use rdtsc for profiler timestamps

src/base/base.cgh

@@ -226,7 +226,7 @@
   #define ThreadLocal __thread
 #endif
 
-//- Compiler memory barriers
+//- Compiler reordering barriers
 #if IsCompilerMsvc
   #define CompilerBarrier() _ReadWriteBarrier()
 #elif IsArchX64
@@ -263,12 +263,6 @@
 #define UsFromNs(ns) ((f64)((ns) / 1000.0))
 #define SecondsFromNs(ns) ((f64)(ns) / 1000000000.0)
 
-//- Busy-wait
-#if IsArchX64
-  #define YieldCpu()
-#else
-#endif
-
 ////////////////////////////////////////////////////////////
 //~ Linked list helper macros
 
@@ -556,7 +550,7 @@ Inline b32 MatchU128(u128 a, u128 b) { return a.lo == b.lo && a.hi == b.hi; }
   AlignedStruct(IsolatedAtomic32, IsolationSize) { Atomic32 v; };
   AlignedStruct(IsolatedAtomic64, IsolationSize) { Atomic64 v; };
 
-  #define Unatomic(atomic) (&(atomic)->_atom)
+  #define NonAtomic(atomic) (&(atomic)->_atom)
 
   #if IsPlatformWindows && IsArchX64
     //- 8 bit atomic ops
@@ -822,11 +816,14 @@ Inline u64 MixU64s(u64 seed_a, u64 seed_b)
 //~ @hookdecl Core api
 
 #if IsCpu
+  //- Application context
   StringList GetRawCommandline(void);
   String GetAppName(void);
   String GetEngineDirectory(void);
   String GetLibsDirectory(void);
   String GetAppDirectory(void);
+
+  //- Program context
   void LoadApi(ApiDesc api);
   void Echo(String msg);
   b32 Panic(String msg);
@@ -834,7 +831,15 @@ Inline u64 MixU64s(u64 seed_a, u64 seed_b)
   Callstack CaptureCallstack(u64 skip_frames);
   b32 IsRunningInDebugger(void);
   b32 IsRunningInWine(void);
+
+  //- Time
   i64 TimeNs(void);
+  u64 BeginTsc(void);
+  u64 EndTsc(void);
+  u64 GetStartupTsc(void);
+  f64 GetNsPerTsc(void);
+
+  //- Misc
   void TrueRand(String buffer);
   CpuTopologyInfo GetCpuTopologyInfo(void);
   void SleepSeconds(f64 seconds);

src/base/base_prof.c

@@ -1,7 +1,39 @@
 ////////////////////////////////////////////////////////////
 //~ Profile zone
 
-void PushProfZoneEx(char *name_cstr_lit, b32 end_zone)
+// void PushProfZoneEx(char *name_cstr_lit, b32 end_zone)
+// {
+//   ProfTrack *track = Base_tl.prof.thread_track;
+//   if (!track)
+//   {
+//     Arena *perm = PermArena();
+//     {
+//       track = PushStruct(perm, ProfTrack);
+//       Base_tl.prof.thread_track = track;
+//     }
+//     LockTicketMutex(&Base.prof.register_track_tm);
+//     {
+//       Base_tl.prof.track_idx = Atomic32Fetch(&Base.prof.tracks_count);
+//       Base.prof.tracks[Base_tl.prof.track_idx] = track;
+//       Atomic32FetchAdd(&Base.prof.tracks_count, 1);
+//     }
+//     UnlockTicketMutex(&Base.prof.register_track_tm);
+//   }
+
+//   u64 sample_seq = track->top_sample_seq;
+//   ProfSample *sample = &track->samples[sample_seq % ProfTrackSamplesCap];
+//   {
+//     // TODO: Wrap rdtsc for cpu-relative <-> program-relative timestamp conversion
+//     sample->stamp = __rdtsc() | ((u64)(!!end_zone) << 63);
+//     // sample->stamp = TimeNs() | ((u64)(!!end_zone) << 63);
+//     sample->name_cstr_lit = name_cstr_lit;
+//   }
+//   track->top_sample_seq = sample_seq + 1;
+// }
+
+
+
+void BeginProfZone(char *name_cstr_lit)
 {
   ProfTrack *track = Base_tl.prof.thread_track;
   if (!track)
@@ -19,16 +51,28 @@ void PushProfZoneEx(char *name_cstr_lit, b32 end_zone)
     }
     UnlockTicketMutex(&Base.prof.register_track_tm);
   }
-
+  u64 sample_seq = *NonAtomic(&track->top_sample_seq);
-  u64 sample_seq = track->top_sample_seq;
   ProfSample *sample = &track->samples[sample_seq % ProfTrackSamplesCap];
   {
-    // TODO: Wrap rdtsc for cpu-relative <-> program-relative timestamp conversion
+    sample->flags = 0;
-    // sample->stamp = __rdtsc() | ((u64)end_zone << 63);
-    sample->stamp = TimeNs() | ((u64)end_zone << 63);
     sample->name_cstr_lit = name_cstr_lit;
+    sample->tsc = BeginTsc();
   }
-  track->top_sample_seq = sample_seq + 1;
+  Atomic64Set(&track->top_sample_seq, sample_seq + 1);
+}
+
+void EndProfZone(void)
+{
+  i64 tsc = EndTsc();
+  ProfTrack *track = Base_tl.prof.thread_track;
+  u64 sample_seq = *NonAtomic(&track->top_sample_seq);
+  ProfSample *sample = &track->samples[sample_seq % ProfTrackSamplesCap];
+  {
+    sample->flags = ProfSampleFlag_EndZone;
+    sample->tsc = tsc;
+    sample->name_cstr_lit = 0;
+  }
+  Atomic64Set(&track->top_sample_seq, sample_seq + 1);
 }
 
 ////////////////////////////////////////////////////////////

src/base/base_prof.h

@@ -4,18 +4,23 @@
 #define ProfTrackSamplesCap Kibi(256)
 #define MaxRegisteredProfTracks Kibi(32)
 
-#define NsFromProfStamp(stamp) (stamp & ~((u64)1 << 63))
+Enum(ProfSampleFlag)
+{
+  ProfSampleFlag_None = 0,
+  ProfSampleFlag_EndZone = (1 << 0)
+};
 
 Struct(ProfSample)
 {
-  // Bit 63: zone-end, Bits 62-0: timestamp
+  ProfSampleFlag flags;
-  u64 stamp;
+  u8 _pad[4];
+  u64 tsc;
   char *name_cstr_lit;
 };
 
 Struct(ProfTrack)
 {
-  u64 top_sample_seq;
+  Atomic64 top_sample_seq;
   ProfSample samples[ProfTrackSamplesCap];
 };
 
@@ -45,10 +50,11 @@ Struct(ThreadLocalProfCtx)
 ////////////////////////////////////////////////////////////
 //~ Profile zone
 
-void PushProfZoneEx(char *name_cstr_lit, b32 end_zone);
+void BeginProfZone(char *name_cstr_lit);
+void EndProfZone(void);
 
 #if PROFILING_ENABLED
-  #define ProfZoneDF(name_cstr_lit)  DeferFor(PushProfZoneEx(name_cstr_lit"\0", 0), PushProfZoneEx(name_cstr_lit"\0", 1))
+  #define ProfZoneDF(name_cstr_lit)  DeferFor(BeginProfZone(name_cstr_lit"\0"), EndProfZone())
 #else
   #define ProfZoneDF(...)
 #endif

src/base/base_string.c

@@ -217,12 +217,17 @@ String StringFromTimeNs(Arena *arena, i64 time_ns, u32 precision)
 {
   String result = Zi;
   u64 abs_time_ns = AbsI64(time_ns);
-  if (abs_time_ns <= 999 * 1000)
+  if (abs_time_ns < 1000)
+  {
+    // Nanoseconds
+    result = StringF(arena, "%F ns", FmtSint(time_ns));
+  }
+  else if (abs_time_ns <= 1000 * 1000)
   {
     // Microseconds
     result = StringF(arena, "%F µs", FmtFloat(UsFromNs(time_ns), .p = precision));
   }
-  else if (abs_time_ns <= 999 * 1000000)
+  else if (abs_time_ns <= 1000 * 1000 * 1000)
   {
     // Milliseconds
     result = StringF(arena, "%F ms", FmtFloat(MsFromNs(time_ns), .p = precision));

src/base/base_win32/base_win32.c

@@ -199,9 +199,36 @@ b32 IsRunningInWine(void)
 i64 TimeNs(void)
 {
   LARGE_INTEGER qpc;
+  CompilerBarrier();
+  {
     QueryPerformanceCounter(&qpc);
-  i64 result = (qpc.QuadPart - W32.timer_start_qpc) * W32.ns_per_qpc;
+  }
-  return result;
+  CompilerBarrier();
+  return (qpc.QuadPart - W32.startup_qpc) * W32.ns_per_qpc;
+}
+
+u64 BeginTsc(void)
+{
+  _mm_lfence();
+  return __rdtsc();
+}
+
+u64 EndTsc(void)
+{
+  u32 aux;
+  u64 tsc = __rdtscp(&aux);
+  _mm_lfence();
+  return tsc;
+}
+
+u64 GetStartupTsc(void)
+{
+  return W32.startup_tsc;
+}
+
+f64 GetNsPerTsc(void)
+{
+  return W32.ns_per_tsc;
 }
 
 void TrueRand(String buffer)
@@ -557,16 +584,39 @@ i32 W32_Main(void)
   }
   #endif
 
-  // Init time
+  // Calibrate QPC
   {
     LARGE_INTEGER qpf = Zi;
     QueryPerformanceFrequency(&qpf);
-    W32.ns_per_qpc = 1000000000 / qpf.QuadPart;
+    W32.ns_per_qpc = 1000000000.0 / qpf.QuadPart;
   }
   {
     LARGE_INTEGER qpc = Zi;
     QueryPerformanceCounter(&qpc);
-    W32.timer_start_qpc = qpc.QuadPart;
+    W32.startup_qpc = qpc.QuadPart;
+  }
+
+  // Calibrate TSC
+  {
+    LARGE_INTEGER start_qpc;
+    LARGE_INTEGER end_qpc;
+    i64 start_tsc;
+    i64 end_tsc;
+    // Collect samples
+    u32 tsc_sleep_calibration_ms = 20;
+    {
+      start_tsc = __rdtsc();
+      QueryPerformanceCounter(&start_qpc);
+      {
+        Sleep(tsc_sleep_calibration_ms);
+      }
+      QueryPerformanceCounter(&end_qpc);
+      end_tsc = __rdtsc();
+    }
+    f64 elapsed_ns = (end_qpc.QuadPart - start_qpc.QuadPart) * W32.ns_per_qpc;
+    f64 elapsed_tsc = end_tsc - start_tsc;
+    W32.ns_per_tsc = elapsed_ns / elapsed_tsc;
+    W32.startup_tsc = end_tsc;
   }
 
   // Load module funcs

src/base/base_win32/base_win32.h

@@ -88,8 +88,10 @@ Struct(W32_Ctx)
   Atomic32 shutdown;
   Atomic32 exit_code;
 
-  i64 timer_start_qpc;
+  u64 startup_qpc;
-  i64 ns_per_qpc;
+  u64 startup_tsc;
+  u64 ns_per_qpc;
+  f64 ns_per_tsc;
 
   StringList raw_command_line;
   String app_name;

src/gpu/gpu_dx12/gpu_dx12_core.c

@@ -1412,26 +1412,6 @@ G_BaseDescriptorIndex G_PushMemory(G_CommandListHandle cl_handle, G_ArenaHandle
       }
     }
 
-
-
-
-
-
-
-
-    // FIXME: Remove this
-    if (memory_desc.buffer.stride == 52)
-    {
-      DEBUGBREAKABLE;
-    }
-
-
-
-
-
-
-
-
     // TODO: Less stringent reuse constraints. We could even create textures as placed resources and reset their underlying heaps.
     can_reuse = can_reuse && MatchStruct(&compare_d3d_desc, &d3d_desc);
     b32 can_reuse_descriptor = can_reuse && (!is_buffer || (MaxU64(memory_desc.buffer.stride, 1) == old_buffer_stride));

src/platform/platform_win32/platform_win32.c

@@ -67,7 +67,7 @@ void PLT_W32_SyncTimerForever(WaveLaneCtx *lane)
   for (;;)
   {
     {
-      // TODO: Minimum timer frequency in case timers ever become ultra precise in the future
+      // TODO: Clamp timer frequency in case timers ever become ultra precise in the future
       LARGE_INTEGER due = Zi;
       due.QuadPart = -1;
       //due.QuadPart = -10000;

src/pp/pp_sim/pp_sim_core.c

@@ -85,8 +85,7 @@ void S_TickForever(WaveLaneCtx *lane)
   //- Sim loop
 
   b32 shutdown = 0;
-  while (!shutdown)
+  while (!shutdown) ProfZoneDF("Sim tick")
-  ProfZoneDF("Sim tick")
   {
     shutdown = Atomic32Fetch(&S.shutdown);
     P_tl.debug_draw_enabled = TweakBool("Simulation debug draw", 0);
@@ -982,6 +981,7 @@ void S_TickForever(WaveLaneCtx *lane)
     //- Sleep
 
     if (!shutdown)
+    ProfZoneDF("Sleep sim")
     {
       i64 step_dt_ns = SIM_TICK_INTERVAL_NS;
       PLT_SleepFrame(begin_time_ns, step_dt_ns);

src/pp/pp_vis/pp_vis_core.c

@@ -871,8 +871,7 @@ void V_TickForever(WaveLaneCtx *lane)
   //- Begin vis loop
 
   b32 shutdown = 0;
-  while (!shutdown)
+  while (!shutdown) ProfZoneDF("Vis tick")
-  ProfZoneDF("Vis tick")
   {
     shutdown = Atomic32Fetch(&V.shutdown);
     P_tl.debug_draw_enabled = TweakBool("Vis debug draw", 0);
@@ -3632,7 +3631,7 @@ void V_TickForever(WaveLaneCtx *lane)
     //////////////////////////////
     //- Build panels
 
-    UI_PushDF(Parent, vis_game_box)
+    UI_SetDF(Parent, vis_game_box)
     // if (show_editor_ui && frame->is_editing)
     {
       Struct(PanelDfsNode) { PanelDfsNode *next; b32 visited; V_Panel *panel; UI_Checkpoint cp; };
@@ -4322,7 +4321,7 @@ void V_TickForever(WaveLaneCtx *lane)
     //- Build command palette
 
     V_Palette *palette = &frame->palette;
-    UI_PushDF(Parent, UI_RootKey)
+    UI_SetDF(Parent, UI_RootKey)
     {
       UI_Push(Tag, HashF("developer command palette"));
 
@@ -5393,7 +5392,7 @@ void V_TickForever(WaveLaneCtx *lane)
     profiler->ns_per_px = ClampF64(profiler->ns_per_px, min_profiler_ns_per_px, max_profiler_ns_per_px);
 
     UI_Key profiler_graph_box = UI_KeyF("graph");
-    UI_PushDF(OmitFlags, UI_BoxFlag_CaptureMouse * !!(frame->is_looking))
+    UI_SetDF(OmitFlags, UI_BoxFlag_CaptureMouse * !!(frame->is_looking))
     if (TweakBool("Show profiler", 1))
     {
       profiler->is_showing = 1;
@@ -5472,8 +5471,8 @@ void V_TickForever(WaveLaneCtx *lane)
 
       //- Build profiler
       UI_Key profiler_box = UI_KeyF("profiler");
-      UI_PushDF(Tag, profiler_box.v)
+      UI_SetDF(Tag, profiler_box.v)
-      UI_PushDF(Parent, profiler_box)
+      UI_SetDF(Parent, profiler_box)
       {
         UI_Key main_box = UI_KeyF("main");
 
@@ -5628,22 +5627,22 @@ void V_TickForever(WaveLaneCtx *lane)
         UI_SetNext(Height, profiler_height);
         UI_SetNext(Flags, UI_BoxFlag_CaptureMouse);
         UI_SetNext(Opacity, profiler_opacity);
-        UI_PushDF(Width, UI_Grow(1, 0))
+        UI_SetDF(Width, UI_Grow(1, 0))
-        UI_PushDF(Parent, UI_BuildColumnEx(profiler_box))
+        UI_SetDF(Parent, UI_BuildColumnEx(profiler_box))
-        UI_PushDF(Parent, UI_BuildRow())
+        UI_SetDF(Parent, UI_BuildRow())
         {
           UI_BuildSpacer(window_padding, Axis_X);
-          UI_PushDF(Parent, UI_BuildColumn())
+          UI_SetDF(Parent, UI_BuildColumn())
           {
             //- Header
-            // UI_PushDF(BackgroundColor, Color_Red)
+            // UI_SetDF(BackgroundColor, Color_Red)
             // UI_SetNext(BorderColor, Color_Red);
             // UI_SetNext(BorderSize, 1);
             UI_SetNext(ChildAlignment, UI_Region_Center);
-            UI_PushDF(Height, header_height)
+            UI_SetDF(Height, header_height)
-            UI_PushDF(Parent, UI_BuildBoxEx(UI_KeyF("profiler header")))
+            UI_SetDF(Parent, UI_BuildBoxEx(UI_KeyF("profiler header")))
             {
-              UI_PushDF(TextColor, theme.col.hint)
+              UI_SetDF(TextColor, theme.col.hint)
               UI_BuildLabelF("Profiler");
             }
             UI_BuildDivider(UI_Px(1, 1), theme.col.divider, Axis_Y);
@@ -5651,7 +5650,7 @@ void V_TickForever(WaveLaneCtx *lane)
             //- Graph
             // UI_SetNext(BackgroundColor, Color_Cyan);
             UI_SetNext(Height, graph_height);
-            UI_PushDF(Parent, UI_BuildBoxEx(profiler_graph_box))
+            UI_SetDF(Parent, UI_BuildBoxEx(profiler_graph_box))
             {
             }
             // UI_BuildDivider(UI_Px(1, 1), theme.col.divider, Axis_Y);
@@ -5667,16 +5666,19 @@ void V_TickForever(WaveLaneCtx *lane)
             UI_SetNext(Rounding, UI_Rpx(4));
             UI_SetNext(Flags, UI_BoxFlag_CaptureMouse | UI_BoxFlag_CaptureThroughChildren | UI_BoxFlag_Scissor);
             // UI_SetNext(Flags, UI_BoxFlag_CaptureMouse);
-            UI_PushDF(Parent, UI_BuildColumnEx(main_box))
+            UI_SetDF(Parent, UI_BuildColumnEx(main_box))
             {
               RegisteredProfTracksArray tracks = FetchRegisteredProfTracks(frame->arena);
               for (u64 track_idx = 0; track_idx < tracks.count; ++track_idx)
               {
+                // FIXME: Remove this
+                // if (track_idx != 0) continue;
+
                 UI_Key track_box = UI_KeyF("track %F", FmtUint(track_idx));
                 UI_SetNext(Height, track_height);
                 // UI_SetNext(Flags, UI_BoxFlag_CaptureMouse);
-                UI_PushDF(Tag, track_box.v)
+                UI_SetDF(Tag, track_box.v)
-                UI_PushDF(Parent, UI_BuildColumnEx(track_box))
+                UI_SetDF(Parent, UI_BuildColumnEx(track_box))
                 {
                   ProfTrack *track = tracks.tracks[track_idx];
 
@@ -5728,14 +5730,15 @@ void V_TickForever(WaveLaneCtx *lane)
 
 
                   // FIXME: Atomic
+                  u64 track_top_sample_seq = Atomic64Fetch(&track->top_sample_seq);
                   ProfSample *samples = track->samples;
-                  u64 samples_start_seq = track->top_sample_seq - MinU64(track->top_sample_seq, ProfTrackSamplesCap);
+                  u64 samples_start_seq = track_top_sample_seq - MinU64(track_top_sample_seq, ProfTrackSamplesCap);
-                  u64 samples_end_seq = track->top_sample_seq;
+                  u64 samples_end_seq = track_top_sample_seq;
 
 
 
-                  // samples_end_seq = MinU64(samples_end_seq, samples_start_seq + 512);
+                  samples_end_seq = MinU64(samples_end_seq, samples_start_seq + 512);
-                  samples_end_seq = MinU64(samples_end_seq, samples_start_seq + 64);
+                  // samples_end_seq = MinU64(samples_end_seq, samples_start_seq + 64);
                   // samples_end_seq = MinU64(samples_end_seq, samples_start_seq + Kibi(16));
 
 
@@ -5758,9 +5761,11 @@ void V_TickForever(WaveLaneCtx *lane)
                     {
                       ProfSample *sample = &samples[sample_seq % ProfTrackSamplesCap];
 
-                      i64 time_ns = NsFromProfStamp(sample->stamp);
 
-                      b32 is_end = !!(sample->stamp & ((u64)1 << 63));
+                      // u64 elapsed_ns = NsFromTsc
+                      i64 time_ns = (sample->tsc - GetStartupTsc()) * GetNsPerTsc();
+
+                      b32 is_end = !!(sample->flags & ProfSampleFlag_EndZone);
                       if (is_end)
                       {
                         depth -= 1;
@@ -5832,7 +5837,7 @@ void V_TickForever(WaveLaneCtx *lane)
                       //- Zone row
                       // UI_SetNext(Height, zone_height);
                       UI_SetNext(Height, zone_height);
-                      UI_PushDF(Parent, UI_BuildRow())
+                      UI_SetDF(Parent, UI_BuildRow())
                       {
                         //- Zones in row
                         u64 zone_idx = 0;
@@ -5848,14 +5853,18 @@ void V_TickForever(WaveLaneCtx *lane)
                             }
 
                             // TODO: Dim in HSV space
-                            Vec4 zone_color_dim = VEC4(0.15, 0.15, 0.15, 0);
 
                             Vec4 zone_color = zone->color;
                             // Vec4 zone_color_bd = Zi;
                             Vec4 zone_color_bd = zone_color;
-                            zone_color_bd.r -= zone_color_dim.r;
+
-                            zone_color_bd.g -= zone_color_dim.g;
+                            // Vec4 zone_color_dim = VEC4(0.15, 0.15, 0.15, 0);
-                            zone_color_bd.b -= zone_color_dim.b;
+                            // zone_color_bd.r -= zone_color_dim.r;
+                            // zone_color_bd.g -= zone_color_dim.g;
+                            // zone_color_bd.b -= zone_color_dim.b;
+
+                            zone_color_bd = MulVec4(zone_color_bd, 0.65);
+                            zone_color_bd.a = 1;
 
                             Vec4 zone_text_color = UI_Top(TextColor);
                             zone_text_color.a *= 0.75;
@@ -5890,7 +5899,7 @@ void V_TickForever(WaveLaneCtx *lane)
                             UI_SetNext(ChildAlignment, UI_Region_Center);
                             UI_SetNext(Flags, UI_BoxFlag_Floating | UI_BoxFlag_DontClampFloatingX | UI_BoxFlag_DontClampFloatingY | UI_BoxFlag_CaptureMouse | UI_BoxFlag_Scissor);
                             // UI_SetNext(Flags, UI_BoxFlag_Floating | UI_BoxFlag_DontClampFloatingX | UI_BoxFlag_DontClampFloatingY);
-                            UI_PushDF(Parent, UI_BuildRowEx(zone_box))
+                            UI_SetDF(Parent, UI_BuildRowEx(zone_box))
                             {
                               UI_BuildSpacer(UI_Px(zone_text_padding_px, 0.25), Axis_X);
 
@@ -5949,7 +5958,7 @@ void V_TickForever(WaveLaneCtx *lane)
                   //   UI_SetNext(FloatingPos, zone_pos);
                   //   UI_SetNext(Flags, UI_BoxFlag_Floating | UI_BoxFlag_DontClampFloatingX | UI_BoxFlag_DontClampFloatingY | UI_BoxFlag_CaptureMouse);
                   //   // UI_SetNext(Flags, UI_BoxFlag_Floating | UI_BoxFlag_DontClampFloatingX | UI_BoxFlag_DontClampFloatingY);
-                  //   UI_PushDF(Parent, UI_BuildBoxEx(zone_box))
+                  //   UI_SetDF(Parent, UI_BuildBoxEx(zone_box))
                   //   {
                   //   }
                   // }
@@ -6008,13 +6017,13 @@ void V_TickForever(WaveLaneCtx *lane)
                 UI_SetNext(Opacity, tooltip_opacity);
                 UI_SetNext(FloatingPos, tooltip_pos);
                 UI_SetNext(Flags, UI_BoxFlag_Floating);
-                UI_PushDF(Parent, UI_BuildRowEx(tooltip_box))
+                UI_SetDF(Parent, UI_BuildRowEx(tooltip_box))
                 {
                   UI_BuildSpacer(window_padding, Axis_X);
 
                   UI_SetNext(Width, UI_Shrink(0, 1));
                   UI_SetNext(Height, UI_Shrink(0, 1));
-                  UI_PushDF(Parent, UI_BuildColumn())
+                  UI_SetDF(Parent, UI_BuildColumn())
                   {
                     if (hovered_zone)
                     {
@@ -6040,9 +6049,9 @@ void V_TickForever(WaveLaneCtx *lane)
             // UI_BuildSpacer(window_padding, Axis_Y);
 
             //- Footer
-            UI_PushDF(Height, footer_height)
+            UI_SetDF(Height, footer_height)
-            UI_PushDF(ChildAlignment, UI_Region_Center)
+            UI_SetDF(ChildAlignment, UI_Region_Center)
-            UI_PushDF(Parent, UI_BuildRow())
+            UI_SetDF(Parent, UI_BuildRow())
             {
               UI_BuildSpacer(UI_Grow(1, 0), Axis_X);
 

src/ui/ui_core.c

@@ -534,8 +534,6 @@ void UI_SetRawTexture(UI_Key key, G_TextureRef tex, Rng2 uv)
 //~ Begin frame
 
 UI_Frame *UI_BeginFrame(UI_FrameFlag frame_flags)
-{
-  ProfZoneDF("Begin UI")
 {
   //////////////////////////////
   //- Init persistent state
@@ -579,6 +577,9 @@ UI_Frame *UI_BeginFrame(UI_FrameFlag frame_flags)
     frame->rects_arena = old_rects_arena;
   }
   frame->window_frame = WND_BeginFrame(G_Format_R16G16B16A16_Float, WND_BackbufferSizeMode_MatchMonitor);
+
+  ProfZoneDF("Begin UI")
+  {
     UI.cl = G_PrepareCommandList(G_QueueKind_Direct);
     ResetArena(frame->arena);
     ResetArena(frame->rects_arena);

src/ui/ui_core.h

@@ -510,7 +510,7 @@ UI_Style UI_PopStyle(UI_StyleDesc desc);
 #define UI_PeekTop(name, ...)       UI_PopStyle(UI_STYLEDESC(name, __VA_ARGS__)).name
 #define UI_Top(name, ...)           UI_PopStyle(UI_STYLEDESC(name, .use = 1, __VA_ARGS__)).name
 
-#define UI_PushDF(name, ...)        DeferFor(UI_Push(name, __VA_ARGS__), UI_Pop(name))
+#define UI_SetDF(name, ...)         DeferFor(UI_Push(name, __VA_ARGS__), UI_Pop(name))
 
 #define UI_PushCopy(name, src, ...) do {       \
   UI_StyleDesc _new = src;                     \

src/window/window_win32/window_win32.c

@@ -472,7 +472,8 @@ WND_Frame WND_BeginFrame(G_Format backbuffer_format, WND_BackbufferSizeMode back
 {
   WND_W32_Window *window = &WND_W32.window;
   WND_Frame result = Zi;
-
+  ProfZoneDF("Begin window")
+  {
     while (!Atomic32Fetch(&window->is_ready))
     {
       _mm_pause();
@@ -513,7 +514,11 @@ WND_Frame WND_BeginFrame(G_Format backbuffer_format, WND_BackbufferSizeMode back
     {
       backbuffer_size = result.monitor_size;
     }
+
+    ProfZoneDF("Prepare backbuffer")
+    {
       result.backbuffer = G_PrepareBackbuffer(window->swapchain, backbuffer_format, backbuffer_size);
+    }
 
     // Reset per-frame data
     if (!window->frame_arena)
@@ -587,13 +592,17 @@ WND_Frame WND_BeginFrame(G_Format backbuffer_format, WND_BackbufferSizeMode back
       result.restore = PushString(window->frame_arena, StringFromStruct(&restore));
     }
 
+  }
   return result;
 }
 
+
 void WND_EndFrame(WND_Frame frame, Vec2I32 backbuffer_pos, G_TextureRef src, Rng2I32 src_range, i32 vsync)
 {
   TempArena scratch = BeginScratchNoConflict();
   WND_W32_Window *window = WND_W32_WindowFromHandle(frame.window);
+  ProfZoneDF("End window")
+  {
     HWND hwnd = window->hwnd;
 
     // Process cmds
@@ -835,8 +844,12 @@ void WND_EndFrame(WND_Frame frame, Vec2I32 backbuffer_pos, G_TextureRef src, Rng
     }
 
     // Commit backbuffer
+    ProfZoneDF("Commit backbuffer")
+    {
       G_CommitBackbuffer(frame.backbuffer, backbuffer_pos, src, src_range, vsync);
+    }
 
     ++window->frame_gen;
+  }
   EndScratch(scratch);
 }