Validate Windows native M FP8 and reduce operand packing cost

README.md

@@ -14,13 +14,17 @@ submission, and returns output in the same frame without CPU image copies.
 RGBA16/32F, R11G11B10, RGBA/BGRA8, depth, exposure and motion formats have
 independent D3D12 typed SRV/UAV conversion tests. Four-frame K/M runs with packed
 output match the FP16 baseline exactly after resource-format quantisation.
-The build incorporates upstream RDNA4 commit `dbef4b2`. Patched OptiScaler now
+The build incorporates upstream updates through `f0d1a65`. Patched OptiScaler now
 passes standalone K/M native-network tests with exact RGB agreement against
 the direct-d4r baseline; see [the source patch](patches/optiscaler/README.md).
 Silent Hill 2 now renders K through the native Windows backend; M also renders
 a saved gameplay level with GPU output NaN/Inf checks. Startup depth/stencil
 and swap-chain lifetime issues found in the game have regression checks.
-Performance and wider resolution coverage remain in progress. See
+Both K and M now pass 3840x2160 game output scans. K reaches a user-reported
+49-51 FPS in the measured scene after native output stores and command-hook
+lookup caching. Native M FP8 is separately replay/network validated; the
+default keeps the FP16-equivalent baseline based on measured performance.
+All sixteen CTest gates pass with the final native Windows ZLUDA runtime. See
 [the reversible game package](docs/windows-game.md) and
 [build instructions and hardware results](docs/windows-rdna4-port.md).
 
@@ -53,7 +57,7 @@ On the same GPU DLSS starts at a disadvantage: its networks were designed for NV
 
 For users who prioritize fidelity over speed, set `[Kernels] PreferAccuracy = true` in `d4r/d4r.ini` and restart the game. It defaults to `false`. This selects accuracy variants of every native kernel and restores conservative translation settings. It aims to match NVIDIA's arithmetic; 1:1 image quality against RTX DLSS is not yet proven. See [native kernel numerics](docs/native-kernels.md#numerics).
 
-## GPU support
+## GPU support (Linux upstream)
 
 d4r builds for RDNA3 and RDNA4. A newly built release compiles native DLSS 4 and 4.5 network kernels for the targets below and selects the matching set at runtime. Only the RX 7700 XT has been tested by this project on a real GPU; an external video reports DLSS 4.5 running through d4r 0.1.2 on an RX 7900 XTX. RDNA4 runtime and performance remain unverified on hardware.
 

docs/windows-game.md

@@ -3,21 +3,27 @@
 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. K also passes 3840x2160 output and M passes 1920x1080.
-Further performance work is in progress. This is a local
+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
 development package; no NVIDIA proprietary DLL is included.
 
 Build the shim/diagnostics, corrected ZLUDA, K/M native objects and patched
 OptiScaler using the scripts in `scripts/windows`, then run:
 
 ```powershell
-.\scripts\windows\package-windows-game.ps1
+.\scripts\windows\package-windows-game.ps1 -ArchivePath "$PWD\dist\windows-rdna4-game.zip"
 ```
 
 The package pins TheRock 10.2.0a20260929 because its Windows external-memory
 mapping lifetime passes the repeated import/release test. Stable HIP 7.2 has
 a reproduced mapped-view leak; see `docs/windows-rdna4-port.md` for the source
 fix, exact dependency versions, validation and complete build commands.
+The ZIP contains only manifest-listed files, including a snapshot of committed
+d4r source. Private NVIDIA-derived objects, caches and diagnostic captures are
+excluded. Original FP8 M and the optional hardware-packing variant pass strict
+replay/network tests but do not establish a full-network speedup, so the game
+package retains its FP16-equivalent M baseline.
 
 Run a D3D12 game without anti-cheat using your locally supplied NVIDIA DLLs:
 

docs/windows-performance.md

@@ -69,9 +69,8 @@ memory, LDS and predicted blocks per multiprocessor. Add
 `--metadata-directory <directory-of-llvm-readobj-notes>` for native wave sizes
 and compiler spill counts. Reports and proprietary workload inputs remain local.
 
-Native FP8 is not the validated baseline and has no performance claim here.
-Replacing the exact per-MMA f16 rounding with f32 FP8 accumulation requires a
-new layer/reference validation; it is not justified by the observed CPU stalls.
+Native FP8 remains a separate optional experiment; the package uses the
+validated FP16-equivalent baseline. Its later checks and timings are below.
 
 Further game coverage (HIP kernel profiling disabled, GPU output checks enabled):
 M at 1920x1080 completes 4833 frames, all finite. K at 3840x2160 completes
@@ -172,5 +171,58 @@ each physical command-backend test and the software indirect-root regression.
 Use `-ProfileCommandHooks -ProfileStages` to collect samples; the sampler is
 disabled for ordinary play.
 
+Native M FP8 was subsequently compiled and tested on this gfx1201. ISA confirms
+`v_wmma_f32_16x16x16_fp8_fp8`, with packed e4m3 weight images; exact FP16
+attention/PV remains. All logical outputs and merges in forty recorded launches
+match the previously NumPy-validated FP16 baseline byte-for-byte. Full DLSS RGB
+also matches exactly: four 512x288 LDR frames and four 3840x2160 HDR frames,
+all RGBA finite. This establishes correctness for those fixtures, not a speedup.
+
+The original software operand packing makes the FP8 network slower in the
+four-frame 4K test: sum of five main-kernel event means (tube weighted six times)
+is 27.789640 ms versus FP16's 23.550875 ms. A paired in-process replay confirms
+the regression without relying solely on separate process timings. Both modules
+and prepared weights remain resident, 512 launches warm the GPU, then sixty-four
+AB/BA pairs time the same restored input using HIP events. Diagnostic CPU input
+restoration is outside the timed event and is unrelated to game VRAM interop.
+
+Added opt-in `D4R_FP8_HW_PACK`: native packed conversion of already quantised
+operands replaces software exponent/subnormal reconstruction. It preserves all
+254 finite OCP e4m3 encodings, including signed zero, in an exhaustive GPU check.
+All forty recorded launches still match exactly. Four complete 4K HDR frames
+also match FP16 RGB exactly with finite RGBA. ISA confirms both packed converts
+and FP8 WMMA. The baseline build's sixteen objects remain byte-identical.
+
+| Paired replay median, ms | FP16 (software-packing test) | FP8 software packing | FP16 (hardware-packing test) | FP8 hardware packing |
+| --- | ---: | ---: | ---: | ---: |
+| enc1 | 0.175450 | 0.217050 | 0.175100 | 0.177700 |
+| enc2 | 0.141350 | 0.164550 | 0.132550 | 0.131550 |
+| enc3 tube | 0.108000 | 0.112000 | 0.107600 | 0.099850 |
+| dec2 | 0.129900 | 0.146000 | 0.143250 | 0.140850 |
+| dec1 | 0.169650 | 0.194450 | 0.154250 | 0.156550 |
+
+These are two paired tests on the captured 512x288 workload. Medians avoid large
+isolated scheduling outliers; raw per-pair ratios are retained. They identify a
+substantial cost in software packing but do not establish an overall FPS gain.
+The separately run four-frame 4K hardware-packing network mean is 24.603463 ms;
+it still does not beat the measured FP16 total. Default packaging retains FP16.
+Hardware counters for bandwidth or WMMA utilization were not measured.
+
+Reproduce the isolated build, encoding/replay checks and paired timings using
+the existing private validated captures; no source editing is required:
+
+```powershell
+.\scripts\windows\build-native-m-fp8.ps1 -DlssDll "$PWD\nvngx_dlss.dll" -HardwarePacking
+.\scripts\windows\test-native-m-fp8.ps1 -ModuleDirectory "$PWD\build\native-m-fp8-hardware-pack" -CaptureDirectory "$PWD\test-results\ngx-m-exact-pv-capture\captures" -BaselineValidation "$PWD\test-results\m-native-exact-pv-temporal-reference\summary.json" -Benchmark
+```
+
+Omit `-HardwarePacking` to compile the original FP8 experiment. The replay
+validator stops at the first mismatch, retains GPU stdout/stderr/exit code and
+hashes every compared logical output. `profile_report.py` accepts replay logs
+and reports per-pair GPU time and candidate/control ratios. This experiment is
+not installed by the game runner or substituted silently for the baseline.
+The conversion builtin is listed in the official
+[Clang AMDGPU builtin reference](https://clang.llvm.org/docs/AMDGPUBuiltinReference.html#builtin-amdgcn-cvt-pk-fp8-f32).
+
 References: [HIP events](https://rocmdocs.amd.com/projects/HIP/en/develop/doxygen/html/group___event.html),
 [HIP occupancy API](https://rocm.docs.amd.com/projects/HIP/en/latest/doxygen/html/group___occupancy.html).

docs/windows-rdna4-port.md

@@ -2,10 +2,11 @@
 
 Branch: `windows-rdna4`. Target: Windows 11 x64, RX 9070 XT, **gfx1201**.
 Priority: correct K, correct M, native Windows, same-frame output, then speed.
-This is a work log for an implementation in progress. Standalone DLSS K/M
+This file records current results followed by the dated development history. Standalone DLSS K/M
 works on the real Windows GPU, including the patched OptiScaler frontend.
 Silent Hill 2 now renders through the Windows backend; gameplay and output
-validation coverage are recorded below. Performance remains in progress.
+validation coverage are recorded below. Performance and game coverage have
+explicit limits; the validated package keeps conservative arithmetic.
 
 ## Milestone and gates
 
@@ -16,9 +17,44 @@ M attention baseline. Upstream's optional denormal override now drops only
 FP32 requirements, preserving FP16/FP64. All sixteen rebuilt native K/M
 objects have the same SHA256 as the previously validated objects. All thirteen
 ZLUDA patches apply to the pinned fresh base; all sixteen CTest gates pass.
-The local ZLUDA source includes that denormal correction at `2378af7`; the
-measured runtime remains the separately recorded `b0161a4` binary, with the
-override unset. Package metadata records the binary build's actual commit.
+Final ZLUDA source and binary are `2378af72e586e1730d15f75e0322daab674f0862`,
+built cleanly into `dist/zluda-windows-final`; nvcuda SHA256 is
+`ccc700974fa5d0b9a5f5c39e617c17ca22f34701d40f745f2f96f039a75b22f8`.
+All sixteen CTest gates pass on it. Three K/4K HDR frames and four M baseline
+frames remain bit-exact to their prior controls, all RGBA finite. Historical
+performance runs below used `b0161a4` with the denormal override unset.
+Package metadata records the binary build's actual commit.
+
+The final-runtime M game test completes 1940 4K frames and 1940 finite GPU
+scans, with 19410 native launches (including a partial final frame), zero
+backend failures/CPU image copies/previous-frame outputs. Native M FP8 also
+passes all logical output/merge values from forty NumPy-validated captures,
+plus four 512x288 and four 4K complete frames. Its original software packing
+regresses performance; the opt-in hardware packing preserves all 254 finite
+FP8 encodings and all forty captures, with exact RGB in four additional 4K
+frames. Paired in-process event tests confirm reduced packing cost, but the
+full-network measurement does not establish a win over FP16. Production keeps
+the baseline. See [windows-performance.md](windows-performance.md) for timings
+and one-command replay/paired profiling.
+
+Current native Windows build, after the pinned tool/source setup described
+below (MSVC v143 / Windows SDK 10.0.26100 required for OptiScaler):
+
+```powershell
+.\scripts\windows\build-zluda-windows.ps1
+.\scripts\windows\build-windows-rdna4.ps1 -RuntimeProfile therock -ZludaRoot "$PWD\dist\zluda-windows-final" -InstallDirectory "$PWD\dist\windows-rdna4-command-list"
+.\scripts\windows\stage-native-k.ps1 -DlssDll "$PWD\nvngx_dlss.dll" -PackageRoot "$PWD\dist\windows-rdna4-command-list"
+.\scripts\windows\stage-native-m.ps1 -DlssDll "$PWD\nvngx_dlss.dll" -PackageRoot "$PWD\dist\windows-rdna4-command-list"
+.\scripts\windows\build-optiscaler-windows.ps1
+& .\.tools\python\cmake\data\bin\ctest.exe --test-dir build/windows-rdna4-therock --output-on-failure
+.\scripts\windows\package-windows-game.ps1 -ArchivePath "$PWD\dist\windows-rdna4-game.zip"
+```
+
+The archive requires committed source. It includes a tracked-source snapshot,
+all runtime file hashes/versions, dependency licenses and source patches, and
+excludes local NVIDIA DLLs, PTX/weights/captures and private texture objects.
+The optional texture optimization is built/validated locally and supplied with
+`-LocalTextureKernels`; it is never folded into the public source snapshot.
 
 Additional production-style runs: M at 1920x1080 completes 4833 frames with
 4833 finite-output GPU scans and no backend failures. K at 3840x2160 completes
@@ -178,14 +214,17 @@ with `-EarlyIndirectProbe` in the standalone OptiScaler runner. This avoids
 the previously documented unknown-signature rejection for observed startup
 objects; signatures created before the frontend is loaded remain untracked.
 
-Current milestone: M1/M2/M4/M5 passed; native NGX Create/Evaluate on the simple
+Current milestone: M0-M9 functional gates pass on the tested Windows GPU/game.
+The package uses FP16-equivalent K/M and conservative denormal handling.
+Native NGX Create/Evaluate on the simple
 requested E path passes with the user's DLLs. All 11 K layers now pass synthetic
 and real-weight NumPy/PTX replay checks; native K Evaluate produces four finite
 frames on Windows. M now passes all five synthetic and real-weight layers against
 NumPy and independent PTX, and four native CUDA frames. The standalone D3D12
 harness now produces four K and four M frames using imported VRAM and shared
 fences, with bit-exact RGB agreement against the corresponding CUDA harnesses.
-Game integration remains in progress. The ordinary NGX D3D12 Evaluate API now
+Silent Hill 2 K/M integration passes, including current-frame 4K VRAM interop.
+The ordinary NGX D3D12 Evaluate API now
 records a boundary in an open command list; a native queue interception backend
 executes prefix, CUDA evaluation, then suffix in the same submitted frame.
 A full temporal M scan
@@ -207,7 +246,7 @@ Do not integrate NGX until the integer PTX workload is stable on the real GPU.
 | M6: K layers, full transformer and image validation | PASS: 11 real-weight layers, 44 native launches, four D3D12 frames bit-exact against native CUDA, 81.85–92.91 dB versus translated K |
 | M7: M FP16-equivalent baseline and full transformer | PASS: all blocks of 40 native launches exactly match NumPy; targeted independent PTX regression matches; four D3D12 frames bit-exact against corrected translated M |
 | M8: standalone Windows D3D12 NGX harness | PASS for ordinary recorded EvaluateFeature and explicit boundaries: K/M, four frames each, no shim CPU image copies; both bit-exact against explicit boundary baselines |
-| M9: OptiScaler integration, profiling, installation | Patched frontend and game K/M PASS, including K 4K and M 1080p finite output; native/translated GPU profiling and reversible install PASS; remaining FPS investigation and final archive in progress |
+| M9: OptiScaler integration, profiling, installation | Patched frontend and game K/M PASS at 4K with finite output; native/translated/FP8 profiling, reversible install and manifest-only public archive PASS; K user reports 49-51 FPS |
 
 ## Public NGX identity and OptiScaler gate (2026-10-01)
 

kernels/m/swin_common.h

@@ -273,11 +273,25 @@ __device__ __forceinline__ sop lds16(const half_t* p)
 // e4m3 bytes of an f16 operand whose values are e4m3 values (K order kept: byte j = element j)
 __device__ __forceinline__ gop to_fp8(sop v)
 {
+#if defined(D4R_FP8_HW_PACK) && defined(__GFX12__)
+    // Inputs already contain exact e4m3 values. gfx12's native packed convert
+    // avoids the software subnormal/sign/exponent reconstruction below. Keep
+    // opt-in until all finite encodings and layer/network outputs pass replay.
+    gop result;
+#pragma unroll
+    for (int word = 0; word < 2; ++word) {
+        int packed = __builtin_amdgcn_cvt_pk_fp8_f32((float)v[4 * word], (float)v[4 * word + 1], 0, false);
+        packed = __builtin_amdgcn_cvt_pk_fp8_f32((float)v[4 * word + 2], (float)v[4 * word + 3], packed, true);
+        result[word] = (uint32_t)packed;
+    }
+    return result;
+#else
     uint32_t c[4];
 #pragma unroll
     for (int j = 0; j < 4; ++j)
         c[j] = codes8x2((hv2){v[2 * j], v[2 * j + 1]});
     return (u2v){__builtin_amdgcn_perm(c[1], c[0], 0x06040200u), __builtin_amdgcn_perm(c[3], c[2], 0x06040200u)};
+#endif
 }
 // f32 += A B over 16 K for e4m3 operands
 __device__ __forceinline__ f8v wmma8(gop a, gop b, f8v c)

scripts/windows/build-native-m-fp8.ps1

@@ -0,0 +1,59 @@
+[CmdletBinding()]
+param([Parameter(Mandatory=$true)][string]$DlssDll, [string]$HipRoot, [string]$OutputDirectory, [switch]$HardwarePacking)
+$ErrorActionPreference='Stop'
+$repo=[IO.Path]::GetFullPath((Join-Path $PSScriptRoot '../..'))
+if (!$HipRoot) { $HipRoot=Join-Path $repo '.tools/therock-10.2.0a20260929/_rocm_sdk_core' }
+if (!$OutputDirectory) { $OutputDirectory=Join-Path $repo $(if ($HardwarePacking) { 'build/native-m-fp8-hardware-pack' } else { 'build/native-m-fp8-experimental' }) }
+$OutputDirectory=[IO.Path]::GetFullPath($OutputDirectory)
+if ($OutputDirectory -eq [IO.Path]::GetFullPath((Join-Path $repo 'build/native-m-gfx1201'))) {
+    throw 'Keep the validated FP16 baseline separate from the experimental FP8 build.'
+}
+$DlssDll=(Get-Item -LiteralPath $DlssDll -ErrorAction Stop).FullName
+$compiler=Join-Path $HipRoot 'lib/llvm/bin/clang++.exe'
+if (!(Test-Path -LiteralPath $compiler)) { $compiler=Join-Path $HipRoot 'bin/clang++.exe' }
+if (!(Test-Path -LiteralPath $compiler)) { throw "HIP compiler missing under $HipRoot" }
+New-Item -ItemType Directory -Force $OutputDirectory | Out-Null
+$disassembler=Join-Path (Split-Path $compiler) 'llvm-objdump.exe'
+if (!(Test-Path -LiteralPath $disassembler)) { throw 'HIP llvm-objdump is required to confirm native FP8 instructions.' }
+foreach ($source in Get-ChildItem -LiteralPath (Join-Path $repo 'kernels/m') -Filter '*.hip' | Sort-Object Name) {
+    $flags=@()
+    if ((Get-Content -LiteralPath $source.FullName -Raw) -match '// d4r-build-flags: ([^\r\n]+)') { $flags=$Matches[1].Trim().Split(' ') }
+    if ($HardwarePacking) { $flags+='-DD4R_FP8_HW_PACK' }
+    $object=Join-Path $OutputDirectory ($source.BaseName+'.hsaco')
+    $ErrorActionPreference='Continue'
+    & $compiler -x hip -std=c++17 --offload-arch=gfx1201 --offload-device-only --no-gpu-bundle-output `
+        -mno-wavefrontsize64 -nogpuinc -nogpulib -O3 -DD4R_DEVICE_ONLY_MINIMAL `
+        -DSWIN_EXACT -DSWIN_EXACT_PV -DD4R_FP8_WMMA @flags $source.FullName `
+        -o $object > (Join-Path $OutputDirectory ($source.BaseName+'.build.stdout.log')) `
+        2> (Join-Path $OutputDirectory ($source.BaseName+'.build.stderr.log'))
+    $ErrorActionPreference='Stop'
+    if ($LASTEXITCODE) { throw "Experimental FP8 compile failed: $($source.Name)" }
+    $isa=Join-Path $OutputDirectory ($source.BaseName+'.isa.txt')
+    & $disassembler --disassemble --mcpu=gfx1201 $object > $isa
+    if ($LASTEXITCODE -or !(Select-String -LiteralPath $isa -Pattern 'v_wmma_f32_16x16x16_fp8_fp8')) {
+        throw "Native FP8 WMMA is absent from $($source.Name); no FP16 fallback is accepted."
+    }
+    if ($HardwarePacking -and !(Select-String -LiteralPath $isa -Pattern 'v_cvt_pk_fp8_f32')) { throw 'Native FP8 operand packing is absent from ISA.' }
+    Write-Host "Compiled gfx1201 FP8 WMMA: $($source.BaseName)"
+}
+if ($HardwarePacking) {
+    $probeDirectory=Join-Path $OutputDirectory 'packing-probe'
+    New-Item -ItemType Directory -Force $probeDirectory | Out-Null
+    & $compiler -x hip -std=c++17 --offload-arch=gfx1201 --offload-device-only --no-gpu-bundle-output `
+        -mno-wavefrontsize64 -nogpuinc -nogpulib -O3 -DD4R_DEVICE_ONLY_MINIMAL `
+        -DD4R_FP8_WMMA -DD4R_FP8_HW_PACK -DSWIN_EXACT -DSWIN_EXACT_PV `
+        (Join-Path $repo 'tools/windows/fp8_pack_probe.hip') -o (Join-Path $probeDirectory 'fp8_pack_probe.hsaco')
+    if ($LASTEXITCODE) { throw 'Native hardware packing diagnostic compile failed.' }
+}
+& python (Join-Path $repo 'kernels/tools/kernel_manifest.py') $OutputDirectory $DlssDll
+if ($LASTEXITCODE) { throw 'Experimental FP8 manifest generation failed.' }
+@{architecture='gfx1201'; family='M'; experimental=$true; nativeFp8=$true; accuracy=$true;
+    numerics='SWIN_EXACT; SWIN_EXACT_PV; FP8 WMMA for e4m3 GEMMs; exact FP16 attention';
+    validation='not validated; do not install before layer/full-network checks';
+    compiler=$compiler; compilerVersion=(& $compiler --version | Select-Object -First 1);
+    sourceCommit=(& git -C $repo rev-parse HEAD).Trim(); nativeFp8IsaVerified=$true; hardwarePacking=[bool]$HardwarePacking;
+    sourceWorkingTreeDirty=[bool](& git -C $repo status --porcelain);
+    dlssSha256=(Get-FileHash -LiteralPath $DlssDll -Algorithm SHA256).Hash;
+    modules=@(Get-ChildItem -LiteralPath $OutputDirectory -Filter '*.hsaco' | Get-FileHash -Algorithm SHA256)} |
+    ConvertTo-Json -Depth 5 | Set-Content -LiteralPath (Join-Path $OutputDirectory 'build-info.json') -Encoding UTF8
+Write-Host "Experimental native FP8 M objects: $OutputDirectory (baseline unchanged)"

scripts/windows/build-native-texture.ps1

@@ -8,7 +8,7 @@ param(
 )
 $ErrorActionPreference='Stop'
 $repo=[IO.Path]::GetFullPath((Join-Path $PSScriptRoot '../..'))
-if (!$ZludaRoot) { $ZludaRoot=Join-Path $repo 'dist/zluda-windows-profile' }
+if (!$ZludaRoot) { $ZludaRoot=Join-Path $repo 'dist/zluda-windows-final' }
 if (!$OutputDirectory) { $OutputDirectory=Join-Path $repo 'build/private-textures-gfx1201' }
 & python (Join-Path $repo 'tools/windows/build_native_texture.py') --dlss-dll $DlssDll --hip-root $HipRoot `
     --zluda-root $ZludaRoot --kernel $Kernel --output-directory $OutputDirectory

scripts/windows/build-zluda-windows.ps1

@@ -4,7 +4,7 @@ $ErrorActionPreference = 'Stop'
 $repo = [IO.Path]::GetFullPath((Join-Path $PSScriptRoot '../..'))
 if (!$SourceRoot) { $SourceRoot = Join-Path $repo 'external/ZLUDA' }
 if (!$LlvmBuildRoot) { $LlvmBuildRoot = Join-Path $repo 'build/zluda-llvm-windows-gnu' }
-if (!$InstallRoot) { $InstallRoot = Join-Path $repo 'dist/zluda-windows-native' }
+if (!$InstallRoot) { $InstallRoot = Join-Path $repo 'dist/zluda-windows-final' }
 $rust = Join-Path $repo '.tools/rust-1.98.1-gnu'
 $toolchain = Join-Path $repo '.tools/llvm-mingw-20260922-ucrt-x86_64'
 $target = Join-Path $repo 'build/zluda-rust-windows-gnu'

scripts/windows/package-windows-game.ps1

@@ -7,13 +7,16 @@ param(
 $ErrorActionPreference = 'Stop'
 $repo = [IO.Path]::GetFullPath((Join-Path $PSScriptRoot '../..'))
 $defaults = @{PackageRoot='dist/windows-rdna4-game'; DiagnosticRoot='dist/windows-rdna4-command-list';
-    OptiScalerRoot='dist/optiscaler-windows-d4r'; ZludaRoot='dist/zluda-windows-profile';
+    OptiScalerRoot='dist/optiscaler-windows-d4r'; ZludaRoot='dist/zluda-windows-final';
     HipRoot='.tools/therock-10.2.0a20260929/_rocm_sdk_core';
     KernelKRoot='build/native-k-gfx1201'; KernelMRoot='build/native-m-gfx1201'}
 foreach ($name in $defaults.Keys) {
     if (!(Get-Variable -Name $name -ValueOnly)) { Set-Variable -Name $name -Value (Join-Path $repo $defaults[$name]) }
 }
 $PackageRoot = [IO.Path]::GetFullPath($PackageRoot)
+$sourceCommit=(& git -C $repo rev-parse HEAD).Trim()
+$sourceDirty=[bool](& git -C $repo status --porcelain)
+if ($ArchivePath -and $sourceDirty) { throw 'Commit source changes before creating the distributable archive.' }
 New-Item -ItemType Directory -Force $PackageRoot | Out-Null
 $stagedFiles = [Collections.Generic.List[string]]::new()
 function Stage([string]$source, [string]$relative) {
@@ -61,10 +64,15 @@ Stage (Join-Path $HipRoot 'share/doc/amd_comgr/LICENSE.txt') 'licenses/amd_comgr
 Stage (Join-Path $repo 'external/clr/LICENSE.md') 'licenses/HIP-CLR.txt'
 foreach ($patch in Get-ChildItem -LiteralPath (Join-Path $repo 'patches/optiscaler') -Filter '*.patch') { Stage $patch.FullName ('source-patches/optiscaler/' + $patch.Name) }
 foreach ($patch in Get-ChildItem -LiteralPath (Join-Path $repo 'patches/zluda') -Filter '*.patch') { Stage $patch.FullName ('source-patches/zluda/' + $patch.Name) }
+$sourceArchive=Join-Path $repo 'build/d4r-windows-rdna4-source.zip'
+New-Item -ItemType Directory -Force (Split-Path $sourceArchive) | Out-Null
+& git -C $repo archive --format=zip --output $sourceArchive $sourceCommit
+if ($LASTEXITCODE) { throw 'Cannot archive the committed d4r source.' }
+Stage $sourceArchive 'source/d4r-windows-rdna4-source.zip'
 $metadata = [ordered]@{
     architecture='gfx1201'; nativeKernels=16; dlssSha256='3975567b8943c53acce397f2b72380092f84f162d00b0d2c7d08a1025c563983';
-    d4rCommit=(& git -C $repo rev-parse HEAD).Trim();
-    d4rWorkingTreeDirty=[bool](& git -C $repo status --porcelain);
+    d4rCommit=$sourceCommit;
+    d4rWorkingTreeDirty=$sourceDirty;
     optiScaler=(Get-Content -LiteralPath (Join-Path $OptiScalerRoot 'build-info.json') -Raw | ConvertFrom-Json);
     zludaBuild=(Get-Content -LiteralPath (Join-Path $ZludaRoot 'build-info.json') -Raw | ConvertFrom-Json);
     hipRuntime='TheRock 10.2.0a20260929 / HIP 7.17.26386';

scripts/windows/profile-windows-rdna4.ps1

@@ -7,7 +7,7 @@ param(
 )
 $ErrorActionPreference='Stop'
 $repo=[IO.Path]::GetFullPath((Join-Path $PSScriptRoot '../..'))
-if (!$ZludaRoot) { $ZludaRoot=Join-Path $repo 'dist/zluda-windows-profile' }
+if (!$ZludaRoot) { $ZludaRoot=Join-Path $repo 'dist/zluda-windows-final' }
 if (!$PackageRoot) { $PackageRoot=Join-Path $repo 'dist/windows-rdna4-command-list' }
 if (!$OutputDirectory) { $OutputDirectory=Join-Path $repo ('test-results/profile-' + $Preset + '-' + (Get-Date -Format 'yyyyMMdd-HHmmss-fff')) }
 $settings=@{D4R_ZLUDA_PROFILE='1'; D4R_PROFILE_STAGES='1'; D4R_VALIDATE_OUTPUT='1';

scripts/windows/test-native-m-fp8.ps1

@@ -0,0 +1,28 @@
+[CmdletBinding()]
+param([Parameter(Mandatory=$true)][string]$CaptureDirectory,
+    [Parameter(Mandatory=$true)][string]$BaselineValidation,
+    [string]$ModuleDirectory, [string]$HipRoot, [string]$OutputDirectory, [switch]$Benchmark)
+$ErrorActionPreference='Stop'
+$repo=[IO.Path]::GetFullPath((Join-Path $PSScriptRoot '../..'))
+if (!$ModuleDirectory) { $ModuleDirectory=Join-Path $repo 'build/native-m-fp8-experimental' }
+if (!$HipRoot) { $HipRoot=Join-Path $repo '.tools/therock-10.2.0a20260929/_rocm_sdk_core' }
+if (!$OutputDirectory) { $OutputDirectory=Join-Path $repo 'test-results/m-native-fp8-replay' }
+$previousPython=$env:PYTHONPATH
+try {
+    $env:PYTHONPATH="$(Join-Path $repo '.tools/python/vendor');$previousPython"
+    $metadata=Get-Content -LiteralPath (Join-Path $ModuleDirectory 'build-info.json') -Raw | ConvertFrom-Json
+    if ($metadata.hardwarePacking) {
+        & python -u (Join-Path $repo 'tools/windows/fp8_pack_validate.py') `
+            --probe (Join-Path $repo 'dist/windows-rdna4-command-list/bin/d4r_native_replay_probe.exe') `
+            --module (Join-Path $ModuleDirectory 'packing-probe/fp8_pack_probe.hsaco') `
+            --hip-root $HipRoot --output-dir (Join-Path $OutputDirectory 'packing-check')
+        if ($LASTEXITCODE) { throw 'Hardware packing does not preserve every finite FP8 encoding; stopping before network replay.' }
+    }
+    $benchmarkArguments=@()
+    if ($Benchmark) { $benchmarkArguments=@('--benchmark-control-dir',(Join-Path $repo 'build/native-m-gfx1201')) }
+    & python -u (Join-Path $repo 'tools/windows/m_native_replay_compare.py') `
+        --capture-dir $CaptureDirectory --module-dir $ModuleDirectory --hip-root $HipRoot `
+        --probe (Join-Path $repo 'dist/windows-rdna4-command-list/bin/d4r_native_replay_probe.exe') `
+        --baseline-validation $BaselineValidation --output-dir $OutputDirectory @benchmarkArguments
+    if ($LASTEXITCODE) { throw "Experimental FP8 validation failed; baseline unchanged. Logs: $OutputDirectory" }
+} finally { $env:PYTHONPATH=$previousPython }

scripts/windows/test-native-overrides.ps1

@@ -4,7 +4,7 @@ $ErrorActionPreference = 'Stop'
 $repo = [IO.Path]::GetFullPath((Join-Path $PSScriptRoot '../..'))
 if (!$HipRoot) { $HipRoot = Join-Path $repo '.tools/therock-10.2.0a20260929/_rocm_sdk_core' }
 if (!$PackageRoot) { $PackageRoot = Join-Path $repo 'dist/windows-rdna4-therock' }
-if (!$ZludaRoot) { $ZludaRoot = Join-Path $repo 'dist/zluda-windows-native' }
+if (!$ZludaRoot) { $ZludaRoot = Join-Path $repo 'dist/zluda-windows-final' }
 if (!$OutputDirectory) { $OutputDirectory = Join-Path $repo ('test-results/native-identity-' + (Get-Date -Format 'yyyyMMdd-HHmmss')) }
 New-Item -ItemType Directory -Force $OutputDirectory | Out-Null
 $previous = @{D4R_ZLUDA_NATIVE_DIR=$env:D4R_ZLUDA_NATIVE_DIR; D4R_ZLUDA_VERBOSE=$env:D4R_ZLUDA_VERBOSE; D4R_DIAG_DIR=$env:D4R_DIAG_DIR}

scripts/windows/test-native-texture.ps1

@@ -22,7 +22,7 @@ try {
     foreach ($flags in @(0,11)) {
         foreach ($mode in @('control','candidate')) {
             $env:D4R_ZLUDA_NATIVE_DIR=if ($mode -eq 'control') { Join-Path $repo 'build/native-k-gfx1201' } else { $combined }
-            $arguments=@{RuntimeProfile='therock'; ZludaRoot=(Join-Path $repo 'dist/zluda-windows-profile');
+            $arguments=@{RuntimeProfile='therock'; ZludaRoot=(Join-Path $repo 'dist/zluda-windows-final');
                 PackageRoot=(Join-Path $repo 'dist/windows-rdna4-command-list'); NgxCore=(Join-Path $repo '_nvngx.dll');
                 DlssDll=(Join-Path $repo 'nvngx_dlss.dll'); NgxMode='d3d12'; NgxAbi='project-legacy'; Preset=11;
                 NgxCreateFlags=$flags; NgxOutputResolution=$OutputResolution; NgxOnly=$true; RequireNativeNetwork=$true; CommandListBackend=$true;

scripts/windows/test-windows-rdna4.ps1

@@ -42,7 +42,7 @@ $PackageRoot = [IO.Path]::GetFullPath($PackageRoot)
 if (!$HipRoot) { $HipRoot = 'C:\Program Files\AMD\ROCm\7.2' }
 $HipRoot = [IO.Path]::GetFullPath($HipRoot)
 if (!$ZludaRoot) {
-    $builtZluda = Join-Path $repoRoot 'dist/zluda-windows-native'
+    $builtZluda = Join-Path $repoRoot 'dist/zluda-windows-final'
     $localZluda = [IO.Path]::GetFullPath((Join-Path $PackageRoot '../../.tools/zluda/zluda'))
     if (Test-Path (Join-Path $builtZluda 'build-info.json')) { $ZludaRoot = $builtZluda }
     elseif (Test-Path (Join-Path $localZluda 'nvcuda.dll')) { $ZludaRoot = $localZluda }

tools/windows/diagnostic.h

@@ -128,6 +128,7 @@ struct Args {
     unsigned ngx_create_flags = 0;
     unsigned output_width = 512, output_height = 288;
     std::string fixture_dir;
+    std::string benchmark_module;
     std::string output_dir;
     std::string kernel_name = "enc1";
     unsigned iterations = 32;
@@ -184,6 +185,7 @@ struct Args {
                     throw std::runtime_error("preset must be 5 (E), 11 (K) or 13 (M)");
             }
             else if (key == "--fixture-dir") fixture_dir = value;
+            else if (key == "--benchmark-module") benchmark_module = value;
             else if (key == "--output-dir") output_dir = value;
             else if (key == "--kernel-name") kernel_name = value;
             else if (key == "--iterations") {

tools/windows/fp8_pack_probe.hip

@@ -0,0 +1,9 @@
+// Diagnostic only: all finite OCP e4m3 encodings, including signed zero.
+#include "../../kernels/m/swin_common.h"
+struct PackParams { const half_t* input; uint32_t* output; };
+extern "C" __global__ void fp8_pack_probe(PackParams p) {
+    const unsigned lane = blockIdx.x * 32 + threadIdx.x;
+    if (lane >= 32) return;
+    const auto packed = to_fp8(*(const sop*)(p.input + lane * 8));
+    p.output[lane * 2] = packed[0]; p.output[lane * 2 + 1] = packed[1];
+}

tools/windows/fp8_pack_validate.py

@@ -0,0 +1,46 @@
+"""Check gfx1201 hardware packing of every finite OCP e4m3 value and signed zero."""
+import argparse
+import hashlib
+import json
+import os
+import pathlib
+import struct
+import subprocess
+
+import numpy as np
+from m_replay_validate import me
+
+
+def main(args):
+    root = args.output_dir.resolve()
+    root.mkdir(parents=True, exist_ok=True)
+    codes = np.arange(256, dtype=np.uint8)
+    codes[(codes & 127) == 127] = 0  # NaN inputs are outside this finite-value packing contract.
+    inputs = me.E4M3[codes].astype(np.float16).tobytes()
+    (root / 'alloc-0.bin').write_bytes(inputs)
+    (root / 'alloc-1.bin').write_bytes(bytes(256))
+    (root / 'args.bin').write_bytes(struct.pack('<QQ', 0x100000000, 0x200000000))
+    (root / 'manifest.txt').write_text('kernel fp8_pack_probe\nlaunch 1 1 1 32 1 1 0\nargs 16\n'
+        'alloc 0 0x100000000 512\npointer 0 0 0\nalloc 1 0x200000000 256\npointer 8 1 0\n')
+    with (root / 'gpu.stdout.log').open('wb') as stdout, (root / 'gpu.stderr.log').open('wb') as stderr:
+        result = subprocess.run([str(args.probe.resolve()), '--hip-root', str(args.hip_root.resolve()),
+            '--module', str(args.module.resolve()), '--fixture-dir', str(root), '--output-dir', str(root / 'output'),
+            '--iterations', '1'], stdout=stdout, stderr=stderr, timeout=90,
+            env=dict(os.environ, D4R_DIAG_DIR=str(root)))
+    summary = dict(passed=False, architecture='gfx1201', finiteEncodings=254, signedZero=True,
+        moduleSha256=hashlib.sha256(args.module.read_bytes()).hexdigest(), exitCode=f'0x{result.returncode & 0xffffffff:08x}')
+    if result.returncode == 0:
+        actual = np.fromfile(root / 'output/alloc-1.bin', np.uint8)
+        summary['passed'] = actual.size == 256 and bool(np.array_equal(actual, codes))
+        if actual.size == 256:
+            summary['differingCodes'] = int(np.count_nonzero(actual != codes))
+    (root / 'summary.json').write_text(json.dumps(summary, indent=2) + '\n')
+    print(('PASS' if summary['passed'] else 'FAIL') + ' FP8_HARDWARE_PACK finite_encodings=254 signed_zero=1 architecture=gfx1201', flush=True)
+    return 0 if summary['passed'] else 1
+
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser(description=__doc__)
+    for name in ('probe', 'module', 'hip-root', 'output-dir'):
+        parser.add_argument('--' + name, type=pathlib.Path, required=True)
+    raise SystemExit(main(parser.parse_args()))

tools/windows/m_native_replay_compare.py

@@ -0,0 +1,109 @@
+"""Replay private M captures and require exact logical outputs against a validated baseline."""
+import argparse
+import hashlib
+import json
+import os
+import pathlib
+import subprocess
+import traceback
+
+import numpy as np
+from m_capture_compare import logical, me, LAYERS
+from m_replay_validate import metrics
+
+
+def sha(path):
+    return hashlib.sha256(path.read_bytes()).hexdigest()
+
+
+def main(args):
+    args.output_dir.mkdir(parents=True, exist_ok=True)
+    summary = dict(passed=False, strict=True, architecture='gfx1201', launches=[],
+                   hipRoot=str(args.hip_root), modules={},
+                   baselineValidation=json.loads(args.baseline_validation.read_text(encoding='utf-8-sig')))
+    try:
+        validation = summary['baselineValidation']
+        if not (validation.get('passed') and validation.get('exact') and validation.get('captured_outputs')):
+            raise RuntimeError('An exact captured-output NumPy validation is required for the baseline')
+        audited = {row['capture'] for row in validation['launches']}
+        captures = []
+        for layer in ('enc1', 'enc2', 'enc3_tube', 'dec2', 'dec1'):
+            paths = sorted(args.capture_dir.glob(f'replay-*-rrlite_{layer}_4x4'))
+            if not paths:
+                raise RuntimeError(f'Baseline has no {layer} captures')
+            captures.extend((layer, path) for path in paths)
+        for layer, capture in captures:
+            if capture.name not in audited:
+                raise RuntimeError(f'Capture was not included in exact baseline validation: {capture.name}')
+            module = args.module_dir / f'rrlite_{layer}_4x4.hsaco'
+            summary['modules'][module.name] = sha(module)
+            output = args.output_dir / capture.name
+            output.mkdir(exist_ok=True)
+            env = dict(os.environ, D4R_DIAG_DIR=str(output.resolve()))
+            with (output / 'gpu.stdout.log').open('wb') as stdout, (output / 'gpu.stderr.log').open('wb') as stderr:
+                process = subprocess.run([str(args.probe.resolve()), '--hip-root', str(args.hip_root.resolve()),
+                    '--module', str(module.resolve()), '--fixture-dir', str(capture.resolve()),
+                    '--output-dir', str((output / 'gpu-output').resolve()), '--iterations', '1'],
+                    stdout=stdout, stderr=stderr, env=env, timeout=90)
+            row = dict(capture=capture.name, layer=layer, exitCode=f'0x{process.returncode & 0xffffffff:08x}', fields=[])
+            summary['launches'].append(row)
+            if process.returncode:
+                raise RuntimeError(f'GPU replay failed: {capture.name} ({row["exitCode"]})')
+            p = me.Params(str(capture))
+            channels, _, merged, _ = LAYERS[layer]
+            fields = [('output', p.out, p.W * p.H * channels)]
+            if merged:
+                fields.append(('merge', p.p48, (p.W // 2) * (p.H // 2) * merged))
+            for name, pointer, count in fields:
+                reference = logical(p, pointer, count, True, capture)
+                actual = logical(p, pointer, count, True, output)
+                if reference.size != count or actual.size != count:
+                    raise RuntimeError(f'Truncated logical {name}: {capture.name}')
+                different = int(np.count_nonzero(reference != actual))
+                x, y = me.E4M3[reference], me.E4M3[actual]
+                maximum, _, psnr, _ = metrics(x, y, f'M_NATIVE_COMPARE capture={capture.name} field={name}')
+                row['fields'].append(dict(name=name, elements=count, differingCodes=different,
+                    maxAbs=maximum, referenceSha256=hashlib.sha256(reference.tobytes()).hexdigest(),
+                    actualSha256=hashlib.sha256(actual.tobytes()).hexdigest()))
+                if different:
+                    raise RuntimeError(f'Exact baseline mismatch: {capture.name} {name}, codes={different}, PSNR={psnr}')
+            row['passed'] = True
+        if args.benchmark_control_dir:
+            summary['benchmarks'] = []
+            for layer in ('enc1', 'enc2', 'enc3_tube', 'dec2', 'dec1'):
+                capture = next(path for item, path in captures if item == layer)
+                output = args.output_dir / ('paired-' + layer)
+                output.mkdir(exist_ok=True)
+                control = args.benchmark_control_dir / f'rrlite_{layer}_4x4.hsaco'
+                candidate = args.module_dir / control.name
+                with (output / 'paired.stdout.log').open('wb') as stdout, (output / 'paired.stderr.log').open('wb') as stderr:
+                    process = subprocess.run([str(args.probe.resolve()), '--hip-root', str(args.hip_root.resolve()),
+                        '--module', str(control.resolve()), '--benchmark-module', str(candidate.resolve()),
+                        '--fixture-dir', str(capture.resolve()), '--output-dir', str((output / 'output').resolve()),
+                        '--iterations', '64'], stdout=stdout, stderr=stderr, timeout=90,
+                        env=dict(os.environ, D4R_DIAG_DIR=str(output.resolve())))
+                if process.returncode:
+                    raise RuntimeError(f'Paired benchmark failed: {layer}, exit={process.returncode:#x}')
+                from profile_report import report
+                measured = report(output)
+                if len(measured['replayPairs']) != 1 or measured['replayPairs'][0]['controlGpuMs']['samples'] != 64:
+                    raise RuntimeError(f'Incomplete paired measurements: {layer}')
+                summary['benchmarks'].extend(measured['replayPairs'])
+                print(f'PAIRED M_NATIVE_COMPARE layer={layer} samples=64 logs={output}', flush=True)
+        summary['passed'] = True
+        print(f'PASS M_NATIVE_COMPARE launches={len(captures)} exact=1 finite=1 architecture=gfx1201', flush=True)
+        return 0
+    except Exception as error:
+        summary['error'] = str(error)
+        traceback.print_exc()
+        return 1
+    finally:
+        (args.output_dir / 'summary.json').write_text(json.dumps(summary, indent=2) + '\n', encoding='utf-8')
+
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser(description=__doc__)
+    for name in ('capture-dir', 'module-dir', 'probe', 'hip-root', 'baseline-validation', 'output-dir'):
+        parser.add_argument('--' + name, type=pathlib.Path, required=True)
+    parser.add_argument('--benchmark-control-dir', type=pathlib.Path)
+    raise SystemExit(main(parser.parse_args()))

tools/windows/native_replay_probe.cpp

@@ -14,8 +14,9 @@ struct Allocation {
     int id;
     size_t size;
     void* device = nullptr;
+    std::vector<uint8_t> initial;
     Allocation(d4r::diag::HipApi& api, int identifier, const std::vector<uint8_t>& bytes)
-        : hip(api), id(identifier), size(bytes.size()) {
+        : hip(api), id(identifier), size(bytes.size()), initial(bytes) {
         hip.check(hip.hipMalloc(&device, size), "hipMalloc(replay)");
         try { hip.check(hip.hipMemcpy(device, bytes.data(), size, hipMemcpyHostToDevice), "hipMemcpy(replay upload)"); }
         catch (...) { (void)hip.hipFree(device); device = nullptr; throw; }
@@ -90,14 +91,15 @@ int main(int argc, char** argv) {
             if (size != sizeof(coordinates)) throw std::runtime_error("Swin diagnostic coordinates ABI");
             hip.check(hip.hipMemcpy(address, coordinates, size, hipMemcpyHostToDevice), "Set Swin diagnostic block");
         }
-        auto scalar = [&](const char* name, unsigned fallback) {
+        auto module_scalar = [&](hipModule_t owner, const char* name, unsigned fallback) {
             hipDeviceptr_t address = nullptr; size_t size = 0;
-            if (hip.hipModuleGetGlobal(&address, &size, module, name) != hipSuccess) return fallback;
+            if (hip.hipModuleGetGlobal(&address, &size, owner, name) != hipSuccess) return fallback;
             if (size != 4) throw std::runtime_error(std::string("Invalid native metadata ") + name);
             unsigned value = 0;
             hip.check(hip.hipMemcpy(&value, address, 4, hipMemcpyDeviceToHost), name);
             return value ? value : fallback;
         };
+        auto scalar = [&](const char* name, unsigned fallback) { return module_scalar(module, name, fallback); };
         block[2] = scalar("d4r_block_z", block[2]);
         const unsigned persistentGrid = scalar("d4r_grid_x", 0);
         if (persistentGrid) { grid[0] = persistentGrid; grid[1] = grid[2] = 1; }
@@ -111,7 +113,65 @@ int main(int argc, char** argv) {
         }
         std::printf("REPLAY kernel=%s grid=%u,%u,%u block=%u,%u,%u allocations=%zu arguments=%zu\n",
             kernel.c_str(), grid[0], grid[1], grid[2], block[0], block[1], block[2], allocations.size(), parameters.size());
-        for (unsigned iteration = 0; iteration < args.iterations; ++iteration)
+        if (!args.benchmark_module.empty()) {
+            hipModule_t candidate = nullptr;
+            hip.check(hip.hipModuleLoad(&candidate, args.benchmark_module.c_str()), "Load paired benchmark candidate");
+            Cleanup candidate_cleanup{hip, candidate};
+            if (module_scalar(candidate, "d4r_block_z", block[2]) != block[2] ||
+                module_scalar(candidate, "d4r_grid_x", 0) != persistentGrid)
+                throw std::runtime_error("Paired modules require identical launch geometry");
+            hipFunction_t other = nullptr, other_prep = nullptr;
+            hip.check(hip.hipModuleGetFunction(&other, candidate, kernel.c_str()), "Find paired benchmark function");
+            if (hip.hipModuleGetFunction(&other_prep, candidate, (kernel + "_prep").c_str()) == hipSuccess) {
+                const auto blocks = module_scalar(candidate, "d4r_prep_blocks", 0);
+                if (!blocks) throw std::runtime_error("Paired candidate prep metadata is missing");
+                hip.check(hip.hipModuleLaunchKernel(other_prep, blocks, 1, 1, 128, 1, 1, 0, nullptr,
+                    launchArguments, nullptr), "Paired candidate prep");
+            }
+            auto create = hip.library.symbol<decltype(&::hipEventCreate)>("hipEventCreate");
+            auto destroy = hip.library.symbol<decltype(&::hipEventDestroy)>("hipEventDestroy");
+            auto record = hip.library.symbol<decltype(&::hipEventRecord)>("hipEventRecord");
+            auto synchronize = hip.library.symbol<decltype(&::hipEventSynchronize)>("hipEventSynchronize");
+            auto elapsed = hip.library.symbol<decltype(&::hipEventElapsedTime)>("hipEventElapsedTime");
+            struct Event {
+                decltype(destroy) release; hipEvent_t value = nullptr;
+                ~Event() { if (value) (void)release(value); }
+            } begin{destroy}, end{destroy};
+            hip.check(create(&begin.value), "Create paired start event");
+            hip.check(create(&end.value), "Create paired end event");
+            auto launch = [&](hipFunction_t fn) {
+                hip.check(hip.hipModuleLaunchKernel(fn, grid[0], grid[1], grid[2], block[0], block[1], block[2],
+                    shared, nullptr, launchArguments, nullptr), "Paired benchmark launch");
+            };
+            // Both modules and weight images stay resident on the same GPU.
+            // Warm continuous work first, then alternate AB/BA order. Restore
+            // captured inputs outside the event so neither module consumes
+            // modified temporal inputs from a preceding sample.
+            const auto verbose = hip.verbose; hip.verbose = false;
+            for (unsigned warm = 0; warm < 16; ++warm)
+                for (auto fn : {function, other}) for (unsigned batch = 0; batch < 16; ++batch) launch(fn);
+            hip.check(hip.hipDeviceSynchronize(), "Paired benchmark warmup");
+            for (unsigned pair = 0; pair < args.iterations; ++pair) for (unsigned order = 0; order < 2; ++order) {
+                const unsigned variant = order ^ (pair & 1u);
+                for (const auto& allocation : allocations)
+                    hip.check(hip.hipMemcpy(allocation->device, allocation->initial.data(), allocation->size,
+                        hipMemcpyHostToDevice), "Restore paired captured input");
+                hip.check(record(begin.value, nullptr), "Record paired start");
+                launch(variant ? other : function);
+                hip.check(record(end.value, nullptr), "Record paired end");
+                hip.check(synchronize(end.value), "Complete paired sample");
+                float ms = 0; hip.check(elapsed(&ms, begin.value, end.value), "Paired GPU elapsed time");
+                std::printf("D4R_REPLAY_PROFILE kernel=%s variant=%s pair=%u gpu_ms=%.6f warmup_launches=512 paired=1 input_restore_outside_event=1\n",
+                    kernel.c_str(), variant ? "candidate" : "control", pair, double(ms));
+            }
+            hip.verbose = verbose;
+            // Leave a single ordinary control replay for numerical checking.
+            for (const auto& allocation : allocations)
+                hip.check(hip.hipMemcpy(allocation->device, allocation->initial.data(), allocation->size,
+                    hipMemcpyHostToDevice), "Restore final reference input");
+            launch(function);
+        }
+        for (unsigned iteration = 0; args.benchmark_module.empty() && iteration < args.iterations; ++iteration)
             hip.check(hip.hipModuleLaunchKernel(function, grid[0], grid[1], grid[2], block[0], block[1], block[2], shared,
                 nullptr, launchArguments, nullptr), "hipModuleLaunchKernel(replay)");
         hip.check(hip.hipDeviceSynchronize(), "hipDeviceSynchronize(replay)");
@@ -132,7 +192,8 @@ int main(int argc, char** argv) {
             file.write(reinterpret_cast<const char*>(bytes.data()), std::streamsize(bytes.size()));
             if (!file) throw std::runtime_error("Cannot save replay allocation");
         }
-        std::printf("PASS NATIVE_REPLAY architecture=gfx1201 kernel=%s iterations=%u numerical_reference=pending\n", kernel.c_str(), args.iterations);
+        std::printf("PASS NATIVE_REPLAY architecture=gfx1201 kernel=%s iterations=%u benchmark_pairs=%u numerical_reference=pending\n",
+            kernel.c_str(), args.benchmark_module.empty() ? args.iterations : 1, args.benchmark_module.empty() ? 0 : args.iterations);
         return 0;
     } catch (const std::exception& error) {
         std::fprintf(stderr, "FAIL NATIVE_REPLAY %s\n", error.what()); loaded_modules(); return 4;

tools/windows/profile_report.py

@@ -42,8 +42,13 @@ def report(directory, metadata_directory=None):
     commands = collections.defaultdict(list)
     kernels = collections.defaultdict(list)
     hook_threads = collections.defaultdict(list)
+    replay = collections.defaultdict(lambda: collections.defaultdict(dict))
     for path in sorted(directory.glob('*.log')):
-        text = path.read_text(encoding='utf-8-sig', errors='replace')
+        data = path.read_bytes()
+        text = data.decode('utf-16' if data.startswith((b'\xff\xfe', b'\xfe\xff')) else 'utf-8-sig', errors='replace')
+        for name, variant, pair, value in re.findall(
+                r'D4R_REPLAY_PROFILE kernel=(\S+) variant=(control|candidate) pair=([0-9]+) gpu_ms=([0-9.]+)', text):
+            replay[name][int(pair)][variant] = float(value)
         for name, value in re.findall(r'D4R_STAGE name=(\S+) cpu_ms=([0-9.]+)', text):
             stages[name].append(float(value))
         for kind, fields in re.findall(r'(?m)^D4R_COMMAND_(RECORD|SUBMIT) ([^\r\n]+)', text):
@@ -64,6 +69,7 @@ def report(directory, metadata_directory=None):
     result = dict(cpuStages={name: stats(values) for name, values in stages.items()},
                   commandStages={name: stats(values) for name, values in commands.items()},
                   commandHookSamples=[],
+                  replayPairs=[],
                   kernels=[], notes=[
                       'CPU stage times include waits and host work; they are not isolated GPU timings.',
                       'Command record intervals are between NGX recording calls on one thread, not Present/FPS.',
@@ -72,6 +78,13 @@ def report(directory, metadata_directory=None):
                       'HIP-event profiling synchronizes every sampled launch and changes scheduling.',
                       'Occupancy is an API prediction, not a measured hardware counter.',
                       'Bandwidth and WMMA utilization require additional supported hardware tooling.'])
+    for name, pairs in replay.items():
+        complete = [row for row in pairs.values() if set(row) == {'control', 'candidate'}]
+        if complete:
+            result['replayPairs'].append(dict(kernel=name,
+                controlGpuMs=stats([row['control'] for row in complete]),
+                candidateGpuMs=stats([row['candidate'] for row in complete]),
+                candidateToControl=stats([row['candidate'] / row['control'] for row in complete if row['control'] > 0])))
     for thread, rows in hook_threads.items():
         sums = {name: sum(float(row[name]) for row in rows) for name in
                 ('period_ms', 'calls', 'samples', 'marked', 'access_sum_ms', 'capture_sum_ms', 'driver_sum_ms')}
@@ -105,7 +118,7 @@ def main():
     parser.add_argument('--metadata-directory', type=pathlib.Path, help='llvm-readobj --notes output for native code objects')
     args = parser.parse_args()
     result = report(args.directory, args.metadata_directory)
-    if not result['cpuStages'] and not result['kernels']:
+    if not any(result[key] for key in ('cpuStages', 'kernels', 'commandStages', 'commandHookSamples', 'replayPairs')):
         parser.error('no complete stage or HIP-event records found')
     if args.output:
         args.output.write_text(json.dumps(result, indent=2) + '\n', encoding='utf-8')
@@ -116,6 +129,9 @@ def main():
     for row in result['commandHookSamples']:
         print(f"HOOK thread={row['thread']} samples={int(row['samples'])} access={row['accessMeanUs']:.3f} us "
               f"capture={row['captureMeanUsPerSample']:.3f} us estimated_cpu={row['estimatedAccessCaptureMsPerSecond']:.3f} ms/s")
+    for row in result['replayPairs']:
+        print(f"PAIRED {row['kernel']}: n={row['controlGpuMs']['samples']} control_median={row['controlGpuMs']['median']:.6f} ms "
+              f"candidate_median={row['candidateGpuMs']['median']:.6f} ms ratio_median={row['candidateToControl']['median']:.6f}")
     for row in result['kernels']:
         times = row['gpuMs']
         print(f"GPU {row['kernel']} {row['backend']} {row['phase']}: n={times['samples']} mean={times['mean']:.6f} ms total={times['total']:.3f} ms")