and a local
+D3D12 proxy identifying itself as vkd3d/Wine, producing one diagnostic ZIP.
+RX 9060 XT needs gfx1200: the old gfx1201-only package fails before DLSS
+initialization on it. Use native D3D12. On an untested GPU, start with
+synchronous K and output/exception diagnostics.
+
 Build the shim/diagnostics, corrected ZLUDA, K/M native objects and patched
 OptiScaler using the scripts in `scripts/windows`, then run:
 

docs/windows-gpu-support.md

@@ -1,40 +1,80 @@
-# Windows GPU targets
+# Experimental Windows RDNA3 / RDNA4 coverage
 
-The native Windows build accepts **gfx1200** and **gfx1201**. It builds a
-separate shim, diagnostics and code objects for each target. The default stays
-gfx1201. GPU selection uses HIP's real `gcnArchName`, followed by D3D12/HIP
-adapter LUID matching; it does not depend on a marketing name or overridden
-DXGI vendor ID.
+The shim, diagnostics, native K/M kernels and game installer accept the
+targets below, with one package per actual ISA. **Only RX 9070 XT has physical
+validation here.** All other targets are experimental and compile-tested;
+working DLSS, interop and performance on them remain unverified.
 
-| Target | Examples in LLVM's processor table | This project's validation |
+| Family | Target | Examples / scope |
 | --- | --- | --- |
-| gfx1201 | RX 9070, RX 9070 XT | RX 9070 XT: hardware, K/M and game tests. Other cards untested. |
-| gfx1200 | RX 9060, RX 9060 XT | Host/device compilation and software tests only. No physical gfx1200 available. |
+| RDNA3 | gfx1100 | RX 7900 family; same-target workstation GPUs |
+| RDNA3 | gfx1101 | RX 7700 / 7700 XT / 7800 XT; same-target mobile/workstation GPUs |
+| RDNA3 | gfx1102 | RX 7600 / 7600 XT; same-target mobile/workstation GPUs |
+| RDNA3 | gfx1103 | RDNA3 integrated graphics |
+| RDNA3.5 | gfx1150, gfx1151, gfx1152, gfx1153, gfx1154 | Corresponding integrated graphics targets |
+| RDNA4 | gfx1200 | RX 9060 / 9060 XT; same-target GPUs |
+| RDNA4 | gfx1201 | RX 9070 / 9070 XT / 9070 GRE; same-target GPUs |
 
-Sources: [LLVM AMDGPU processors](https://llvm.org/docs/AMDGPUUsage.html#processors)
-and [AMD's Windows compatibility matrix](https://rocm.docs.amd.com/projects/radeon-ryzen/en/latest/docs/compatibility/compatibilityrad/windows/windows_compatibility.html).
-AMD's matrix is dependency information, not evidence that d4r's K/M or external
-memory path works on an untested card. Other RDNA4 SKUs are eligible only when
-HIP reports one of these exact targets and exposes the required Windows APIs.
-Device-specific drivers, performance and memory capacity still need testing.
-gfx1250/gfx1251 are a separate ISA group and are not enabled by this change.
+Names and IDs follow [LLVM processors](https://llvm.org/docs/AMDGPUUsage.html#processors)
+and [ELF ABI](https://llvm.org/docs/AMDGPUUsage.html#amdgpu-elf-header-e-flags).
+`tools/windows/gpu-targets.json` supplies CMake-generated C++, PowerShell and
+the private texture builder. Selection uses HIP's real `gcnArchName` and the
+D3D12 adapter LUID, including systems with several GPUs sharing a target.
+Marketing names do not control eligibility.
 
-Do not rename a gfx1201 package to gfx1200. `gpu-target.json`, the compiled
-host target and each ELF code-object target must agree. Runtime selection
-rejects another architecture before launching kernels. Staging, packaging and
-installation inspect the documented AMDGPU ELF machine ID and reject mixed
-objects. They never rewrite a binary target or set `HSA_OVERRIDE_GFX_VERSION`.
-Legacy packages without target metadata remain gfx1201 only.
+All listed RDNA3/RDNA3.5/RDNA4 targets with existing gfx11/gfx12 WMMA contracts
+are included. TBD RDNA4m gfx1170..gfx1172 and gfx12.5 gfx1250/gfx1251 are
+different ISA groups and excluded. Unknown future targets need explicit review.
+Check [AMD's Windows matrix](https://rocm.docs.amd.com/projects/radeon-ryzen/en/latest/docs/compatibility/compatibilityrad/windows/windows_compatibility.html)
+for driver/runtime eligibility. Compiler support cannot supply missing Windows
+driver support or external-memory APIs, especially on APUs.
 
-## Build and package
+## Installation and diagnostics
 
-Prepare the pinned dependencies as described in `windows-rdna4-port.md`.
-Select a target explicitly; this example creates the experimental gfx1200
-package. Use gfx1201 for a 9070/9070 XT. Keep separate output directories when
-building both targets.
+Use the actual target's archive: **gfx1200 for RX 9060 XT**, for example.
+Metadata, host target and every ELF target must agree. Runtime/scripts reject
+mixed targets and never rename targets, rewrite ELF IDs or set
+`HSA_OVERRIDE_GFX_VERSION`. Legacy metadata-free packages remain gfx1201 only.
+
+Before changing game files, `windows-game.ps1` now runs a read-only HIP
+inventory and checks the target and wave32. Inventory allocates no GPU buffers
+and runs no kernels; success is discovery, not validation. Failure prints one
+ZIP containing stdout/stderr, exit code, runtime/DLL and driver details. A local
+`d3d12.dll` identifying itself as vkd3d/Wine is rejected before install. Restore
+the game's native D3D12 installation; the script never loads or removes that
+proxy itself. Begin synchronously on an untested GPU:
 
 ```powershell
-$arch = 'gfx1200'
+.\windows-game.ps1 -GameExe "D:\Games\Example\Game.exe" -NgxCore "C:\LocalDLLs\_nvngx.dll" -DlssDll "C:\LocalDLLs\nvngx_dlss.dll" -Preset 11 -ValidateOutput -CaptureExceptions
+```
+
+The local DLLs must match the documented NGX version and DLSS 310.9.1 SHA256
+in `package.json`. NVIDIA DLLs are never distributed. Retain the printed ZIP
+on failure. Async K and performance are subsequent gates; gfx1201-specific
+FP8, arithmetic and tuning experiments remain separate.
+
+## Issue #10 review, 2026-10-02
+
+Reviewed [comment 5952822115](https://github.com/countervolts/d4r/issues/10#issuecomment-5952822115),
+including all 32 attached text logs from eight attempts. HIP discovers
+**RX 9060 XT / gfx1200** successfully. Every attempt fails with the old shim's
+**"No selected gfx1201 device"** rejection; OptiScaler reports `support=false`
+/ `BAD00002`. None reaches `D4R_RUNTIME` initialization or native transformer
+execution. These logs do not demonstrate a K/M kernel crash. Use the gfx1200
+package instead of overriding architecture. Three attempts also load
+vkd3d-proton 3.1.0 from a local D3D12 proxy, outside this Windows backend.
+
+The new packages/preflight address the identified failures. They do not prove
+those games work on RX 9060 XT; a new native-D3D12 run on that GPU is required.
+Raw community logs and personal directory paths stay local.
+
+## Build
+
+Prepare dependencies following `windows-rdna4-port.md`. Default: gfx1201.
+Select another target explicitly and keep separate directories:
+
+```powershell
+$arch = 'gfx1101'  # for example RX 7800 XT
 $diag = "$PWD\dist\windows-$arch-diagnostics"
 .\scripts\windows\build-windows-rdna4.ps1 -RuntimeProfile therock -GpuArch $arch -ZludaRoot "$PWD\dist\zluda-windows-final" -BuildDirectory "$PWD\build\windows-$arch" -InstallDirectory $diag
 .\scripts\windows\stage-native-k.ps1 -DlssDll "$PWD\nvngx_dlss.dll" -PackageRoot $diag
@@ -42,89 +82,60 @@ $diag = "$PWD\dist\windows-$arch-diagnostics"
 .\scripts\windows\package-windows-game.ps1 -DiagnosticRoot $diag -GpuArch $arch -PackageRoot "$PWD\dist\windows-$arch-game" -ArchivePath "$PWD\dist\windows-$arch-game.zip"
 ```
 
-ZLUDA, HIP and OptiScaler must already be built/prepared. The package contains
-the strict K FP16 and M FP16-equivalent baseline; native FP8 stays disabled.
-The architecture does not change arithmetic or the existing WMMA layout.
-Private NVIDIA-derived texture overrides remain optional local experiments;
-their builder accepts `-GpuArch`, and their validation must match the target
-package before installation. Existing gfx1201 performance measurements are
-not measurements of gfx1200.
-The separate experimental FP8/F16 arithmetic and tuning recipes remain
-gfx1201-specific; this change enables the conservative production baseline
-on both build targets.
+Rebuild diagnostics for the new inventory executable. The package keeps strict
+K FP16 and M FP16-equivalent baselines; native M FP8 remains disabled. CMake:
+`-DD4R_GPU_ARCH=gfx1101`. Some source/executable names retain gfx1201/gfx12 for
+compatibility; actual targets follow the selected architecture.
 
-Equivalent CMake selection is `-DD4R_GPU_ARCH=gfx1200`. Unknown targets fail
-configuration. Some diagnostic executable names and source filenames retain
-`gfx1201` for compatibility; the logged architecture and compiled code objects
-use the selected target.
-
-## First hardware gate on another card
-
-From the repository root, with the matching diagnostic package above and the
-pinned dependencies, run this single command:
+One command compiles all eleven targets without GPU workloads:
 
 ```powershell
-.\scripts\windows\test-windows-rdna4.ps1 -RuntimeProfile therock -PackageRoot "$PWD\dist\windows-gfx1200-diagnostics" -ZludaRoot "$PWD\dist\zluda-windows-final"
+.\scripts\windows\build-windows-gpu-matrix.ps1 -ZludaRoot "$PWD\dist\zluda-windows-final"
 ```
 
-Success requires exit 0, all selected tests passed in `summary.json`, the
-correct actual GPU/target in stdout, exact HIP/PTX/WMMA/pixel-format checks and
-stable interop lifetime. On failure, retain the single printed diagnostic ZIP;
-it includes stdout/stderr, exit code, runtime/driver inventory and DLL identity.
-No NVIDIA DLL is required for this first gate. Full K/M reference, temporal,
-standalone DLSS and game validation are subsequent gates on that hardware;
-passing the first gate alone does not close them.
+Add `-PackageGames -DlssDll "$PWD\nvngx_dlss.dll" -ArchiveDirectory "$PWD\dist\windows-gpu-archives"`
+to stage/archive all packages after committing source and preparing OptiScaler.
+Use `-GpuArchitectures gfx1100,gfx1101,gfx1102,gfx1200,gfx1201` for a subset.
+Per-target logs and `matrix-build.json` distinguish compilation from execution.
+The supplied DLL generates manifests locally and is never copied into archives.
 
-Without that GPU, run `ctest --test-dir build/windows-gfx1200 -L software
---output-on-failure`. These CPU/WARP checks do not execute gfx1200 kernels.
+## Fork adaptation
 
-## Fork review
+Reviewed [realdody's branch](https://github.com/realdody/d4r/tree/windows-rdna3)
+at `ae530b0f53ed5c0bb87558dd7c20f619e014e4fd` and the newer
+[`1897b4a` correction](https://github.com/realdody/d4r/commit/1897b4afa79832bd879234cfb37e7c1468db5cde).
+Earlier MSVC binary16 and generated `BOOL` fixes remain. The earlier forced
+gfx12 fragment layout is not used.
 
-Reviewed [realdody's windows-rdna3 branch](https://github.com/realdody/d4r/tree/windows-rdna3)
-at `ae530b0f53ed5c0bb87558dd7c20f619e014e4fd`, on 2026-10-02. Adapted:
+The newer fork reports Windows gfx1101 `permlanex16` returning its own lane,
+duplicating even K values into odd slots. This port adapts that finding into
+`kernels/common/wave32_exchange.h`: a Windows/gfx11-only per-wave LDS exchange,
+with native gfx11 layout, generic x/y/z indexing, 2 KiB scratch (512 threads
+maximum), volatile accesses and wave fences/barriers protecting scratch reuse.
+It covers K/M fragment conversions and other K half-wave exchanges. gfx12 and
+Linux keep the original instruction path. The raw gfx11 probe uses its actual
+intrinsic/layout, without a forced gfx12 shim.
 
-- Per-target build configuration and actual-target diagnostics, narrowed here
-  to the two documented RDNA4 subtargets, with package/ELF guards added.
-- MSVC binary16 conversion in the synthetic CUDA MMA reference. An exhaustive
-  CPU test covers 63,490 finite/infinity encodings and 31,743 rounding
-  boundaries; LLVM also checks it against native `_Float16` conversion.
-- `WINBOOL` to `BOOL` normalization in generated D3D12 hooks, so the generated
-  header works with the Microsoft SDK as well as MinGW headers.
+The fork reports RX 7800 XT enc1/enc2 identity/nonzero/reference checks. Those
+are **the author's results**, not validation of this adaptation or full K/M
+here. LDS exchange, full network, translated CUDA, accuracy, interop and speed
+still need physical gfx11 testing. Credit: realdody <dodobozicek@gmail.com>.
 
-The fork's RDNA3 implementation forces `D4R_WMMA_LAYOUT=12` for gfx11. That
-kernel-layout change, the raw gfx11 probe and the expanded gfx11 runtime
-allowlist are not imported. Its reported RX 7800 XT/K result has not been
-independently reproduced here; its M/translated-path limitations still require
-separate checks. This change does not advertise native Windows RDNA3 readiness.
-The existing upstream Linux/RDNA3 implementation is unchanged. Credit for
-the adapted portability fixes: realdody <dodobozicek@gmail.com>.
+## Results and next hardware gate
 
-## Local results, 2026-10-02
+All eleven targets compile the complete MinGW host build and 19 device objects
+with TheRock `10.2.0a20260929`, HIP `7.17.26386`, Clang 24. Software target
+contracts inspect actual ELF objects without other-target GPU launches.
+RX 9070 XT passes all 20 CTest gates after these changes. K/M full-frame and
+package checks are recorded in `windows-rdna4-port.md`. Local logs are in
+`build/windows-gpu-coverage` and `test-results/windows-gpu-coverage`.
 
-- Both targets: complete MinGW host build and 19 device code objects using
-  TheRock `10.2.0a20260929`, HIP `7.17.26386`, Clang 24. The runtime, driver,
-  patched ZLUDA and OptiScaler versions are unchanged from the validated port.
-- gfx1201/RX 9070 XT: 20/20 CTest gates, including 16 physical-GPU gates.
-- gfx1200: 4/4 software gates; no gfx1200 kernel execution.
-- MSVC v143 `14.44.35207`: binary16 CPU test passes; CUDA MMA probe and D3D12
-  hook translation units compile. This is not a complete MSVC build claim.
-- K: 12 finite 4K frames at 2259x1271 input, burst and feature recreation;
-  all six async RGB images exactly match their synchronous controls.
-- M: eight finite 4K control/candidate frames at the same input dimensions;
-  all four new-build RGB images exactly match the prior validated build.
-  Both presets log every required native layer, zero CPU image copies and
-  current-frame output.
-- Original gfx1201 objects remain byte-identical on disk. In fresh builds,
-  `.text` and `.rodata*` match the prior gfx1201 objects for all 19 kernels;
-  fresh ELF files have different overall hashes. gfx1200 and gfx1201 have
-  distinct correct ELF targets. Matching instruction sections are compile
-  evidence, not a substitute for testing gfx1200 hardware.
-- A gfx1200 HIP probe on the physical gfx1201 fails cleanly with exit 4 before
-  any GPU allocation/kernel execution. This tests refusal, not gfx1200 support.
-- Both target packages pass install/restore checks in isolated workspace
-  fixtures; no game is launched. Only manifest-listed runtime files install.
-  Packaging rejects a mismatched destination or native kernel directory before
-  writing files. Every packaged code object has the expected ELF target.
+The first hardware gate on another GPU needs no NVIDIA DLL:
 
-Logs are local in `test-results/rdna4-targets-*` and
-`test-results/rdna4-target-objects.json`; they contain no public NVIDIA DLLs.
+```powershell
+.\scripts\windows\test-windows-rdna4.ps1 -RuntimeProfile therock -PackageRoot "$PWD\dist\windows-gfx1101-diagnostics" -ZludaRoot "$PWD\dist\zluda-windows-final"
+```
+
+Success requires exit 0, correct actual target, exact HIP/PTX/WMMA/pixel checks
+and stable interop lifetime. Failure prints one ZIP. Full K/M reference,
+temporal, DLSS and game validation remain subsequent gates.

docs/windows-rdna4-port.md

@@ -1,6 +1,7 @@
 # Windows / RDNA4 port
 
-Branch: `windows-rdna4`. Native Windows build targets: **gfx1200 / gfx1201**.
+Branch: `windows-rdna4`. Experimental Windows build targets: **gfx1100..gfx1103,
+gfx1150..gfx1154, gfx1200 / gfx1201**.
 Validated physical hardware: Windows 11 x64, RX 9070 XT, **gfx1201**.
 Priority: correct K, correct M, native Windows, same-frame output, then speed.
 Upstream integration: [PR #11](https://github.com/countervolts/d4r/pull/11)
@@ -18,15 +19,38 @@ explicit limits; the validated package keeps conservative arithmetic.
 
 ## Milestone and gates
 
-2026-10-02 target coverage: gfx1200 and gfx1201 now have separate host/device
-builds and guarded packages. Both compile; gfx1201 passes 20 CTest gates on
-RX 9070 XT, and gfx1200 passes four software-only gates. No physical gfx1200
-test is available. K's new host build passes twelve finite 4K burst/recreation
-frames with all six async RGB images exact. M's four new-build 4K RGB images
-exactly match the previous build. The fork audit, adopted MSVC fixes,
-per-target commands, validation limits and first hardware gate are maintained
-in [windows-gpu-support.md](windows-gpu-support.md). RDNA3 runtime/layout changes
-from the reviewed fork are not imported. K performance remains open.
+2026-10-02 coverage: all eleven listed RDNA3 / RDNA3.5 / RDNA4 targets compile
+with separate host/device objects and guarded packages. Only RX 9070 XT has
+physical validation here; no other-target GPU workloads are executed. The
+newest realdody correction informs a Windows-only gfx11 LDS half-wave exchange
+preserving native gfx11 WMMA layout. gfx12 and Linux keep their original
+instruction path. A shared registry drives build, runtime, ELF checks and
+scripts. HIP selection matches the D3D12 LUID before selecting an ordinal,
+including same-target multi-GPU systems.
+
+Issue #10 comment 5952822115 contains 32 logs / eight attempts: all discover
+RX 9060 XT / gfx1200 but reject the old gfx1201 package before NGX runtime
+initialization. Three also load a vkd3d-proton D3D12 proxy. The installer now
+catches these conditions before changing game files and prints one diagnostic
+ZIP. The next community gate is the gfx1200 package on native D3D12; those logs
+do not establish a native K/M crash. Coverage, sources, fork credit, commands
+and hardware limits are in [windows-gpu-support.md](windows-gpu-support.md).
+
+The rebuilt gfx1201 passes 20/20 CTest gates on RX 9070 XT. K completes twelve
+finite 4K burst/recreation frames at 2259x1271 input, with all six async RGB
+images exactly matching sync. M completes eight finite control/candidate 4K
+frames, with all four candidate RGB images matching the prior build exactly.
+Both log every required native layer, zero CPU frame copies and frame age 0.
+All 209 compiled objects have the correct ELF target. For all nineteen gfx1201
+objects, instruction/constant sections match the previous validated build.
+
+All eleven game directories stage strict K/M baselines without NVIDIA DLLs.
+An isolated gfx1201 install/restore fixture preserves original bytes; wrong
+gfx1200-on-gfx1201 and dummy vkd3d-proxy fixtures refuse before game changes
+and produce one diagnostic ZIP each. The real game installation is untouched.
+Results: `test-results/windows-gpu-coverage`; build logs and the compile-only
+matrix report: `build/windows-gpu-coverage`. No physical multi-GPU, APU or
+other-target validation is claimed. K performance remains open.
 
 **Current milestone: K performance remains open.** The latest unprofiled 4K
 Silent Hill 2 scene capture records 65.250 FPS. Functional K/M and the

kernels/common/wave32_exchange.h

@@ -0,0 +1,31 @@
+#pragma once
+#include <stdint.h>
+
+// Windows gfx11 portability workaround informed by realdody/d4r 1897b4a.
+// The reported permlanex16 failure duplicates the even K value into the odd
+// slot. Retain the native gfx11 layout and exchange through per-wave LDS.
+// Other platforms and gfx12 keep their original instruction sequence.
+__attribute__((device, always_inline)) static inline uint32_t d4r_wave32_other_half(uint32_t value)
+{
+#if defined(D4R_WINDOWS_GFX11_LANE_EXCHANGE) && defined(__GFX11__)
+    __attribute__((shared)) volatile uint32_t scratch[512];
+    const uint32_t tid = __builtin_amdgcn_workitem_id_x() + __builtin_amdgcn_workgroup_size_x() *
+        (__builtin_amdgcn_workitem_id_y() + __builtin_amdgcn_workgroup_size_y() * __builtin_amdgcn_workitem_id_z());
+    // All native K/M launch bounds are <=512. Do not silently corrupt LDS if
+    // this helper is later reused by a larger workgroup.
+    if (tid >= 512) __builtin_trap();
+    scratch[tid] = value;
+    __builtin_amdgcn_fence(__ATOMIC_RELEASE, "wavefront");
+    __builtin_amdgcn_wave_barrier();
+    __builtin_amdgcn_fence(__ATOMIC_ACQUIRE, "wavefront");
+    const uint32_t result = scratch[tid ^ 16u];
+    // Protect scratch reuse by the next exchange in this wave. No block-wide
+    // rendezvous: different waves can reach GEMMs in a different order.
+    __builtin_amdgcn_fence(__ATOMIC_RELEASE, "wavefront");
+    __builtin_amdgcn_wave_barrier();
+    __builtin_amdgcn_fence(__ATOMIC_ACQUIRE, "wavefront");
+    return result;
+#else
+    return __builtin_amdgcn_permlanex16(value, value, 0x76543210u, 0xfedcba98u, false, false);
+#endif
+}

kernels/common/wmma_backend.h

@@ -1,5 +1,6 @@
 #pragma once
 #include <stdint.h>
+#include "wave32_exchange.h"
 
 // d4r's gfx11 interface: each lane supplies 16 FP16 K operands. The 8 FP32
 // outputs have logical row 2*i + lane/16 and logical column lane%16.
@@ -13,8 +14,7 @@ using d4r_wmma_f8 = float __attribute__((ext_vector_type(8)));
 __attribute__((device, always_inline)) static inline float d4r_wmma_other_half(float value)
 {
     const uint32_t bits = __builtin_bit_cast(uint32_t, value);
-    const uint32_t swapped = __builtin_amdgcn_permlanex16(
-        bits, bits, 0x76543210u, 0xfedcba98u, false, false);
+    const uint32_t swapped = d4r_wave32_other_half(bits);
     return __builtin_bit_cast(float, swapped);
 }
 
@@ -49,7 +49,7 @@ d4r_wmma_legacy_layout(d4r_wmma_h16 a, d4r_wmma_h16 b, d4r_wmma_f8 c)
         result[i] = (row >> 3) == group ? d12[new_index] : remote_d[new_index];
     }
     return result;
-#elif defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__) || defined(__gfx1103__)
+#elif defined(__GFX11__)
     return __builtin_amdgcn_wmma_f32_16x16x16_f16_w32(a, b, c);
 #else
 #error "d4r WMMA backend requires a validated gfx11 or gfx12 target"

kernels/common/wmma_layout.h

@@ -18,6 +18,7 @@
 // Only clang builtins are used (the texture tails compile with -nogpuinc -nogpulib).
 #pragma once
 #include <stdint.h>
+#include "wave32_exchange.h"
 
 #ifndef D4R_WMMA_LAYOUT
 #if defined(__GFX12__)
@@ -56,7 +57,7 @@ WM_FN uint32_t wm_half()
 // value of the lane with the same index in the other half of the wave
 WM_FN uint32_t wm_other_half(uint32_t v)
 {
-    return __builtin_amdgcn_permlanex16(v, v, 0x76543210u, 0xfedcba98u, false, false);
+    return d4r_wave32_other_half(v);
 }
 WM_FN float wm_other_half_f(float v)
 {

kernels/k/pwin_common.h

@@ -122,7 +122,7 @@ __device__ __forceinline__ acc8v splat8(float v)
 
 __device__ __forceinline__ uint32_t other_half(uint32_t v)
 {
-    return __builtin_amdgcn_permlanex16(v, v, 0x76543210u, 0xfedcba98u, false, false);
+    return wm_other_half(v);
 }
 
 __device__ __forceinline__ uint32_t pack2(half_t lo, half_t hi)

scripts/windows/build-native-texture.ps1

@@ -5,10 +5,12 @@ param(
     [string]$HipRoot = 'C:\Program Files\AMD\ROCm\7.2',
     [string]$ZludaRoot,
     [ValidateSet('cu','wgp')][string]$ShaderMode='cu',
-    [ValidateSet('gfx1200','gfx1201')][string]$GpuArch='gfx1201',
+    [string]$GpuArch='gfx1201',
     [string]$OutputDirectory
 )
 $ErrorActionPreference='Stop'
+. (Join-Path $PSScriptRoot 'gpu-target.ps1')
+[void](Get-D4RGpuTargetInfo $GpuArch)
 $repo=[IO.Path]::GetFullPath((Join-Path $PSScriptRoot '../..'))
 if (!$ZludaRoot) { $ZludaRoot=Join-Path $repo 'dist/zluda-windows-final' }
 if (!$OutputDirectory) { $OutputDirectory=Join-Path $repo "build/private-textures-$GpuArch" }

scripts/windows/build-windows-gpu-matrix.ps1

@@ -0,0 +1,44 @@
+[CmdletBinding()]
+param(
+    [string[]]$GpuArchitectures,
+    [ValidateSet('stable','therock')][string]$RuntimeProfile='therock',
+    [string]$ZludaRoot, [string]$DlssDll,
+    [switch]$PackageGames,
+    [string]$BuildRoot, [string]$OutputRoot, [string]$ArchiveDirectory
+)
+$ErrorActionPreference='Stop'
+. (Join-Path $PSScriptRoot 'gpu-target.ps1')
+$repo=[IO.Path]::GetFullPath((Join-Path $PSScriptRoot '../..'))
+if (!$GpuArchitectures) { $GpuArchitectures=@(Get-D4RGpuTargets | ForEach-Object { $_.architecture }) }
+foreach ($arch in $GpuArchitectures) { [void](Get-D4RGpuTargetInfo $arch) }
+if (@($GpuArchitectures | Select-Object -Unique).Count -ne $GpuArchitectures.Count) { throw 'Duplicate GPU target.' }
+if (!$BuildRoot) { $BuildRoot=Join-Path $repo 'build/windows-gpu-coverage' }
+if (!$OutputRoot) { $OutputRoot=Join-Path $repo 'dist/windows-gpu-coverage' }
+if (!$ZludaRoot) { $ZludaRoot=Join-Path $repo 'dist/zluda-windows-final' }
+if ($ArchiveDirectory -and !$PackageGames) { throw 'ArchiveDirectory requires PackageGames.' }
+if ($PackageGames -and (!$DlssDll -or ![IO.Path]::IsPathRooted($DlssDll) -or !(Test-Path -LiteralPath $DlssDll))) {
+    throw 'PackageGames requires an absolute path to your local nvngx_dlss.dll; it will not be redistributed.'
+}
+New-Item -ItemType Directory -Force $BuildRoot | Out-Null
+if ($ArchiveDirectory) { New-Item -ItemType Directory -Force $ArchiveDirectory | Out-Null }
+$results=[Collections.Generic.List[object]]::new()
+foreach ($arch in $GpuArchitectures) {
+    $build="$BuildRoot-$arch"; $diag="$OutputRoot-$arch"
+    $log=Join-Path $BuildRoot "$arch-build.log"
+    Write-Host "Building $arch (compile/package only; no GPU workloads). Log: $log"
+    & (Join-Path $PSScriptRoot 'build-windows-rdna4.ps1') -RuntimeProfile $RuntimeProfile -GpuArch $arch `
+        -ZludaRoot $ZludaRoot -BuildDirectory $build -InstallDirectory $diag *> $log
+    $row=[ordered]@{architecture=$arch; compiled=$true; hardwareTestExecuted=$false; buildDirectory=$build; diagnosticRoot=$diag; log=$log}
+    if ($PackageGames) {
+        $k="$BuildRoot-native-k-$arch"; $m="$BuildRoot-native-m-$arch"; $game="$OutputRoot-game-$arch"
+        & (Join-Path $PSScriptRoot 'stage-native-k.ps1') -DlssDll $DlssDll -PackageRoot $diag -GpuArch $arch -OutputDirectory $k
+        & (Join-Path $PSScriptRoot 'stage-native-m.ps1') -DlssDll $DlssDll -PackageRoot $diag -GpuArch $arch -OutputDirectory $m
+        $packageArgs=@{DiagnosticRoot=$diag; GpuArch=$arch; PackageRoot=$game; ZludaRoot=$ZludaRoot; KernelKRoot=$k; KernelMRoot=$m}
+        if ($ArchiveDirectory) { $packageArgs.ArchivePath=Join-Path $ArchiveDirectory "d4r-windows-$arch.zip" }
+        & (Join-Path $PSScriptRoot 'package-windows-game.ps1') @packageArgs
+        $row.gamePackage=$game
+    }
+    $results.Add($row)
+    $results | ConvertTo-Json -Depth 5 | Set-Content (Join-Path $BuildRoot 'matrix-build.json') -Encoding UTF8
+    Write-Host "Compiled $arch; hardware validation remains separate."
+}

scripts/windows/build-windows-rdna4.ps1

@@ -2,13 +2,15 @@
 param(
     [string]$HipRoot = $env:HIP_PATH,
     [ValidateSet('stable', 'therock')][string]$RuntimeProfile = 'stable',
-    [ValidateSet('gfx1200','gfx1201')][string]$GpuArch = 'gfx1201',
+    [string]$GpuArch = 'gfx1201',
     [string]$ZludaRoot,
     [string]$ToolchainRoot,
     [string]$BuildDirectory,
     [string]$InstallDirectory
 )
 $ErrorActionPreference = 'Stop'
+. (Join-Path $PSScriptRoot 'gpu-target.ps1')
+[void](Get-D4RGpuTargetInfo $GpuArch)
 $repo = [IO.Path]::GetFullPath((Join-Path $PSScriptRoot '../..'))
 if ($RuntimeProfile -eq 'therock') {
     if (!$PSBoundParameters.ContainsKey('HipRoot')) { $HipRoot = Join-Path $repo '.tools/therock-10.2.0a20260929/_rocm_sdk_core' }

scripts/windows/gpu-preflight.ps1

@@ -0,0 +1,58 @@
+# Read-only inventory before touching game files. It allocates no GPU memory
+# and executes no kernels; successful discovery is not hardware validation.
+function Invoke-D4RGpuPreflight([string]$PackageRoot, [string]$GameDirectory, [string]$Target, [string]$OutputDirectory) {
+    New-Item -ItemType Directory -Force $OutputDirectory | Out-Null
+    $result=[ordered]@{passed=$false; target=$Target; gpuWorkExecuted=$false; failure=$null}
+    try {
+        $proxy=Join-Path $GameDirectory 'd3d12.dll'
+        if (Test-Path -LiteralPath $proxy) {
+            $item=Get-Item -LiteralPath $proxy
+            $result.localD3D12=@{path=$proxy; sha256=(Get-FileHash -LiteralPath $proxy).Hash; product=$item.VersionInfo.ProductName}
+            # Identify known Vulkan/Wine proxies by their own strings. Do not
+            # load, remove or replace a third-party game DLL.
+            if ($item.Length -le 64MB -and [Text.Encoding]::ASCII.GetString([IO.File]::ReadAllBytes($proxy)) -match '(?i)vkd3d|wined3d') {
+                throw 'Local d3d12.dll identifies as vkd3d/Wine. Restore the game native D3D12 installation before using this Windows backend.'
+            }
+        }
+        $info=[Diagnostics.ProcessStartInfo]::new()
+        $info.FileName=Join-Path $PackageRoot 'd4r/d4r_gpu_inventory.exe'
+        $hip=Join-Path $PackageRoot 'd4r/hip'
+        foreach ($value in @($hip,$OutputDirectory)) { if ($value.Contains('"')) { throw 'Unsupported quote in inventory path.' } }
+        $info.Arguments='--hip-root "' + $hip + '" --output-dir "' + $OutputDirectory + '"'
+        $info.UseShellExecute=$false; $info.CreateNoWindow=$true
+        $info.EnvironmentVariables['D4R_DIAG_DIR']=$OutputDirectory
+        $info.RedirectStandardOutput=$true; $info.RedirectStandardError=$true
+        $process=[Diagnostics.Process]::new(); $process.StartInfo=$info
+        try {
+            if (!$process.Start()) { throw 'Cannot start GPU inventory.' }
+            $stdout=$process.StandardOutput.ReadToEndAsync(); $stderr=$process.StandardError.ReadToEndAsync()
+            $timedOut=!$process.WaitForExit(30000)
+            if ($timedOut) { $process.Kill(); $process.WaitForExit() }
+            $stdout.GetAwaiter().GetResult() | Set-Content (Join-Path $OutputDirectory 'gpu-inventory.stdout.log') -Encoding UTF8
+            $stderr.GetAwaiter().GetResult() | Set-Content (Join-Path $OutputDirectory 'gpu-inventory.stderr.log') -Encoding UTF8
+            $result.exitCode=$process.ExitCode
+            $result.exceptionOrExitHex='0x{0:X8}' -f [BitConverter]::ToUInt32([BitConverter]::GetBytes([int]$process.ExitCode),0)
+            if ($timedOut) { $result.timedOut=$true; throw 'HIP inventory timed out after 30 seconds.' }
+            if ($process.ExitCode) { throw "HIP inventory failed (exit $($process.ExitCode))." }
+        } finally { $process.Dispose() }
+        $inventory=Get-Content (Join-Path $OutputDirectory 'gpu-inventory.json') -Raw | ConvertFrom-Json
+        $result.inventory=$inventory
+        if ($inventory.compiledTarget -ne $Target) { throw "Inventory executable targets $($inventory.compiledTarget); package requires $Target." }
+        $matching=@($inventory.devices | Where-Object { $_.architecture -eq $Target -and $_.wavefrontSize -eq 32 })
+        if (!$matching.Count) {
+            $actual=@($inventory.devices | ForEach-Object { "$($_.architecture) ($($_.name))" }) -join ', '
+            throw "GPU target mismatch: detected $actual; package requires $Target. Use the package for the actual HIP architecture; do not rename code objects."
+        }
+        $result.passed=$true
+        Write-Host "D4R_PREFLIGHT target=$Target devices=$($matching.Count) gpu_work_executed=0"
+    } catch {
+        $result.failure=$_.Exception.Message
+        $result | ConvertTo-Json -Depth 8 | Set-Content (Join-Path $OutputDirectory 'preflight.json') -Encoding UTF8
+        try { Get-CimInstance Win32_VideoController | Select-Object Name,DriverVersion,DriverDate | ConvertTo-Json |
+            Set-Content (Join-Path $OutputDirectory 'driver.json') -Encoding UTF8 } catch {}
+        $bundle=$OutputDirectory+'.zip'
+        Compress-Archive -Path (Join-Path $OutputDirectory '*') -DestinationPath $bundle -Force
+        throw "$($result.failure) Diagnostic bundle: $bundle"
+    }
+    $result | ConvertTo-Json -Depth 8 | Set-Content (Join-Path $OutputDirectory 'preflight.json') -Encoding UTF8
+}

scripts/windows/gpu-target.ps1

@@ -1,19 +1,31 @@
 # Shared by the build, diagnostics, staging and game package scripts.
+function Get-D4RGpuTargets {
+    $registry=Join-Path $PSScriptRoot 'gpu-targets.json'
+    if (!(Test-Path -LiteralPath $registry)) { $registry=Join-Path $PSScriptRoot '../../tools/windows/gpu-targets.json' }
+    $data=Get-Content -LiteralPath $registry -Raw | ConvertFrom-Json
+    if ($data.schema -ne 1) { throw 'Unsupported GPU target registry schema.' }
+    return $data.targets
+}
+function Get-D4RGpuTargetInfo([string]$Target) {
+    $info=@(Get-D4RGpuTargets | Where-Object { $_.architecture -eq $Target })
+    if ($info.Count -ne 1) { throw "Unsupported GPU target: $Target" }
+    return $info[0]
+}
 function Get-D4RGpuTarget([string]$Root, [string]$Requested) {
     $configured=$null
     $targetFile=Join-Path $Root 'gpu-target.json'
     if (Test-Path -LiteralPath $targetFile) {
         $configured=(Get-Content -LiteralPath $targetFile -Raw | ConvertFrom-Json).architecture
-        if ($configured -notin @('gfx1200','gfx1201')) { throw "Unsupported package GPU target: $configured" }
+        [void](Get-D4RGpuTargetInfo $configured)
     }
     $target=if ($Requested) { $Requested } elseif ($configured) { $configured } else { 'gfx1201' }
-    if ($target -notin @('gfx1200','gfx1201')) { throw "Unsupported GPU target: $target" }
-    if (!$configured -and $target -ne 'gfx1201') { throw 'A gfx1200 build requires gpu-target.json; rebuild with -GpuArch gfx1200.' }
+    [void](Get-D4RGpuTargetInfo $target)
+    if (!$configured -and $target -ne 'gfx1201') { throw "A $target build requires gpu-target.json; rebuild with -GpuArch $target." }
     if ($configured -and $target -ne $configured) { throw "Requested $target differs from package target $configured" }
     return $target
 }
 function Assert-D4RCodeObjectTarget([string]$Path, [string]$Target) {
-    if ($Target -notin @('gfx1200','gfx1201')) { throw "Unsupported GPU target: $Target" }
+    $info=Get-D4RGpuTargetInfo $Target
     $reader=[IO.BinaryReader]::new([IO.File]::OpenRead($Path))
     try { $header=$reader.ReadBytes(64) } finally { $reader.Dispose() }
     if ($header.Length -ne 64 -or $header[0] -ne 127 -or $header[1] -ne 69 -or $header[2] -ne 76 -or $header[3] -ne 70 -or
@@ -21,6 +33,6 @@ function Assert-D4RCodeObjectTarget([string]$Path, [string]$Target) {
         throw "Invalid AMDGPU ELF64 code object: $Path"
     }
     # LLVM AMDGPU e_flags machine IDs; inspect, never patch or spoof them.
-    $expected=if ($Target -eq 'gfx1200') { 0x48 } else { 0x4e }
+    $expected=$info.elfMachine
     if ($header[48] -ne $expected) { throw "Code object does not target ${Target}: $Path (ELF machine=$($header[48]))" }
 }

scripts/windows/package-windows-game.ps1

@@ -2,7 +2,7 @@
 param(
     [string]$PackageRoot, [string]$DiagnosticRoot, [string]$OptiScalerRoot,
     [string]$ZludaRoot, [string]$HipRoot, [string]$KernelKRoot, [string]$KernelMRoot,
-    [ValidateSet('gfx1200','gfx1201')][string]$GpuArch,
+    [string]$GpuArch,
     [string]$ArchivePath
 )
 $ErrorActionPreference = 'Stop'
@@ -15,7 +15,7 @@ foreach ($name in $defaults.Keys) {
     if (!(Get-Variable -Name $name -ValueOnly)) { Set-Variable -Name $name -Value (Join-Path $repo $defaults[$name]) }
 }
 $GpuArch=Get-D4RGpuTarget $DiagnosticRoot $GpuArch
-if ($GpuArch -eq 'gfx1200' -and !$PSBoundParameters.ContainsKey('PackageRoot')) { $PackageRoot += '-gfx1200' }
+if ($GpuArch -ne 'gfx1201' -and !$PSBoundParameters.ContainsKey('PackageRoot')) { $PackageRoot += "-$GpuArch" }
 if (!$KernelKRoot) { $KernelKRoot=Join-Path $repo "build/native-k-$GpuArch" }
 if (!$KernelMRoot) { $KernelMRoot=Join-Path $repo "build/native-m-$GpuArch" }
 foreach ($directory in @($KernelKRoot,$KernelMRoot)) {
@@ -46,6 +46,7 @@ Stage (Join-Path $DiagnosticRoot 'bin/d4r_nvngx.dll') 'd4r/_nvngx.dll'
 Stage (Join-Path $DiagnosticRoot "bin/pixel_convert_${GpuArch}.hsaco") "d4r/pixel_convert_${GpuArch}.hsaco"
 Stage (Join-Path $DiagnosticRoot 'nvapi-compat/nvapi64.dll') 'd4r/nvapi/nvapi64.dll'
 Stage (Join-Path $DiagnosticRoot 'bin/d4r_debug_launcher.exe') 'd4r/d4r_debug_launcher.exe'
+Stage (Join-Path $DiagnosticRoot 'bin/d4r_gpu_inventory.exe') 'd4r/d4r_gpu_inventory.exe'
 foreach ($file in @('nvcuda.dll','nvapi64.dll','libc++.dll','libunwind.dll')) { Stage (Join-Path $ZludaRoot $file) "d4r/zluda/$file" }
 foreach ($file in @('amdhip64_7.dll','amd_comgr.dll','hiprtc0717.dll','hiprtc-builtins0717.dll','rocm_kpack.dll','AMD.ROCM.Comgr.MANIFEST','.hipVersion')) {
     Stage (Join-Path $HipRoot "bin/$file") "d4r/hip/bin/$file"
@@ -66,6 +67,8 @@ if ($entries.Count -ne 16) { throw "Expected 11 K and 5 M native objects; found
 $stagedFiles.Add('d4r/native/d4r-kernels.txt')
 Stage (Join-Path $repo 'scripts/windows/windows-game.ps1') 'windows-game.ps1'
 Stage (Join-Path $repo 'scripts/windows/gpu-target.ps1') 'gpu-target.ps1'
+Stage (Join-Path $repo 'scripts/windows/gpu-preflight.ps1') 'gpu-preflight.ps1'
+Stage (Join-Path $repo 'tools/windows/gpu-targets.json') 'gpu-targets.json'
 if (Test-Path -LiteralPath (Join-Path $DiagnosticRoot 'gpu-target.json')) {
     Stage (Join-Path $DiagnosticRoot 'gpu-target.json') 'gpu-target.json'
 } else {

scripts/windows/stage-native-k.ps1

@@ -1,5 +1,5 @@
 [CmdletBinding()]
-param([Parameter(Mandatory=$true)][string]$DlssDll, [string]$PackageRoot, [string]$OutputDirectory, [ValidateSet('gfx1200','gfx1201')][string]$GpuArch)
+param([Parameter(Mandatory=$true)][string]$DlssDll, [string]$PackageRoot, [string]$OutputDirectory, [string]$GpuArch)
 $ErrorActionPreference = 'Stop'
 . (Join-Path $PSScriptRoot 'gpu-target.ps1')
 $repo = [IO.Path]::GetFullPath((Join-Path $PSScriptRoot '../..'))

scripts/windows/stage-native-m.ps1

@@ -1,5 +1,5 @@
 [CmdletBinding()]
-param([Parameter(Mandatory=$true)][string]$DlssDll, [string]$PackageRoot, [string]$OutputDirectory, [ValidateSet('gfx1200','gfx1201')][string]$GpuArch)
+param([Parameter(Mandatory=$true)][string]$DlssDll, [string]$PackageRoot, [string]$OutputDirectory, [string]$GpuArch)
 $ErrorActionPreference = 'Stop'
 . (Join-Path $PSScriptRoot 'gpu-target.ps1')
 $repo = [IO.Path]::GetFullPath((Join-Path $PSScriptRoot '../..'))

scripts/windows/test-windows-rdna4.ps1

@@ -4,7 +4,7 @@ param(
     [ValidateSet('stable', 'therock')][string]$RuntimeProfile = 'stable',
     [string]$ZludaRoot,
     [string]$PackageRoot,
-    [ValidateSet('gfx1200','gfx1201')][string]$GpuArch,
+    [string]$GpuArch,
     [string]$OutputDirectory,
     [string]$NgxCore = $env:D4R_NGX_CORE,
     [string]$DlssDll = $env:D4R_DLSS_DLL,

scripts/windows/windows-game.ps1

@@ -27,6 +27,7 @@ param(
 )
 $ErrorActionPreference = 'Stop'
 . (Join-Path $PSScriptRoot 'gpu-target.ps1')
+. (Join-Path $PSScriptRoot 'gpu-preflight.ps1')
 if ($ProfileGpuBoundary -and !$AsyncInterop) { throw '-ProfileGpuBoundary requires -AsyncInterop.' }
 if ($ProfileKernels -and $ProfileKernelsDeferred) { throw 'Choose either serializing or deferred kernel profiling.' }
 if ($ProfileLegacyStream -and !$ProfileKernelsDeferred) { throw '-ProfileLegacyStream requires -ProfileKernelsDeferred.' }
@@ -74,6 +75,7 @@ foreach ($file in $metadata.files) {
     if ((Get-FileHash -LiteralPath (Join-Path $package $file.path) -Algorithm SHA256).Hash -ne $file.sha256) { throw "Package file changed; rebuild package: $($file.path)" }
 }
 if ((Get-FileHash -LiteralPath $DlssDll -Algorithm SHA256).Hash -ne $metadata.dlssSha256) { throw 'DLSS DLL does not match the validated 310.9.1 native manifest. Rebuild and validate native manifests for this DLL first.' }
+Invoke-D4RGpuPreflight $package $game $GpuArch (Join-Path $package ('results/preflight-' + (Get-Date -Format 'yyyyMMdd-HHmmss-fff')))
 $textureFiles=@(); $textureManifest=@()
 if ($LocalTextureKernels) {
     $LocalTextureKernels=(Get-Item -LiteralPath $LocalTextureKernels -ErrorAction Stop).FullName
@@ -193,6 +195,7 @@ $settings = @{
     # false have the same numerical effect; unset reuses the validated cache.
     D4R_ZLUDA_IGNORE_DENORMAL=$null; D4R_ZLUDA_FAST_MATH=$null;
     D4R_ZLUDA_WMMA_F32ACC=$null; D4R_ZLUDA_WAVE64=$null;
+    D4R_ZLUDA_WMMA_LAYOUT=$null;
     D4R_QUIET_API=$(if ($VerboseRuntime) { $null } else { '1' });
     D4R_VALIDATE_OUTPUT=$(if ($ValidateOutput) { '1' } else { $null });
     D4R_PROFILE_STAGES=$(if ($ProfileStages) { '1' } else { $null });

tools/windows/build_native_texture.py

@@ -1,4 +1,4 @@
-"""Build a private RDNA4 texture override from the user's local DLSS DLL.
+"""Build a private target-specific texture override from the user's local DLSS DLL.
 
 The resulting PTX/code object contains NVIDIA code and must remain local.
 No Wine, shell script, GPU, CUDA SDK or system installation is required.
@@ -13,13 +13,14 @@ import re
 import shutil
 import subprocess
 import sys
+from gpu_targets import TARGETS
 
 
 def require_code_object_target(path, target):
     with path.open('rb') as file:
         header = file.read(64)
     if (len(header) != 64 or header[:6] != b'\x7fELF\x02\x01' or
-            header[18:20] != b'\xe0\x00' or header[48] != {'gfx1200': 0x48, 'gfx1201': 0x4e}[target]):
+            header[18:20] != b'\xe0\x00' or header[48] != TARGETS[target]['elfMachine']):
         raise RuntimeError(f'Code object does not target {target}: {path}')
 
 
@@ -30,7 +31,7 @@ def main():
     parser.add_argument('--zluda-root', type=pathlib.Path, required=True)
     parser.add_argument('--kernel', required=True)
     parser.add_argument('--output-directory', type=pathlib.Path, required=True)
-    parser.add_argument('--gpu-arch', choices=['gfx1200', 'gfx1201'], default='gfx1201')
+    parser.add_argument('--gpu-arch', choices=tuple(TARGETS), default='gfx1201')
     parser.add_argument('--shader-mode', choices=['cu', 'wgp'], default='cu')
     args = parser.parse_args()
     if not re.fullmatch(r'hiluma_engine_(?:output_depth(?:inv|reg)_mv(?:hi|lo)_(?:hdr|ldr)(?:_max)?|'

tools/windows/d4r_nvngx_windows.cpp

@@ -87,7 +87,7 @@ API unsigned NVSDK_NGX_D3D12_GetFeatureRequirements(IDXGIAdapter* adapter, const
         DXGI_ADAPTER_DESC desc{}; d4r::win::dx(adapter->GetDesc(&desc), "NGX requirements adapter");
         // OptiScaler can override the DXGI vendor ID. The physical HIP
         // architecture and full adapter LUID are the authoritative identity.
-        d4r::diag::HipApi hip(d4r::win::env_path("D4R_HIP_ROOT")); hipDeviceProp_t props{}; hip.select_architecture(-1,props);
+        d4r::diag::HipApi hip(d4r::win::env_path("D4R_HIP_ROOT")); hipDeviceProp_t props{}; hip.select_architecture(-1,props,&desc.AdapterLuid);
         requirements->FeatureSupported = std::memcmp(&desc.AdapterLuid,props.luid,sizeof(LUID)) ? 4u : 0u;
         std::strcpy(requirements->MinOSVersion,"10.0.22000.0");
         std::printf("D4R_REQUIREMENTS feature=%u support=%u architecture=%s\n",info->FeatureID,requirements->FeatureSupported,d4r::diag::HipApi::target_arch());

tools/windows/gpu-targets.json

@@ -0,0 +1,16 @@
+{
+  "schema": 1,
+  "targets": [
+    {"architecture":"gfx1100","elfMachine":65,"isaMajor":11,"family":"RDNA3"},
+    {"architecture":"gfx1101","elfMachine":70,"isaMajor":11,"family":"RDNA3"},
+    {"architecture":"gfx1102","elfMachine":71,"isaMajor":11,"family":"RDNA3"},
+    {"architecture":"gfx1103","elfMachine":68,"isaMajor":11,"family":"RDNA3"},
+    {"architecture":"gfx1150","elfMachine":67,"isaMajor":11,"family":"RDNA3.5"},
+    {"architecture":"gfx1151","elfMachine":74,"isaMajor":11,"family":"RDNA3.5"},
+    {"architecture":"gfx1152","elfMachine":85,"isaMajor":11,"family":"RDNA3.5"},
+    {"architecture":"gfx1153","elfMachine":88,"isaMajor":11,"family":"RDNA3.5"},
+    {"architecture":"gfx1154","elfMachine":87,"isaMajor":11,"family":"RDNA3.5"},
+    {"architecture":"gfx1200","elfMachine":72,"isaMajor":12,"family":"RDNA4"},
+    {"architecture":"gfx1201","elfMachine":78,"isaMajor":12,"family":"RDNA4"}
+  ]
+}

tools/windows/gpu_inventory.cpp

@@ -0,0 +1,57 @@
+#include "hip_api.h"
+#include <fstream>
+#include <sstream>
+#include <iomanip>
+
+static std::string json_string(const std::string& value) {
+    std::ostringstream out; out << '"';
+    for (unsigned char c : value) {
+        if (c == '"' || c == '\\') out << '\\' << char(c);
+        else if (c < 32) out << "\\u" << std::hex << std::setw(4) << std::setfill('0') << unsigned(c);
+        else out << char(c);
+    }
+    out << '"'; return out.str();
+}
+int main(int argc, char** argv) {
+    using namespace d4r::diag;
+    start();
+    try {
+        Args args(argc, argv);
+        if (args.output_dir.empty()) throw std::runtime_error("--output-dir is required");
+        for (const char* name : {"HSA_OVERRIDE_GFX_VERSION", "HSA_OVERRIDE_GFX_VERSION_0"})
+            if (std::getenv(name)) throw std::runtime_error(std::string("Unset architecture override ") + name);
+        HipApi hip(args.hip_root);
+        hip.check(hip.hipInit(0), "hipInit(inventory)");
+        int count=0, version=0, driver=0;
+        hip.check(hip.hipRuntimeGetVersion(&version), "hipRuntimeGetVersion");
+        hip.check(hip.hipDriverGetVersion(&driver), "hipDriverGetVersion");
+        std::printf("GPU_INVENTORY_RUNTIME runtime=%d driver_api=%d headers=%d.%d.%d sizeof(properties)=%zu\n",
+                    version,driver,HIP_VERSION_MAJOR,HIP_VERSION_MINOR,HIP_VERSION_PATCH,sizeof(hipDeviceProp_t));
+        hip.check(hip.hipGetDeviceCount(&count), "hipGetDeviceCount");
+        std::ostringstream json;
+        json << "{\"schema\":1,\"runtimeVersion\":" << version << ",\"driverApiVersion\":" << driver << ",\"compiledTarget\":"
+             << json_string(D4R_TARGET_ARCH) << ",\"gpuWorkExecuted\":false,\"devices\":[";
+        for (int i=0; i<count; ++i) {
+            hipDeviceProp_t p{}; hip.check(hip.hipGetDeviceProperties(&p,i), "hipGetDevicePropertiesR0600");
+            const auto arch=gpu_architecture(p.gcnArchName);
+            std::printf("GPU_INVENTORY ordinal=%d name=%s gcnArchName=%s wave=%d build_eligible=%d\n",
+                        i,p.name,p.gcnArchName,p.warpSize,supported_gpu_target(arch));
+            if (i) json << ',';
+            json << "{\"ordinal\":" << i << ",\"name\":" << json_string(p.name)
+                 << ",\"gcnArchName\":" << json_string(p.gcnArchName) << ",\"architecture\":" << json_string(arch)
+                 << ",\"wavefrontSize\":" << p.warpSize << ",\"vramBytes\":" << p.totalGlobalMem
+                 << ",\"buildEligible\":" << (supported_gpu_target(arch) ? "true" : "false") << '}';
+        }
+        json << "]}\n";
+        std::filesystem::create_directories(wide(args.output_dir));
+        std::ofstream file(std::filesystem::path(wide(args.output_dir))/L"gpu-inventory.json",std::ios::binary);
+        file << json.str(); file.close();
+        if (!file) throw std::runtime_error("Cannot write GPU inventory");
+        if (!count) throw std::runtime_error("HIP reports no devices");
+        loaded_modules();
+        std::printf("PASS GPU_INVENTORY devices=%d gpu_work_executed=0\n",count);
+        return 0;
+    } catch (const std::exception& error) {
+        std::fprintf(stderr,"FAIL GPU_INVENTORY %s\n",error.what()); loaded_modules(); return 4;
+    }
+}

tools/windows/gpu_target.h

@@ -4,6 +4,7 @@
 #include <fstream>
 #include <stdexcept>
 #include <string>
+#include "gpu_targets.generated.h"
 
 #ifndef D4R_TARGET_ARCH
 #define D4R_TARGET_ARCH "gfx1201"
@@ -11,7 +12,8 @@
 
 namespace d4r::diag {
 inline bool supported_gpu_target(const std::string& target) {
-    return target == "gfx1200" || target == "gfx1201";
+    for (const auto& row : gpu_targets) if (target == row.architecture) return true;
+    return false;
 }
 inline std::string gpu_architecture(const std::string& name) {
     return name.substr(0, name.find(':'));
@@ -29,11 +31,8 @@ inline std::string code_object_target(const std::filesystem::path& path) {
         header[0] != 0x7f || header[1] != 'E' || header[2] != 'L' || header[3] != 'F' ||
         header[4] != 2 || header[5] != 1 || header[18] != 0xe0 || header[19] != 0)
         throw std::runtime_error("Invalid AMDGPU ELF64 code object: " + path.string());
-    switch (header[48]) {
-        case 0x48: return "gfx1200";
-        case 0x4e: return "gfx1201";
-        default: throw std::runtime_error("Unsupported AMDGPU code-object target: " + path.string());
-    }
+    for (const auto& row : gpu_targets) if (header[48] == row.elf_machine) return row.architecture;
+    throw std::runtime_error("Unsupported AMDGPU code-object target: " + path.string());
 }
 inline void require_code_object_target(const std::filesystem::path& path, const std::string& expected) {
     require_gpu_target(code_object_target(path), expected);

tools/windows/gpu_target_probe.cpp

@@ -5,10 +5,11 @@ int main(int argc, char** argv) {
     using namespace d4r::diag;
     try {
         unsigned rejected = 0, objects = 0;
-        for (const char* target : {"gfx1200", "gfx1201"}) {
+        for (const auto& info : gpu_targets) {
+            const char* target = info.architecture;
             require_gpu_target(target, target);
             require_gpu_target(std::string(target) + ":xnack-:sramecc-", target);
-            for (const char* invalid : {"gfx1101", "gfx12000", "gfx1202", "gfx1250", ""}) {
+            for (const char* invalid : {"gfx1030", "gfx1104", "gfx1155", "gfx1170", "gfx12000", "gfx1202", "gfx1250", ""}) {
                 try { require_gpu_target(invalid, target); }
                 catch (const std::exception&) { ++rejected; continue; }
                 throw std::runtime_error("Invalid GPU target accepted");

tools/windows/gpu_targets.generated.h.in

@@ -0,0 +1,7 @@
+#pragma once
+#include <array>
+namespace d4r::diag {
+struct GpuTarget { const char* architecture; unsigned elf_machine; unsigned isa_major; const char* family; };
+inline constexpr std::array<GpuTarget, @d4r_target_count@> gpu_targets = {{
+@d4r_target_cpp_entries@}};
+}

tools/windows/gpu_targets.py

@@ -0,0 +1,10 @@
+"""Canonical Windows build target registry, shared with CMake and PowerShell."""
+import json
+import pathlib
+
+data = json.loads(pathlib.Path(__file__).with_name('gpu-targets.json').read_text())
+if data['schema'] != 1:
+    raise RuntimeError('Unsupported GPU target registry schema')
+TARGETS = {row['architecture']: row for row in data['targets']}
+if len(TARGETS) != len(data['targets']):
+    raise RuntimeError('Duplicate GPU target')

tools/windows/hip_api.h

@@ -42,7 +42,7 @@ struct HipApi {
         if (result != hipSuccess) throw std::runtime_error(std::string(call) + ": " + hipGetErrorString(result));
     }
     static const char* target_arch() { return D4R_TARGET_ARCH; }
-    int select_architecture(int requested, hipDeviceProp_t& selected) const {
+    int select_architecture(int requested, hipDeviceProp_t& selected, const void* required_luid = nullptr) const {
         // Architecture spoofing invalidates a gfx12 correctness test.
         for (const char* name : {"HSA_OVERRIDE_GFX_VERSION", "HSA_OVERRIDE_GFX_VERSION_0"})
             if (std::getenv(name)) throw std::runtime_error(std::string("Unset architecture override ") + name);
@@ -61,12 +61,14 @@ struct HipApi {
                 i, props.name, props.gcnArchName, props.warpSize, props.totalGlobalMem,
                 props.pciDomainID, props.pciBusID, props.pciDeviceID);
             const std::string arch = props.gcnArchName;
-            if (gpu_architecture(arch) == target_arch() && (requested < 0 || requested == i) && found < 0) {
+            if (gpu_architecture(arch) == target_arch() && (requested < 0 || requested == i) && found < 0 &&
+                (!required_luid || std::memcmp(props.luid, required_luid, sizeof(props.luid)) == 0)) {
                 found = i;
                 selected = props;
             }
         }
-        if (found < 0) throw std::runtime_error(std::string("No selected ") + target_arch() + " device; refusing architecture substitution");
+        if (found < 0) throw std::runtime_error(std::string("No selected ") + target_arch() +
+            " device for this adapter; use the package matching the actual HIP architecture (see GPU log)");
         require_gpu_target(selected.gcnArchName, target_arch());
         check(hipSetDevice(found), "hipSetDevice");
         std::printf("SELECTED HIP ordinal=%d architecture=%s validation=%s\n", found, target_arch(),

tools/windows/k_layer_validate.py

@@ -1,5 +1,5 @@
 #!/usr/bin/env python3
-"""Compare RDNA4 K enc1/enc2 launches with the existing numpy reference."""
+"""Compare native K enc1/enc2 launches with the existing numpy reference."""
 import argparse
 import math
 import pathlib
@@ -73,7 +73,7 @@ def main():
     # after each K stage has been compared on captured DLSS weights/activations.
     if full_abs > 0.001 or full_psnr < 60.0 or merged_abs > 0.001 or merged_psnr < 60.0:
         raise RuntimeError(f"K {args.kernel_name} numerical mismatch against pwin_model.py")
-    print(f"PASS K_REFERENCE reference_layout=gfx12 kernel={args.kernel_name} "
+    print(f"PASS K_REFERENCE reference=logical kernel={args.kernel_name} "
           "nonzero_weights=1 transformer_executed=1 nan_inf=0")
 
 

tools/windows/k_replay_validate.py

@@ -139,7 +139,7 @@ def validate(layer, directory, output, window_count=1, real_capture=False):
     if maximum > absolute_limit or psnr < 60:
         worst = int(difference.argmax())
         raise RuntimeError(f'K mismatch at {worst}: reference={expected[worst]}, gpu={actual[worst]}')
-    print(f'PASS K_REPLAY_REFERENCE kernel={layer} reference_layout=gfx12 nan_inf=0')
+    print(f'PASS K_REPLAY_REFERENCE kernel={layer} reference=logical nan_inf=0')
 
 
 if __name__ == '__main__':

tools/windows/m_replay_validate.py

@@ -153,7 +153,7 @@ def validate(layer, directory, output, block_count, real_capture, exact=False):
     maximum, peak, psnr, steps = metrics(expected, actual, f'M_REPLAY_REFERENCE kernel={layer}')
     if ((exact or not real_capture) and maximum != 0) or (real_capture and (steps.max() > 1 or psnr < 50)):
         raise RuntimeError(f'M mismatch: exact synthetic / <=1 FP8 step and >=50 dB real gate; peak={peak}')
-    print(f'PASS M_REPLAY_REFERENCE kernel={layer} blocks={len(pairs)} reference_layout=gfx12 nan_inf=0')
+    print(f'PASS M_REPLAY_REFERENCE kernel={layer} blocks={len(pairs)} reference=logical nan_inf=0')
 
 
 if __name__ == '__main__':

tools/windows/ngx_windows_runtime.h

@@ -94,9 +94,9 @@ struct Runtime {
         if (!d3d) throw std::runtime_error("Null D3D12 device");
         try {
             hipDeviceProp_t props{};
-            hip.select_architecture(-1, props);
-            hip.verbose = cuda.verbose = !std::getenv("D4R_QUIET_API");
             const auto luid = adapter_luid(d3d);
+            hip.select_architecture(-1, props, &luid);
+            hip.verbose = cuda.verbose = !std::getenv("D4R_QUIET_API");
             if (std::memcmp(&luid, props.luid, sizeof(luid))) throw std::runtime_error("D3D12/HIP LUID mismatch");
             cuda.check(cuda.cuInit(0), "cuInit(Windows runtime)");
             int count = 0; cuda.check(cuda.cuDeviceGetCount(&count), "cuDeviceGetCount");

tools/windows/wmma_gfx1201.hip

@@ -25,6 +25,7 @@ extern "C" __attribute__((global)) void d4r_wmma_gfx1201(
     const unsigned int group = lane >> 4;
     const _Float16* a = reinterpret_cast<const _Float16*>(a_bits);
     const _Float16* b = reinterpret_cast<const _Float16*>(b_bits);
+#if defined(__GFX12__)
     h8 a_frag, b_frag;
     f8 c_frag;
     for (unsigned int i = 0; i < 8u; ++i) {
@@ -39,6 +40,23 @@ extern "C" __attribute__((global)) void d4r_wmma_gfx1201(
             b_frag, a_frag, result);
     for (unsigned int i = 0; i < 8u; ++i)
         output[row * 16u + group * 8u + i] = result[i];
+#else
+    // Adapted from realdody/d4r ae530b0: real gfx11 operands/accumulators,
+    // independent of the shared layout abstraction.
+    d4r_wmma_h16 a_frag, b_frag;
+    d4r_wmma_f8 c_frag;
+    for (unsigned int i = 0; i < 16u; ++i) {
+        a_frag[i] = a[row * 16u + i];
+        b_frag[i] = b[row * 16u + i];
+    }
+    for (unsigned int i = 0; i < 8u; ++i)
+        c_frag[i] = initial_c[(2u * i + group) * 16u + row];
+    d4r_wmma_f8 result = __builtin_amdgcn_wmma_f32_16x16x16_f16_w32(a_frag, b_frag, c_frag);
+    if (steps == 2u)
+        result = __builtin_amdgcn_wmma_f32_16x16x16_f16_w32(a_frag, b_frag, result);
+    for (unsigned int i = 0; i < 8u; ++i)
+        output[(2u * i + group) * 16u + row] = result[i];
+#endif
 }
 
 // d4r's original gfx11 fragment contract, exercised independently of K/M.