Initialize Windows NGX CUDA with HIP-backed NVAPI topology

CMakeLists.txt

@@ -21,7 +21,7 @@ add_custom_command(OUTPUT "${probe_object}"
     -o "${probe_object}"
   DEPENDS tools/windows/probe_kernel.hip VERBATIM)
 add_custom_target(d4r_probe_code_object ALL DEPENDS "${probe_object}")
-foreach(probe hip_gfx1201_probe cuda_driver_probe d3d12_hip_interop_probe hip_stream_lifecycle_probe)
+foreach(probe hip_gfx1201_probe cuda_driver_probe d3d12_hip_interop_probe hip_stream_lifecycle_probe ngx_cuda_init_probe)
   add_executable(d4r_${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")
@@ -37,6 +37,39 @@ foreach(probe hip_gfx1201_probe cuda_driver_probe d3d12_hip_interop_probe hip_st
   endif()
   install(TARGETS d4r_${probe} RUNTIME DESTINATION bin)
 endforeach()
+if(MSVC)
+  target_sources(d4r_ngx_cuda_init_probe PRIVATE tools/d4r_ngx_param_msvc.cpp tools/d4r_ngx_param_host.cpp)
+else()
+  find_program(D4R_CLANG_CL NAMES clang-cl.exe PATHS "${D4R_HIP_ROOT}/bin" "${D4R_HIP_ROOT}/lib/llvm/bin" NO_DEFAULT_PATH REQUIRED)
+  set(ngx_parameter_object "${CMAKE_CURRENT_BINARY_DIR}/ngx_parameter_msvc.obj")
+  add_custom_command(OUTPUT "${ngx_parameter_object}"
+    COMMAND "${D4R_CLANG_CL}" /c /O2 /Zl /GS- /GR- /EHs-c-
+      "${CMAKE_CURRENT_SOURCE_DIR}/tools/d4r_ngx_param_msvc.cpp" "/Fo${ngx_parameter_object}"
+    DEPENDS tools/d4r_ngx_param_msvc.cpp VERBATIM)
+  set_source_files_properties("${ngx_parameter_object}" PROPERTIES GENERATED TRUE EXTERNAL_OBJECT TRUE)
+  target_sources(d4r_ngx_cuda_init_probe PRIVATE "${ngx_parameter_object}" tools/d4r_ngx_param_host.cpp)
+endif()
+add_library(d4r_nvapi_trace SHARED tools/windows/nvapi_trace.cpp)
+set_target_properties(d4r_nvapi_trace PROPERTIES PREFIX "" OUTPUT_NAME "nvapi64"
+  ARCHIVE_OUTPUT_DIRECTORY "${CMAKE_CURRENT_BINARY_DIR}/nvapi-trace"
+  RUNTIME_OUTPUT_DIRECTORY "${CMAKE_CURRENT_BINARY_DIR}/nvapi-trace")
+target_compile_definitions(d4r_nvapi_trace PRIVATE WIN32_LEAN_AND_MEAN NOMINMAX)
+target_link_libraries(d4r_nvapi_trace PRIVATE version dbghelp)
+if(MINGW)
+  target_link_options(d4r_nvapi_trace PRIVATE -static)
+endif()
+install(TARGETS d4r_nvapi_trace RUNTIME DESTINATION nvapi-trace)
+add_library(d4r_nvapi_compat SHARED tools/windows/nvapi_trace.cpp)
+set_target_properties(d4r_nvapi_compat PROPERTIES PREFIX "" OUTPUT_NAME "nvapi64"
+  ARCHIVE_OUTPUT_DIRECTORY "${CMAKE_CURRENT_BINARY_DIR}/nvapi-compat"
+  RUNTIME_OUTPUT_DIRECTORY "${CMAKE_CURRENT_BINARY_DIR}/nvapi-compat")
+target_compile_definitions(d4r_nvapi_compat PRIVATE WIN32_LEAN_AND_MEAN NOMINMAX __HIP_PLATFORM_AMD__ D4R_NVAPI_COMPAT)
+target_include_directories(d4r_nvapi_compat PRIVATE "${D4R_HIP_ROOT}/include")
+target_link_libraries(d4r_nvapi_compat PRIVATE version dbghelp)
+if(MINGW)
+  target_link_options(d4r_nvapi_compat PRIVATE -static)
+endif()
+install(TARGETS d4r_nvapi_compat RUNTIME DESTINATION nvapi-compat)
 target_link_libraries(d4r_d3d12_hip_interop_probe PRIVATE d3d12 dxgi dxguid)
 add_dependencies(d4r_hip_gfx1201_probe d4r_probe_code_object)
 install(FILES "${probe_object}" DESTINATION bin)

README.md

@@ -7,7 +7,8 @@ d4r runs NVIDIA's official DLSS Super Resolution library (`nvngx_dlss.dll`) in W
 It is an experiment. It has been tested on one GPU in a handful of games, it depends on unreleased patches to ZLUDA and vkd3d-proton.
 
 **Windows / RDNA4 development:** the `windows-rdna4` branch adds native Windows
-gfx1201 HIP and ZLUDA diagnostics. Build commands, hardware results and remaining
+gfx1201 HIP/ZLUDA diagnostics, validated D3D12/HIP interop and an NGX init probe.
+Build commands, hardware results and remaining
 porting gates are in [docs/windows-rdna4-port.md](docs/windows-rdna4-port.md).
 This branch does not yet provide a Windows DLSS runtime.
 

docs/windows-rdna4-port.md

@@ -6,8 +6,8 @@ This is a work log for an implementation in progress, not a claim of DLSS suppor
 
 ## Milestone and gates
 
-Current milestone: M1/M2 and M4 passed on the physical RX 9070 XT with the
+Current milestone: M1/M2/M4 passed; NGX init + DLSS CUDA capability discovery
-isolated TheRock profile. Next: NGX initialization and validated gfx12 WMMA.
+passed with the user's DLLs. Next: validated gfx12 WMMA and native K kernels.
 Do not integrate NGX until the integer PTX workload is stable on the real GPU.
 
 | Gate | Status |
@@ -15,7 +15,7 @@ Do not integrate NGX until the integer PTX workload is stable on the real GPU.
 | M0: branch, audit, Windows build and diagnostics | Initial implementation complete |
 | M1: native HIP gfx1201 allocation, kernel, CPU verification | PASS, 32 iterations + guard verification |
 | M2: Windows ZLUDA integer PTX, primary and created contexts | PASS, 32 iterations each + guard verification |
-| M3: Windows NGX initialization and simple DLSS path | Pending |
+| M3: Windows NGX initialization and simple DLSS path | Init + SR capabilities PASS; Evaluate/transformer pending |
 | M4: D3D12 / HIP external memory and fence round trip | PASS with TheRock; stable 7.2 has mapped-view leak |
 | M5: independently validated gfx12 WMMA backend | Pending |
 | M6: K layers, full transformer and image validation | Pending |
@@ -148,6 +148,43 @@ the 64-cycle map lifetime test and the 32-cycle synchronized texture round trip.
 The runner automatically isolates resource/import/stream lifetimes on failure.
 Do not declare stable SDK 7.2 acceptable for the leak-free fast path.
 
+## Windows NGX initialization
+
+`tools/windows/ngx_cuda_init_probe.cpp` retains the physical HIP/PCI device
+check, initializes a ZLUDA primary context, loads the user-supplied DLLs,
+discovers DLSS CUDA capability and tests the existing MSVC parameter accessor.
+This does not create/evaluate a DLSS feature or assert transformer correctness.
+
+The supplied driver `_nvngx.dll` 32.0.16.1714 uses the **driver/snippet**
+three-argument `NVSDK_NGX_CUDA_Init(appId, dataPath, sdkVersion)` ABI. The
+four-argument SDK client interface used by the upstream shim passes a pointer
+as the version and returns `0xBAD0000C` (OutOfDate) for this binary. The two
+interfaces are distinguished in NVIDIA's public `nvsdk_ngx.h` by
+`NGX_SNIPPET_BUILD`. The Windows backend must keep this distinction. The probe
+has explicit `--ngx-abi driver` (default) / `sdk`; it does not guess and retry.
+
+Upstream Windows ZLUDA NVAPI lacks physical GPU enumeration, architecture and
+logical GPU/LUID functions used by NGX. The compatibility target
+`nvapi-compat/nvapi64.dll` implements those **public NVAPI** APIs through HIP;
+unknown queries are logged and forwarded to the supplied ZLUDA NVAPI backend.
+It validates version/size and includes `NV_LOGICAL_GPU_DATA_V1.reserved[8]`
+(568-byte x64 ABI). Concurrent initialization is synchronized. The AD102/AD100
+identity is explicitly a CUDA/NGX network profile matching d4r's sm_89 kernels;
+physical detection and native compilation remain gfx1201. It is currently a
+probe backend, not a complete game-wide NVAPI replacement. A separate
+`nvapi-trace` build forwards every response unchanged for diagnosis.
+
+Driver init's default feature search needs the DLSS DLL next to the executable.
+The runner creates a unique private `%TEMP%/d4r-ngx-<GUID>` directory containing
+the probe and copies of the two supplied DLLs. Paths and original hashes/versions
+are logged; the directory is removed after the subprocess ends. Neither DLL is
+placed in dist/, git, or the diagnostics ZIP. No private NVAPI function is
+invented and no NVIDIA binary is patched.
+
+References: [NGX API signatures](https://github.com/NVIDIA/DLSS/blob/main/include/nvsdk_ngx.h),
+[public NVAPI ABI](https://github.com/NVIDIA/nvapi/blob/main/nvapi.h),
+[public NVAPI interface IDs](https://github.com/NVIDIA/nvapi/blob/main/nvapi_interface.h).
+
 ## Authoritative references
 
 * [AMD HIP SDK Windows support matrix](https://rocm.docs.amd.com/projects/install-on-windows/en/latest/reference/system-requirements.html)
@@ -186,6 +223,16 @@ This builds into `build/windows-rdna4-therock` and installs into
 `dist/windows-rdna4-therock`. The stable profile stays available for reproducing
 the runtime bug. It is expected to return failure at the lifetime gate.
 
+One command adds the local NGX initialization gate after the five hardware gates:
+
+```powershell
+powershell -NoProfile -File scripts/windows/test-windows-rdna4.ps1 -RuntimeProfile therock -NgxCore "C:/Users/Administrator/d4r/_nvngx.dll" -DlssDll "C:/Users/Administrator/d4r/nvngx_dlss.dll"
+```
+
+Success requires all six tests and `PASS NGX_INIT ... sr_available=1`;
+`transformer_executed=0` is explicit. On NGX failure, CUDA trace and NVAPI query
+logs are collected automatically. NVIDIA DLLs remain local test inputs.
+
 The build accepts `-HipRoot`, `-ZludaRoot`, `-ToolchainRoot`, `-BuildDirectory`
 and `-InstallDirectory`. MSVC users can invoke CMake from a Developer Shell
 without the portable MinGW compiler. Equivalent direct build:
@@ -242,6 +289,12 @@ stable runtime mapped-view bug; the fix is the runtime ownership correction.
 Local proprietary inputs supplied by the user: `nvngx_dlss.dll` 310.9.1.0,
 `_nvngx.dll` 32.0.16.1714. Neither is packaged or tracked.
 
+2026-09-29 NGX initialization: PASS on gfx1201 with those two DLLs and the native
+NVAPI topology backend. SR Available=1, FeatureInitResult=0x1, NeedsUpdatedDriver=0,
+MSVC parameter roundtrip Width=640, checked parameter/context/NGX cleanup.
+Local logs: `test-results/ngx-init-private-runtime.zip`. This is only the first
+part of M3; feature execution, K/M correctness and the Windows shim remain open.
+
 Audit found an upstream validation blocker: `swin_model.py` imports
 `model_enc3`, which is absent from the repository. M reference validation needs
 that dependency restored or an independently checked replacement.

scripts/windows/test-windows-rdna4.ps1

@@ -5,6 +5,8 @@ param(
     [string]$ZludaRoot,
     [string]$PackageRoot,
     [string]$OutputDirectory,
+    [string]$NgxCore = $env:D4R_NGX_CORE,
+    [string]$DlssDll = $env:D4R_DLSS_DLL,
     [int]$Iterations = 32,
     [int]$TimeoutSeconds = 180
 )
@@ -39,7 +41,9 @@ $originalHip = $env:HIP_PATH
 $originalDiag = $env:D4R_DIAG_DIR
 $originalLog = $env:ZLUDA_LOG_DIR
 $originalCuda = $env:ZLUDA_CUDA_LIB
+$originalNvapi = $env:D4R_NVAPI_BACKEND
 $exitStatus = 1
+$ngxRuntimeDirectory = $null
 $summary = [ordered]@{utc=[DateTime]::UtcNow.ToString('o'); profile=$RuntimeProfile; hipRoot=$HipRoot; zludaRoot=$ZludaRoot; tests=@()}
 function Quote-Argument([string]$Value) {
     # All generated paths are absolute file/directory paths, with no trailing backslash.
@@ -86,10 +90,18 @@ try {
     $env:D4R_DIAG_DIR = $OutputDirectory
     $env:ZLUDA_LOG_DIR = Join-Path $OutputDirectory 'zluda-trace'
     $env:ZLUDA_CUDA_LIB = Join-Path $ZludaRoot 'nvcuda.dll'
+    $env:D4R_NVAPI_BACKEND = Join-Path $ZludaRoot 'nvapi64.dll'
     $comgr = Join-Path $HipRoot 'bin/amd_comgr_3.dll'
     if (!(Test-Path -LiteralPath $comgr)) { $comgr = Join-Path $HipRoot 'bin/amd_comgr.dll' }
     $files = @((Join-Path $HipRoot 'bin/amdhip64_7.dll'), $comgr,
         (Join-Path $ZludaRoot 'nvcuda.dll'), (Join-Path $PackageRoot 'bin/probe_gfx1201.hsaco'))
+    if ($NgxCore -or $DlssDll) {
+        if (!$NgxCore -or !$DlssDll) { throw 'Supply both -NgxCore and -DlssDll absolute paths.' }
+        if (![IO.Path]::IsPathRooted($NgxCore) -or ![IO.Path]::IsPathRooted($DlssDll)) {
+            throw 'NVIDIA DLL paths must be absolute.'
+        }
+        $files += @($NgxCore, $DlssDll)
+    }
     $summary.files = @($files | ForEach-Object {
         if (!(Test-Path -LiteralPath $_)) { throw "Required file missing: $_" }
         $file = Get-Item -LiteralPath $_
@@ -132,12 +144,46 @@ try {
                     '--hip-root', $HipRoot, '--module', (Join-Path $bin 'probe_gfx1201.hsaco'), '--iterations', '32') | Out-Null
             }
             if ($mapOk -and $interopOk) { $exitStatus = 0 }
+            if ($exitStatus -eq 0 -and $NgxCore) {
+                # The driver Init ABI searches the executable directory. Keep the
+                # supplied binaries in a private temporary runtime, outside dist/ZIP.
+                $ngxRuntimeDirectory = Join-Path ([IO.Path]::GetTempPath()) ('d4r-ngx-' + [Guid]::NewGuid().ToString('N'))
+                New-Item -ItemType Directory -Path $ngxRuntimeDirectory | Out-Null
+                $ngxExe = Join-Path $ngxRuntimeDirectory 'd4r_ngx_cuda_init_probe.exe'
+                $localCore = Join-Path $ngxRuntimeDirectory '_nvngx.dll'
+                $localDlss = Join-Path $ngxRuntimeDirectory 'nvngx_dlss.dll'
+                Copy-Item -LiteralPath (Join-Path $bin 'd4r_ngx_cuda_init_probe.exe') -Destination $ngxExe
+                Copy-Item -LiteralPath $NgxCore -Destination $localCore
+                Copy-Item -LiteralPath $DlssDll -Destination $localDlss
+                $ngxArguments = @('--hip-root', $HipRoot, '--cuda-dll', (Join-Path $ZludaRoot 'nvcuda.dll'),
+                    '--ngx-core', $localCore, '--dlss-dll', $localDlss,
+                    '--nvapi-dll', (Join-Path $PackageRoot 'nvapi-compat/nvapi64.dll'))
+                $ngxOk = Invoke-Probe 'ngx-init' $ngxExe $ngxArguments
+                if (!$ngxOk) {
+                    $exitStatus = 1
+                    if (Test-Path (Join-Path $ZludaRoot 'trace/nvcuda.dll')) {
+                        Invoke-Probe 'ngx-init-trace' $ngxExe @(
+                            '--hip-root', $HipRoot, '--cuda-dll', (Join-Path $ZludaRoot 'trace/nvcuda.dll'),
+                            '--ngx-core', $localCore, '--dlss-dll', $localDlss,
+                            '--nvapi-dll', (Join-Path $PackageRoot 'nvapi-compat/nvapi64.dll')) | Out-Null
+                    }
+                }
+            }
         }
     }
 } catch {
     $summary.error = $_.Exception.Message
     Write-Warning $summary.error
 } finally {
+    if ($ngxRuntimeDirectory -and (Test-Path -LiteralPath $ngxRuntimeDirectory)) {
+        $resolvedRuntime = [IO.Path]::GetFullPath($ngxRuntimeDirectory)
+        $tempRoot = [IO.Path]::GetFullPath([IO.Path]::GetTempPath()).TrimEnd('\') + '\'
+        if (!$resolvedRuntime.StartsWith($tempRoot, [StringComparison]::OrdinalIgnoreCase) -or
+            [IO.Path]::GetFileName($resolvedRuntime) -notmatch '^d4r-ngx-[0-9a-f]{32}$') {
+            throw 'Refusing to remove a path outside the private NGX temporary runtime.'
+        }
+        Remove-Item -LiteralPath $resolvedRuntime -Recurse -Force
+    }
     $summary.passed = $exitStatus -eq 0
     $summary | ConvertTo-Json -Depth 8 | Set-Content -LiteralPath (Join-Path $OutputDirectory 'summary.json') -Encoding UTF8
     $env:PATH = $originalPath
@@ -145,6 +191,7 @@ try {
     $env:D4R_DIAG_DIR = $originalDiag
     $env:ZLUDA_LOG_DIR = $originalLog
     $env:ZLUDA_CUDA_LIB = $originalCuda
+    $env:D4R_NVAPI_BACKEND = $originalNvapi
     $archive = "$OutputDirectory.zip"
     Compress-Archive -LiteralPath $OutputDirectory -DestinationPath $archive -Force
     Write-Host "Diagnostic bundle: $archive"

tools/windows/cuda_api.h

@@ -0,0 +1,44 @@
+#pragma once
+#include "diagnostic.h"
+// CUDA Driver API ABI subset. No CUDA SDK, CUDA runtime or NVIDIA binary is needed.
+using CUdevice = int;
+using CUdeviceptr = uint64_t;
+using CUcontext = void*;
+using CUmodule = void*;
+using CUfunction = void*;
+struct CudaApi {
+    d4r::diag::Library library;
+#define D4R_CU(name, ret, ...) using name##_fn = ret(WINAPI*)(__VA_ARGS__); name##_fn name = library.symbol<name##_fn>(#name)
+    D4R_CU(cuInit, int, unsigned);
+    D4R_CU(cuDriverGetVersion, int, int*);
+    D4R_CU(cuDeviceGetCount, int, int*);
+    D4R_CU(cuDeviceGet, int, CUdevice*, int);
+    D4R_CU(cuDeviceGetName, int, char*, int, CUdevice);
+    D4R_CU(cuDeviceGetAttribute, int, int*, int, CUdevice);
+    D4R_CU(cuDevicePrimaryCtxRetain, int, CUcontext*, CUdevice);
+    D4R_CU(cuDevicePrimaryCtxRelease_v2, int, CUdevice);
+    D4R_CU(cuCtxSetCurrent, int, CUcontext);
+    D4R_CU(cuCtxCreate_v2, int, CUcontext*, unsigned, CUdevice);
+    D4R_CU(cuCtxDestroy_v2, int, CUcontext);
+    D4R_CU(cuCtxSynchronize, int);
+    D4R_CU(cuMemAlloc_v2, int, CUdeviceptr*, size_t);
+    D4R_CU(cuMemFree_v2, int, CUdeviceptr);
+    D4R_CU(cuMemcpyHtoD_v2, int, CUdeviceptr, const void*, size_t);
+    D4R_CU(cuMemcpyDtoH_v2, int, void*, CUdeviceptr, size_t);
+    D4R_CU(cuModuleLoadData, int, CUmodule*, const void*);
+    D4R_CU(cuModuleUnload, int, CUmodule);
+    D4R_CU(cuModuleGetFunction, int, CUfunction*, CUmodule, const char*);
+    D4R_CU(cuLaunchKernel, int, CUfunction, unsigned, unsigned, unsigned, unsigned, unsigned, unsigned,
+        unsigned, void*, void**, void**);
+    D4R_CU(cuGetErrorName, int, int, const char**);
+    D4R_CU(cuGetErrorString, int, int, const char**);
+#undef D4R_CU
+    explicit CudaApi(const std::string& path) : library(d4r::diag::wide(path)) {}
+    void check(int result, const char* call) {
+        const char* name = "unknown", *message = "unknown";
+        if (result) { cuGetErrorName(result, &name); cuGetErrorString(result, &message); }
+        else name = "CUDA_SUCCESS";
+        std::printf("CUDA %s -> %d (%s)\n", call, result, name ? name : "unknown");
+        if (result) throw std::runtime_error(std::string(call) + ": " + (message ? message : "unknown"));
+    }
+};

tools/windows/cuda_driver_probe.cpp

@@ -1,48 +1,7 @@
 #include "hip_api.h"
 #include <algorithm>
 
-// CUDA Driver API ABI subset. No CUDA SDK, CUDA runtime or NVIDIA binary is needed.
+#include "cuda_api.h"
-using CUdevice = int;
-using CUdeviceptr = uint64_t;
-using CUcontext = void*;
-using CUmodule = void*;
-using CUfunction = void*;
-struct CudaApi {
-    d4r::diag::Library library;
-#define D4R_CU(name, ret, ...) using name##_fn = ret(WINAPI*)(__VA_ARGS__); name##_fn name = library.symbol<name##_fn>(#name)
-    D4R_CU(cuInit, int, unsigned);
-    D4R_CU(cuDriverGetVersion, int, int*);
-    D4R_CU(cuDeviceGetCount, int, int*);
-    D4R_CU(cuDeviceGet, int, CUdevice*, int);
-    D4R_CU(cuDeviceGetName, int, char*, int, CUdevice);
-    D4R_CU(cuDeviceGetAttribute, int, int*, int, CUdevice);
-    D4R_CU(cuDevicePrimaryCtxRetain, int, CUcontext*, CUdevice);
-    D4R_CU(cuDevicePrimaryCtxRelease_v2, int, CUdevice);
-    D4R_CU(cuCtxSetCurrent, int, CUcontext);
-    D4R_CU(cuCtxCreate_v2, int, CUcontext*, unsigned, CUdevice);
-    D4R_CU(cuCtxDestroy_v2, int, CUcontext);
-    D4R_CU(cuCtxSynchronize, int);
-    D4R_CU(cuMemAlloc_v2, int, CUdeviceptr*, size_t);
-    D4R_CU(cuMemFree_v2, int, CUdeviceptr);
-    D4R_CU(cuMemcpyHtoD_v2, int, CUdeviceptr, const void*, size_t);
-    D4R_CU(cuMemcpyDtoH_v2, int, void*, CUdeviceptr, size_t);
-    D4R_CU(cuModuleLoadData, int, CUmodule*, const void*);
-    D4R_CU(cuModuleUnload, int, CUmodule);
-    D4R_CU(cuModuleGetFunction, int, CUfunction*, CUmodule, const char*);
-    D4R_CU(cuLaunchKernel, int, CUfunction, unsigned, unsigned, unsigned, unsigned, unsigned, unsigned,
-        unsigned, void*, void**, void**);
-    D4R_CU(cuGetErrorName, int, int, const char**);
-    D4R_CU(cuGetErrorString, int, int, const char**);
-#undef D4R_CU
-    explicit CudaApi(const std::string& path) : library(d4r::diag::wide(path)) {}
-    void check(int result, const char* call) {
-        const char* name = "unknown", *message = "unknown";
-        cuGetErrorName(result, &name);
-        cuGetErrorString(result, &message);
-        std::printf("CUDA %s -> %d (%s)\n", call, result, name ? name : "unknown");
-        if (result) throw std::runtime_error(std::string(call) + ": " + (message ? message : "unknown"));
-    }
-};
 
 static constexpr char ptx[] = R"PTX(
 .version 7.0

tools/windows/diagnostic.h

@@ -86,6 +86,9 @@ struct Args {
     std::string hip_root, cuda_dll, module, context = "primary";
     std::string stream_lifetime = "persistent";
     std::string interop_mode = "roundtrip";
+    std::string ngx_core, dlss_dll;
+    std::string nvapi_dll;
+    std::string ngx_abi = "driver";
     unsigned iterations = 32;
     int device = -1;
     Args(int argc, char** argv) {
@@ -99,6 +102,10 @@ struct Args {
             else if (key == "--context") context = value;
             else if (key == "--stream-lifetime") stream_lifetime = value;
             else if (key == "--interop-mode") interop_mode = value;
+            else if (key == "--ngx-core") ngx_core = value;
+            else if (key == "--dlss-dll") dlss_dll = value;
+            else if (key == "--nvapi-dll") nvapi_dll = value;
+            else if (key == "--ngx-abi") ngx_abi = value;
             else if (key == "--iterations") {
                 size_t end = 0;
                 iterations = static_cast<unsigned>(std::stoul(value, &end));

tools/windows/ngx_cuda_init_probe.cpp

@@ -0,0 +1,127 @@
+#include "hip_api.h"
+#include "cuda_api.h"
+
+// Same NGX C ABI and freestanding MSVC parameter accessors as the upstream shim.
+// NVIDIA SDK headers and proprietary binaries are not build dependencies.
+struct NgxPathListInfo { const wchar_t* const* Path; unsigned Length; };
+struct NgxLoggingInfo { void* LoggingCallback; int MinimumLoggingLevel; bool DisableOtherLoggingSinks; };
+struct NgxFeatureCommonInfo { NgxPathListInfo PathListInfo; void* InternalData; NgxLoggingInfo LoggingInfo; };
+static_assert(sizeof(NgxFeatureCommonInfo) == 40);
+extern "C" void d4r_ngx_set_uint(void*, const char*, unsigned);
+extern "C" unsigned d4r_ngx_get_uint(void*, const char*, unsigned*);
+extern "C" unsigned d4r_ngx_get_int(void*, const char*, int*);
+
+static void ngx_log(const char* message, int level, unsigned component)
+{
+    if (message) std::printf("NGX_LOG level=%d component=%u %s\n", level, component, message);
+}
+static void check(unsigned result, const char* call)
+{
+    std::printf("NGX %s -> 0x%08x\n", call, result);
+    if (result != 1) throw std::runtime_error(std::string(call) + " NGX result=" + std::to_string(result));
+}
+
+int main(int argc, char** argv)
+{
+    using namespace d4r::diag;
+    start();
+    try {
+        Args args(argc, argv);
+        if (args.cuda_dll.empty() || args.ngx_core.empty() || args.dlss_dll.empty())
+            throw std::runtime_error("--cuda-dll, --ngx-core and --dlss-dll absolute paths required");
+        const std::filesystem::path core_path(wide(args.ngx_core)), dlss_path(wide(args.dlss_dll));
+        if (dlss_path.filename() != L"nvngx_dlss.dll") throw std::runtime_error("DLSS file must be named nvngx_dlss.dll");
+        HipApi hip(args.hip_root);
+        hipDeviceProp_t props{};
+        hip.select_gfx1201(args.device, props);
+        SearchDirectory cuda_search(std::filesystem::path(wide(args.cuda_dll)).parent_path());
+        CudaApi cuda(args.cuda_dll);
+        cuda.check(cuda.cuInit(0), "cuInit");
+        int count = 0;
+        cuda.check(cuda.cuDeviceGetCount(&count), "cuDeviceGetCount");
+        CUdevice device = -1;
+        for (int i = 0; i < count; ++i) {
+            CUdevice candidate = -1;
+            int bus = -1, pci_device = -1, domain = -1;
+            cuda.check(cuda.cuDeviceGet(&candidate, i), "cuDeviceGet");
+            cuda.check(cuda.cuDeviceGetAttribute(&bus, 33, candidate), "cuDeviceGetAttribute(PCI_BUS_ID)");
+            cuda.check(cuda.cuDeviceGetAttribute(&pci_device, 34, candidate), "cuDeviceGetAttribute(PCI_DEVICE_ID)");
+            cuda.check(cuda.cuDeviceGetAttribute(&domain, 50, candidate), "cuDeviceGetAttribute(PCI_DOMAIN_ID)");
+            if (bus == props.pciBusID && pci_device == props.pciDeviceID && domain == props.pciDomainID) device = candidate;
+        }
+        if (device < 0) throw std::runtime_error("NGX CUDA/HIP PCI identity mismatch");
+        CUcontext context = nullptr;
+        cuda.check(cuda.cuDevicePrimaryCtxRetain(&context, device), "cuDevicePrimaryCtxRetain");
+        struct ContextCleanup {
+            CudaApi& cuda; CUcontext& context; CUdevice device;
+            ~ContextCleanup() { if (context) { (void)cuda.cuCtxSetCurrent(nullptr); (void)cuda.cuDevicePrimaryCtxRelease_v2(device); } }
+        } cc{cuda, context, device};
+        cuda.check(cuda.cuCtxSetCurrent(context), "cuCtxSetCurrent");
+        auto nvapi_path = std::filesystem::path(wide(args.cuda_dll)).parent_path() / L"nvapi64.dll";
+        if (!std::filesystem::exists(nvapi_path)) nvapi_path = nvapi_path.parent_path().parent_path() / L"nvapi64.dll";
+        if (!args.nvapi_dll.empty()) nvapi_path = wide(args.nvapi_dll);
+        Library nvapi(nvapi_path);
+        SearchDirectory feature_search(dlss_path.parent_path());
+        // Preload the supplied absolute file so a system copy cannot replace it.
+        Library dlss(dlss_path), core(core_path);
+        using Init = unsigned(*)(unsigned long long, const wchar_t*, const NgxFeatureCommonInfo*, unsigned);
+        using DriverInit = unsigned(*)(unsigned long long, const wchar_t*, unsigned);
+        using Shutdown = unsigned(*)();
+        using Parameters = unsigned(*)(void**);
+        using Destroy = unsigned(*)(void*);
+        const auto init = core.symbol<Init>("NVSDK_NGX_CUDA_Init");
+        const auto driver_init = core.symbol<DriverInit>("NVSDK_NGX_CUDA_Init");
+        const auto shutdown = core.symbol<Shutdown>("NVSDK_NGX_CUDA_Shutdown");
+        const auto capabilities = core.symbol<Parameters>("NVSDK_NGX_CUDA_GetCapabilityParameters");
+        const auto allocate = core.symbol<Parameters>("NVSDK_NGX_CUDA_AllocateParameters");
+        const auto destroy = core.symbol<Destroy>("NVSDK_NGX_CUDA_DestroyParameters");
+        const std::wstring directory = dlss_path.parent_path().wstring();
+        const wchar_t* paths[] = {directory.c_str()};
+        const char* diagnostic_dir = std::getenv("D4R_DIAG_DIR");
+        const auto data_path = diagnostic_dir ? std::filesystem::path(wide(diagnostic_dir)) :
+            std::filesystem::current_path() / L"test-results" / L"ngx-init";
+        std::filesystem::create_directories(data_path);
+        NgxFeatureCommonInfo common{};
+        common.PathListInfo = {paths, 1};
+        common.LoggingInfo = {reinterpret_cast<void*>(&ngx_log), 2, false};
+        loaded_modules();
+        std::printf("STAGE NVSDK_NGX_CUDA_Init sdk=0x15 core=%s dlss=%s\n", args.ngx_core.c_str(), args.dlss_dll.c_str());
+        if (args.ngx_abi == "driver") check(driver_init(241534723ull, data_path.c_str(), 0x15), "NVSDK_NGX_CUDA_Init(driver ABI)");
+        else if (args.ngx_abi == "sdk") check(init(241534723ull, data_path.c_str(), &common, 0x15), "NVSDK_NGX_CUDA_Init(SDK ABI)");
+        else throw std::runtime_error("--ngx-abi must be driver or sdk");
+        bool initialized = true;
+        struct NgxCleanup { Shutdown shutdown; bool& active; ~NgxCleanup() { if (active) (void)shutdown(); } } nc{shutdown, initialized};
+        void* caps = nullptr;
+        check(capabilities(&caps), "NVSDK_NGX_CUDA_GetCapabilityParameters");
+        if (!caps) throw std::runtime_error("NGX returned null capability parameters");
+        struct ParameterCleanup { Destroy destroy; void*& p; ~ParameterCleanup() { if (p) (void)destroy(p); } } pc{destroy, caps};
+        int available = -1, init_result = -1, needs_driver = -1;
+        check(d4r_ngx_get_int(caps, "SuperSampling.Available", &available), "Get(SuperSampling.Available)");
+        check(d4r_ngx_get_int(caps, "SuperSampling.FeatureInitResult", &init_result), "Get(SuperSampling.FeatureInitResult)");
+        check(d4r_ngx_get_int(caps, "SuperSampling.NeedsUpdatedDriver", &needs_driver), "Get(SuperSampling.NeedsUpdatedDriver)");
+        std::printf("CAPABILITIES available=%d feature_init=0x%08x needs_driver=%d\n", available, static_cast<unsigned>(init_result), needs_driver);
+        if (available != 1 || init_result != 1 || needs_driver != 0)
+            throw std::runtime_error("DLSS CUDA feature is unavailable; initialization alone is insufficient");
+        void* parameters = nullptr;
+        check(allocate(&parameters), "NVSDK_NGX_CUDA_AllocateParameters");
+        if (!parameters) throw std::runtime_error("Null NGX parameters");
+        ParameterCleanup ac{destroy, parameters};
+        d4r_ngx_set_uint(parameters, "Width", 640);
+        unsigned roundtrip = 0;
+        check(d4r_ngx_get_uint(parameters, "Width", &roundtrip), "MSVC parameter ABI roundtrip");
+        if (roundtrip != 640) throw std::runtime_error("NGX parameter ABI mismatch");
+        check(destroy(parameters), "DestroyParameters(allocated)"); parameters = nullptr;
+        check(destroy(caps), "DestroyParameters(capabilities)"); caps = nullptr;
+        loaded_modules();
+        check(shutdown(), "NVSDK_NGX_CUDA_Shutdown"); initialized = false;
+        cuda.check(cuda.cuCtxSynchronize(), "cuCtxSynchronize");
+        cuda.check(cuda.cuCtxSetCurrent(nullptr), "cuCtxSetCurrent(NULL)");
+        cuda.check(cuda.cuDevicePrimaryCtxRelease_v2(device), "cuDevicePrimaryCtxRelease_v2"); context = nullptr;
+        std::printf("PASS NGX_INIT architecture=gfx1201 sr_available=1 transformer_executed=0\n");
+        return 0;
+    } catch (const std::exception& e) {
+        std::fprintf(stderr, "FAIL NGX_INIT %s\n", e.what());
+        loaded_modules();
+        return 4;
+    }
+}

tools/windows/nvapi_trace.cpp

@@ -0,0 +1,113 @@
+#include "diagnostic.h"
+#if defined(D4R_NVAPI_COMPAT)
+#include "hip_api.h"
+#include <memory>
+#include <mutex>
+#include <atomic>
+
+namespace {
+// Public NVAPI ABI: NVIDIA/nvapi nvapi.h and nvapi_interface.h.
+// AD100 describes the CUDA/NGX model profile, not the physical HIP target.
+struct ArchInfo { unsigned version, architecture, implementation, revision; };
+struct LogicalGpuData { unsigned version; void* os_adapter_id; unsigned count; void* physical[64]; unsigned reserved[8]; };
+static_assert(sizeof(LogicalGpuData) == 568);
+std::unique_ptr<d4r::diag::HipApi> hip;
+hipDeviceProp_t properties{};
+int device_ordinal = -1;
+int physical_handle;
+int logical_handle;
+std::mutex init_mutex;
+std::atomic<bool> runtime_ready{false};
+bool valid(void* handle) { return runtime_ready.load(std::memory_order_acquire) && handle == &physical_handle; }
+int __cdecl initialize()
+{
+    try {
+        std::lock_guard<std::mutex> lock(init_mutex);
+        if (!runtime_ready.load(std::memory_order_acquire)) {
+            const char* root = std::getenv("HIP_PATH");
+            if (!root) return -5;
+            auto runtime = std::make_unique<d4r::diag::HipApi>(root);
+            device_ordinal = runtime->select_gfx1201(-1, properties);
+            hip = std::move(runtime);
+            runtime_ready.store(true, std::memory_order_release);
+        }
+        std::printf("NVAPI_COMPAT initialize physical=gfx1201 CUDA_model_profile=AD100\n");
+        return 0;
+    } catch (const std::exception& e) { std::fprintf(stderr, "NVAPI_COMPAT init failed: %s\n", e.what()); return -1; }
+}
+int __cdecl enumerate(void** handles, unsigned* count)
+{
+    if (!handles || !count) return -5;
+    if (!runtime_ready.load(std::memory_order_acquire)) return -4;
+    handles[0] = &physical_handle; *count = 1;
+    std::printf("NVAPI_COMPAT EnumPhysicalGPUs count=1 HIP_ordinal=%d\n", device_ordinal);
+    return 0;
+}
+int __cdecl architecture(void* handle, ArchInfo* info)
+{
+    if (!valid(handle) || !info) return -5;
+    if (info->version != (sizeof(ArchInfo) | 1u << 16) && info->version != (sizeof(ArchInfo) | 2u << 16)) return -9;
+    info->architecture = 0x190; info->implementation = 2; info->revision = 0x11;
+    std::printf("NVAPI_COMPAT GetArchInfo profile=AD102 physical=gfx1201\n");
+    return 0;
+}
+int __cdecl adapter_id(void* handle, void* luid)
+{
+    if (!valid(handle) || !luid) return -5;
+    std::memcpy(luid, properties.luid, sizeof(LUID));
+    std::printf("NVAPI_COMPAT GetAdapterIdFromPhysicalGpu matched_HIP_LUID=1\n");
+    return 0;
+}
+int __cdecl logical_from_physical(void* handle, void** logical)
+{
+    if (!valid(handle) || !logical) return -5;
+    *logical = &logical_handle;
+    std::printf("NVAPI_COMPAT GetLogicalGPUFromPhysicalGPU count=1\n");
+    return 0;
+}
+int __cdecl logical_info(void* logical, LogicalGpuData* info)
+{
+    std::printf("NVAPI_COMPAT GetLogicalGpuInfo handle_valid=%d info=%p version=0x%08x os_adapter_buffer=%p\n",
+        runtime_ready.load(std::memory_order_acquire) && logical == &logical_handle, static_cast<void*>(info), info ? info->version : 0,
+        info ? info->os_adapter_id : nullptr);
+    if (!runtime_ready.load(std::memory_order_acquire) || logical != &logical_handle || !info || !info->os_adapter_id) return -5;
+    if (info->version != (sizeof(LogicalGpuData) | 1u << 16)) return -9;
+    std::memcpy(info->os_adapter_id, properties.luid, sizeof(LUID));
+    info->count = 1;
+    std::memset(info->physical, 0, sizeof(info->physical));
+    std::memset(info->reserved, 0, sizeof(info->reserved));
+    info->physical[0] = &physical_handle;
+    std::printf("NVAPI_COMPAT GetLogicalGpuInfo matched_HIP_LUID=1\n");
+    return 0;
+}
+}
+#endif
+
+// Trace build forwards responses unchanged. The compatibility build adds the
+// public topology APIs above; other queries still go to upstream ZLUDA NVAPI.
+extern "C" __declspec(dllexport) void* __cdecl nvapi_QueryInterface(unsigned id)
+{
+    try {
+#if defined(D4R_NVAPI_COMPAT)
+        switch (id) {
+        case 0x0150e828: return reinterpret_cast<void*>(&initialize);
+        case 0xe5ac921f: return reinterpret_cast<void*>(&enumerate);
+        case 0xd8265d24: return reinterpret_cast<void*>(&architecture);
+        case 0x0ff07fde: return reinterpret_cast<void*>(&adapter_id);
+        case 0xadd604d1: return reinterpret_cast<void*>(&logical_from_physical);
+        case 0x842b066e: return reinterpret_cast<void*>(&logical_info);
+        }
+#endif
+        const char* path = std::getenv("D4R_NVAPI_BACKEND");
+        if (!path || !*path) throw std::runtime_error("D4R_NVAPI_BACKEND absolute path required");
+        static d4r::diag::Library backend(d4r::diag::wide(path));
+        using Query = void* (__cdecl*)(unsigned);
+        static Query query = backend.symbol<Query>("nvapi_QueryInterface");
+        void* function = query(id);
+        std::printf("NVAPI_QUERY id=0x%08x supported=%d\n", id, function != nullptr);
+        return function;
+    } catch (const std::exception& e) {
+        std::fprintf(stderr, "NVAPI_TRACE_ERROR id=0x%08x %s\n", id, e.what());
+        return nullptr;
+    }
+}