Remove game log backpressure and measure Windows K/M kernels

docs/windows-game.md

@@ -46,6 +46,18 @@ the GPU before returning it to D3D12. Only two diagnostic counters (8 bytes)
 are read by the CPU; image data remains in VRAM. Validation failures are
 reported explicitly. Disable this diagnostic when measuring production speed.
 
+`-ProfileStages` records CPU completion time for seven interop/NGX stages.
+The runner drains stdout/stderr concurrently as raw bytes; the earlier
+line-oriented PowerShell collector imposed large per-frame delays. The optional
+ZLUDA kernel profiler is for diagnostics only. Measurements and reproduction
+commands are in [windows-performance.md](windows-performance.md).
+
+Ordinary runs use OptiScaler info logging and suppress successful HIP/CUDA API
+messages. Errors, GPU identity, imports, frame records and native/translated
+kernel launches remain logged. `-VerboseRuntime` enables all API and frontend
+trace messages when investigating a failure. Numerical relaxation flags are
+cleared in the child process; FP16 rounding and subnormals stay enabled.
+
 Original files are backed up before replacement. Restore with:
 
 ```powershell

docs/windows-performance.md

@@ -0,0 +1,77 @@
+# Windows RDNA4 measurements
+
+Hardware: RX 9070 XT / gfx1201, Windows 11, display driver 32.0.31041.1004.
+Runtime: TheRock 10.2.0a20260929, HIP 7.17.26386; ZLUDA b0161a4 with
+`D4R_ZLUDA_WMMA=1`, FP8 widening, native FP8 disabled, f16 reference rounding.
+
+The optional profiler uses HIP events on the actual launch stream, including
+separate measurements of one-time weight preparation. It synchronizes each
+sampled launch and changes scheduling: these numbers are not application FPS.
+HIP occupancy predictions are not measured hardware occupancy. The native
+objects' actual wave32, LDS, VGPR/SGPR and spill counts are available in
+`llvm-readobj --notes` and can be attached to the JSON report. Memory bandwidth
+and WMMA utilization were not measured; no utilization is inferred from timing.
+
+The identical four-frame standalone D3D12/OptiScaler workloads output 512x288.
+Summing the measured main-kernel event times and dividing by four gives:
+
+| Preset | Fully translated, ms/frame | Native FP16-equivalent, ms/frame |
+| --- | ---: | ---: |
+| K | 157.745441 | 0.684675 |
+| M | 7.428538 | 1.566500 |
+
+Both rows include translated input/output tails. Native overrides cover all
+11 K layers and five M layers; M's tube layer executes six times per frame.
+M native and translated RGB are bit-exact, with max absolute/relative error 0.
+K native remains bit-exact against its independently validated native control;
+fully translated K has the previously documented arithmetic differences.
+All outputs are finite. These are reference-correct codegen settings, not
+measurements of a relaxed fast-math translated configuration.
+
+M's first profiled Evaluate includes about 0.9 seconds of host initialization;
+the four-sample CPU mean must not be treated as steady-state latency. Its
+median Evaluate host completion is 4.855 ms. GPU events and prep phases are
+reported separately in the raw JSON.
+
+The initial game runner used PowerShell line-oriented output redirection.
+It heavily throttled verbose stdout/stderr. The runner now concurrently drains
+raw byte streams using the same approach as the standalone diagnostics.
+
+| Silent Hill 2 stage, CPU mean | Initial capture, ms | Raw-byte capture with HIP profiler, ms |
+| --- | ---: | ---: |
+| Input fence wait | 5.641 | 0.833 |
+| Input conversion and CUDA array upload | 13.196 | 0.668 |
+| NGX Evaluate host call | 25.715 | 3.350 |
+| NGX all-stream completion | 1.846 | 0.033 |
+| Output CUDA array download | 4.288 | 0.167 |
+| Output conversion and shared fence signal | 1.500 | 0.033 |
+
+The first run additionally scanned every output (10.142 ms in that throttled
+capture); the kernel-profiling run disables that scan. Both use a temporary
+1280x720 window and current-frame VRAM interop. These are separate menu/game
+runs, not a controlled whole-game FPS comparison. No synchronization was
+removed to obtain the improvement. The raw-byte profiling run completed
+2345 K frames, with no d4r failure, CPU image copy or previous-frame output.
+
+Reproduce native or translated measurements with locally supplied NVIDIA DLLs
+at the repository root and the separately built profiling runtime:
+
+```powershell
+.\scripts\windows\profile-windows-rdna4.ps1 -Preset 11
+.\scripts\windows\profile-windows-rdna4.ps1 -Preset 11 -Translated
+.\scripts\windows\profile-windows-rdna4.ps1 -Preset 13
+.\scripts\windows\profile-windows-rdna4.ps1 -Preset 13 -Translated
+```
+
+`profile_report.py <results-directory> --output profile.json` summarizes raw
+CPU stages and every kernel's event time, launch geometry, registers, private
+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.
+
+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

@@ -9,6 +9,23 @@ validation coverage are recorded below. Performance remains in progress.
 
 ## Milestone and gates
 
+Performance diagnostics (2026-10-01): the separate ZLUDA `b0161a4` runtime
+passes 32 PTX launches in each context and all four native K/M frames with
+bit-exact RGB against their controls. All thirteen exported patches (0002
+through 0014) apply to the pinned fresh base. The full Windows build and all
+16 CTest gates pass. Actual K profiling completes 2345 frames with zero backend
+errors. Replacing PowerShell's line-oriented stdout/stderr collection with
+concurrent raw byte copies removes substantial diagnostic backpressure:
+mean Evaluate host time drops from 25.715 to 3.350 ms despite per-kernel HIP
+event synchronization. Native/translated GPU times, scope and limitations are
+in [windows-performance.md](windows-performance.md). Quiet successful API
+logging is separately selectable via `D4R_QUIET_API=1`; error checks stay active.
+
+An additional K game run (`silent-hill2-k-output-stages.zip`) completes 1688
+frames / 18568 native launches / 1688 GPU output scans, all finite, with no
+backend failure or previous-frame output. The user confirms visible imagery
+after the depth-plane fix; low FPS is the current investigation.
+
 Actual K rendering: `silent-hill2-k-depth-plane.zip` completes **1794 DLSS
 frames / 19734 native K launches / zero d4r failures** at 753x424 -> 1280x720
 with the game's depth-inverted HDR path and DXGI-19 depth plane. The captured

patches/zluda/0014-windows-kernel-profiling.patch

@@ -0,0 +1,148 @@
+From b0161a4166652eee86a8113a38d042b02403db6e Mon Sep 17 00:00:00 2001
+From: Codex <codex@localhost>
+Date: Thu, 1 Oct 2026 18:07:07 +0300
+Subject: [PATCH] Add opt-in Windows HIP event and occupancy diagnostics
+
+---
+ zluda/src/impl/d4r_profile.rs | 60 +++++++++++++++++++++++++++++++++++
+ zluda/src/impl/function.rs    |  9 ++++++
+ zluda/src/impl/mod.rs         |  1 +
+ zluda/src/impl/module.rs      |  2 +-
+ 4 files changed, 71 insertions(+), 1 deletion(-)
+ create mode 100644 zluda/src/impl/d4r_profile.rs
+
+diff --git a/zluda/src/impl/d4r_profile.rs b/zluda/src/impl/d4r_profile.rs
+new file mode 100644
+index 0000000..33ba5f6
+--- /dev/null
++++ b/zluda/src/impl/d4r_profile.rs
+@@ -0,0 +1,60 @@
++//! Optional, deliberately serializing HIP-event diagnostics. Not a production
++//! scheduler or a claim about application FPS / measured hardware occupancy.
++use hip_runtime_sys::*;
++use std::{ffi::CStr, ptr, time::Instant};
++
++pub(super) struct Launch {
++    start: hipEvent_t,
++    end: hipEvent_t,
++    host: Instant,
++}
++
++impl Drop for Launch {
++    fn drop(&mut self) {
++        unsafe {
++            if !self.end.is_null() { let _ = hipEventDestroy(self.end); }
++            if !self.start.is_null() { let _ = hipEventDestroy(self.start); }
++        }
++    }
++}
++
++impl Launch {
++    pub(super) unsafe fn begin(stream: hipStream_t) -> Result<Option<Self>, hipErrorCode_t> {
++        if std::env::var_os("D4R_ZLUDA_PROFILE").is_none() { return Ok(None); }
++        let mut result = Self { start: ptr::null_mut(), end: ptr::null_mut(), host: Instant::now() };
++        hipEventCreateWithFlags(&mut result.start, 0)?;
++        hipEventCreateWithFlags(&mut result.end, 0)?;
++        hipEventRecord(result.start, stream)?;
++        result.host = Instant::now();
++        Ok(Some(result))
++    }
++
++    pub(super) unsafe fn finish(self, function: hipFunction_t, name: &CStr, backend: &str,
++        phase: &str, grid: [u32;3], block: [u32;3], shared: u32, stream: hipStream_t) -> hipError_t {
++        let enqueue_ms = self.host.elapsed().as_secs_f64() * 1000.;
++        hipEventRecord(self.end, stream)?;
++        hipEventSynchronize(self.end)?;
++        let completion_ms = self.host.elapsed().as_secs_f64() * 1000.;
++        let mut gpu_ms = 0.;
++        hipEventElapsedTime(&mut gpu_ms, self.start, self.end)?;
++        // Module function handles use the Driver API attribute query. The
++        // plural Runtime API expects a registered host function address.
++        let mut lds = 0;
++        let mut regs = 0;
++        let mut private = 0;
++        let mut max_threads = 0;
++        hipFuncGetAttribute(&mut lds, hipFunction_attribute::HIP_FUNC_ATTRIBUTE_SHARED_SIZE_BYTES, function)?;
++        hipFuncGetAttribute(&mut regs, hipFunction_attribute::HIP_FUNC_ATTRIBUTE_NUM_REGS, function)?;
++        hipFuncGetAttribute(&mut private, hipFunction_attribute::HIP_FUNC_ATTRIBUTE_LOCAL_SIZE_BYTES, function)?;
++        hipFuncGetAttribute(&mut max_threads, hipFunction_attribute::HIP_FUNC_ATTRIBUTE_MAX_THREADS_PER_BLOCK, function)?;
++        let threads = block.iter().try_fold(1u32, |a, b| a.checked_mul(*b))
++            .filter(|n| *n <= i32::MAX as u32).ok_or(hipErrorCode_t::InvalidValue)?;
++        let mut blocks_per_multiprocessor = 0;
++        hipModuleOccupancyMaxActiveBlocksPerMultiprocessor(&mut blocks_per_multiprocessor,
++            function, threads as i32, shared as usize)?;
++        eprintln!("D4R_KERNEL_PROFILE kernel={name:?} backend={backend} phase={phase} gpu_ms={gpu_ms:.6} enqueue_ms={enqueue_ms:.6} completion_ms={completion_ms:.6} grid={},{},{} block={},{},{} static_lds_bytes={} dynamic_lds_bytes={shared} regs={} private_bytes={} max_threads={} predicted_blocks_per_multiprocessor={blocks_per_multiprocessor} serializing=1",
++            grid[0], grid[1], grid[2], block[0], block[1], block[2], lds,
++            regs, private, max_threads);
++        Ok(())
++    }
++}
+diff --git a/zluda/src/impl/function.rs b/zluda/src/impl/function.rs
+index 7a7470a..c456dfe 100644
+--- a/zluda/src/impl/function.rs
++++ b/zluda/src/impl/function.rs
+@@ -191,6 +191,7 @@ pub(crate) fn launch_kernel(
+         }
+     };
+     if let (Some((prep, blocks)), true) = (f.prep, run_prep) {
++        let profile = unsafe { super::d4r_profile::Launch::begin(stream) }?;
+         unsafe {
+             hipModuleLaunchKernel(
+                 prep,
+@@ -206,6 +207,9 @@ pub(crate) fn launch_kernel(
+                 extra_for_hip,
+             )
+         }?;
++        if let Some(profile) = profile {
++            unsafe { profile.finish(prep, &f.name, "native", "prep", [blocks,1,1], [128,1,1], 0, stream) }?;
++        }
+         f.prep_last.store(prep_key, std::sync::atomic::Ordering::Relaxed);
+     }
+     let block_dim_z = if f.block_z != 0 { f.block_z } else { block_dim_z };
+@@ -214,6 +218,7 @@ pub(crate) fn launch_kernel(
+         eprintln!("[d4r-launch] kernel={:?} backend={} grid={grid_dim_x},{grid_dim_y},{grid_dim_z} block={block_dim_x},{block_dim_y},{block_dim_z}",
+             f.name, if f.native_owner.is_some() { "native" } else { "translated" });
+     }
++    let profile = unsafe { super::d4r_profile::Launch::begin(stream) }?;
+     let result = unsafe {
+         hipModuleLaunchKernel(
+             f.base,
+@@ -230,6 +235,10 @@ pub(crate) fn launch_kernel(
+         )
+     };
+     result?;
++    if let Some(profile) = profile {
++        unsafe { profile.finish(f.base, &f.name, if f.native_owner.is_some() { "native" } else { "translated" },
++            "main", [grid_dim_x, grid_dim_y, grid_dim_z], [block_dim_x, block_dim_y, block_dim_z], shared_mem_bytes, stream) }?;
++    }
+     if let Some(replay) = replay { replay.finish(stream)?; }
+     Ok(())
+ }
+diff --git a/zluda/src/impl/mod.rs b/zluda/src/impl/mod.rs
+index 14e0aee..380eb6e 100644
+--- a/zluda/src/impl/mod.rs
++++ b/zluda/src/impl/mod.rs
+@@ -6,6 +6,7 @@ pub(super) mod context;
+ pub(super) mod device;
+ pub(super) mod driver;
+ pub(super) mod d4r_replay;
++pub(super) mod d4r_profile;
+ pub(super) mod event;
+ pub(super) mod function;
+ pub(super) mod graph;
+diff --git a/zluda/src/impl/module.rs b/zluda/src/impl/module.rs
+index 5a79368..232f7d4 100644
+--- a/zluda/src/impl/module.rs
++++ b/zluda/src/impl/module.rs
+@@ -502,7 +502,7 @@ fn zluda_build_version() -> &'static str {
+                 key.starts_with("D4R_ZLUDA_")
+                     && !matches!(
+                         key.as_str(),
+-                        "D4R_ZLUDA_NATIVE_DIR" | "D4R_ZLUDA_DIR" | "D4R_ZLUDA_DUMP_DIR" | "D4R_ZLUDA_DUMP_PTX_DIR" | "D4R_ZLUDA_LIBCUDA" | "D4R_ZLUDA_VERBOSE"
++                        "D4R_ZLUDA_NATIVE_DIR" | "D4R_ZLUDA_DIR" | "D4R_ZLUDA_DUMP_DIR" | "D4R_ZLUDA_DUMP_PTX_DIR" | "D4R_ZLUDA_LIBCUDA" | "D4R_ZLUDA_VERBOSE" | "D4R_ZLUDA_PROFILE"
+                             | "D4R_ZLUDA_WMMA_FP8_NATIVE" // keyed by effect (ExtraCacheAttributes)
+                     )
+             })
+-- 
+2.55.0.windows.5
+

scripts/windows/capture-game-window.ps1

@@ -15,6 +15,7 @@ public static class D4RWindowCapture {
     [DllImport("user32.dll")] public static extern bool GetWindowRect(IntPtr window, out Rect rect);
     [DllImport("user32.dll",CharSet=CharSet.Unicode)] public static extern int GetWindowText(IntPtr window, StringBuilder text, int count);
     [DllImport("user32.dll")] public static extern bool IsWindowVisible(IntPtr window);
+    [DllImport("user32.dll")] public static extern IntPtr GetForegroundWindow();
     [DllImport("user32.dll")] public static extern IntPtr SetThreadDpiAwarenessContext(IntPtr context);
 }
 '@
@@ -39,6 +40,9 @@ $callback = [D4RWindowCapture+Callback]{ param($window,$data)
 $windows | Select-Object handle,title,visible,area | Format-Table
 $chosen = $windows | Where-Object { $_.visible -and $_.area -gt 4096 } | Sort-Object area -Descending | Select-Object -First 1
 if (!$chosen) { throw "No visible game window for PID $GameProcessId" }
+if ([D4RWindowCapture]::GetForegroundWindow() -ne $chosen.handle) {
+    throw 'Game window is not foreground; capture refused to avoid including unrelated overlapping windows.'
+}
 $rect = $chosen.rect
 $bitmap = [Drawing.Bitmap]::new($rect.Right-$rect.Left,$rect.Bottom-$rect.Top)
 $graphics = [Drawing.Graphics]::FromImage($bitmap)

scripts/windows/package-windows-game.ps1

@@ -49,6 +49,7 @@ Stage (Join-Path $repo 'scripts/windows/windows-game.ps1') 'windows-game.ps1'
 Stage (Join-Path $repo 'scripts/windows/capture-game-window.ps1') 'capture-game-window.ps1'
 Stage (Join-Path $repo 'scripts/windows/send-game-key.ps1') 'send-game-key.ps1'
 Stage (Join-Path $repo 'docs/windows-game.md') 'README.md'
+Stage (Join-Path $repo 'docs/windows-performance.md') 'windows-performance.md'
 Stage (Join-Path $repo 'LICENSE') 'licenses/d4r.txt'
 Stage (Join-Path $OptiScalerRoot 'GPL-3.0.txt') 'licenses/OptiScaler.txt'
 Stage (Join-Path $repo 'external/ZLUDA/LICENSE-MIT') 'licenses/ZLUDA-MIT.txt'
@@ -62,8 +63,9 @@ foreach ($patch in Get-ChildItem -LiteralPath (Join-Path $repo 'patches/zluda')
 $metadata = [ordered]@{
     architecture='gfx1201'; nativeKernels=16; dlssSha256='3975567b8943c53acce397f2b72380092f84f162d00b0d2c7d08a1025c563983';
     d4rCommit=(& git -C $repo rev-parse HEAD).Trim();
+    d4rWorkingTreeDirty=[bool](& git -C $repo status --porcelain);
     optiScaler=(Get-Content -LiteralPath (Join-Path $OptiScalerRoot 'build-info.json') -Raw | ConvertFrom-Json);
-    zludaCommit=(& git -C (Join-Path $repo 'external/ZLUDA') rev-parse HEAD).Trim();
+    zludaBuild=(Get-Content -LiteralPath (Join-Path $ZludaRoot 'build-info.json') -Raw | ConvertFrom-Json);
     hipRuntime='TheRock 10.2.0a20260929 / HIP 7.17.26386';
     files=@($stagedFiles | ForEach-Object { Get-Item -LiteralPath (Join-Path $PackageRoot $_) } | ForEach-Object {
         @{path=$_.FullName.Substring($PackageRoot.Length+1); sha256=(Get-FileHash -LiteralPath $_.FullName -Algorithm SHA256).Hash; version=$_.VersionInfo.FileVersion}

scripts/windows/profile-windows-rdna4.ps1

@@ -0,0 +1,30 @@
+[CmdletBinding()]
+param(
+    [ValidateSet(11,13)][int]$Preset = 11,
+    [string]$ZludaRoot, [string]$PackageRoot, [string]$OutputDirectory,
+    [int]$Iterations = 4,
+    [switch]$Translated
+)
+$ErrorActionPreference='Stop'
+$repo=[IO.Path]::GetFullPath((Join-Path $PSScriptRoot '../..'))
+if (!$ZludaRoot) { $ZludaRoot=Join-Path $repo 'dist/zluda-windows-profile' }
+if (!$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';
+    D4R_ZLUDA_NATIVE_DIR=$(if ($Translated) { $null } elseif ($Preset -eq 11) { Join-Path $repo 'build/native-k-gfx1201' } else { Join-Path $repo 'build/native-m-gfx1201' })}
+$old=@{}
+try {
+    foreach ($key in $settings.Keys) { $old[$key]=[Environment]::GetEnvironmentVariable($key,'Process'); [Environment]::SetEnvironmentVariable($key,$settings[$key],'Process') }
+    $arguments=@{RuntimeProfile='therock'; ZludaRoot=$ZludaRoot; PackageRoot=$PackageRoot; OutputDirectory=$OutputDirectory;
+        NgxCore=(Join-Path $repo '_nvngx.dll'); DlssDll=(Join-Path $repo 'nvngx_dlss.dll'); NgxMode='d3d12';
+        NgxAbi='project-legacy'; Preset=$Preset; NgxOnly=$true; RequireNativeNetwork=(!$Translated); CommandListBackend=$true;
+        CaptureExceptions=$true; EarlyIndirectProbe=$true; PixelProfile='depth-stencil'; BarrierMode='inherited-legacy';
+        Iterations=$Iterations; TimeoutSeconds=300; OptiScalerDll=(Join-Path $repo 'dist/optiscaler-windows-d4r/OptiScaler.dll')}
+    & (Join-Path $PSScriptRoot 'test-windows-rdna4.ps1') @arguments
+    if ($LASTEXITCODE) { throw 'Profiling workload failed; see diagnostic bundle.' }
+    $python=(Get-Command python.exe -ErrorAction Stop).Source
+    & $python (Join-Path $repo 'tools/windows/profile_report.py') $OutputDirectory --output (Join-Path $OutputDirectory 'profile.json')
+    if ($LASTEXITCODE) { throw 'No usable profile records.' }
+} finally {
+    foreach ($key in $old.Keys) { [Environment]::SetEnvironmentVariable($key,$old[$key],'Process') }
+}

scripts/windows/windows-game.ps1

@@ -8,6 +8,8 @@ param(
     [string]$DiagnosticResolution,
     [string]$CacheDirectory = $env:ZLUDA_CACHE_DIR,
     [switch]$ValidateOutput,
+    [switch]$ProfileStages,
+    [switch]$VerboseRuntime,
     [string[]]$GameArguments = @('-dx12','-windowed','-ResX=1280','-ResY=720'),
     [string]$OutputDirectory
 )
@@ -108,7 +110,7 @@ RestoreGraphicSignature=false
 ExtendedStateRestore=false
 [Log]
 LogToFile=true
-LogLevel=0
+LogLevel=$(if ($VerboseRuntime) { 0 } else { 2 })
 [Menu]
 DisableSplash=true
 "@
@@ -129,11 +131,18 @@ $settings = @{
     D4R_FORMAT_MODULE=(GamePath 'd4r/pixel_convert_gfx1201.hsaco');
     D4R_ZLUDA_NATIVE_DIR=(GamePath 'd4r/native'); D4R_D3D12_COMMAND_BACKEND='1'; D4R_ZLUDA_VERBOSE='1';
     D4R_ZLUDA_WMMA='1'; D4R_ZLUDA_WMMA_FP8='1'; D4R_ZLUDA_WMMA_FP8_NATIVE='0'; D4R_ZLUDA_WMMA_F16_REFERENCE='1';
+    # Do not inherit relaxation flags from unrelated experiments. Unset and
+    # false have the same numerical effect; unset reuses the validated cache.
+    D4R_ZLUDA_IGNORE_DENORMAL=$null; D4R_ZLUDA_FAST_MATH=$null;
+    D4R_ZLUDA_WMMA_F32ACC=$null; D4R_ZLUDA_WAVE64=$null;
+    D4R_QUIET_API=$(if ($VerboseRuntime) { $null } else { '1' });
     D4R_DIAG_DIR=$OutputDirectory; ZLUDA_LOG_DIR=(Join-Path $OutputDirectory 'zluda-trace');
     ZLUDA_CACHE_DIR=$CacheDirectory; PATH=((GamePath 'd4r/hip/bin') + ';' + (GamePath 'd4r/zluda') + ';' + $env:PATH)
 }
 $old = @{}; $process = $null; $originalSettings = @()
+$outFile = $null; $errFile = $null; $outTask = $null; $errTask = $null
 if ($ValidateOutput) { $settings['D4R_VALIDATE_OUTPUT']='1' }
+if ($ProfileStages) { $settings['D4R_PROFILE_STAGES']='1' }
 try {
     if ($DiagnosticResolution) {
         if ([IO.Path]::GetFileName($GameExe) -ne 'SHProto-Win64-Shipping.exe' -or $DiagnosticResolution -notmatch '^([0-9]{3,4})x([0-9]{3,4})$') { throw 'DiagnosticResolution currently supports Silent Hill 2 only; use e.g. 1280x720.' }
@@ -168,14 +177,28 @@ try {
         ConvertTo-Json -Depth 12 | Set-Content -LiteralPath (Join-Path $OutputDirectory 'environment.json') -Encoding UTF8
     function Quote([string]$value) { if ($value.Contains('"')) { throw 'A command argument contains an unsupported quote.' }; return '"' + $value + '"' }
     $arguments = @('--output-directory', (Quote $OutputDirectory), '--', (Quote $GameExe)) + @($GameArguments | ForEach-Object { Quote $_ })
-    $process = Start-Process -FilePath (GamePath 'd4r/d4r_debug_launcher.exe') -ArgumentList $arguments -WorkingDirectory $game -WindowStyle Hidden -PassThru `
-        -RedirectStandardOutput (Join-Path $OutputDirectory 'd4r.stdout.log') -RedirectStandardError (Join-Path $OutputDirectory 'd4r.stderr.log')
+    # Drain raw bytes concurrently. PowerShell's line-oriented redirection can
+    # throttle a verbose game and distort the very timings we are collecting.
+    $info = [Diagnostics.ProcessStartInfo]::new()
+    $info.FileName = GamePath 'd4r/d4r_debug_launcher.exe'
+    $info.Arguments = $arguments -join ' '
+    $info.WorkingDirectory = $game
+    $info.UseShellExecute = $false; $info.CreateNoWindow = $true
+    $info.RedirectStandardOutput = $true; $info.RedirectStandardError = $true
+    $process = [Diagnostics.Process]::new(); $process.StartInfo = $info
+    if (!$process.Start()) { throw 'Game debugger launch failed.' }
+    $outFile = [IO.File]::Open((Join-Path $OutputDirectory 'd4r.stdout.log'), [IO.FileMode]::Create, [IO.FileAccess]::Write, [IO.FileShare]::Read)
+    $errFile = [IO.File]::Open((Join-Path $OutputDirectory 'd4r.stderr.log'), [IO.FileMode]::Create, [IO.FileAccess]::Write, [IO.FileShare]::Read)
+    $outTask = $process.StandardOutput.BaseStream.CopyToAsync($outFile)
+    $errTask = $process.StandardError.BaseStream.CopyToAsync($errFile)
     Write-Host "Game diagnostic started (preset $Preset). Logs: $OutputDirectory"
     $timer = [Diagnostics.Stopwatch]::StartNew()
     while (!$process.WaitForExit(1000)) {
         if ($RunSeconds -gt 0 -and $timer.Elapsed.TotalSeconds -ge $RunSeconds) { $process.Kill(); $process.WaitForExit(); break }
     }
     $process.WaitForExit()
+    [void]$outTask.GetAwaiter().GetResult(); [void]$errTask.GetAwaiter().GetResult()
+    $outFile.Dispose(); $outFile = $null; $errFile.Dispose(); $errFile = $null
     foreach ($file in Get-ChildItem -LiteralPath $game -Filter 'OptiScaler*.log' -File) { Copy-Item -LiteralPath $file.FullName -Destination $OutputDirectory -Force }
     $stderr = Get-Content -LiteralPath (Join-Path $OutputDirectory 'd4r.stderr.log') -Raw
     $stdout = Get-Content -LiteralPath (Join-Path $OutputDirectory 'd4r.stdout.log') -Raw
@@ -201,6 +224,10 @@ try {
     Write-Host "Game diagnostic bundle: $OutputDirectory.zip"
 } finally {
     if ($process -and !$process.HasExited) { $process.Kill(); $process.WaitForExit() }
+    if ($outTask) { try { [void]$outTask.GetAwaiter().GetResult() } catch {} }
+    if ($errTask) { try { [void]$errTask.GetAwaiter().GetResult() } catch {} }
+    if ($outFile) { $outFile.Dispose() }; if ($errFile) { $errFile.Dispose() }
+    if ($process) { $process.Dispose() }
     foreach ($key in $old.Keys) { [Environment]::SetEnvironmentVariable($key,$old[$key],'Process') }
     foreach ($entry in $originalSettings) { Copy-Item -LiteralPath $entry.saved -Destination $entry.path -Force }
 }

tools/windows/cuda_api.h

@@ -7,6 +7,7 @@ using CUcontext = void*;
 using CUmodule = void*;
 using CUfunction = void*;
 struct CudaApi {
+    bool verbose = true;
     d4r::diag::SearchDirectory search;
     d4r::diag::Library library;
 #define D4R_CU(name, ret, ...) using name##_fn = ret(WINAPI*)(__VA_ARGS__); name##_fn name = library.symbol<name##_fn>(#name)
@@ -41,7 +42,7 @@ struct CudaApi {
         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 (verbose || result) 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/d4r_nvngx_windows.cpp

@@ -154,7 +154,7 @@ API unsigned NVSDK_NGX_D3D12_EvaluateFeature(ID3D12GraphicsCommandList* list, co
             d4r_ngx_set_uint(snapshot->parameters, (prefix + ".Enhanced").c_str(), state.enhanced);
             d4r_ngx_set_uint(snapshot->parameters, (prefix + ".Layout").c_str(), unsigned(state.layout));
             d4r_ngx_set_uint(snapshot->parameters, (prefix + ".Access").c_str(), unsigned(state.access));
-            std::printf("D4R_RESOURCE_ACCESS plane=%s enhanced=%u legacy=0x%x layout=%u access=0x%x\n",
+            if (!std::getenv("D4R_QUIET_API")) std::printf("D4R_RESOURCE_ACCESS plane=%s enhanced=%u legacy=0x%x layout=%u access=0x%x\n",
                 names[i], state.enhanced, unsigned(state.legacy), unsigned(state.layout), unsigned(state.access));
         }
         d4r::win::commands::record_boundary(list, [feature, snapshot](ID3D12CommandQueue* queue) {
@@ -168,7 +168,7 @@ API unsigned NVSDK_NGX_D3D12_EvaluateFeature(ID3D12GraphicsCommandList* list, co
                 d4r_ngx_set_uint(snapshot->parameters, (prefix + ".Enhanced").c_str(), resolved.enhanced);
                 d4r_ngx_set_uint(snapshot->parameters, (prefix + ".Layout").c_str(), unsigned(resolved.layout));
                 d4r_ngx_set_uint(snapshot->parameters, (prefix + ".Access").c_str(), unsigned(resolved.access));
-                std::printf("D4R_INHERITED_ACCESS plane=%s tracked=%u enhanced=%u layout=%u legacy=0x%x\n",
+                if (!std::getenv("D4R_QUIET_API")) std::printf("D4R_INHERITED_ACCESS plane=%s tracked=%u enhanced=%u layout=%u legacy=0x%x\n",
                     names[i], !resolved.inherited, resolved.enhanced, unsigned(resolved.layout), unsigned(resolved.legacy));
             }
             feature->evaluate_boundary(queue, snapshot->parameters);

tools/windows/hip_api.h

@@ -4,6 +4,7 @@
 
 namespace d4r::diag {
 struct HipApi {
+    bool verbose = true; // Probes log every call; the game can suppress successful API calls.
     SearchDirectory search;
     Library library;
 #define D4R_HIP_FUNCTION(name) decltype(&::name) name = library.symbol<decltype(&::name)>(#name)
@@ -36,7 +37,7 @@ struct HipApi {
         search(std::filesystem::path(wide(root)) / L"bin"),
         library(std::filesystem::path(wide(root)) / L"bin" / L"amdhip64_7.dll") {}
     void check(hipError_t result, const char* call) const {
-        std::printf("HIP %s -> %d (%s)\n", call, static_cast<int>(result), hipGetErrorName(result));
+        if (verbose || result != hipSuccess) std::printf("HIP %s -> %d (%s)\n", call, static_cast<int>(result), hipGetErrorName(result));
         if (result != hipSuccess) throw std::runtime_error(std::string(call) + ": " + hipGetErrorString(result));
     }
     int select_gfx1201(int requested, hipDeviceProp_t& selected) const {

tools/windows/ngx_windows_runtime.h

@@ -17,7 +17,7 @@ inline std::string env_path(const char* name) {
     return value;
 }
 inline void ngx_check(unsigned result, const char* operation) {
-    std::printf("D4R_NGX %s result=0x%08x\n", operation, result);
+    if (result != 1 || !std::getenv("D4R_QUIET_API")) std::printf("D4R_NGX %s result=0x%08x\n", operation, result);
     if (result != 1) throw std::runtime_error(std::string(operation) + " NGX=" + std::to_string(result));
 }
 inline ResourceAccess ngx_resource_access(void* parameters, const char* name, D3D12_RESOURCE_STATES fallback) {
@@ -82,6 +82,7 @@ struct Runtime {
         try {
             hipDeviceProp_t props{};
             hip.select_gfx1201(-1, props);
+            hip.verbose = cuda.verbose = !std::getenv("D4R_QUIET_API");
             const auto luid = adapter_luid(d3d);
             if (std::memcmp(&luid, props.luid, sizeof(luid))) throw std::runtime_error("D3D12/HIP LUID mismatch");
             cuda.check(cuda.cuInit(0), "cuInit(Windows runtime)");
@@ -289,8 +290,18 @@ private:
     }
     void execute_cuda(ID3D12CommandQueue* queue) {
         const auto begin = std::chrono::steady_clock::now();
+        auto stage = begin;
+        const bool profileStages = std::getenv("D4R_PROFILE_STAGES") != nullptr;
+        auto mark = [&](const char* name) {
+            if (!profileStages) return;
+            const auto now = std::chrono::steady_clock::now();
+            std::printf("D4R_STAGE name=%s cpu_ms=%.6f\n", name,
+                std::chrono::duration<double, std::milli>(now - stage).count());
+            stage = std::chrono::steady_clock::now();
+        };
         rt_->timeline->wait_input(queue);
         rt_->cuda.check(rt_->cuda.cuCtxSynchronize(), "Interop input producer completion");
+        mark("input_fence_wait");
         for (unsigned i : {0u, 1u, 2u, 4u}) if (planes_[i].texture) {
             auto& plane = planes_[i];
             void* data = plane.shared->mapped; uint64_t pitch = plane.shared->footprint.Footprint.RowPitch;
@@ -301,6 +312,7 @@ private:
             }
             plane.image->upload_device(reinterpret_cast<uintptr_t>(data), pitch);
         }
+        mark("input_conversion_array_upload");
         if (std::getenv("D4R_INTEROP_VERIFY")) {
             for (unsigned i : {0u, 1u, 2u, 4u}) if (planes_[i].texture) {
                 uint32_t sample[8]{};
@@ -315,13 +327,16 @@ private:
             }
         }
         ngx_check(rt_->evaluate(handle_, parameters_, nullptr), "CUDA EvaluateFeature");
+        mark("ngx_evaluate_host");
         // NGX can use internal nonblocking streams. The default stream alone
         // does not establish completion of those streams.
         rt_->cuda.check(rt_->cuda.cuCtxSynchronize(), "NGX all-stream completion");
+        mark("ngx_all_stream_completion");
         auto& output = planes_[3];
         output.image->download_device(reinterpret_cast<uintptr_t>(output.canonical ? output.canonical->data : output.shared->mapped),
             output.canonical ? output.canonical->pitch : output.shared->footprint.Footprint.RowPitch);
         rt_->cuda.check(rt_->cuda.cuCtxSynchronize(), "VRAM output completion");
+        mark("output_array_download");
         if (std::getenv("D4R_VALIDATE_OUTPUT")) {
             if (!rt_->pixels) rt_->pixels = std::make_unique<PixelProgram>(rt_->hip, pixel_module_path());
             const auto counts = rt_->pixels->validate(output.canonical ? output.canonical->data : output.shared->mapped,
@@ -329,9 +344,11 @@ private:
             std::printf("D4R_OUTPUT_VALIDATION elements=%llu nan=%u inf=%u diagnostics_cpu_bytes=8\n", output.desc.Width * output.desc.Height * 4, counts[0], counts[1]);
             if (counts[0] || counts[1]) throw std::runtime_error("DLSS output contains NaN/Inf");
         }
+        mark("output_diagnostic");
         if (output.canonical) rt_->pixels->convert(true, output.canonical->data, output.canonical->pitch, output.shared->mapped,
             output.shared->footprint.Footprint.RowPitch, unsigned(output.desc.Width), output.desc.Height, output.spec.storage, 3);
         rt_->timeline->signal_output(queue);
+        mark("output_conversion_fence_signal");
         std::printf("D4R_FRAME cpu_frame_copies=%u frame_age=0 interop_ngx_ms=%.3f\n", std::getenv("D4R_INTEROP_VERIFY") ? 1u : 0u,
             std::chrono::duration<double, std::milli>(std::chrono::steady_clock::now() - begin).count());
     }

tools/windows/profile_report.py

@@ -0,0 +1,92 @@
+"""Summarize native Windows d4r CPU stages and optional serializing HIP events.
+
+CPU completion time is not GPU execution time. HIP occupancy API predictions
+are not hardware counters. No bandwidth / WMMA utilization is inferred here.
+"""
+import argparse
+import collections
+import json
+import pathlib
+import re
+import statistics
+
+
+def stats(values):
+    values = sorted(values)
+    def quantile(q):
+        index = (len(values) - 1) * q
+        lower = int(index)
+        return values[lower] + (values[min(lower + 1, len(values) - 1)] - values[lower]) * (index - lower)
+    return dict(samples=len(values), total=sum(values), mean=statistics.mean(values),
+                median=statistics.median(values), p95=quantile(.95), maximum=values[-1])
+
+
+def metadata(directory):
+    result = {}
+    if directory is None:
+        return result
+    for path in sorted(directory.glob('*.txt')):
+        text = path.read_text(encoding='utf-8-sig', errors='replace')
+        for block in re.split(r'(?m)^  - \.args:', text)[1:]:
+            fields = dict(re.findall(r'(?m)^    \.(\w+):\s*([^\r\n]+)', block))
+            if 'name' in fields:
+                result[fields['name'].strip()] = {key: int(fields[key]) for key in (
+                    'wavefront_size', 'vgpr_count', 'sgpr_count', 'vgpr_spill_count', 'sgpr_spill_count',
+                    'group_segment_fixed_size', 'private_segment_fixed_size') if key in fields}
+    return result
+
+
+def report(directory, metadata_directory=None):
+    kernel_metadata = metadata(metadata_directory)
+    stages = collections.defaultdict(list)
+    kernels = collections.defaultdict(list)
+    for path in sorted(directory.glob('*.log')):
+        text = path.read_text(encoding='utf-8-sig', errors='replace')
+        for name, value in re.findall(r'D4R_STAGE name=(\S+) cpu_ms=([0-9.]+)', text):
+            stages[name].append(float(value))
+        for line in text.splitlines():
+            if not line.startswith('D4R_KERNEL_PROFILE '):
+                continue
+            fields = dict(re.findall(r'(\w+)=("[^"]+"|\S+)', line))
+            kernels[(fields['kernel'].strip('"'), fields['backend'], fields['phase'])].append(fields)
+    result = dict(cpuStages={name: stats(values) for name, values in stages.items()},
+                  kernels=[], notes=[
+                      'CPU stage times include waits and host work; they are not isolated GPU timings.',
+                      '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, backend, phase), rows in kernels.items():
+        item = dict(kernel=name, backend=backend, phase=phase,
+                    gpuMs=stats([float(row['gpu_ms']) for row in rows]),
+                    enqueueMs=stats([float(row['enqueue_ms']) for row in rows]),
+                    completionMs=stats([float(row['completion_ms']) for row in rows]),
+                    configurations=[dict(config) for config in sorted({tuple((key, row[key]) for key in
+                        ('grid', 'block', 'static_lds_bytes', 'dynamic_lds_bytes', 'regs', 'private_bytes',
+                         'max_threads', 'predicted_blocks_per_multiprocessor')) for row in rows})])
+        result['kernels'].append(item)
+        if backend == 'native':
+            item['elfMetadata'] = kernel_metadata.get(name + ('_prep' if phase == 'prep' else ''), {})
+    result['kernels'].sort(key=lambda row: row['gpuMs']['total'], reverse=True)
+    return result
+
+
+def main():
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument('directory', type=pathlib.Path)
+    parser.add_argument('--output', type=pathlib.Path)
+    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']:
+        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')
+    for name, row in result['cpuStages'].items():
+        print(f"CPU {name}: n={row['samples']} mean={row['mean']:.3f} ms median={row['median']:.3f} ms p95={row['p95']:.3f} ms")
+    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")
+
+
+if __name__ == '__main__':
+    main()