SL-17287 Update Tracy to 0.8.1. Clean up GPU instrumentation.

12 files changed, 49 insertions, 38 deletions

diff --git a/autobuild.xml b/autobuild.xml
index fa3f7b2743..4c69ff71d4 100644
--- a/autobuild.xml
+++ b/autobuild.xml
@@ -3142,11 +3142,11 @@ Copyright (c) 2012, 2014, 2015, 2016 nghttp2 contributors</string>
             <key>archive</key>
             <map>
               <key>hash</key>
-              <string>b649ee6591e67d2341e886b3fc3484a7</string>
+              <string>dff3dd1db6a4cb0b73f60a5db3ee0de2</string>
               <key>hash_algorithm</key>
               <string>md5</string>
               <key>url</key>
-              <string>https://automated-builds-secondlife-com.s3.amazonaws.com/ct2/86974/801642/tracy-v0.7.8.563351-windows64-563351.tar.bz2</string>
+              <string>https://automated-builds-secondlife-com.s3.amazonaws.com/ct2/100101/882119/tracy-v0.8.1.571983-windows64-571983.tar.bz2</string>
             </map>
             <key>name</key>
             <string>windows64</string>
diff --git a/indra/cmake/Tracy.cmake b/indra/cmake/Tracy.cmake
index cfff956bcf..dba57eb920 100644
--- a/indra/cmake/Tracy.cmake
+++ b/indra/cmake/Tracy.cmake
@@ -12,6 +12,7 @@ if (USE_TRACY)
 
   if (WINDOWS)
     MESSAGE(STATUS "Including Tracy for Windows: '${TRACY_INCLUDE_DIR}'")
+    set(TRACY_LIBRARY "TracyClient")
   endif (WINDOWS)
 
   if (DARWIN)
diff --git a/indra/llcommon/llframetimer.cpp b/indra/llcommon/llframetimer.cpp
index c54029e8b4..58af32f1af 100644
--- a/indra/llcommon/llframetimer.cpp
+++ b/indra/llcommon/llframetimer.cpp
@@ -30,9 +30,9 @@
 #include "llframetimer.h"
 
 // We don't bother building a stand alone lib; we just need to include the one source file for Tracy support
-#if LL_PROFILER_CONFIGURATION == LL_PROFILER_CONFIG_TRACY || LL_PROFILER_CONFIGURATION == LL_PROFILER_CONFIG_TRACY_FAST_TIMER
-	#include "TracyClient.cpp"
-#endif // LL_PROFILER_CONFIGURATION
+//#if LL_PROFILER_CONFIGURATION == LL_PROFILER_CONFIG_TRACY || LL_PROFILER_CONFIGURATION == LL_PROFILER_CONFIG_TRACY_FAST_TIMER
+//	#include "TracyClient.cpp"
+//#endif // LL_PROFILER_CONFIGURATION
 
 // Static members
 //LLTimer	LLFrameTimer::sInternalTimer;
diff --git a/indra/llcommon/llprofiler.h b/indra/llcommon/llprofiler.h
index f9d7ae7ce4..bc46128d21 100644
--- a/indra/llcommon/llprofiler.h
+++ b/indra/llcommon/llprofiler.h
@@ -86,8 +86,8 @@ extern thread_local bool gProfilerEnabled;
         #define TRACY_ONLY_IPV4      1
         #include "Tracy.hpp"
 
-        // Mutually exclusive with detailed memory tracing
-        #define LL_PROFILER_ENABLE_TRACY_OPENGL 0
+        // Disable memory tracing when enabled, but enabled 
+        #define LL_PROFILER_ENABLE_TRACY_OPENGL 1
     #endif
 
     #if LL_PROFILER_CONFIGURATION == LL_PROFILER_CONFIG_TRACY
@@ -104,8 +104,6 @@ extern thread_local bool gProfilerEnabled;
         #define LL_PROFILE_ZONE_ERR(name)               LL_PROFILE_ZONE_NAMED_COLOR( name, 0XFF0000  )  // RGB yellow
         #define LL_PROFILE_ZONE_INFO(name)              LL_PROFILE_ZONE_NAMED_COLOR( name, 0X00FFFF  )  // RGB cyan
         #define LL_PROFILE_ZONE_WARN(name)              LL_PROFILE_ZONE_NAMED_COLOR( name, 0x0FFFF00 )  // RGB red
-        #define LL_PROFILE_ALLOC(ptr, size)             TracyAlloc(ptr, size)
-        #define LL_PROFILE_FREE(ptr)                    TracyFree(ptr)
     #endif
     #if LL_PROFILER_CONFIGURATION == LL_PROFILER_CONFIG_FAST_TIMER
         #define LL_PROFILER_FRAME_END
@@ -121,8 +119,6 @@ extern thread_local bool gProfilerEnabled;
         #define LL_PROFILE_ZONE_ERR(name)               (void)(name); // Not supported
         #define LL_PROFILE_ZONE_INFO(name)              (void)(name); // Not supported
         #define LL_PROFILE_ZONE_WARN(name)              (void)(name); // Not supported
-        #define LL_PROFILE_ALLOC(ptr, size)             (void)(ptr); (void)(size);
-        #define LL_PROFILE_FREE(ptr)                    (void)(ptr);
     #endif
     #if LL_PROFILER_CONFIGURATION == LL_PROFILER_CONFIG_TRACY_FAST_TIMER
         #define LL_PROFILER_FRAME_END                   FrameMark
@@ -138,14 +134,37 @@ extern thread_local bool gProfilerEnabled;
         #define LL_PROFILE_ZONE_ERR(name)               LL_PROFILE_ZONE_NAMED_COLOR( name, 0XFF0000  )  // RGB yellow
         #define LL_PROFILE_ZONE_INFO(name)              LL_PROFILE_ZONE_NAMED_COLOR( name, 0X00FFFF  )  // RGB cyan
         #define LL_PROFILE_ZONE_WARN(name)              LL_PROFILE_ZONE_NAMED_COLOR( name, 0x0FFFF00 )  // RGB red
-        #define LL_PROFILE_ALLOC(ptr, size)             TracyAlloc(ptr, size)
-        #define LL_PROFILE_FREE(ptr)                    TracyFree(ptr)
     #endif
 #else
     #define LL_PROFILER_FRAME_END
     #define LL_PROFILER_SET_THREAD_NAME( name ) (void)(name)
 #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_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);
+#else
+#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
+
+#if LL_PROFILER_CONFIG > 1
+#define LL_PROFILE_ALLOC(ptr, size)             TracyAlloc(ptr, size)
+#define LL_PROFILE_FREE(ptr)                    TracyFree(ptr)
+#else
+#define LL_PROFILE_ALLOC(ptr, size)             (void)(ptr); (void)(size);
+#define LL_PROFILE_FREE(ptr)                    (void)(ptr);
+#endif
+
+#endif
+
 #include "llprofilercategories.h"
 
 #endif // LL_PROFILER_H
diff --git a/indra/llrender/llglheaders.h b/indra/llrender/llglheaders.h
index aa429505c1..c53631dc7b 100644
--- a/indra/llrender/llglheaders.h
+++ b/indra/llrender/llglheaders.h
@@ -830,14 +830,7 @@ extern void glGetBufferPointervARB (GLenum, GLenum, GLvoid* *);
     #define glGetQueryiv        glGetQueryivARB
     #define glGetQueryObjectiv  glGetQueryObjectivARB
     #include <tracy/TracyOpenGL.hpp>
-
-    #define LL_PROFILER_GPU_ZONEC(name,color) TracyGpuZoneC(name,color);
-    #define LL_PROFILER_GPU_COLLECT           TracyGpuCollect
-    #define LL_PROFILER_GPU_CONTEXT           TracyGpuContext
-#else
-    #define LL_PROFILER_GPU_ZONEC(name,color) (void)name;(void)color;
-    #define LL_PROFILER_GPU_COLLECT
-    #define LL_PROFILER_GPU_CONTEXT
 #endif
+    
 
 #endif // LL_LLGLHEADERS_H
diff --git a/indra/llrender/llrendertarget.cpp b/indra/llrender/llrendertarget.cpp
index 0a4cc1bcf5..85c2a48279 100644
--- a/indra/llrender/llrendertarget.cpp
+++ b/indra/llrender/llrendertarget.cpp
@@ -488,13 +488,11 @@ void LLRenderTarget::bindTarget()
 									GL_COLOR_ATTACHMENT1,
 									GL_COLOR_ATTACHMENT2,
 									GL_COLOR_ATTACHMENT3};
-			LL_PROFILER_GPU_ZONEC( "gl.DrawBuffersARB", 0x4000FF )
 			glDrawBuffersARB(mTex.size(), drawbuffers);
 		}
 			
 		if (mTex.empty())
 		{ //no color buffer to draw to
-			LL_PROFILER_GPU_ZONEC( "gl.DrawBuffer", 0x0000FF )
 			glDrawBuffer(GL_NONE);
 			glReadBuffer(GL_NONE);
 		}
diff --git a/indra/llrender/llvertexbuffer.cpp b/indra/llrender/llvertexbuffer.cpp
index 6338cab96a..b1b69f1508 100644
--- a/indra/llrender/llvertexbuffer.cpp
+++ b/indra/llrender/llvertexbuffer.cpp
@@ -624,7 +624,6 @@ void LLVertexBuffer::drawRange(U32 mode, U32 start, U32 end, U32 count, U32 indi
 
 	stop_glerror();
 	LLGLSLShader::startProfile();
-    LL_PROFILER_GPU_ZONEC( "gl.DrawRangeElements", 0xFFFF00 )
 	glDrawRangeElements(sGLMode[mode], start, end, count, GL_UNSIGNED_SHORT, 
 		idx);
 	LLGLSLShader::stopProfile(count, mode);
@@ -642,7 +641,6 @@ void LLVertexBuffer::drawRangeFast(U32 mode, U32 start, U32 end, U32 count, U32
 
     U16* idx = ((U16*)getIndicesPointer()) + indices_offset;
 
-    LL_PROFILER_GPU_ZONEC("gl.DrawRangeElements", 0xFFFF00)
         glDrawRangeElements(sGLMode[mode], start, end, count, GL_UNSIGNED_SHORT,
             idx);
 }
@@ -688,8 +686,7 @@ void LLVertexBuffer::draw(U32 mode, U32 count, U32 indices_offset) const
 
 	stop_glerror();
 	LLGLSLShader::startProfile();
-    LL_PROFILER_GPU_ZONEC( "gl.DrawElements", 0xA0FFA0 )
-	glDrawElements(sGLMode[mode], count, GL_UNSIGNED_SHORT,
+    glDrawElements(sGLMode[mode], count, GL_UNSIGNED_SHORT,
 		((U16*) getIndicesPointer()) + indices_offset);
 	LLGLSLShader::stopProfile(count, mode);
 	stop_glerror();
@@ -736,7 +733,6 @@ void LLVertexBuffer::drawArrays(U32 mode, U32 first, U32 count) const
 
     LLGLSLShader::startProfile();
     {
-        LL_PROFILER_GPU_ZONEC("gl.DrawArrays", 0xFF4040)
             glDrawArrays(sGLMode[mode], first, count);
     }
     LLGLSLShader::stopProfile(count, mode);
diff --git a/indra/llwindow/llwindowwin32.cpp b/indra/llwindow/llwindowwin32.cpp
index 1f3823509c..02b3c25129 100644
--- a/indra/llwindow/llwindowwin32.cpp
+++ b/indra/llwindow/llwindowwin32.cpp
@@ -1623,8 +1623,6 @@ const	S32   max_format  = (S32)num_formats - 1;
 		return FALSE;
 	}
 
-	LL_PROFILER_GPU_CONTEXT
-
 	if (!gGLManager.initGL())
 	{
 		OSMessageBox(mCallbacks->translateString("MBVideoDrvErr"), mCallbacks->translateString("MBError"), OSMB_OK);
@@ -1656,6 +1654,8 @@ const	S32   max_format  = (S32)num_formats - 1;
 		swapBuffers();
 	}
 
+    LL_PROFILER_GPU_CONTEXT;
+
 	return TRUE;
 }
 
@@ -1809,6 +1809,7 @@ void* LLWindowWin32::createSharedContext()
 void LLWindowWin32::makeContextCurrent(void* contextPtr)
 {
     wglMakeCurrent(mhDC, (HGLRC) contextPtr);
+    LL_PROFILER_GPU_CONTEXT;
 }
 
 void LLWindowWin32::destroySharedContext(void* contextPtr)
@@ -3577,7 +3578,7 @@ void LLWindowWin32::swapBuffers()
     glFlush(); //superstitious flush for maybe frame stall removal?
 	SwapBuffers(mhDC);
 
-    LL_PROFILER_GPU_COLLECT
+    LL_PROFILER_GPU_COLLECT;
 }
 
 
diff --git a/indra/newview/llface.cpp b/indra/newview/llface.cpp
index d0c26bc43b..a2d9c0ba2b 100644
--- a/indra/newview/llface.cpp
+++ b/indra/newview/llface.cpp
@@ -587,7 +587,6 @@ void LLFace::renderSelected(LLViewerTexture *imagep, const LLColor4& color)
 						glTexCoordPointer(2, GL_FLOAT, 8, vol_face.mTexCoords);
 					}
 					gGL.syncMatrices();
-					LL_PROFILER_GPU_ZONEC( "gl.DrawElements", 0x00FF00 );
 					glDrawElements(GL_TRIANGLES, vol_face.mNumIndices, GL_UNSIGNED_SHORT, vol_face.mIndices);
 					glDisableClientState(GL_TEXTURE_COORD_ARRAY);
 				}
diff --git a/indra/newview/llspatialpartition.cpp b/indra/newview/llspatialpartition.cpp
index b6bd07085b..39bb46e6fa 100644
--- a/indra/newview/llspatialpartition.cpp
+++ b/indra/newview/llspatialpartition.cpp
@@ -2632,14 +2632,12 @@ void renderPhysicsShape(LLDrawable* drawable, LLVOVolume* volume)
 			gGL.diffuseColor4fv(line_color.mV);
 			gGL.syncMatrices();
 			{
-				LL_PROFILER_GPU_ZONEC( "gl.DrawElements", 0x20FF20 )
 				glDrawElements(GL_TRIANGLES, phys_volume->mNumHullIndices, GL_UNSIGNED_SHORT, phys_volume->mHullIndices);
 			}
 			
 			gGL.diffuseColor4fv(color.mV);
 			glPolygonMode(GL_FRONT_AND_BACK, GL_FILL);
 			{
-				LL_PROFILER_GPU_ZONEC( "gl.DrawElements", 0x40FF40 )
 				glDrawElements(GL_TRIANGLES, phys_volume->mNumHullIndices, GL_UNSIGNED_SHORT, phys_volume->mHullIndices);
 			}
 		}
@@ -3178,7 +3176,6 @@ void renderRaycast(LLDrawable* drawablep)
 						gGL.diffuseColor4f(0,1,1,0.5f);
 						glVertexPointer(3, GL_FLOAT, sizeof(LLVector4a), face.mPositions);
 						gGL.syncMatrices();
-						LL_PROFILER_GPU_ZONEC( "gl.DrawElements", 0x60FF60 );
 						glDrawElements(GL_TRIANGLES, face.mNumIndices, GL_UNSIGNED_SHORT, face.mIndices);
 					}
 					
diff --git a/indra/newview/llviewerwindow.cpp b/indra/newview/llviewerwindow.cpp
index 9c9c30ad95..ad57fcd802 100644
--- a/indra/newview/llviewerwindow.cpp
+++ b/indra/newview/llviewerwindow.cpp
@@ -5274,6 +5274,7 @@ BOOL LLViewerWindow::cubeSnapshot(const LLVector3& origin, LLCubeMapArray* cubea
 {
     // NOTE: implementation derived from LLFloater360Capture::capture360Images() and simpleSnapshot
     LL_PROFILE_ZONE_SCOPED_CATEGORY_APP;
+    LL_PROFILE_GPU_ZONE("cubeSnapshot");
     llassert(LLPipeline::sRenderDeferred);
     llassert(!gCubeSnapshot); //assert a snapshot isn't already in progress
     
diff --git a/indra/newview/pipeline.cpp b/indra/newview/pipeline.cpp
index 6ea519c71b..8f38f7a8f4 100644
--- a/indra/newview/pipeline.cpp
+++ b/indra/newview/pipeline.cpp
@@ -4251,7 +4251,7 @@ U32 LLPipeline::sCurRenderPoolType = 0 ;
 void LLPipeline::renderGeom(LLCamera& camera, bool forceVBOUpdate)
 {
 	LL_PROFILE_ZONE_SCOPED_CATEGORY_PIPELINE; //LL_RECORD_BLOCK_TIME(FTM_RENDER_GEOMETRY);
-
+    LL_PROFILE_GPU_ZONE("renderGeom");
 	assertInitialized();
 
 	F32 saved_modelview[16];
@@ -4501,8 +4501,9 @@ void LLPipeline::renderGeom(LLCamera& camera, bool forceVBOUpdate)
 void LLPipeline::renderGeomDeferred(LLCamera& camera)
 {
 	LLAppViewer::instance()->pingMainloopTimeout("Pipeline:RenderGeomDeferred");
-
 	LL_PROFILE_ZONE_SCOPED_CATEGORY_DRAWPOOL; //LL_RECORD_BLOCK_TIME(FTM_RENDER_GEOMETRY);
+    LL_PROFILE_GPU_ZONE("renderGeomDeferred");
+
 	{
 		LL_PROFILE_ZONE_NAMED_CATEGORY_DRAWPOOL("deferred pools"); //LL_RECORD_BLOCK_TIME(FTM_DEFERRED_POOLS);
 
@@ -4601,6 +4602,8 @@ void LLPipeline::renderGeomDeferred(LLCamera& camera)
 void LLPipeline::renderGeomPostDeferred(LLCamera& camera, bool do_occlusion)
 {
 	LL_PROFILE_ZONE_SCOPED_CATEGORY_DRAWPOOL; //LL_RECORD_BLOCK_TIME(FTM_POST_DEFERRED_POOLS);
+    LL_PROFILE_GPU_ZONE("renderGeomPostDeferred");
+
 	U32 cur_type = 0;
 
 	LLGLEnable cull(GL_CULL_FACE);
@@ -4697,6 +4700,7 @@ void LLPipeline::renderGeomPostDeferred(LLCamera& camera, bool do_occlusion)
 void LLPipeline::renderGeomShadow(LLCamera& camera)
 {
     LL_PROFILE_ZONE_SCOPED_CATEGORY_PIPELINE;
+    LL_PROFILE_GPU_ZONE("renderGeomShadow");
     U32 cur_type = 0;
 	
 	LLGLEnable cull(GL_CULL_FACE);
@@ -9202,6 +9206,7 @@ inline float sgn(float a)
 void LLPipeline::generateWaterReflection(LLCamera& camera_in)
 {
     LL_PROFILE_ZONE_SCOPED_CATEGORY_PIPELINE;
+    LL_PROFILE_GPU_ZONE("generateWaterReflection");
 
     if (!assertInitialized() || gCubeSnapshot)
     {
@@ -10082,6 +10087,7 @@ void LLPipeline::generateSunShadow(LLCamera& camera)
 	}
 
 	LL_PROFILE_ZONE_SCOPED_CATEGORY_PIPELINE; //LL_RECORD_BLOCK_TIME(FTM_GEN_SUN_SHADOW);
+    LL_PROFILE_GPU_ZONE("generateSunShadow");
 
 	bool skip_avatar_update = false;
 	if (!isAgentAvatarValid() || gAgentCamera.getCameraAnimating() || gAgentCamera.getCameraMode() != CAMERA_MODE_MOUSELOOK || !LLVOAvatar::sVisibleInFirstPerson)
@@ -10932,6 +10938,7 @@ static LLTrace::BlockTimerStatHandle FTM_GENERATE_IMPOSTOR("Generate Impostor");
 void LLPipeline::generateImpostor(LLVOAvatar* avatar, bool preview_avatar)
 {
     LL_RECORD_BLOCK_TIME(FTM_GENERATE_IMPOSTOR);
+    LL_PROFILE_GPU_ZONE("generateImpostor");
 	LLGLState::checkStates();
 	LLGLState::checkTextureChannels();
 
@@ -11175,7 +11182,6 @@ void LLPipeline::generateImpostor(LLVOAvatar* avatar, bool preview_avatar)
 		if (LLPipeline::sRenderDeferred)
 		{
 			GLuint buff = GL_COLOR_ATTACHMENT0;
-			LL_PROFILER_GPU_ZONEC( "gl.DrawBuffersARB", 0x8000FF );
 			glDrawBuffersARB(1, &buff);
 		}