Merge pull request #3016 from secondlife/rye/tracyfixes

Tracy Improvements

15 files changed, 155 insertions, 112 deletions

diff --git a/indra/cmake/CMakeLists.txt b/indra/cmake/CMakeLists.txt
index 9d95a23a59..1083d7f2c4 100644
--- a/indra/cmake/CMakeLists.txt
+++ b/indra/cmake/CMakeLists.txt
@@ -57,6 +57,7 @@ set(cmake_SOURCE_FILES
         TemplateCheck.cmake
         TinyEXR.cmake
         TinyGLTF.cmake
+        Tracy.cmake
         Tut.cmake
         UI.cmake
         UnixInstall.cmake
diff --git a/indra/cmake/Tracy.cmake b/indra/cmake/Tracy.cmake
index 9eee9eb57f..e3f46cca9b 100644
--- a/indra/cmake/Tracy.cmake
+++ b/indra/cmake/Tracy.cmake
@@ -20,6 +20,7 @@ endif(LINUX)
 if (USE_TRACY)
   option(USE_TRACY_ON_DEMAND "Use on-demand Tracy profiling." ON)
   option(USE_TRACY_LOCAL_ONLY "Disallow remote Tracy profiling." OFF)
+  option(USE_TRACY_GPU "Use Tracy GPU profiling" OFF)
 
   use_system_binary(tracy)
   use_prebuilt_binary(tracy)
@@ -36,9 +37,8 @@ if (USE_TRACY)
     target_compile_definitions(ll::tracy INTERFACE -DTRACY_NO_BROADCAST=1 -DTRACY_ONLY_LOCALHOST=1)
   endif ()
 
-  # GHA runners don't always provide invariant TSC support, but always build with LL_TESTS enabled
-  if (DARWIN AND LL_TESTS)
-    target_compile_definitions(ll::tracy INTERFACE -DTRACY_TIMER_FALLBACK=1)
+  if (USE_TRACY_GPU AND NOT DARWIN) # Tracy OpenGL mode is incompatible with macOS/iOS
+    target_compile_definitions(ll::tracy INTERFACE -DLL_PROFILER_ENABLE_TRACY_OPENGL=1)
   endif ()
 
   # See: indra/llcommon/llprofiler.h
diff --git a/indra/llcommon/linden_common.h b/indra/llcommon/linden_common.h
index a918caa2e8..a41af153fe 100644
--- a/indra/llcommon/linden_common.h
+++ b/indra/llcommon/linden_common.h
@@ -28,12 +28,6 @@
 #define LL_LINDEN_COMMON_H
 
 #include "llprofiler.h"
-#if TRACY_ENABLE && !defined(LL_PROFILER_ENABLE_TRACY_OPENGL)  // hooks for memory profiling
-void *tracy_aligned_malloc(size_t size, size_t alignment);
-void  tracy_aligned_free(void *memblock);
-#define _aligned_malloc(X, Y) tracy_aligned_malloc((X), (Y))
-#define _aligned_free(X)      tracy_aligned_free((X))
-#endif
 
 // *NOTE:  Please keep includes here to a minimum!
 //
diff --git a/indra/llcommon/llcommon.cpp b/indra/llcommon/llcommon.cpp
index f1f3958fe0..110a1c32a1 100644
--- a/indra/llcommon/llcommon.cpp
+++ b/indra/llcommon/llcommon.cpp
@@ -33,23 +33,23 @@
 #include "lltracethreadrecorder.h"
 #include "llcleanup.h"
 
-thread_local bool gProfilerEnabled = false;
-
-#if (TRACY_ENABLE)
+#if LL_PROFILER_CONFIGURATION > 1 && TRACY_ENABLE
 // Override new/delete for tracy memory profiling
 
 void* ll_tracy_new(size_t size)
 {
-    void* ptr;
-    if (gProfilerEnabled)
-    {
-        //LL_PROFILE_ZONE_SCOPED_CATEGORY_MEMORY;
-        ptr = (malloc)(size);
-    }
-    else
+    void* ptr = (malloc)(size);
+    if (!ptr)
     {
-        ptr = (malloc)(size);
+        throw std::bad_alloc();
     }
+    TracyAlloc(ptr, size);
+    return ptr;
+}
+
+void* ll_tracy_aligned_new(size_t size, size_t alignment)
+{
+    void* ptr = ll_aligned_malloc_fallback(size, alignment);
     if (!ptr)
     {
         throw std::bad_alloc();
@@ -58,6 +58,18 @@ void* ll_tracy_new(size_t size)
     return ptr;
 }
 
+void ll_tracy_delete(void* ptr)
+{
+    TracyFree(ptr);
+    (free)(ptr);
+}
+
+void ll_tracy_aligned_delete(void* ptr)
+{
+    TracyFree(ptr);
+    ll_aligned_free_fallback(ptr);
+}
+
 void* operator new(size_t size)
 {
     return ll_tracy_new(size);
@@ -68,18 +80,14 @@ void* operator new[](std::size_t count)
     return ll_tracy_new(count);
 }
 
-void ll_tracy_delete(void* ptr)
+void* operator new(size_t size, std::align_val_t align)
 {
-    TracyFree(ptr);
-    if (gProfilerEnabled)
-    {
-        //LL_PROFILE_ZONE_SCOPED_CATEGORY_MEMORY;
-        (free)(ptr);
-    }
-    else
-    {
-        (free)(ptr);
-    }
+    return ll_tracy_aligned_new(size, (size_t)align);
+}
+
+void* operator new[](std::size_t count, std::align_val_t align)
+{
+    return ll_tracy_aligned_new(count, (size_t)align);
 }
 
 void operator delete(void *ptr) noexcept
@@ -92,27 +100,17 @@ void operator delete[](void* ptr) noexcept
     ll_tracy_delete(ptr);
 }
 
-// C-style malloc/free can't be so easily overridden, so we define tracy versions and use
-// a pre-processor #define in linden_common.h to redirect to them. The parens around the native
-// functions below prevents recursive substitution by the preprocessor.
-//
-// Unaligned mallocs are rare in LL code but hooking them causes problems in 3p lib code (looking at
-// you, Havok), so we'll only capture the aligned version.
-
-void *tracy_aligned_malloc(size_t size, size_t alignment)
+void operator delete(void *ptr, std::align_val_t align) noexcept
 {
-    auto ptr = ll_aligned_malloc_fallback(size, alignment);
-    if (ptr) TracyAlloc(ptr, size);
-    return ptr;
+    ll_tracy_aligned_delete(ptr);
 }
 
-void tracy_aligned_free(void *memblock)
+void operator delete[](void* ptr, std::align_val_t align) noexcept
 {
-    TracyFree(memblock);
-    ll_aligned_free_fallback(memblock);
+    ll_tracy_aligned_delete(ptr);
 }
 
-#endif
+#endif // TRACY_ENABLE && !LL_PROFILER_ENABLE_TRACY_OPENGL
 
 //static
 bool LLCommon::sAprInitialized = false;
diff --git a/indra/llcommon/llmemory.h b/indra/llcommon/llmemory.h
index 80cfe554c4..72aec57080 100644
--- a/indra/llcommon/llmemory.h
+++ b/indra/llcommon/llmemory.h
@@ -222,7 +222,7 @@ inline void* ll_aligned_realloc_16(void* ptr, size_t size, size_t old_size) // r
         ll_aligned_free_16(ptr);
     }
 #endif
-    LL_PROFILE_ALLOC(ptr, size);
+    LL_PROFILE_ALLOC(ret, size);
     return ret;
 }
 
@@ -231,8 +231,6 @@ inline void* ll_aligned_malloc_32(size_t size) // returned hunk MUST be freed wi
     LL_PROFILE_ZONE_SCOPED_CATEGORY_MEMORY;
 #if defined(LL_WINDOWS)
     void* ret = _aligned_malloc(size, 32);
-#elif defined(LL_DARWIN)
-    void* ret = ll_aligned_malloc_fallback( size, 32 );
 #else
     void *ret;
     if (0 != posix_memalign(&ret, 32, size))
@@ -248,8 +246,31 @@ inline void ll_aligned_free_32(void *p)
     LL_PROFILE_FREE(p);
 #if defined(LL_WINDOWS)
     _aligned_free(p);
-#elif defined(LL_DARWIN)
-    ll_aligned_free_fallback( p );
+#else
+    free(p); // posix_memalign() is compatible with heap deallocator
+#endif
+}
+
+inline void* ll_aligned_malloc_64(size_t size) // returned hunk MUST be freed with ll_aligned_free_32().
+{
+    LL_PROFILE_ZONE_SCOPED_CATEGORY_MEMORY;
+#if defined(LL_WINDOWS)
+    void* ret = _aligned_malloc(size, 64);
+#else
+    void *ret;
+    if (0 != posix_memalign(&ret, 64, size))
+        return nullptr;
+#endif
+    LL_PROFILE_ALLOC(ret, size);
+    return ret;
+}
+
+inline void ll_aligned_free_64(void *p)
+{
+    LL_PROFILE_ZONE_SCOPED_CATEGORY_MEMORY;
+    LL_PROFILE_FREE(p);
+#if defined(LL_WINDOWS)
+    _aligned_free(p);
 #else
     free(p); // posix_memalign() is compatible with heap deallocator
 #endif
@@ -261,19 +282,23 @@ LL_FORCE_INLINE void* ll_aligned_malloc(size_t size)
 {
     LL_PROFILE_ZONE_SCOPED_CATEGORY_MEMORY;
     void* ret;
-    if (LL_DEFAULT_HEAP_ALIGN % ALIGNMENT == 0)
+    if constexpr (LL_DEFAULT_HEAP_ALIGN % ALIGNMENT == 0)
     {
         ret = malloc(size);
         LL_PROFILE_ALLOC(ret, size);
     }
-    else if (ALIGNMENT == 16)
+    else if constexpr (ALIGNMENT == 16)
     {
         ret = ll_aligned_malloc_16(size);
     }
-    else if (ALIGNMENT == 32)
+    else if constexpr (ALIGNMENT == 32)
     {
         ret = ll_aligned_malloc_32(size);
     }
+    else if constexpr (ALIGNMENT == 64)
+    {
+        ret = ll_aligned_malloc_64(size);
+    }
     else
     {
         ret = ll_aligned_malloc_fallback(size, ALIGNMENT);
@@ -285,16 +310,20 @@ template<size_t ALIGNMENT>
 LL_FORCE_INLINE void ll_aligned_free(void* ptr)
 {
     LL_PROFILE_ZONE_SCOPED_CATEGORY_MEMORY;
-    if (ALIGNMENT == LL_DEFAULT_HEAP_ALIGN)
+    if constexpr (ALIGNMENT == LL_DEFAULT_HEAP_ALIGN)
     {
         LL_PROFILE_FREE(ptr);
         free(ptr);
     }
-    else if (ALIGNMENT == 16)
+    else if constexpr (ALIGNMENT == 16)
     {
         ll_aligned_free_16(ptr);
     }
-    else if (ALIGNMENT == 32)
+    else if constexpr (ALIGNMENT == 32)
+    {
+        return ll_aligned_free_32(ptr);
+    }
+    else if constexpr (ALIGNMENT == 64)
     {
         return ll_aligned_free_32(ptr);
     }
diff --git a/indra/llcommon/llprofiler.h b/indra/llcommon/llprofiler.h
index f6a4d24747..6ff4341b69 100644
--- a/indra/llcommon/llprofiler.h
+++ b/indra/llcommon/llprofiler.h
@@ -74,23 +74,18 @@
 #define LL_PROFILER_CONFIGURATION           LL_PROFILER_CONFIG_FAST_TIMER
 #endif
 
-extern thread_local bool gProfilerEnabled;
-
 #if defined(LL_PROFILER_CONFIGURATION) && (LL_PROFILER_CONFIGURATION > LL_PROFILER_CONFIG_NONE)
     #if LL_PROFILER_CONFIGURATION == LL_PROFILER_CONFIG_TRACY || LL_PROFILER_CONFIGURATION == LL_PROFILER_CONFIG_TRACY_FAST_TIMER
         #include "tracy/Tracy.hpp"
 
-        // Enable OpenGL profiling
-        #define LL_PROFILER_ENABLE_TRACY_OPENGL 0
-
         // Enable RenderDoc labeling
-        #define LL_PROFILER_ENABLE_RENDER_DOC 0
+        //#define LL_PROFILER_ENABLE_RENDER_DOC 0
 
     #endif
 
     #if LL_PROFILER_CONFIGURATION == LL_PROFILER_CONFIG_TRACY
         #define LL_PROFILER_FRAME_END                   FrameMark
-        #define LL_PROFILER_SET_THREAD_NAME( name )     tracy::SetThreadName( name );    gProfilerEnabled = true;
+        #define LL_PROFILER_SET_THREAD_NAME( name )     tracy::SetThreadName( name );
         #define LL_RECORD_BLOCK_TIME(name)              ZoneScoped // Want descriptive names; was: ZoneNamedN( ___tracy_scoped_zone, #name, true );
         #define LL_PROFILE_ZONE_NAMED(name)             ZoneNamedN( ___tracy_scoped_zone, name, true );
         #define LL_PROFILE_ZONE_NAMED_COLOR(name,color) ZoneNamedNC( ___tracy_scopped_zone, name, color, true ) // RGB
@@ -133,7 +128,7 @@ extern thread_local bool gProfilerEnabled;
     #endif
     #if LL_PROFILER_CONFIGURATION == LL_PROFILER_CONFIG_TRACY_FAST_TIMER
         #define LL_PROFILER_FRAME_END                   FrameMark
-        #define LL_PROFILER_SET_THREAD_NAME( name )     tracy::SetThreadName( name );    gProfilerEnabled = true;
+        #define LL_PROFILER_SET_THREAD_NAME( name )     tracy::SetThreadName( name );
         #define LL_RECORD_BLOCK_TIME(name)              ZoneNamedN(___tracy_scoped_zone, #name, true);   const LLTrace::BlockTimer& LL_GLUE_TOKENS(block_time_recorder, __LINE__)(LLTrace::timeThisBlock(name)); (void)LL_GLUE_TOKENS(block_time_recorder, __LINE__);
         #define LL_PROFILE_ZONE_NAMED(name)             ZoneNamedN( ___tracy_scoped_zone, #name, true );
         #define LL_PROFILE_ZONE_NAMED_COLOR(name,color) ZoneNamedNC( ___tracy_scopped_zone, name, color, true ) // RGB
@@ -158,23 +153,20 @@ extern thread_local bool gProfilerEnabled;
 #endif // LL_PROFILER
 
 #if LL_PROFILER_ENABLE_TRACY_OPENGL
-#define LL_PROFILE_GPU_ZONE(name)        TracyGpuZone(name)
-#define LL_PROFILE_GPU_ZONEC(name,color) TracyGpuZoneC(name,color)
+#define LL_PROFILE_GPU_ZONE(name)         TracyGpuZone(name)
+#define LL_PROFILE_GPU_ZONEC(name,color)  TracyGpuZoneC(name,color)
 #define LL_PROFILER_GPU_COLLECT           TracyGpuCollect
 #define LL_PROFILER_GPU_CONTEXT           TracyGpuContext
-
-// disable memory tracking (incompatible with GPU tracing
-#define LL_PROFILE_ALLOC(ptr, size)             (void)(ptr); (void)(size);
-#define LL_PROFILE_FREE(ptr)                    (void)(ptr);
+#define LL_PROFILER_GPU_CONTEXT_NAMED     TracyGpuContextName
 #else
-#define LL_PROFILE_GPU_ZONE(name)        (void)name;
-#define LL_PROFILE_GPU_ZONEC(name,color) (void)name;(void)color;
+#define LL_PROFILE_GPU_ZONE(name)           (void)name;
+#define LL_PROFILE_GPU_ZONEC(name,color)    (void)name;(void)color;
 #define LL_PROFILER_GPU_COLLECT
 #define LL_PROFILER_GPU_CONTEXT
+#define LL_PROFILER_GPU_CONTEXT_NAMED(name) (void)name;
+#endif // LL_PROFILER_ENABLE_TRACY_OPENGL
 
-#define LL_LABEL_OBJECT_GL(type, name, length, label)
-
-#if !LL_DARWIN && LL_PROFILER_CONFIGURATION > 1
+#if LL_PROFILER_CONFIGURATION > 1
 #define LL_PROFILE_ALLOC(ptr, size)             TracyAlloc(ptr, size)
 #define LL_PROFILE_FREE(ptr)                    TracyFree(ptr)
 #else
@@ -182,9 +174,7 @@ extern thread_local bool gProfilerEnabled;
 #define LL_PROFILE_FREE(ptr)                    (void)(ptr);
 #endif
 
-#endif
-
-#if LL_PROFILER_ENABLE_RENDER_DOC
+#ifdef LL_PROFILER_ENABLE_RENDER_DOC
 #define LL_LABEL_OBJECT_GL(type, name, length, label) glObjectLabel(type, name, length, label)
 #else
 #define LL_LABEL_OBJECT_GL(type, name, length, label)
diff --git a/indra/llwindow/llwindowwin32.cpp b/indra/llwindow/llwindowwin32.cpp
index f349c4aea8..f926d106be 100644
--- a/indra/llwindow/llwindowwin32.cpp
+++ b/indra/llwindow/llwindowwin32.cpp
@@ -1661,6 +1661,11 @@ const   S32   max_format  = (S32)num_formats - 1;
         return false;
     }
 
+    // Setup Tracy gpu context
+    {
+        LL_PROFILER_GPU_CONTEXT;
+    }
+
     // Disable vertical sync for swap
     toggleVSync(enable_vsync);
 
@@ -1692,8 +1697,6 @@ const   S32   max_format  = (S32)num_formats - 1;
         swapBuffers();
     }
 
-    LL_PROFILER_GPU_CONTEXT;
-
     return true;
 }
 
diff --git a/indra/newview/llappviewer.cpp b/indra/newview/llappviewer.cpp
index 423fe2469a..fda4f043b0 100644
--- a/indra/newview/llappviewer.cpp
+++ b/indra/newview/llappviewer.cpp
@@ -1473,6 +1473,7 @@ void sendGameControlInput()
 bool LLAppViewer::doFrame()
 {
     LL_RECORD_BLOCK_TIME(FTM_FRAME);
+    LL_PROFILE_GPU_ZONE("Frame");
     {
     // and now adjust the visuals from previous frame.
     if(LLPerfStats::tunables.userAutoTuneEnabled && LLPerfStats::tunables.tuningFlag != LLPerfStats::Tunables::Nothing)
@@ -1562,24 +1563,26 @@ bool LLAppViewer::doFrame()
 
         if (!LLApp::isExiting())
         {
-            LL_PROFILE_ZONE_NAMED_CATEGORY_APP("df JoystickKeyboard");
-            pingMainloopTimeout("Main:JoystickKeyboard");
-
-            // Scan keyboard for movement keys.  Command keys and typing
-            // are handled by windows callbacks.  Don't do this until we're
-            // done initializing.  JC
-            if (gViewerWindow
-                && (gHeadlessClient || gViewerWindow->getWindow()->getVisible())
-                && gViewerWindow->getActive()
-                && !gViewerWindow->getWindow()->getMinimized()
-                && LLStartUp::getStartupState() == STATE_STARTED
-                && (gHeadlessClient || !gViewerWindow->getShowProgress())
-                && !gFocusMgr.focusLocked())
             {
-                LLPerfStats::RecordSceneTime T (LLPerfStats::StatType_t::RENDER_IDLE);
-                joystick->scanJoystick();
-                gKeyboard->scanKeyboard();
-                gViewerInput.scanMouse();
+                LL_PROFILE_ZONE_NAMED_CATEGORY_APP("df JoystickKeyboard");
+                pingMainloopTimeout("Main:JoystickKeyboard");
+
+                // Scan keyboard for movement keys.  Command keys and typing
+                // are handled by windows callbacks.  Don't do this until we're
+                // done initializing.  JC
+                if (gViewerWindow
+                    && (gHeadlessClient || gViewerWindow->getWindow()->getVisible())
+                    && gViewerWindow->getActive()
+                    && !gViewerWindow->getWindow()->getMinimized()
+                    && LLStartUp::getStartupState() == STATE_STARTED
+                    && (gHeadlessClient || !gViewerWindow->getShowProgress())
+                    && !gFocusMgr.focusLocked())
+                {
+                    LLPerfStats::RecordSceneTime T(LLPerfStats::StatType_t::RENDER_IDLE);
+                    joystick->scanJoystick();
+                    gKeyboard->scanKeyboard();
+                    gViewerInput.scanMouse();
+                }
             }
 
             // Update state based on messages, user input, object idle.
diff --git a/indra/newview/llappviewerlinux.cpp b/indra/newview/llappviewerlinux.cpp
index 3dbbda11f9..febe2d1e41 100644
--- a/indra/newview/llappviewerlinux.cpp
+++ b/indra/newview/llappviewerlinux.cpp
@@ -113,6 +113,11 @@ static void exceptionTerminateHandler()
 
 int main( int argc, char **argv )
 {
+    // Call Tracy first thing to have it allocate memory
+    // https://github.com/wolfpld/tracy/issues/196
+    LL_PROFILER_FRAME_END;
+    LL_PROFILER_SET_THREAD_NAME("App");
+
     gArgC = argc;
     gArgV = argv;
 
diff --git a/indra/newview/llappviewermacosx.cpp b/indra/newview/llappviewermacosx.cpp
index 4162c0479a..f497a3cdf3 100644
--- a/indra/newview/llappviewermacosx.cpp
+++ b/indra/newview/llappviewermacosx.cpp
@@ -231,6 +231,11 @@ void infos(const std::string& message)
 
 int main( int argc, char **argv )
 {
+    // Call Tracy first thing to have it allocate memory
+    // https://github.com/wolfpld/tracy/issues/196
+    LL_PROFILER_FRAME_END;
+    LL_PROFILER_SET_THREAD_NAME("App");
+
     // Store off the command line args for use later.
     gArgC = argc;
     gArgV = argv;
diff --git a/indra/newview/lldrawpool.h b/indra/newview/lldrawpool.h
index bc412214c7..34257f98f8 100644
--- a/indra/newview/lldrawpool.h
+++ b/indra/newview/lldrawpool.h
@@ -339,7 +339,7 @@ public:
         }
     }
     #else
-    static inline const char* lookupPass(U32 pass) { return ""; }
+    static inline const char* lookupPassName(U32 pass) { return ""; }
     #endif
 
     LLRenderPass(const U32 type);
diff --git a/indra/newview/llheroprobemanager.cpp b/indra/newview/llheroprobemanager.cpp
index ce419498cf..8e3cd0626f 100644
--- a/indra/newview/llheroprobemanager.cpp
+++ b/indra/newview/llheroprobemanager.cpp
@@ -81,6 +81,7 @@ void LLHeroProbeManager::update()
     }
 
     LL_PROFILE_ZONE_SCOPED_CATEGORY_DISPLAY;
+    LL_PROFILE_GPU_ZONE("hero manager update");
     llassert(!gCubeSnapshot); // assert a snapshot is not in progress
     if (LLAppViewer::instance()->logoutRequestSent())
     {
@@ -279,6 +280,9 @@ void LLHeroProbeManager::renderProbes()
 // In effect this simulates single-bounce lighting.
 void LLHeroProbeManager::updateProbeFace(LLReflectionMap* probe, U32 face, bool is_dynamic, F32 near_clip)
 {
+    LL_PROFILE_ZONE_SCOPED_CATEGORY_DISPLAY;
+    LL_PROFILE_GPU_ZONE("hero probe update");
+
     // hacky hot-swap of camera specific render targets
     gPipeline.mRT = &gPipeline.mHeroProbeRT;
 
@@ -349,7 +353,7 @@ void LLHeroProbeManager::updateProbeFace(LLReflectionMap* probe, U32 face, bool
 
         for (int i = 0; i < mMipChain.size(); ++i)
         {
-            LL_PROFILE_GPU_ZONE("probe mip");
+            LL_PROFILE_GPU_ZONE("hero probe mip");
             mMipChain[i].bindTarget();
             if (i == 0)
             {
@@ -376,7 +380,7 @@ void LLHeroProbeManager::updateProbeFace(LLReflectionMap* probe, U32 face, bool
 
             if (mip >= 0)
             {
-                LL_PROFILE_GPU_ZONE("probe mip copy");
+                LL_PROFILE_GPU_ZONE("hero probe mip copy");
                 mTexture->bind(0);
 
                 glCopyTexSubImage3D(GL_TEXTURE_CUBE_MAP_ARRAY, mip, 0, 0, sourceIdx * 6 + face, 0, 0, res, res);
@@ -424,7 +428,7 @@ void LLHeroProbeManager::generateRadiance(LLReflectionMap* probe)
 
             for (int i = 0; i < mMipChain.size() / 4; ++i)
             {
-                LL_PROFILE_GPU_ZONE("probe radiance gen");
+                LL_PROFILE_GPU_ZONE("hero probe radiance gen");
                 static LLStaticHashedString sMipLevel("mipLevel");
                 static LLStaticHashedString sRoughness("roughness");
                 static LLStaticHashedString sWidth("u_width");
@@ -471,6 +475,7 @@ void LLHeroProbeManager::updateUniforms()
     }
 
     LL_PROFILE_ZONE_SCOPED_CATEGORY_DISPLAY;
+    LL_PROFILE_GPU_ZONE("hpmu - uniforms")
 
     LLMatrix4a modelview;
     modelview.loadu(gGLModelView);
diff --git a/indra/newview/llreflectionmapmanager.cpp b/indra/newview/llreflectionmapmanager.cpp
index f083747bfe..4b79ce03eb 100644
--- a/indra/newview/llreflectionmapmanager.cpp
+++ b/indra/newview/llreflectionmapmanager.cpp
@@ -210,6 +210,7 @@ void LLReflectionMapManager::update()
     }
 
     LL_PROFILE_ZONE_SCOPED_CATEGORY_DISPLAY;
+    LL_PROFILE_GPU_ZONE("reflection manager update");
     llassert(!gCubeSnapshot); // assert a snapshot is not in progress
     if (LLAppViewer::instance()->logoutRequestSent())
     {
@@ -677,6 +678,8 @@ void LLReflectionMapManager::doProbeUpdate()
 // In effect this simulates single-bounce lighting.
 void LLReflectionMapManager::updateProbeFace(LLReflectionMap* probe, U32 face)
 {
+    LL_PROFILE_ZONE_SCOPED_CATEGORY_DISPLAY;
+    LL_PROFILE_GPU_ZONE("probe update");
     // hacky hot-swap of camera specific render targets
     gPipeline.mRT = &gPipeline.mAuxillaryRT;
 
@@ -991,6 +994,7 @@ void LLReflectionMapManager::updateUniforms()
     }
 
     LL_PROFILE_ZONE_SCOPED_CATEGORY_DISPLAY;
+    LL_PROFILE_GPU_ZONE("rmmu - uniforms")
 
     // structure for packing uniform buffer object
     // see class3/deferred/reflectionProbeF.glsl
diff --git a/indra/newview/llviewerdisplay.cpp b/indra/newview/llviewerdisplay.cpp
index b0e07d8ea3..94ebca77e2 100644
--- a/indra/newview/llviewerdisplay.cpp
+++ b/indra/newview/llviewerdisplay.cpp
@@ -406,6 +406,7 @@ static void update_tp_display(bool minimized)
 void display(bool rebuild, F32 zoom_factor, int subfield, bool for_snapshot)
 {
     LL_PROFILE_ZONE_NAMED_CATEGORY_DISPLAY("Render");
+    LL_PROFILE_GPU_ZONE("Render");
 
     LLPerfStats::RecordSceneTime T (LLPerfStats::StatType_t::RENDER_DISPLAY); // render time capture - This is the main stat for overall rendering.
 
@@ -710,6 +711,7 @@ void display(bool rebuild, F32 zoom_factor, int subfield, bool for_snapshot)
         if (gPipeline.RenderMirrors && !gSnapshot)
         {
             LL_PROFILE_ZONE_NAMED_CATEGORY_DISPLAY("Update hero probes");
+            LL_PROFILE_GPU_ZONE("hero manager")
             gPipeline.mHeroProbeManager.update();
             gPipeline.mHeroProbeManager.renderProbes();
         }
diff --git a/indra/newview/pipeline.cpp b/indra/newview/pipeline.cpp
index 65f7c1d70f..1edbf584d6 100644
--- a/indra/newview/pipeline.cpp
+++ b/indra/newview/pipeline.cpp
@@ -7058,11 +7058,12 @@ extern LLPointer<LLImageGL> gEXRImage;
 
 void LLPipeline::tonemap(LLRenderTarget* src, LLRenderTarget* dst, bool gamma_correct/* = true*/)
 {
-    dst->bindTarget();
     // gamma correct lighting
     {
         LL_PROFILE_GPU_ZONE("tonemap");
 
+        dst->bindTarget();
+
         static LLCachedControl<bool> buildNoPost(gSavedSettings, "RenderDisablePostProcessing", false);
 
         LLGLDepthTest depth(GL_FALSE, GL_FALSE);
@@ -7115,13 +7116,13 @@ void LLPipeline::tonemap(LLRenderTarget* src, LLRenderTarget* dst, bool gamma_co
 
         gGL.getTexUnit(channel)->unbind(src->getUsage());
         shader->unbind();
+
+        dst->flush();
     }
-    dst->flush();
 }
 
 void LLPipeline::copyScreenSpaceReflections(LLRenderTarget* src, LLRenderTarget* dst)
 {
-
     if (RenderScreenSpaceReflections && !gCubeSnapshot)
     {
         LL_PROFILE_GPU_ZONE("ssr copy");
@@ -7148,9 +7149,9 @@ void LLPipeline::copyScreenSpaceReflections(LLRenderTarget* src, LLRenderTarget*
 
 void LLPipeline::generateGlow(LLRenderTarget* src)
 {
+    LL_PROFILE_GPU_ZONE("glow generate");
     if (sRenderGlow)
     {
-        LL_PROFILE_GPU_ZONE("glow");
         mGlow[2].bindTarget();
         mGlow[2].clear();
 
@@ -7261,6 +7262,8 @@ void LLPipeline::applyCAS(LLRenderTarget* src, LLRenderTarget* dst)
     static LLCachedControl<F32> cas_sharpness(gSavedSettings, "RenderCASSharpness", 0.4f);
     static LLCachedControl<bool> should_auto_adjust(gSavedSettings, "RenderSkyAutoAdjustLegacy", true);
 
+    LL_PROFILE_GPU_ZONE("cas");
+
     LLSettingsSky::ptr_t psky = LLEnvironment::instance().getCurrentSky();
     LLGLSLShader* sharpen_shader = psky->getReflectionProbeAmbiance(should_auto_adjust) == 0.f ? &gCASLegacyGammaProgram : &gCASProgram;
 
@@ -7303,7 +7306,7 @@ void LLPipeline::applyFXAA(LLRenderTarget* src, LLRenderTarget* dst, bool combin
     llassert(!gCubeSnapshot);
     LLGLSLShader* shader = &gGlowCombineProgram;
 
-    LL_PROFILE_GPU_ZONE("aa");
+    LL_PROFILE_GPU_ZONE("FXAA");
     S32 width = src->getWidth();
     S32 height = src->getHeight();
 
@@ -7388,7 +7391,7 @@ void LLPipeline::generateSMAABuffers(LLRenderTarget* src)
     // Present everything.
     if (multisample)
     {
-        LL_PROFILE_GPU_ZONE("aa");
+        LL_PROFILE_GPU_ZONE("SMAA Edge");
         static LLCachedControl<U32> aa_quality(gSavedSettings, "RenderFSAASamples", 0U);
         U32 fsaa_quality = std::clamp(aa_quality(), 0U, 3U);
 
@@ -7503,7 +7506,7 @@ void LLPipeline::applySMAA(LLRenderTarget* src, LLRenderTarget* dst, bool combin
 {
     llassert(!gCubeSnapshot);
 
-    LL_PROFILE_GPU_ZONE("aa");
+    LL_PROFILE_GPU_ZONE("SMAA");
     static LLCachedControl<U32> aa_quality(gSavedSettings, "RenderFSAASamples", 0U);
     U32 fsaa_quality = std::clamp(aa_quality(), 0U, 3U);
 
@@ -7566,7 +7569,6 @@ void LLPipeline::applySMAA(LLRenderTarget* src, LLRenderTarget* dst, bool combin
 
 void LLPipeline::copyRenderTarget(LLRenderTarget* src, LLRenderTarget* dst)
 {
-
     LL_PROFILE_GPU_ZONE("copyRenderTarget");
     dst->bindTarget();
 
@@ -7587,8 +7589,9 @@ void LLPipeline::copyRenderTarget(LLRenderTarget* src, LLRenderTarget* dst)
 
 void LLPipeline::combineGlow(LLRenderTarget* src, LLRenderTarget* dst)
 {
-    // Go ahead and do our glow combine here in our destination.  We blit this later into the front buffer.
+    LL_PROFILE_GPU_ZONE("glow combine");
 
+    // Go ahead and do our glow combine here in our destination.  We blit this later into the front buffer.
     if (dst)
     {
         dst->bindTarget();
@@ -7604,6 +7607,7 @@ void LLPipeline::combineGlow(LLRenderTarget* src, LLRenderTarget* dst)
         mScreenTriangleVB->setBuffer();
         mScreenTriangleVB->drawArrays(LLRender::TRIANGLES, 0, 3);
     }
+
     if (dst)
     {
         dst->flush();
@@ -7612,9 +7616,9 @@ void LLPipeline::combineGlow(LLRenderTarget* src, LLRenderTarget* dst)
 
 void LLPipeline::renderDoF(LLRenderTarget* src, LLRenderTarget* dst)
 {
+    LL_PROFILE_GPU_ZONE("dof");
     gViewerWindow->setup3DViewport();
 
-    LL_PROFILE_GPU_ZONE("dof");
     LLGLDisable blend(GL_BLEND);
 
     // depth of field focal plane calculations