Record controlled K timings and add read-only AMD GPU telemetry

docs/windows-performance.md

@@ -249,8 +249,8 @@ completion. Every image stays in VRAM. The ZLUDA implementation adds
 `cuMemcpy2DAsync_v2` using `hipMemcpyParam2DAsync`; format-80 array emulation
 is explicitly rejected because its synchronous implementation uses temporary
 host storage. The Windows path supplies canonical FP16/FP32 arrays instead.
-This experiment is disabled by default until the correctness and game timing
-gates below pass; there is no claimed FPS improvement yet.
+This experiment remains disabled by default. Correctness checks pass, but the
+controlled game comparison below shows only a small frame-time difference.
 
 ```powershell
 .\scripts\windows\test-deferred-kernel-profile.ps1
@@ -277,6 +277,74 @@ four-way K/4K gate passes again with exact candidate RGB and finite RGBA
 (`test-results/k-batch-committed-source-4k`). This warms the committed runtime's
 JIT cache before the game comparison.
 
+The following user-confirmed 30-second foreground-scene capture contains 1925
+presents at **64.265 FPS**, versus the previous unbatched run's 63.574 FPS.
+Both use 2259x1271 -> 3840x2160, Independent Flip and SyncInterval 0, without
+kernel-event profiling, output scans or attached debuggers. Mean present time
+is 15.561 versus 15.730 ms, a 0.169 ms difference. This single pair does not
+establish repeatability or a large performance improvement. The containing
+ten-minute candidate run completes 35673 K frames with zero backend failures,
+CPU image copies, previous-frame outputs or recorded game crashes. Whole-run
+median input preparation drops from 0.482 to 0.332 ms; median external span
+is 5.438 versus 5.399 ms. Whole-run stages are not the selected PresentMon
+window. No GPU utilization conclusion is drawn from ETW estimates. Results:
+`test-results/silent-hill2-k-batch-inputs-4k`. The K performance gate stays open.
+
+### AMD driver telemetry
+
+A separate read-only diagnostic polls the installed driver's ADLX base C
+interfaces for device-wide usage, core/VRAM clocks, board power and temperatures.
+It adds no work to game queues and changes no tuning setting. Build it locally
+using the official SDK pinned at `32b5a740d42295c5dfe9026b9f52683da0f3af91`:
+
+```powershell
+.\scripts\windows\build-gpu-telemetry.ps1
+.\scripts\windows\capture-presentmon.ps1 -Seconds 30 -GpuTelemetry
+```
+
+The SDK retains its own AMD license; SDK headers, implementations, driver DLLs
+and the SDK-dependent executable are excluded from public game packages.
+The probe is independently written and uses the documented C ABI, including
+with LLVM-MinGW. It loads `amdadlx64.dll` only from Windows System32 and
+releases acquired interfaces before ADLX termination/unload. GPU selection
+must match exactly one device (`-GpuName` selects a different name substring).
+The process exits at its bound or target process exit. The capture retains
+CSV, raw stderr, exit/timeout, driver DLL version/hash and SDK build identity.
+No optional telemetry failure is silently converted to a valid measurement.
+
+CSV records independent host UTC/QPC timestamps and the driver's raw sample
+timestamp. On this installed ADLX 1.5.0.124, the latter is uptime despite the
+SDK description of epoch time; do not interpret it as UTC. Unsupported metrics
+are blank, zero utilization/clock remains valid, and the report uses each
+driver timestamp once. Device-wide usage is distinct from PresentMon's
+per-process frame attribution and is not measured WMMA occupancy. ADLX API:
+[AMD GPU metrics sample](https://gpuopen.com/manuals/adlx/adlx-c__perf_g_p_u_metrics/).
+
+The initial idle-device smoke test captures six valid polls, selects the real
+RX 9070 XT and terminates ADLX successfully. Board power/core clock/usage
+and temperatures are available; `GPUPower` is unsupported and stays absent
+from the report. The first poll costs roughly 10 ms for initialization;
+subsequent polls are about 0.08 ms at 500 ms intervals. This idle smoke test
+does not establish game clocks, utilization or an overhead-free benchmark.
+Results: `test-results/gpu-telemetry-smoke`.
+
+The subsequent combined-capture smoke test collects 246 presents and ten
+ADLX polls and exits cleanly (`test-results/gpu-telemetry-capture-smoke`).
+It includes startup/scene changes and is excluded from FPS comparisons.
+After the user confirms the foreground comparison scene, the 30-second run
+with detailed d4r stages/boundary timestamps disabled records **64.661 FPS**
+over 1935 presents. ADLX's sixty polls report mean usage **93.183%**, core
+clock **3124.667 MHz**, board power **316.717 W**, temperature 57 C and hotspot
+82.9 C. Mean PresentMon GPU busy/wait is 14.551/0.800 ms. All presents retain
+Independent Flip / SyncInterval 0. This records high device-wide driver usage
+in this run; the previous user-reported 63% has not been captured from the same
+sensor/window and its discrepancy is unresolved. Do not infer a physical CPU
+limit, isolated shader occupancy or a FSR4 speed comparison. Query overhead
+is median 0.094 ms per 500 ms poll, with a roughly 12 ms first query. Results:
+`test-results/silent-hill2-k-telemetry-unprofiled-4k`. The game stability
+summary is still pending its existing ten-minute bound. Remaining K work now
+prioritizes transformer/output-kernel cost; frame scheduling remains measured.
+
 Further game coverage (HIP kernel profiling disabled, GPU output checks enabled):
 M at 1920x1080 completes 4833 frames, all finite. K at 3840x2160 completes
 4685 frames, all finite. Neither has a backend failure, CPU image copy or an

docs/windows-rdna4-port.md

@@ -26,6 +26,12 @@ host synchronization; postpone further M/FP8 optimization. These user readings
 are not a controlled same-scene benchmark. Do not infer a hardware CPU limit
 from utilization alone.
 
+A subsequent unprofiled K/4K capture records **64.661 FPS** with direct AMD
+ADLX mean usage **93.183%**, core clock **3124.667 MHz** and board power
+**316.717 W**. The previous 63% reading is from a different sensor/window;
+its discrepancy remains unresolved. This new capture places GPU kernel cost
+at the center of the next K investigation, while retaining queue measurements.
+
 The next diagnostics add `-ProfileCudaApi` (ZLUDA patch 0015): aggregate host
 CUDA API durations without HIP-event waits. Boundary stages now separate
 prepare, copy-list setup, input/output submission and output-copy drain.
@@ -103,7 +109,7 @@ stream, which needs explicit opt-in for timestamp sampling. The depth/stencil
 and packed-format burst/recreation gates pass 24 full K/4K frames: all twelve
 candidate RGB images match their fresh unbatched synchronous controls exactly,
 with finite RGBA. Results: `test-results/k-batch-deferred-4k` and
-`test-results/k-batch-packed-4k`. The game performance gate is pending. The candidate remains
+`test-results/k-batch-packed-4k`. The candidate remains
 disabled by default; see `windows-performance.md` for semantics and commands.
 
 Current candidate source is ZLUDA `67127dde599879ada022d6439383d682e09e0b36`
@@ -118,6 +124,23 @@ well as TheRock; PTX output/guards remain correct. It also affects a native K
 dec4 timing sample, so valid-looking events alone are not a timing-accuracy
 gate. No hardware utilization or kernel speedup is inferred from these samples.
 
+The committed batched-input game run completes **35673 K/4K frames** without
+backend errors, CPU image copies, previous-frame output or recorded crashes.
+The user-confirmed 30-second PresentMon window records **64.265 FPS** (1925
+presents), compared with the prior unbatched 63.574 FPS. Both keep Independent
+Flip / SyncInterval 0 and 2259x1271 -> 3840x2160. This small single-pair
+difference is not a repeatable speedup gate. Whole-run median input preparation
+falls from 0.482 to 0.332 ms, while median external span is 5.438 versus
+5.399 ms. Results: `test-results/silent-hill2-k-batch-inputs-4k`.
+K performance remains open and batching remains opt-in. The next diagnostic
+adds a separate read-only AMD ADLX usage/clock/power capture alongside
+PresentMon. Its local RX 9070 XT smoke test passes, retaining unsupported
+metrics as absent and preserving the driver's raw timestamp. See
+`windows-performance.md` for build/capture commands and SDK packaging limits.
+The combined capture smoke test passes, followed by the user-confirmed
+1935-present/60-poll measurement recorded at the top of this document. Logs
+and runtime/DLL hashes: `test-results/silent-hill2-k-telemetry-unprofiled-4k`.
+
 The public prerelease `windows-rdna4-dev-20261002-a346d76` is published to
 `xdfnx-dev/d4r`, containing source commit `a346d76` and ZLUDA `a1c506f`.
 Archive: `d4r-windows-rdna4-20261002.zip`, 88893567 bytes, SHA256

scripts/windows/build-gpu-telemetry.ps1

@@ -0,0 +1,28 @@
+[CmdletBinding()]
+param([string]$AdlxRoot, [string]$BuildDirectory, [string]$Compiler)
+$ErrorActionPreference='Stop'
+$repo=[IO.Path]::GetFullPath((Join-Path $PSScriptRoot '../..'))
+$commit='32b5a740d42295c5dfe9026b9f52683da0f3af91'
+if (!$AdlxRoot) { $AdlxRoot=Join-Path $repo 'external/ADLX' }
+if (!$BuildDirectory) { $BuildDirectory=Join-Path $repo 'build/gpu-telemetry' }
+if (!$Compiler) { $Compiler=Join-Path $repo '.tools/llvm-mingw-20260922-ucrt-x86_64/bin/x86_64-w64-mingw32-clang.exe' }
+if (!(Test-Path -LiteralPath $AdlxRoot)) {
+    & git clone --no-checkout https://github.com/GPUOpen-LibrariesAndSDKs/ADLX.git $AdlxRoot
+    if ($LASTEXITCODE) { throw 'ADLX SDK download failed.' }
+    & git -C $AdlxRoot checkout --detach $commit
+    if ($LASTEXITCODE) { throw 'ADLX SDK checkout failed.' }
+}
+if ((& git -C $AdlxRoot rev-parse HEAD).Trim() -ne $commit -or (& git -C $AdlxRoot status --porcelain)) {
+    throw "Expected clean ADLX SDK $commit; use a separate AdlxRoot."
+}
+if (!(Test-Path -LiteralPath $Compiler)) { throw 'Set Compiler to an x64 LLVM-MinGW clang.exe.' }
+New-Item -ItemType Directory -Force $BuildDirectory | Out-Null
+$exe=Join-Path $BuildDirectory 'd4r_gpu_telemetry.exe'
+& $Compiler -std=c11 -O2 -Wall -Wextra -Wno-unknown-pragmas -DWIN32_LEAN_AND_MEAN -D_M_AMD64 `
+    -isystem "$AdlxRoot/SDK/Include" (Join-Path $repo 'tools/windows/gpu_telemetry.c') -o $exe
+if ($LASTEXITCODE) { throw "GPU telemetry compile failed: $LASTEXITCODE" }
+@{sdkCommit=$commit; sdkLicense=(Join-Path $AdlxRoot 'ADLX SDK License Agreement.pdf');
+    executable=$exe; compiler=$Compiler; sha256=(Get-FileHash -LiteralPath $exe -Algorithm SHA256).Hash} |
+    ConvertTo-Json | Set-Content -LiteralPath (Join-Path $BuildDirectory 'build-info.json') -Encoding UTF8
+Write-Host "Built local read-only telemetry: $exe"
+# The SDK, driver DLL and SDK-dependent binary are not added to public packages.

scripts/windows/capture-presentmon.ps1

@@ -5,6 +5,9 @@ param(
     [ValidateRange(5,300)][int]$Seconds = 30,
     [ValidateRange(0,300)][int]$DelaySeconds = 0,
     [string]$PresentMonPath,
+    [switch]$GpuTelemetry,
+    [string]$TelemetryPath,
+    [string]$GpuName = 'RX 9070 XT',
     [string]$OutputDirectory
 )
 $ErrorActionPreference='Stop'
@@ -32,16 +35,67 @@ $OutputDirectory=[IO.Path]::GetFullPath($OutputDirectory)
 New-Item -ItemType Directory -Force $OutputDirectory | Out-Null
 $csv=Join-Path $OutputDirectory 'presentmon.csv'
 if (Test-Path -LiteralPath $csv) { throw 'Preserve the existing capture; choose another OutputDirectory.' }
+if ($TelemetryPath) { $GpuTelemetry=$true }
+if ($GpuTelemetry) {
+    if (!$TelemetryPath) { $TelemetryPath=Join-Path $repo 'build/gpu-telemetry/d4r_gpu_telemetry.exe' }
+    $TelemetryPath=(Get-Item -LiteralPath $TelemetryPath -ErrorAction Stop).FullName
+    if ($GpuName.Contains('"')) { throw 'GpuName cannot contain a double quote.' }
+    foreach ($file in @('gpu-telemetry.csv','gpu-telemetry.stderr.log','gpu-telemetry-environment.json')) {
+        if (Test-Path -LiteralPath (Join-Path $OutputDirectory $file)) { throw "Preserve existing $file; choose another OutputDirectory." }
+    }
+}
 @{version='2.6.0'; executable=$PresentMonPath; sha256=(Get-FileHash -LiteralPath $PresentMonPath -Algorithm SHA256).Hash;
     targetProcessId=$target.Id; processName=$target.ProcessName; seconds=$Seconds; delaySeconds=$DelaySeconds;
+    gpuTelemetryRequested=[bool]$GpuTelemetry;
     hwsRegistry=$(Get-ItemPropertyValue -LiteralPath 'HKLM:/SYSTEM/CurrentControlSet/Control/GraphicsDrivers' -Name HwSchMode -ErrorAction SilentlyContinue);
     driver=@(Get-CimInstance Win32_VideoController | Select-Object Name,DriverVersion)} |
     ConvertTo-Json -Depth 4 | Set-Content -LiteralPath (Join-Path $OutputDirectory 'presentmon-environment.json') -Encoding UTF8
 # Unique ETW session: never stop or replace an existing user's trace. The CLI
 # terminates itself at the bound/process exit, without a service installation.
 $session='D4R_' + [Guid]::NewGuid().ToString('N')
-& $PresentMonPath --process_id $target.Id --timed $Seconds --delay $DelaySeconds --terminate_after_timed --terminate_on_proc_exit --session_name $session --no_console_stats --no_track_input --qpc_time_ms --output_file $csv 2>&1 |
+$telemetry=$null; $telemetryStarted=$false; $outStream=$null; $errStream=$null; $outTask=$null; $errTask=$null
+try {
+    if ($GpuTelemetry) {
+        $info=[Diagnostics.ProcessStartInfo]::new()
+        $info.FileName=$TelemetryPath; $info.UseShellExecute=$false; $info.CreateNoWindow=$true
+        $info.RedirectStandardOutput=$true; $info.RedirectStandardError=$true
+        $info.Arguments="--seconds $Seconds --delay-ms $($DelaySeconds*1000) --interval-ms 500 --target-pid $($target.Id) --gpu-name `"$GpuName`""
+        $outStream=[IO.File]::Open((Join-Path $OutputDirectory 'gpu-telemetry.csv'),[IO.FileMode]::CreateNew,[IO.FileAccess]::Write,[IO.FileShare]::Read)
+        $errStream=[IO.File]::Open((Join-Path $OutputDirectory 'gpu-telemetry.stderr.log'),[IO.FileMode]::CreateNew,[IO.FileAccess]::Write,[IO.FileShare]::Read)
+        $telemetry=[Diagnostics.Process]::new(); $telemetry.StartInfo=$info
+        if (!$telemetry.Start()) { throw 'GPU telemetry process failed to start.' }
+        $telemetryStarted=$true
+        $outTask=$telemetry.StandardOutput.BaseStream.CopyToAsync($outStream)
+        $errTask=$telemetry.StandardError.BaseStream.CopyToAsync($errStream)
+    }
+    & $PresentMonPath --process_id $target.Id --timed $Seconds --delay $DelaySeconds --terminate_after_timed --terminate_on_proc_exit --session_name $session --no_console_stats --no_track_input --qpc_time_ms --output_file $csv 2>&1 |
         Tee-Object -FilePath (Join-Path $OutputDirectory 'presentmon-console.log')
-if ($LASTEXITCODE) { throw "PresentMon failed with exit code $LASTEXITCODE; see presentmon-console.log." }
+    if ($LASTEXITCODE) { throw "PresentMon failed with exit code $LASTEXITCODE; see presentmon-console.log." }
+} finally {
+    try { if ($telemetryStarted) {
+        $timedOut=$false
+        if (!$telemetry.HasExited -and !$telemetry.WaitForExit(30000)) { $timedOut=$true; $telemetry.Kill(); $telemetry.WaitForExit() }
+        if ($outTask) { [void]$outTask.GetAwaiter().GetResult() }
+        if ($errTask) { [void]$errTask.GetAwaiter().GetResult() }
+        $driverDll=Join-Path $env:WINDIR 'System32/amdadlx64.dll'
+        $driverInfo=$null
+        if (Test-Path -LiteralPath $driverDll) {
+            $driverInfo=@{path=$driverDll; version=(Get-Item -LiteralPath $driverDll).VersionInfo.FileVersion;
+                sha256=(Get-FileHash -LiteralPath $driverDll -Algorithm SHA256).Hash}
+        }
+        @{executable=$TelemetryPath; sha256=(Get-FileHash -LiteralPath $TelemetryPath -Algorithm SHA256).Hash;
+            exitCode=$telemetry.ExitCode; timeout=$timedOut; targetProcessId=$target.Id; gpuName=$GpuName;
+            intervalMs=500; driverDll=$driverInfo; readOnly=$true;
+            buildInfo=$(if (Test-Path -LiteralPath (Join-Path (Split-Path $TelemetryPath) 'build-info.json')) {
+                Get-Content -LiteralPath (Join-Path (Split-Path $TelemetryPath) 'build-info.json') -Raw | ConvertFrom-Json
+            })} |
+            ConvertTo-Json -Depth 6 | Set-Content -LiteralPath (Join-Path $OutputDirectory 'gpu-telemetry-environment.json') -Encoding UTF8
+        if ($telemetry.ExitCode -ne 0) { Write-Warning 'GPU telemetry failed; retain PresentMon capture and see gpu-telemetry.stderr.log.' }
+    } } finally {
+        if ($telemetry) { $telemetry.Dispose() }
+        if ($outStream) { $outStream.Dispose() }
+        if ($errStream) { $errStream.Dispose() }
+    }
+}
 if (!(Test-Path -LiteralPath $csv) -or (Get-Item -LiteralPath $csv).Length -eq 0) { throw 'PresentMon captured no frames.' }
 Write-Host "PresentMon capture: $csv"

tools/windows/gpu_telemetry.c

@@ -0,0 +1,149 @@
+// Read-only AMD driver telemetry, separate from the game and HIP runtime.
+// Uses the ADLX SDK's C interface; no SDK/sample implementation is vendored.
+#include <windows.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <math.h>
+#include <errno.h>
+#include <ADLX.h>
+#include <IPerformanceMonitoring.h>
+
+static int number(const char* text, unsigned* value) {
+    char* end = NULL;
+    errno = 0;
+    unsigned long parsed = strtoul(text, &end, 10);
+    if (!text[0] || *end || text[0] == '-' || errno == ERANGE) return 0;
+    *value = (unsigned)parsed;
+    return 1;
+}
+
+static void real_value(ADLX_RESULT status, double value) {
+    if (ADLX_SUCCEEDED(status) && isfinite(value) && value >= 0) printf("%.6f", value);
+}
+
+static void int_value(ADLX_RESULT status, int value) {
+    if (ADLX_SUCCEEDED(status) && value >= 0) printf("%d", value);
+}
+
+int main(int argc, char** argv) {
+    unsigned seconds = 30, interval = 500, target_pid = 0, delay = 0;
+    const char* gpu_name = "RX 9070 XT";
+    for (int i = 1; i < argc; ++i) {
+        if (i + 1 == argc) return 2;
+        const char* key = argv[i];
+        const char* value = argv[++i];
+        unsigned* dest = !strcmp(key, "--seconds") ? &seconds :
+            !strcmp(key, "--interval-ms") ? &interval :
+            !strcmp(key, "--target-pid") ? &target_pid :
+            !strcmp(key, "--delay-ms") ? &delay : NULL;
+        if (dest) { if (!number(value, dest)) return 2; }
+        else if (!strcmp(key, "--gpu-name")) gpu_name = value;
+        else return 2;
+    }
+    if (seconds < 1 || seconds > 300 || interval < 100 || interval > 5000 || delay > 300000) return 2;
+    int exit_code = 1, initialized = 0;
+    HMODULE dll = NULL;
+    HANDLE target = NULL;
+    IADLXSystem* system = NULL; // Owned by ADLXInitialize/ADLXTerminate.
+    IADLXGPUList* gpus = NULL;
+    IADLXGPU* selected = NULL;
+    IADLXPerformanceMonitoringServices* service = NULL;
+    ADLXTerminate_Fn terminate = NULL;
+    if (target_pid) {
+        target = OpenProcess(SYNCHRONIZE, FALSE, target_pid);
+        if (!target) { fprintf(stderr, "OpenProcess error=%lu\n", GetLastError()); goto done; }
+    }
+    // Load the installed driver library, never a DLL from the game/current dir.
+    dll = LoadLibraryExW(L"amdadlx64.dll", NULL, LOAD_LIBRARY_SEARCH_SYSTEM32);
+    if (!dll) { fprintf(stderr, "ADLX driver DLL unavailable error=%lu\n", GetLastError()); goto done; }
+    ADLXQueryFullVersion_Fn query_version = (ADLXQueryFullVersion_Fn)GetProcAddress(dll, ADLX_QUERY_FULL_VERSION_FUNCTION_NAME);
+    ADLXInitialize_Fn initialize = (ADLXInitialize_Fn)GetProcAddress(dll, ADLX_INIT_FUNCTION_NAME);
+    terminate = (ADLXTerminate_Fn)GetProcAddress(dll, ADLX_TERMINATE_FUNCTION_NAME);
+    adlx_uint64 runtime_version = 0;
+    if (!query_version || !initialize || !terminate || ADLX_FAILED(query_version(&runtime_version))) goto done;
+    // Only use the base interfaces. Request no version newer than either the
+    // installed runtime or these headers (e.g. SDK 2.0 with driver ADLX 1.5).
+    adlx_uint64 requested = runtime_version < ADLX_FULL_VERSION ? runtime_version : ADLX_FULL_VERSION;
+    ADLX_RESULT result = initialize(requested, &system);
+    fprintf(stderr, "ADLX runtime=%llu requested=%llu initialize=%d interval_ms=%u target_pid=%u\n",
+        (unsigned long long)runtime_version, (unsigned long long)requested, result, interval, target_pid);
+    if (ADLX_FAILED(result) || !system) goto done;
+    initialized = 1;
+    result = system->pVtbl->GetGPUs(system, &gpus);
+    if (ADLX_FAILED(result) || !gpus) goto done;
+    unsigned matches = 0;
+    for (adlx_uint i = gpus->pVtbl->Begin(gpus); i < gpus->pVtbl->End(gpus); ++i) {
+        IADLXGPU* gpu = NULL;
+        if (ADLX_FAILED(gpus->pVtbl->At_GPUList(gpus, i, &gpu)) || !gpu) goto done;
+        const char* name = NULL; const char* pnp = NULL; adlx_int id = -1;
+        gpu->pVtbl->Name(gpu, &name);
+        gpu->pVtbl->PNPString(gpu, &pnp);
+        gpu->pVtbl->UniqueId(gpu, &id);
+        fprintf(stderr, "ADLX GPU index=%u id=%d name=%s pnp=%s\n", i, id, name ? name : "unknown", pnp ? pnp : "unknown");
+        if (name && strstr(name, gpu_name)) {
+            ++matches;
+            if (!selected) { selected = gpu; gpu = NULL; }
+        }
+        if (gpu) gpu->pVtbl->Release(gpu);
+    }
+    if (matches != 1) { fprintf(stderr, "Expected one GPU matching '%s', found %u\n", gpu_name, matches); goto done; }
+    result = system->pVtbl->GetPerformanceMonitoringServices(system, &service);
+    if (ADLX_FAILED(result) || !service) goto done;
+    LARGE_INTEGER frequency, qpc;
+    QueryPerformanceFrequency(&frequency);
+    // Preserve the driver's timestamp verbatim: the installed ADLX 1.5 reports
+    // uptime here despite SDK documentation describing epoch time. Use our
+    // independent QPC/UTC stamps for correlation with PresentMon.
+    printf("qpc_ms,host_unix_ms,sensor_timestamp_ms,gpu_usage_pct,gpu_clock_mhz,vram_clock_mhz,gpu_power_w,board_power_w,temperature_c,hotspot_c,poll_ms\n");
+    fflush(stdout);
+    if (delay) Sleep(delay);
+    ULONGLONG start = GetTickCount64();
+    unsigned samples = 0;
+    while (GetTickCount64() - start < (ULONGLONG)seconds * 1000) {
+        if (target && WaitForSingleObject(target, 0) == WAIT_OBJECT_0) break;
+        LARGE_INTEGER before, after;
+        QueryPerformanceCounter(&before);
+        IADLXGPUMetrics* metrics = NULL;
+        result = service->pVtbl->GetCurrentGPUMetrics(service, selected, &metrics);
+        QueryPerformanceCounter(&after);
+        if (ADLX_FAILED(result) || !metrics) { fprintf(stderr, "GetCurrentGPUMetrics result=%d\n", result); goto done; }
+        FILETIME time; ULARGE_INTEGER unix_time;
+        GetSystemTimePreciseAsFileTime(&time);
+        unix_time.LowPart = time.dwLowDateTime; unix_time.HighPart = time.dwHighDateTime;
+        QueryPerformanceCounter(&qpc);
+        adlx_int64 sensor_time = 0;
+        printf("%.6f,%llu,", qpc.QuadPart * 1000. / frequency.QuadPart,
+            (unsigned long long)((unix_time.QuadPart - 116444736000000000ULL) / 10000));
+        if (ADLX_SUCCEEDED(metrics->pVtbl->TimeStamp(metrics, &sensor_time))) printf("%lld", (long long)sensor_time);
+        double real = -1; int integer = -1;
+        printf(","); result = metrics->pVtbl->GPUUsage(metrics, &real); real_value(result, real);
+        printf(","); result = metrics->pVtbl->GPUClockSpeed(metrics, &integer); int_value(result, integer);
+        printf(","); result = metrics->pVtbl->GPUVRAMClockSpeed(metrics, &integer); int_value(result, integer);
+        printf(","); result = metrics->pVtbl->GPUPower(metrics, &real); real_value(result, real);
+        printf(","); result = metrics->pVtbl->GPUTotalBoardPower(metrics, &real); real_value(result, real);
+        printf(","); result = metrics->pVtbl->GPUTemperature(metrics, &real); real_value(result, real);
+        printf(","); result = metrics->pVtbl->GPUHotspotTemperature(metrics, &real); real_value(result, real);
+        printf(",%.6f\n", (after.QuadPart - before.QuadPart) * 1000. / frequency.QuadPart);
+        metrics->pVtbl->Release(metrics);
+        ++samples; fflush(stdout);
+        ULONGLONG elapsed = GetTickCount64() - start;
+        ULONGLONG next = (ULONGLONG)samples * interval;
+        if (next > elapsed) Sleep((DWORD)(next - elapsed));
+    }
+    fprintf(stderr, "ADLX collected=%u read_only=1\n", samples);
+    exit_code = samples ? 0 : 1;
+done:
+    // All acquired interfaces must be released before unloading ADLX.
+    if (service) service->pVtbl->Release(service);
+    if (selected) selected->pVtbl->Release(selected);
+    if (gpus) gpus->pVtbl->Release(gpus);
+    if (initialized) {
+        ADLX_RESULT status = terminate();
+        fprintf(stderr, "ADLX terminate=%d\n", status);
+        if (ADLX_FAILED(status)) exit_code = 1;
+    }
+    if (dll) FreeLibrary(dll);
+    if (target) CloseHandle(target);
+    return exit_code;
+}

tools/windows/profile_report.py

@@ -76,6 +76,38 @@ def presentmon(directory):
     return captures
 
 
+def gpu_telemetry(directory):
+    captures = []
+    for path in sorted(directory.glob('gpu-telemetry*.csv')):
+        unique = {}
+        with path.open(encoding='utf-8-sig', newline='') as stream:
+            rows = list(csv.DictReader(stream))
+        # Driver telemetry may repeat a cached measurement. Count polls but
+        # use each sensor timestamp once; missing/unsupported values stay absent.
+        for row in rows:
+            try:
+                timestamp = int(row.get('sensor_timestamp_ms', ''))
+            except (ValueError, TypeError):
+                continue
+            if timestamp > 0:
+                unique.setdefault(timestamp, row)
+        metrics = {}
+        for name in ('gpu_usage_pct', 'gpu_clock_mhz', 'vram_clock_mhz', 'gpu_power_w',
+                     'board_power_w', 'temperature_c', 'hotspot_c', 'poll_ms'):
+            values = []
+            for row in unique.values():
+                try:
+                    value = float(row.get(name, ''))
+                except (ValueError, TypeError):
+                    continue
+                if math.isfinite(value) and value >= 0 and (name != 'gpu_usage_pct' or value <= 100):
+                    values.append(value)
+            if values:
+                metrics[name] = stats(values)
+        captures.append(dict(file=path.name, polls=len(rows), uniqueSamples=len(unique), metrics=metrics))
+    return captures
+
+
 def report(directory, metadata_directory=None):
     kernel_metadata = metadata(metadata_directory)
     stages = collections.defaultdict(list)
@@ -138,6 +170,7 @@ def report(directory, metadata_directory=None):
                   invalidKernelSamples=invalid_samples,
                   gpuBoundaryStages={},
                   presentMon=presentmon(directory),
+                  gpuTelemetry=gpu_telemetry(directory),
                   replayPairs=[],
                   kernels=[], notes=[
                       'CPU stage times include waits and host work; they are not isolated GPU timings.',
@@ -145,6 +178,7 @@ def report(directory, metadata_directory=None):
                       'D3D12 boundary timestamps span copies and HIP/scheduling across the external fence; they do not isolate kernel execution or measure Present.',
                       'Boundary timestamps are read after existing completion fences; only 32 timing bytes are read, with no extra completion wait. GPU clock idle behavior can affect intervals.',
                       'PresentMon ETW GPU busy/wait are per-process frame estimates, not hardware sensor utilization. HWS and cross-API context attribution can affect them.',
+                      'ADLX telemetry is driver-reported device-wide utilization/clocks/power, sampled in a separate read-only process. It does not isolate DLSS or WMMA occupancy.',
                       'Command record intervals are between NGX recording calls on one thread, not Present/FPS.',
                       'Hook estimates use random one-in-64 samples; driver timing covers generated forwarding methods only.',
                       'Hook access includes lock waits; summing threads does not measure serial frame latency or CPU execution time.',
@@ -218,7 +252,7 @@ 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 any(result[key] for key in ('cpuStages', 'kernels', 'commandStages', 'commandHookSamples', 'replayPairs', 'cudaApi', 'gpuBoundaryStages', 'presentMon')):
+    if not any(result[key] for key in ('cpuStages', 'kernels', 'commandStages', 'commandHookSamples', 'replayPairs', 'cudaApi', 'gpuBoundaryStages', 'presentMon', 'gpuTelemetry')):
         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')
@@ -232,6 +266,10 @@ def main():
         print(f"PRESENT {capture['file']} process={capture['processId']} swapchain={capture['swapchain']} frames={capture['frames']} fps={capture.get('presentFps', float('nan')):.3f}")
         for name, row in capture['metrics'].items():
             print(f"  {name}: n={row['samples']} mean={row['mean']:.3f} ms median={row['median']:.3f} ms p95={row['p95']:.3f} ms")
+    for capture in result['gpuTelemetry']:
+        print(f"ADLX {capture['file']}: polls={capture['polls']} unique_samples={capture['uniqueSamples']}")
+        for name, row in capture['metrics'].items():
+            print(f"  {name}: n={row['samples']} mean={row['mean']:.3f} median={row['median']:.3f} p95={row['p95']:.3f}")
     for row in result['cudaApi'][:20]:
         print(f"CUDA API {row['api']} {row['thread']}: n={row['calls']} mean={row['meanMs']:.6f} ms total={row['totalMs']:.3f} ms max={row['maxMs']:.3f} ms")
     for row in result['commandHookSamples']: