graphics: attribute native-pass GPU time to per-stage profile buckets

include/rex/graphics/native_rhi.h

@@ -398,6 +398,26 @@ struct Rect {
   int32_t bottom = 0;
 };
 
+// GPU-time attribution stages for ProfileRegion: the renderer marks the start
+// of each of its passes and the command processor's timestamp-bucket profiler
+// (d3d12_gpu_timestamp_buckets / vulkan_gpu_timestamp_buckets) reports the
+// per-stage GPU spans. kTail marks the end of the last stage; on Vulkan the
+// span from kTail to frame end is reported as its own bucket, on D3D12 it
+// stays in the profiler's untimed remainder.
+enum class ProfileStage : uint8_t {
+  kCommit,     // content-store commits, decode uploads, eviction
+  kShadow,     // dynamic caster shadow atlas + blur
+  kStaticSun,  // static sun-shadow cascade renders
+  kMain,       // main scene pass (opaque + transparent)
+  kResolve,    // MSAA resolve + outline composite
+  kAmbientOcclusion,
+  kSsr,
+  kVolumetrics,
+  kBloom,  // bloom pyramid + tonemap
+  k2d,     // photo grab, blur chains, 2D overlay replay
+  kTail,   // after the last stage: backend/present tail
+};
+
 // ---------------------------------------------------------------------------
 // Command recording. Wraps the command processor's deferred command list for
 // the current frame; valid only inside the render callback, render thread
@@ -469,6 +489,11 @@ class Cmd {
                        ResourceState after) = 0;
   virtual void FlushBarriers() = 0;
 
+  // Marks the start of a renderer pass for GPU-time attribution (closing the
+  // previously marked stage). No-op while the backend's timestamp-bucket
+  // profiler is disabled; see ProfileStage for the reporting semantics.
+  virtual void ProfileRegion(ProfileStage stage) = 0;
+
  protected:
   virtual ~Cmd() = default;
 };

include/rex/perf/counter.h

@@ -63,6 +63,20 @@ enum class DrawBucket : uint8_t {
   // The native guest-output renderer's whole GPU pass (scene + 2D replay +
   // resolve), recorded inside IssueSwap. Runs until frame end.
   kNativeScene,
+  // Native-pass stages (nrhi::Cmd::ProfileRegion): successive marks partition
+  // the native pass into its renderer-defined stages; when the renderer marks
+  // them, they replace kNativeScene in the frame's attribution.
+  kNativeCommit,   // content-store commits, decode uploads, eviction
+  kNativeShadow,   // dynamic caster shadow atlas + blur
+  kNativeSun,      // static sun-shadow cascade renders
+  kNativeMain,     // main scene pass (opaque + transparent + SSR G-buffer)
+  kNativeResolve,  // MSAA resolve + outline composite
+  kNativeAo,       // screen-space ambient occlusion
+  kNativeSsr,      // screen-space reflection march + composite
+  kNativeVol,      // volumetric sun shafts + haze
+  kNativeBloom,    // bloom pyramid + tonemap
+  kNative2d,       // photo grab, blur chains, 2D overlay replay
+  kNativeTail,     // after the last stage: backend/present tail
   kCount,
 };
 
@@ -103,6 +117,28 @@ inline const char* DrawBucketName(DrawBucket bucket) {
       return "TexUpCopy";
     case DrawBucket::kNativeScene:
       return "Native";
+    case DrawBucket::kNativeCommit:
+      return "NatCommit";
+    case DrawBucket::kNativeShadow:
+      return "NatShadow";
+    case DrawBucket::kNativeSun:
+      return "NatSun";
+    case DrawBucket::kNativeMain:
+      return "NatMain";
+    case DrawBucket::kNativeResolve:
+      return "NatResolve";
+    case DrawBucket::kNativeAo:
+      return "NatAO";
+    case DrawBucket::kNativeSsr:
+      return "NatSSR";
+    case DrawBucket::kNativeVol:
+      return "NatVol";
+    case DrawBucket::kNativeBloom:
+      return "NatBloom";
+    case DrawBucket::kNative2d:
+      return "Nat2D";
+    case DrawBucket::kNativeTail:
+      return "NatTail";
     default:
       return "Unknown";
   }

src/graphics/d3d12/native_rhi_d3d12.cpp

@@ -398,6 +398,7 @@ class NrCmdD3D12 : public nrhi::Cmd {
   void Barrier(nrhi::Texture* texture, ResourceState before,
                ResourceState after) override;
   void FlushBarriers() override;
+  void ProfileRegion(nrhi::ProfileStage stage) override;
 
   // Called at frame begin and on root-signature changes: the fresh deferred
   // command list / new root signature carries no bindings, so the latched
@@ -407,9 +408,20 @@ class NrCmdD3D12 : public nrhi::Cmd {
     std::memset(last_table_counts_, 0, sizeof(last_table_counts_));
   }
 
+  // Frame begin only (never on root-signature changes): each frame gets a
+  // fresh timestamp-query range, so an open profile region must not carry an
+  // end query across frames.
+  void ResetProfileRegion() { profile_region_query_ = UINT32_MAX; }
+
   NrDeviceD3D12* device = nullptr;
 
  private:
+  // Start query of the open ProfileRegion span, UINT32_MAX when none. Unlike
+  // Vulkan's auto-closing regions, D3D12 buckets are explicit begin/end query
+  // pairs; an unended pair would read stale readback memory, so every opened
+  // region is closed by the next mark (kTail closes without opening).
+  uint32_t profile_region_query_ = UINT32_MAX;
+
   void BindTextureTable(uint32_t param, NrTextureViewD3D12* const* views, uint32_t count);
 
   // Last tuple bound per root param: consecutive draws mostly rebind the
@@ -936,6 +948,7 @@ class NrDeviceD3D12 : public nrhi::Device {
                         uint32_t height, nrhi::Texture** guest_output_out) {
     guest_output_state_ = guest_output_internal_state;
     cmd_.ResetFrameState();
+    cmd_.ResetProfileRegion();
     const auto maint_t0 = std::chrono::steady_clock::now();
     const size_t dissolved_count = dissolved_views_.size();
     FlushDissolvedViews();
@@ -1417,6 +1430,47 @@ void NrCmdD3D12::Barrier(nrhi::Texture* texture, ResourceState before, ResourceS
 
 void NrCmdD3D12::FlushBarriers() { device->cp()->SubmitBarriers(); }
 
+rex::perf::DrawBucket ProfileStageBucket(nrhi::ProfileStage stage) {
+  switch (stage) {
+    case nrhi::ProfileStage::kCommit:
+      return rex::perf::DrawBucket::kNativeCommit;
+    case nrhi::ProfileStage::kShadow:
+      return rex::perf::DrawBucket::kNativeShadow;
+    case nrhi::ProfileStage::kStaticSun:
+      return rex::perf::DrawBucket::kNativeSun;
+    case nrhi::ProfileStage::kMain:
+      return rex::perf::DrawBucket::kNativeMain;
+    case nrhi::ProfileStage::kResolve:
+      return rex::perf::DrawBucket::kNativeResolve;
+    case nrhi::ProfileStage::kAmbientOcclusion:
+      return rex::perf::DrawBucket::kNativeAo;
+    case nrhi::ProfileStage::kSsr:
+      return rex::perf::DrawBucket::kNativeSsr;
+    case nrhi::ProfileStage::kVolumetrics:
+      return rex::perf::DrawBucket::kNativeVol;
+    case nrhi::ProfileStage::kBloom:
+      return rex::perf::DrawBucket::kNativeBloom;
+    case nrhi::ProfileStage::k2d:
+      return rex::perf::DrawBucket::kNative2d;
+    case nrhi::ProfileStage::kTail:
+      return rex::perf::DrawBucket::kNativeTail;
+  }
+  return rex::perf::DrawBucket::kNativeScene;
+}
+
+void NrCmdD3D12::ProfileRegion(nrhi::ProfileStage stage) {
+  if (profile_region_query_ != UINT32_MAX) {
+    device->cp()->EndGpuTimestampedDraw(profile_region_query_);
+    profile_region_query_ = UINT32_MAX;
+  }
+  if (stage == nrhi::ProfileStage::kTail) {
+    // No frame-end query to pair with on D3D12; the tail lands in the
+    // profiler's untimed remainder.
+    return;
+  }
+  profile_region_query_ = device->cp()->BeginGpuTimestampedDraw(ProfileStageBucket(stage));
+}
+
 }  // namespace
 
 nrhi::Device* CreateNativeRhiDevice(D3D12CommandProcessor* command_processor) {

src/graphics/vulkan/native_rhi_vulkan.cpp

@@ -440,6 +440,7 @@ class NrCmdVulkan : public nrhi::Cmd {
                            uint32_t height) override;
   void Barrier(nrhi::Texture* texture, ResourceState before, ResourceState after) override;
   void FlushBarriers() override;
+  void ProfileRegion(nrhi::ProfileStage stage) override;
 
   // --- frame handling (driven by the device) ---
   void ResetFrameState();
@@ -2693,6 +2694,41 @@ void NrCmdVulkan::FlushBarriers() {
   device->cp()->SubmitBarriers(true);
 }
 
+rex::perf::DrawBucket ProfileStageBucket(nrhi::ProfileStage stage) {
+  switch (stage) {
+    case nrhi::ProfileStage::kCommit:
+      return rex::perf::DrawBucket::kNativeCommit;
+    case nrhi::ProfileStage::kShadow:
+      return rex::perf::DrawBucket::kNativeShadow;
+    case nrhi::ProfileStage::kStaticSun:
+      return rex::perf::DrawBucket::kNativeSun;
+    case nrhi::ProfileStage::kMain:
+      return rex::perf::DrawBucket::kNativeMain;
+    case nrhi::ProfileStage::kResolve:
+      return rex::perf::DrawBucket::kNativeResolve;
+    case nrhi::ProfileStage::kAmbientOcclusion:
+      return rex::perf::DrawBucket::kNativeAo;
+    case nrhi::ProfileStage::kSsr:
+      return rex::perf::DrawBucket::kNativeSsr;
+    case nrhi::ProfileStage::kVolumetrics:
+      return rex::perf::DrawBucket::kNativeVol;
+    case nrhi::ProfileStage::kBloom:
+      return rex::perf::DrawBucket::kNativeBloom;
+    case nrhi::ProfileStage::k2d:
+      return rex::perf::DrawBucket::kNative2d;
+    case nrhi::ProfileStage::kTail:
+      return rex::perf::DrawBucket::kNativeTail;
+  }
+  return rex::perf::DrawBucket::kNativeScene;
+}
+
+void NrCmdVulkan::ProfileRegion(nrhi::ProfileStage stage) {
+  // Region semantics: beginning a bucket closes the active one, and the CP
+  // closes the last bucket at frame end - so the kTail mark records the
+  // present/submission tail as its own span.
+  device->cp()->BeginGpuTimestampedRegion(ProfileStageBucket(stage));
+}
+
 }  // namespace
 
 // ---------------------------------------------------------------------------