MAINT-2371 More optimizations.

Reviewed by Graham

7 files changed, 182 insertions, 55 deletions

diff --git a/indra/llrender/llgl.cpp b/indra/llrender/llgl.cpp
index 89f1f36297..58bd346c15 100644
--- a/indra/llrender/llgl.cpp
+++ b/indra/llrender/llgl.cpp
@@ -741,7 +741,7 @@ bool LLGLManager::initGL()
 #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);
+		glDebugMessageControlARB(GL_DONT_CARE, GL_DEBUG_TYPE_DEPRECATED_BEHAVIOR_ARB, GL_DEBUG_SEVERITY_LOW_ARB, 0, NULL, GL_TRUE);
 		glDebugMessageCallbackARB((GLDEBUGPROCARB) gl_debug_callback, NULL);
 		glEnable(GL_DEBUG_OUTPUT_SYNCHRONOUS_ARB);
 	}
@@ -1478,7 +1478,7 @@ void do_assert_glerror()
 
 void assert_glerror()
 {
-	if (!gGLActive)
+/*	if (!gGLActive)
 	{
 		//llwarns << "GL used while not active!" << llendl;
 
@@ -1487,8 +1487,13 @@ void assert_glerror()
 			//ll_fail("GL used while not active");
 		}
 	}
+*/
 
-	if (gDebugGL) 
+	if (!gDebugGL) 
+	{
+		//funny looking if for branch prediction -- gDebugGL is almost always false and assert_glerror is called often
+	}
+	else
 	{
 		do_assert_glerror();
 	}
diff --git a/indra/llrender/llimagegl.cpp b/indra/llrender/llimagegl.cpp
index a4d7872ec2..ef2648ae98 100755
--- a/indra/llrender/llimagegl.cpp
+++ b/indra/llrender/llimagegl.cpp
@@ -709,9 +709,12 @@ void LLImageGL::setImage(const U8* data_in, BOOL data_hasmips)
 
 					mMipLevels = wpo2(llmax(w, h));
 
-					//use legacy mipmap generation mode
-					glTexParameteri(mTarget, GL_GENERATE_MIPMAP, GL_TRUE);
-					
+					if (!gGLManager.mHasFramebufferObject)
+					{
+						//use legacy mipmap generation mode
+						glTexParameteri(mTarget, GL_GENERATE_MIPMAP, GL_TRUE);
+					}
+										
 					LLImageGL::setManualImage(mTarget, 0, mFormatInternal,
 								 w, h, 
 								 mFormatPrimary, mFormatType,
@@ -726,6 +729,11 @@ void LLImageGL::setImage(const U8* data_in, BOOL data_hasmips)
 						glPixelStorei(GL_UNPACK_SWAP_BYTES, 0);
 						stop_glerror();
 					}
+
+					if (gGLManager.mHasFramebufferObject)
+					{
+						glGenerateMipmap(mTarget);
+					}
 				}
 			}
 			else
@@ -1057,6 +1065,16 @@ void LLImageGL::generateTextures(LLTexUnit::eTextureType type, U32 format, S32 n
 {
 	bool empty = true;
 
+	if (LLRender::sGLCoreProfile)
+	{
+		switch (format)
+		{
+			case GL_LUMINANCE8: format = GL_RGB8; break;
+			case GL_LUMINANCE8_ALPHA8:
+			case GL_ALPHA8: format = GL_RGBA8; break;
+		}
+	}
+
 	dead_texturelist_t::iterator iter = sDeadTextureList[type].find(format);
 	
 	if (iter != sDeadTextureList[type].end())
@@ -1084,6 +1102,16 @@ void LLImageGL::deleteTextures(LLTexUnit::eTextureType type, U32 format, S32 mip
 {
 	if (gGLManager.mInited)
 	{
+		if (LLRender::sGLCoreProfile)
+		{
+			switch (format)
+			{
+				case GL_LUMINANCE8: format = GL_RGB8; break;
+				case GL_LUMINANCE8_ALPHA8:
+				case GL_ALPHA8: format = GL_RGBA8; break;
+			}
+		}
+
 		if (format == 0 ||  type == LLTexUnit::TT_CUBE_MAP || mip_levels == -1)
 		{ //unknown internal format or unknown number of mip levels, not safe to reuse
 			glDeleteTextures(numTextures, textures);
diff --git a/indra/llrender/llrender.cpp b/indra/llrender/llrender.cpp
index cb82cbfb74..dada27557e 100644
--- a/indra/llrender/llrender.cpp
+++ b/indra/llrender/llrender.cpp
@@ -1070,6 +1070,16 @@ LLRender::~LLRender()
 
 void LLRender::init()
 {
+	if (sGLCoreProfile && !LLVertexBuffer::sUseVAO)
+	{ //bind a dummy vertex array object so we're core profile compliant
+#ifdef GL_ARB_vertex_array_object
+		U32 ret;
+		glGenVertexArrays(1, &ret);
+		glBindVertexArray(ret);
+#endif
+	}
+
+
 	llassert_always(mBuffer.isNull()) ;
 	stop_glerror();
 	mBuffer = new LLVertexBuffer(immediate_mask, 0);
diff --git a/indra/llrender/llshadermgr.cpp b/indra/llrender/llshadermgr.cpp
index b6a9a6b653..825f80a6dc 100644
--- a/indra/llrender/llshadermgr.cpp
+++ b/indra/llrender/llshadermgr.cpp
@@ -974,7 +974,9 @@ void LLShaderMgr::initAttribsAndUniforms()
 	mReservedUniforms.push_back("texture_matrix1");
 	mReservedUniforms.push_back("texture_matrix2");
 	mReservedUniforms.push_back("texture_matrix3");
-	llassert(mReservedUniforms.size() == LLShaderMgr::TEXTURE_MATRIX3+1);
+	mReservedUniforms.push_back("object_plane_s");
+	mReservedUniforms.push_back("object_plane_t");
+	llassert(mReservedUniforms.size() == LLShaderMgr::OBJECT_PLANE_T+1);
 
 	mReservedUniforms.push_back("viewport");
 
@@ -1116,6 +1118,48 @@ void LLShaderMgr::initAttribsAndUniforms()
 	mReservedUniforms.push_back("bloomMap");
 	mReservedUniforms.push_back("projectionMap");
 
+	mReservedUniforms.push_back("matrixPalette");
+	
+	
+	mReservedUniforms.reserve(12);
+	mReservedUniforms.push_back("screenTex");
+	mReservedUniforms.push_back("screenDepth");
+	mReservedUniforms.push_back("refTex");
+	mReservedUniforms.push_back("eyeVec");
+	mReservedUniforms.push_back("time");
+	mReservedUniforms.push_back("d1");
+	mReservedUniforms.push_back("d2");
+	mReservedUniforms.push_back("lightDir");
+	mReservedUniforms.push_back("specular");
+	mReservedUniforms.push_back("lightExp");
+	mReservedUniforms.push_back("waterFogColor");
+	mReservedUniforms.push_back("waterFogDensity");
+	mReservedUniforms.push_back("waterFogKS");
+	mReservedUniforms.push_back("refScale");
+	mReservedUniforms.push_back("waterHeight");
+	mReservedUniforms.push_back("waterPlane");
+	mReservedUniforms.push_back("normScale");
+	mReservedUniforms.push_back("fresnelScale");
+	mReservedUniforms.push_back("fresnelOffset");
+	mReservedUniforms.push_back("blurMultiplier");
+	mReservedUniforms.push_back("sunAngle");
+	mReservedUniforms.push_back("scaledAngle");
+	mReservedUniforms.push_back("sunAngle2");
+	
+	mReservedUniforms.push_back("camPosLocal");
+
+	mReservedUniforms.push_back("gWindDir");
+	mReservedUniforms.push_back("gSinWaveParams");
+	mReservedUniforms.push_back("gGravity");
+
+	mReservedUniforms.push_back("detail_0");
+	mReservedUniforms.push_back("detail_1");
+	mReservedUniforms.push_back("detail_2");
+	mReservedUniforms.push_back("detail_3");
+	mReservedUniforms.push_back("alpha_ramp");
+
+	mReservedUniforms.push_back("origin");
+
 	llassert(mReservedUniforms.size() == END_RESERVED_UNIFORMS);
 
 	std::set<std::string> dupe_check;
diff --git a/indra/llrender/llshadermgr.h b/indra/llrender/llshadermgr.h
index 7a16b7c20f..77e90372e0 100644
--- a/indra/llrender/llshadermgr.h
+++ b/indra/llrender/llshadermgr.h
@@ -47,6 +47,8 @@ public:
 		TEXTURE_MATRIX1,
 		TEXTURE_MATRIX2,
 		TEXTURE_MATRIX3,
+		OBJECT_PLANE_S,
+		OBJECT_PLANE_T,
 		VIEWPORT,
 		LIGHT_POSITION,
 		LIGHT_DIRECTION,
@@ -164,7 +166,49 @@ public:
 		DEFERRED_LIGHT,
 		DEFERRED_BLOOM,
 		DEFERRED_PROJECTION,
+		
+		AVATAR_MATRIX,
+
+		WATER_SCREENTEX,
+		WATER_SCREENDEPTH,
+		WATER_REFTEX,
+		WATER_EYEVEC,
+		WATER_TIME,
+		WATER_WAVE_DIR1,
+		WATER_WAVE_DIR2,
+		WATER_LIGHT_DIR,
+		WATER_SPECULAR,
+		WATER_SPECULAR_EXP,
+		WATER_FOGCOLOR,
+		WATER_FOGDENSITY,
+		WATER_FOGKS,
+		WATER_REFSCALE,
+		WATER_WATERHEIGHT,
+		WATER_WATERPLANE,
+		WATER_NORM_SCALE,
+		WATER_FRESNEL_SCALE,
+		WATER_FRESNEL_OFFSET,
+		WATER_BLUR_MULTIPLIER,
+		WATER_SUN_ANGLE,
+		WATER_SCALED_ANGLE,
+		WATER_SUN_ANGLE2,
+		
+		WL_CAMPOSLOCAL,
+
+		AVATAR_WIND,
+		AVATAR_SINWAVE,
+		AVATAR_GRAVITY,
+
+		TERRAIN_DETAIL0,
+		TERRAIN_DETAIL1,
+		TERRAIN_DETAIL2,
+		TERRAIN_DETAIL3,
+		TERRAIN_ALPHARAMP,
+		
+		SHINY_ORIGIN,
+
 		END_RESERVED_UNIFORMS
+
 	} eGLSLReservedUniforms;
 
 	// singleton pattern implementation
diff --git a/indra/llrender/llvertexbuffer.cpp b/indra/llrender/llvertexbuffer.cpp
index f152911b24..1d257d8415 100644
--- a/indra/llrender/llvertexbuffer.cpp
+++ b/indra/llrender/llvertexbuffer.cpp
@@ -202,7 +202,7 @@ volatile U8* LLVBOPool::allocate(U32& name, U32 size, bool for_seed)
 			glBufferDataARB(mType, size, 0, mUsage);
 			if (mUsage != GL_DYNAMIC_COPY_ARB)
 			{ //data will be provided by application
-				ret = (U8*) ll_aligned_malloc_16(size);
+				ret = (U8*) ll_aligned_malloc(size, 64);
 			}
 		}
 		else
@@ -256,7 +256,7 @@ void LLVBOPool::release(U32 name, volatile U8* buffer, U32 size)
 	llassert(vbo_block_size(size) == size);
 
 	deleteBuffer(name);
-	ll_aligned_free_16((U8*) buffer);
+	ll_aligned_free((U8*) buffer);
 
 	if (mType == GL_ARRAY_BUFFER_ARB)
 	{
@@ -1298,7 +1298,7 @@ void LLVertexBuffer::allocateBuffer(S32 nverts, S32 nindices, bool create)
 		//actually allocate space for the vertex buffer if using VBO mapping
 		flush();
 
-		if (gGLManager.mHasVertexArrayObject && useVBOs() && (LLRender::sGLCoreProfile || sUseVAO))
+		if (gGLManager.mHasVertexArrayObject && useVBOs() && (sUseVAO))
 		{
 #if GL_ARB_vertex_array_object
 			mGLArray = getVAOName();
@@ -1454,21 +1454,18 @@ bool LLVertexBuffer::useVBOs() const
 
 //----------------------------------------------------------------------------
 
-bool expand_region(LLVertexBuffer::MappedRegion& region, S32 index, S32 count)
+bool expand_region(LLVertexBuffer::MappedRegion& region, S32 start, S32 end)
 {
-	S32 end = index+count;
-	S32 region_end = region.mIndex+region.mCount;
-	
 	if (end < region.mIndex ||
-		index > region_end)
+		start > region.mEnd)
 	{ //gap exists, do not merge
 		return false;
 	}
 
-	S32 new_end = llmax(end, region_end);
-	S32 new_index = llmin(index, region.mIndex);
-	region.mIndex = new_index;
-	region.mCount = new_end-new_index;
+	region.mEnd = llmax(end, region.mEnd);
+	region.mIndex = llmin(start, region.mIndex);
+	region.mCount = region.mEnd-region.mIndex;
+
 	return true;
 }
 
@@ -1478,7 +1475,6 @@ static LLFastTimer::DeclareTimer FTM_VBO_MAP_BUFFER("VBO Map");
 // Map for data access
 volatile U8* LLVertexBuffer::mapVertexBuffer(S32 type, S32 index, S32 count, bool map_range)
 {
-	bindGLBuffer(true);
 	if (mFinal)
 	{
 		llerrs << "LLVertexBuffer::mapVeretxBuffer() called on a finalized buffer." << llendl;
@@ -1499,23 +1495,23 @@ volatile U8* LLVertexBuffer::mapVertexBuffer(S32 type, S32 index, S32 count, boo
 
 			bool mapped = false;
 			//see if range is already mapped
-			for (U32 i = 0; i < mMappedVertexRegions.size(); ++i)
+			S32 start_index = mOffsets[type]+index*sTypeSize[type];
+			S32 end_index = start_index+count*sTypeSize[type];
+
+			for (std::vector<MappedRegion>::iterator iter = mMappedVertexRegions.begin(), end = mMappedVertexRegions.end(); iter != end; ++iter)
 			{
-				MappedRegion& region = mMappedVertexRegions[i];
-				if (region.mType == type)
+				MappedRegion& region = *iter;
+				if (expand_region(region, index, end_index))
 				{
-					if (expand_region(region, index, count))
-					{
-						mapped = true;
-						break;
-					}
+					mapped = true;
+					break;
 				}
 			}
 
 			if (!mapped)
 			{
 				//not already mapped, map new region
-				MappedRegion region(type, mMappable && map_range ? -1 : index, count);
+				MappedRegion region(mMappable && map_range ? -1 : start_index, end_index-start_index);
 				mMappedVertexRegions.push_back(region);
 			}
 		}
@@ -1539,6 +1535,7 @@ volatile U8* LLVertexBuffer::mapVertexBuffer(S32 type, S32 index, S32 count, boo
 			{
 				volatile U8* src = NULL;
 				waitFence();
+				bindGLBuffer();
 				if (gGLManager.mHasMapBufferRange)
 				{
 					if (map_range)
@@ -1657,7 +1654,6 @@ static LLFastTimer::DeclareTimer FTM_VBO_MAP_INDEX("IBO Map");
 
 volatile U8* LLVertexBuffer::mapIndexBuffer(S32 index, S32 count, bool map_range)
 {
-	bindGLIndices(true);
 	if (mFinal)
 	{
 		llerrs << "LLVertexBuffer::mapIndexBuffer() called on a finalized buffer." << llendl;
@@ -1676,12 +1672,14 @@ volatile U8* LLVertexBuffer::mapIndexBuffer(S32 index, S32 count, bool map_range
 				count = mNumIndices-index;
 			}
 
+			S32 end = index+count;
+
 			bool mapped = false;
 			//see if range is already mapped
 			for (U32 i = 0; i < mMappedIndexRegions.size(); ++i)
 			{
 				MappedRegion& region = mMappedIndexRegions[i];
-				if (expand_region(region, index, count))
+				if (expand_region(region, index, end))
 				{
 					mapped = true;
 					break;
@@ -1691,7 +1689,7 @@ volatile U8* LLVertexBuffer::mapIndexBuffer(S32 index, S32 count, bool map_range
 			if (!mapped)
 			{
 				//not already mapped, map new region
-				MappedRegion region(TYPE_INDEX, mMappable && map_range ? -1 : index, count);
+				MappedRegion region(mMappable && map_range ? -1 : index, count);
 				mMappedIndexRegions.push_back(region);
 			}
 		}
@@ -1707,23 +1705,23 @@ volatile U8* LLVertexBuffer::mapIndexBuffer(S32 index, S32 count, bool map_range
 			sMappedCount++;
 			stop_glerror();	
 
-			if (gDebugGL && useVBOs())
-			{
-				GLint elem = 0;
-				glGetIntegerv(GL_ELEMENT_ARRAY_BUFFER_BINDING_ARB, &elem);
-
-				if (elem != mGLIndices)
-				{
-					llerrs << "Wrong index buffer bound!" << llendl;
-				}
-			}
-
 			if(!mMappable)
 			{
 				map_range = false;
 			}
 			else
 			{
+				bindGLIndices();
+				if (gDebugGL && useVBOs())
+				{
+					GLint elem = 0;
+					glGetIntegerv(GL_ELEMENT_ARRAY_BUFFER_BINDING_ARB, &elem);
+
+					if (elem != mGLIndices)
+					{
+						llerrs << "Wrong index buffer bound!" << llendl;
+					}
+				}
 				volatile U8* src = NULL;
 				waitFence();
 				if (gGLManager.mHasMapBufferRange)
@@ -1837,7 +1835,7 @@ void LLVertexBuffer::unmapBuffer()
 		llassert(mUsage != GL_DYNAMIC_COPY_ARB);
 		
 		LLFastTimer t(FTM_VBO_UNMAP);
-		bindGLBuffer(true);
+		bindGLBuffer();
 		updated_all = mIndexLocked; //both vertex and index buffers done updating
 
 		if(!mMappable)
@@ -1848,8 +1846,8 @@ void LLVertexBuffer::unmapBuffer()
 				for (U32 i = 0; i < mMappedVertexRegions.size(); ++i)
 				{
 					const MappedRegion& region = mMappedVertexRegions[i];
-					S32 offset = region.mIndex >= 0 ? mOffsets[region.mType]+sTypeSize[region.mType]*region.mIndex : 0;
-					S32 length = sTypeSize[region.mType]*region.mCount;
+					S32 offset = region.mIndex;
+					S32 length = region.mCount;
 					glBufferSubDataARB(GL_ARRAY_BUFFER_ARB, offset, length, (U8*) mMappedData+offset);
 					stop_glerror();
 				}
@@ -1873,8 +1871,8 @@ void LLVertexBuffer::unmapBuffer()
 					for (U32 i = 0; i < mMappedVertexRegions.size(); ++i)
 					{
 						const MappedRegion& region = mMappedVertexRegions[i];
-						S32 offset = region.mIndex >= 0 ? mOffsets[region.mType]+sTypeSize[region.mType]*region.mIndex : 0;
-						S32 length = sTypeSize[region.mType]*region.mCount;
+						S32 offset = region.mIndex;
+						S32 length = region.mCount;
 						if (gGLManager.mHasMapBufferRange)
 						{
 							LLFastTimer t(FTM_VBO_FLUSH_RANGE);
@@ -2083,7 +2081,6 @@ bool LLVertexBuffer::bindGLArray()
 	if (mGLArray && sGLRenderArray != mGLArray)
 	{
 		{
-			LLFastTimer t(FTM_BIND_GL_ARRAY);
 #if GL_ARB_vertex_array_object
 			glBindVertexArray(mGLArray);
 #endif
@@ -2453,11 +2450,10 @@ void LLVertexBuffer::setupVertexBuffer(U32 data_mask)
 	llglassertok();
 }
 
-LLVertexBuffer::MappedRegion::MappedRegion(S32 type, S32 index, S32 count)
-: mType(type), mIndex(index), mCount(count)
+LLVertexBuffer::MappedRegion::MappedRegion(S32 index, S32 count)
+: mIndex(index), mCount(count)
 { 
-	llassert(mType == LLVertexBuffer::TYPE_INDEX || 
-			mType < LLVertexBuffer::TYPE_TEXTURE_INDEX);
+	mEnd = mIndex+mCount;	
 }	
 
 
diff --git a/indra/llrender/llvertexbuffer.h b/indra/llrender/llvertexbuffer.h
index a3400ae80c..52559d3505 100644
--- a/indra/llrender/llvertexbuffer.h
+++ b/indra/llrender/llvertexbuffer.h
@@ -104,11 +104,11 @@ public:
 	class MappedRegion
 	{
 	public:
-		S32 mType;
 		S32 mIndex;
 		S32 mCount;
+		S32 mEnd;
 		
-		MappedRegion(S32 type, S32 index, S32 count);
+		MappedRegion(S32 index, S32 count);
 	};
 
 	LLVertexBuffer(const LLVertexBuffer& rhs)