SL-16418 Media texture update stall fix. Make media texture updates use LLImageGL thread to update, fix AMD sync issue on ImageGL thread and install debug callbacks on LLImageGL thread when debug gl enabled.

6 files changed, 140 insertions, 70 deletions

diff --git a/indra/llrender/llgl.cpp b/indra/llrender/llgl.cpp
index 922fcb867b..50a5972d28 100644
--- a/indra/llrender/llgl.cpp
+++ b/indra/llrender/llgl.cpp
@@ -757,17 +757,6 @@ bool LLGLManager::initGL()
 
 	stop_glerror();
 
-#if LL_WINDOWS
-	if (mHasDebugOutput && gDebugGL)
-	{ //setup debug output callback
-		//glDebugMessageControlARB(GL_DONT_CARE, GL_DONT_CARE, GL_DEBUG_SEVERITY_LOW_ARB, 0, NULL, GL_TRUE);
-		glDebugMessageCallbackARB((GLDEBUGPROCARB) gl_debug_callback, NULL);
-		glEnable(GL_DEBUG_OUTPUT_SYNCHRONOUS_ARB);
-	}
-#endif
-
-	stop_glerror();
-
 	//HACK always disable texture multisample, use FXAA instead
 	mHasTextureMultisample = FALSE;
 #if LL_WINDOWS
diff --git a/indra/llrender/llgltexture.cpp b/indra/llrender/llgltexture.cpp
index b6a02f1c0a..c7de8bc338 100644
--- a/indra/llrender/llgltexture.cpp
+++ b/indra/llrender/llgltexture.cpp
@@ -164,11 +164,11 @@ BOOL LLGLTexture::createGLTexture()
 	return mGLTexturep->createGLTexture() ;
 }
 
-BOOL LLGLTexture::createGLTexture(S32 discard_level, const LLImageRaw* imageraw, S32 usename, BOOL to_create, S32 category)
+BOOL LLGLTexture::createGLTexture(S32 discard_level, const LLImageRaw* imageraw, S32 usename, BOOL to_create, S32 category, bool defer_copy)
 {
 	llassert(mGLTexturep.notNull()) ;	
 
-	BOOL ret = mGLTexturep->createGLTexture(discard_level, imageraw, usename, to_create, category) ;
+	BOOL ret = mGLTexturep->createGLTexture(discard_level, imageraw, usename, to_create, category, defer_copy) ;
 
 	if(ret)
 	{
@@ -260,20 +260,20 @@ LLTexUnit::eTextureType LLGLTexture::getTarget(void) const
 	return mGLTexturep->getTarget() ;
 }
 
-BOOL LLGLTexture::setSubImage(const LLImageRaw* imageraw, S32 x_pos, S32 y_pos, S32 width, S32 height)
+BOOL LLGLTexture::setSubImage(const LLImageRaw* imageraw, S32 x_pos, S32 y_pos, S32 width, S32 height, bool use_new_name /* = false */)
 {
     LL_PROFILE_ZONE_SCOPED_CATEGORY_TEXTURE;
 	llassert(mGLTexturep.notNull()) ;
 
-	return mGLTexturep->setSubImage(imageraw, x_pos, y_pos, width, height) ;
+	return mGLTexturep->setSubImage(imageraw, x_pos, y_pos, width, height, 0, use_new_name) ;
 }
 
-BOOL LLGLTexture::setSubImage(const U8* datap, S32 data_width, S32 data_height, S32 x_pos, S32 y_pos, S32 width, S32 height)
+BOOL LLGLTexture::setSubImage(const U8* datap, S32 data_width, S32 data_height, S32 x_pos, S32 y_pos, S32 width, S32 height, bool use_new_name /* = false */)
 {
     LL_PROFILE_ZONE_SCOPED_CATEGORY_TEXTURE;
 	llassert(mGLTexturep.notNull()) ;
 
-	return mGLTexturep->setSubImage(datap, data_width, data_height, x_pos, y_pos, width, height) ;
+	return mGLTexturep->setSubImage(datap, data_width, data_height, x_pos, y_pos, width, height, 0, use_new_name) ;
 }
 
 void LLGLTexture::setGLTextureCreated (bool initialized)
diff --git a/indra/llrender/llgltexture.h b/indra/llrender/llgltexture.h
index 028457c510..216e7f640d 100644
--- a/indra/llrender/llgltexture.h
+++ b/indra/llrender/llgltexture.h
@@ -124,13 +124,21 @@ public:
 	BOOL       hasGLTexture() const ;
 	LLGLuint   getTexName() const ;		
 	BOOL       createGLTexture() ;
-	BOOL       createGLTexture(S32 discard_level, const LLImageRaw* imageraw, S32 usename = 0, BOOL to_create = TRUE, S32 category = LLGLTexture::OTHER);
+	
+    // Create a GL Texture from an image raw
+    // discard_level - mip level, 0 for highest resultion mip
+    // imageraw - the image to copy from
+    // usename - explicit GL name override
+    // to_create - set to FALSE to force gl texture to not be created
+    // category - LLGLTexture category for this LLGLTexture
+    // defer_copy - set to true to allocate GL texture but NOT initialize with imageraw data
+    BOOL       createGLTexture(S32 discard_level, const LLImageRaw* imageraw, S32 usename = 0, BOOL to_create = TRUE, S32 category = LLGLTexture::OTHER, bool defer_copy = false);
 
 	void       setFilteringOption(LLTexUnit::eTextureFilterOptions option);
 	void       setExplicitFormat(LLGLint internal_format, LLGLenum primary_format, LLGLenum type_format = 0, BOOL swap_bytes = FALSE);
 	void       setAddressMode(LLTexUnit::eTextureAddressMode mode);
-	BOOL       setSubImage(const LLImageRaw* imageraw, S32 x_pos, S32 y_pos, S32 width, S32 height);
-	BOOL       setSubImage(const U8* datap, S32 data_width, S32 data_height, S32 x_pos, S32 y_pos, S32 width, S32 height);
+	BOOL       setSubImage(const LLImageRaw* imageraw, S32 x_pos, S32 y_pos, S32 width, S32 height, bool use_new_name = false);
+	BOOL       setSubImage(const U8* datap, S32 data_width, S32 data_height, S32 x_pos, S32 y_pos, S32 width, S32 height, bool use_new_name = false);
 	void       setGLTextureCreated (bool initialized);
 	void       setCategory(S32 category) ;
     void       setTexName(LLGLuint); // for forcing w/ externally created textures only
diff --git a/indra/llrender/llimagegl.cpp b/indra/llrender/llimagegl.cpp
index 1e9b9f642e..694c4a7b06 100644
--- a/indra/llrender/llimagegl.cpp
+++ b/indra/llrender/llimagegl.cpp
@@ -550,7 +550,7 @@ bool LLImageGL::setSize(S32 width, S32 height, S32 ncomponents, S32 discard_leve
 	if (width != mWidth || height != mHeight || ncomponents != mComponents)
 	{
 		// Check if dimensions are a power of two!
-		if (!checkSize(width,height))
+		if (!checkSize(width, height))
 		{
 			LL_WARNS() << llformat("Texture has non power of two dimension: %dx%d",width,height) << LL_ENDL;
 			return false;
@@ -674,7 +674,7 @@ void LLImageGL::setImage(const LLImageRaw* imageraw)
 	setImage(rawdata, FALSE);
 }
 
-BOOL LLImageGL::setImage(const U8* data_in, BOOL data_hasmips, S32 usename)
+BOOL LLImageGL::setImage(const U8* data_in, BOOL data_hasmips /* = FALSE */, S32 usename /* = 0 */)
 {
     LL_PROFILE_ZONE_SCOPED_CATEGORY_TEXTURE;
 	bool is_compressed = false;
@@ -709,8 +709,14 @@ BOOL LLImageGL::setImage(const U8* data_in, BOOL data_hasmips, S32 usename)
 	
     gGL.getTexUnit(0)->bind(this, false, false, usename);
 
-
-    if (mUseMipMaps)
+    if (data_in == nullptr)
+    {
+        S32 w = getWidth();
+        S32 h = getHeight();
+        LLImageGL::setManualImage(mTarget, 0, mFormatInternal, w, h,
+            mFormatPrimary, mFormatType, (GLvoid*)data_in, mAllowCompression);
+    }
+    else if (mUseMipMaps)
 	{
 		if (data_hasmips)
 		{
@@ -1069,14 +1075,14 @@ void LLImageGL::postAddToAtlas()
 	stop_glerror();	
 }
 
-BOOL LLImageGL::setSubImage(const U8* datap, S32 data_width, S32 data_height, S32 x_pos, S32 y_pos, S32 width, S32 height, BOOL force_fast_update)
+BOOL LLImageGL::setSubImage(const U8* datap, S32 data_width, S32 data_height, S32 x_pos, S32 y_pos, S32 width, S32 height, BOOL force_fast_update /* = FALSE */, bool use_new_name /* = false */)
 {
     LL_PROFILE_ZONE_SCOPED_CATEGORY_TEXTURE;
 	if (!width || !height)
 	{
 		return TRUE;
 	}
-	if (mTexName == 0)
+	if (0 == (use_new_name ? mNewTexName : mTexName))
 	{
 		// *TODO: Re-enable warning?  Ran into thread locking issues? DK 2011-02-18
 		//LL_WARNS() << "Setting subimage on image without GL texture" << LL_ENDL;
@@ -1092,7 +1098,7 @@ BOOL LLImageGL::setSubImage(const U8* datap, S32 data_width, S32 data_height, S3
 	// HACK: allow the caller to explicitly force the fast path (i.e. using glTexSubImage2D here instead of calling setImage) even when updating the full texture.
 	if (!force_fast_update && x_pos == 0 && y_pos == 0 && width == getWidth() && height == getHeight() && data_width == width && data_height == height)
 	{
-		setImage(datap, FALSE);
+		setImage(datap, FALSE, use_new_name ? mNewTexName : mTexName);
 	}
 	else
 	{
@@ -1144,12 +1150,11 @@ BOOL LLImageGL::setSubImage(const U8* datap, S32 data_width, S32 data_height, S3
 
 		datap += (y_pos * data_width + x_pos) * getComponents();
 		// Update the GL texture
-		BOOL res = gGL.getTexUnit(0)->bindManual(mBindTarget, mTexName);
+		BOOL res = gGL.getTexUnit(0)->bindManual(mBindTarget, use_new_name ? mNewTexName : mTexName);
 		if (!res) LL_ERRS() << "LLImageGL::setSubImage(): bindTexture failed" << LL_ENDL;
 		stop_glerror();
 
-		glTexSubImage2D(mTarget, 0, x_pos, y_pos, 
-						width, height, mFormatPrimary, mFormatType, datap);
+		glTexSubImage2D(mTarget, 0, x_pos, y_pos, width, height, mFormatPrimary, mFormatType, datap);
 		gGL.getTexUnit(0)->disable();
 		stop_glerror();
 
@@ -1166,10 +1171,10 @@ BOOL LLImageGL::setSubImage(const U8* datap, S32 data_width, S32 data_height, S3
 	return TRUE;
 }
 
-BOOL LLImageGL::setSubImage(const LLImageRaw* imageraw, S32 x_pos, S32 y_pos, S32 width, S32 height, BOOL force_fast_update)
+BOOL LLImageGL::setSubImage(const LLImageRaw* imageraw, S32 x_pos, S32 y_pos, S32 width, S32 height, BOOL force_fast_update /* = FALSE */, bool use_new_name /* = false */)
 {
     LL_PROFILE_ZONE_SCOPED_CATEGORY_TEXTURE;
-	return setSubImage(imageraw->getData(), imageraw->getWidth(), imageraw->getHeight(), x_pos, y_pos, width, height, force_fast_update);
+	return setSubImage(imageraw->getData(), imageraw->getWidth(), imageraw->getHeight(), x_pos, y_pos, width, height, force_fast_update, use_new_name);
 }
 
 // Copy sub image from frame buffer
@@ -1356,7 +1361,7 @@ BOOL LLImageGL::createGLTexture()
 	return TRUE ;
 }
 
-BOOL LLImageGL::createGLTexture(S32 discard_level, const LLImageRaw* imageraw, S32 usename/*=0*/, BOOL to_create, S32 category)
+BOOL LLImageGL::createGLTexture(S32 discard_level, const LLImageRaw* imageraw, S32 usename/*=0*/, BOOL to_create, S32 category, bool defer_copy)
 {
     LL_PROFILE_ZONE_SCOPED_CATEGORY_TEXTURE;
     checkActiveThread();
@@ -1386,6 +1391,7 @@ BOOL LLImageGL::createGLTexture(S32 discard_level, const LLImageRaw* imageraw, S
 	// Actual image width/height = raw image width/height * 2^discard_level
 	S32 raw_w = imageraw->getWidth() ;
 	S32 raw_h = imageraw->getHeight() ;
+
 	S32 w = raw_w << discard_level;
 	S32 h = raw_h << discard_level;
 
@@ -1468,15 +1474,24 @@ BOOL LLImageGL::createGLTexture(S32 discard_level, const LLImageRaw* imageraw, S
 
 	setCategory(category);
  	const U8* rawdata = imageraw->getData();
-	return createGLTexture(discard_level, rawdata, FALSE, usename);
+	return createGLTexture(discard_level, rawdata, FALSE, usename, defer_copy);
 }
 
-BOOL LLImageGL::createGLTexture(S32 discard_level, const U8* data_in, BOOL data_hasmips, S32 usename)
+BOOL LLImageGL::createGLTexture(S32 discard_level, const U8* data_in, BOOL data_hasmips, S32 usename, bool defer_copy)
+    // Call with void data, vmem is allocated but unitialized
 {
     LL_PROFILE_ZONE_SCOPED_CATEGORY_TEXTURE;
     checkActiveThread();
 
-    llassert(data_in);
+    if (defer_copy)
+    {
+        data_in = nullptr;
+    }
+    else
+    {
+        llassert(data_in);
+    }
+
     stop_glerror();
 
     if (discard_level < 0)
@@ -1486,7 +1501,8 @@ BOOL LLImageGL::createGLTexture(S32 discard_level, const U8* data_in, BOOL data_
     }
     discard_level = llclamp(discard_level, 0, (S32)mMaxDiscardLevel);
 
-    if (mTexName != 0 && discard_level == mCurrentDiscardLevel)
+    if (!defer_copy // <--- hacky way to force creation of mNewTexName from media texture update
+        && mTexName != 0 && discard_level == mCurrentDiscardLevel)
     {
         // This will only be true if the size has not changed
         return setImage(data_in, data_hasmips);
@@ -1534,36 +1550,12 @@ BOOL LLImageGL::createGLTexture(S32 discard_level, const U8* data_in, BOOL data_
     }
 
     //if we're on the image loading thread, be sure to delete old_texname and update mTexName on the main thread
-    if (! on_main_thread())
+    if (!on_main_thread())
     {
+        if (!defer_copy)
         {
-            LL_PROFILE_ZONE_NAMED("cglt - sync");
-            if (gGLManager.mHasSync)
-            {
-                auto sync = glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, 0);
-                glClientWaitSync(sync, 0, 0);
-                glDeleteSync(sync);
-            }
-            else
-            {
-                glFinish();
-            }
+            syncToMainThread();
         }
-
-        ref();
-        LL::WorkQueue::postMaybe(
-            mMainQueue,
-            [=]()
-            {
-                LL_PROFILE_ZONE_NAMED("cglt - delete callback");
-                if (old_texname != 0)
-                {
-                    LLImageGL::deleteTextures(1, &old_texname);
-                }
-                mTexName = mNewTexName;
-                mNewTexName = 0;
-                unref();
-            });
     }
     else 
     {
@@ -1587,6 +1579,55 @@ BOOL LLImageGL::createGLTexture(S32 discard_level, const U8* data_in, BOOL data_
     return TRUE;
 }
 
+void LLImageGL::syncToMainThread()
+{
+    {
+        LL_PROFILE_ZONE_NAMED("cglt - sync");
+        if (gGLManager.mHasSync)
+        {
+            // post a sync to the main thread (will execute before tex name swap lambda below)
+            auto sync = glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, 0);
+            glFlush();
+            LL::WorkQueue::postMaybe(
+                mMainQueue,
+                [=]()
+                {
+                    LL_PROFILE_ZONE_NAMED("cglt - wait sync");
+                    {
+                        LL_PROFILE_ZONE_NAMED("glWaitSync");
+                        glWaitSync(sync, 0, GL_TIMEOUT_IGNORED);
+                    }
+                    {
+                        LL_PROFILE_ZONE_NAMED("glDeleteSync");
+                        glDeleteSync(sync);
+                    }
+                });
+        }
+        else
+        {
+            glFinish();
+        }
+    }
+
+    ref();
+    LL::WorkQueue::postMaybe(
+        mMainQueue,
+        [=]()
+        {
+            LL_PROFILE_ZONE_NAMED("cglt - delete callback");
+            if (mNewTexName != 0)
+            {
+                if (mTexName != 0)
+                {
+                    LLImageGL::deleteTextures(1, &mTexName);
+                }
+                mTexName = mNewTexName;
+                mNewTexName = 0;
+                unref();
+            }
+        });
+}
+
 BOOL LLImageGL::readBackRaw(S32 discard_level, LLImageRaw* imageraw, bool compressed_ok) const
 {
 	llassert_always(sAllowReadBackRaw) ;
@@ -1714,7 +1755,15 @@ void LLImageGL::destroyGLTexture()
 		mCurrentDiscardLevel = -1 ; //invalidate mCurrentDiscardLevel.
 		mTexName = 0;		
 		mGLTextureCreated = FALSE ;
-	}	
+	}
+
+    // clean up any in-flight name change
+    if (0 != mNewTexName)
+    {
+        // Memory is transient, not tracked by sGlobalTextuerMemory
+        LLImageGL::deleteTextures(1, &mNewTexName);
+        mNewTexName = 0;
+    }
 }
 
 //force to invalidate the gl texture, most likely a sculpty texture
diff --git a/indra/llrender/llimagegl.h b/indra/llrender/llimagegl.h
index d6f4b13a51..6c6e327e42 100644
--- a/indra/llrender/llimagegl.h
+++ b/indra/llrender/llimagegl.h
@@ -110,14 +110,18 @@ public:
     
 	BOOL createGLTexture() ;
 	BOOL createGLTexture(S32 discard_level, const LLImageRaw* imageraw, S32 usename = 0, BOOL to_create = TRUE,
-		S32 category = sMaxCategories-1);
-	BOOL createGLTexture(S32 discard_level, const U8* data, BOOL data_hasmips = FALSE, S32 usename = 0);
+		S32 category = sMaxCategories-1, bool defer_copy = false);
+	BOOL createGLTexture(S32 discard_level, const U8* data, BOOL data_hasmips = FALSE, S32 usename = 0, bool defer_copy = false);
 	void setImage(const LLImageRaw* imageraw);
 	BOOL setImage(const U8* data_in, BOOL data_hasmips = FALSE, S32 usename = 0);
-	BOOL setSubImage(const LLImageRaw* imageraw, S32 x_pos, S32 y_pos, S32 width, S32 height, BOOL force_fast_update = FALSE);
-	BOOL setSubImage(const U8* datap, S32 data_width, S32 data_height, S32 x_pos, S32 y_pos, S32 width, S32 height, BOOL force_fast_update = FALSE);
+	BOOL setSubImage(const LLImageRaw* imageraw, S32 x_pos, S32 y_pos, S32 width, S32 height, BOOL force_fast_update = FALSE, bool use_new_name = false);
+	BOOL setSubImage(const U8* datap, S32 data_width, S32 data_height, S32 x_pos, S32 y_pos, S32 width, S32 height, BOOL force_fast_update = FALSE, bool use_new_name = false);
 	BOOL setSubImageFromFrameBuffer(S32 fb_x, S32 fb_y, S32 x_pos, S32 y_pos, S32 width, S32 height);
-	
+
+    // wait for gl commands to finish on current thread and push
+    // a lambda to main thread to swap mNewTexName and mTexName
+    void syncToMainThread();
+
 	// Read back a raw image for this discard level, if it exists
 	BOOL readBackRaw(S32 discard_level, LLImageRaw* imageraw, bool compressed_ok) const;
 	void destroyGLTexture();
diff --git a/indra/llrender/llrender.cpp b/indra/llrender/llrender.cpp
index 92d8e6193f..a03a27cf94 100644
--- a/indra/llrender/llrender.cpp
+++ b/indra/llrender/llrender.cpp
@@ -36,6 +36,17 @@
 #include "lltexture.h"
 #include "llshadermgr.h"
 
+#if LL_WINDOWS
+extern void APIENTRY gl_debug_callback(GLenum source,
+                                GLenum type,
+                                GLuint id,
+                                GLenum severity,
+                                GLsizei length,
+                                const GLchar* message,
+                                GLvoid* userParam)
+;
+#endif
+
 thread_local LLRender gGL;
 
 // Handy copies of last good GL matrices
@@ -860,6 +871,15 @@ LLRender::~LLRender()
 
 void LLRender::init()
 {
+#if LL_WINDOWS
+    if (gGLManager.mHasDebugOutput && gDebugGL)
+    { //setup debug output callback
+        //glDebugMessageControlARB(GL_DONT_CARE, GL_DONT_CARE, GL_DEBUG_SEVERITY_LOW_ARB, 0, NULL, GL_TRUE);
+        glDebugMessageCallbackARB((GLDEBUGPROCARB) gl_debug_callback, NULL);
+        glEnable(GL_DEBUG_OUTPUT_SYNCHRONOUS_ARB);
+    }
+#endif
+
     glPixelStorei(GL_PACK_ALIGNMENT, 1);
     glPixelStorei(GL_UNPACK_ALIGNMENT, 1);