diff --git a/indra/llrender/llfontfreetype.cpp b/indra/llrender/llfontfreetype.cpp
index 9b1a4b431..30299ba9d 100644
--- a/indra/llrender/llfontfreetype.cpp
+++ b/indra/llrender/llfontfreetype.cpp
@@ -487,10 +487,20 @@ void LLFontFreetype::renderGlyph(const U32 glyph_index) const
 	if (mFTFace == NULL)
 		return;
 
-	llassert_always(! FT_Load_Glyph(mFTFace, glyph_index, FT_LOAD_DEFAULT));
-
-	llassert_always(! FT_Render_Glyph(mFTFace->glyph, gFontRenderMode) );
-
+	FT_Error error = FT_Load_Glyph(mFTFace, glyph_index, FT_LOAD_DEFAULT);
+#ifdef SHOW_ASSERT
+	if (error)
+	{
+		LL_ERRS() << "FT_Load_Glyph returned " << error << LL_ENDL;
+	}
+#endif
+	error = FT_Render_Glyph(mFTFace->glyph, gFontRenderMode);
+#ifdef SHOW_ASSERT
+	if (error)
+	{
+		LL_ERRS() << "FT_Render_Glyph returned " << error << LL_ENDL;
+	}
+#endif
 	mRenderGlyphCount++;
 }
 
diff --git a/indra/llrender/llrender.cpp b/indra/llrender/llrender.cpp
index 52f4951e6..00df47b1c 100644
--- a/indra/llrender/llrender.cpp
+++ b/indra/llrender/llrender.cpp
@@ -59,9 +59,7 @@ static const U32 LL_NUM_LIGHT_UNITS = 8;
 static const GLenum sGLTextureType[] =
 {
 	GL_TEXTURE_2D,
-	GL_TEXTURE_RECTANGLE_ARB,
 	GL_TEXTURE_CUBE_MAP_ARB
-	//,GL_TEXTURE_2D_MULTISAMPLE  Don't use.
 };
 
 static const GLint sGLAddressMode[] =
@@ -171,7 +169,7 @@ void LLTexUnit::activate(void)
 
 	if ((S32)gGL.mCurrTextureUnitIndex != mIndex || gGL.mDirty)
 	{
-		gGL.flush();
+		//gGL.flush();
 		glActiveTextureARB(GL_TEXTURE0_ARB + mIndex);
 		gGL.mCurrTextureUnitIndex = mIndex;
 	}
@@ -230,7 +228,7 @@ bool LLTexUnit::bind(LLTexture* texture, bool for_rendering, bool forceBind)
 	stop_glerror();
 	if (mIndex >= 0)
 	{
-		gGL.flush();
+		//gGL.flush();
 
 		LLImageGL* gl_tex = NULL ;
 
@@ -344,8 +342,6 @@ bool LLTexUnit::bind(LLCubeMap* cubeMap)
 {
 	if (mIndex < 0) return false;
 
-	gGL.flush();
-
 	if (cubeMap == NULL)
 	{
 		LL_WARNS() << "NULL LLTexUnit::bind cubemap" << LL_ENDL;
@@ -361,6 +357,7 @@ bool LLTexUnit::bind(LLCubeMap* cubeMap)
 	{
 		if (gGLManager.mHasCubeMap && LLCubeMap::sUseCubeMaps)
 		{
+			gGL.flush();
 			activate();
 			enable(LLTexUnit::TT_CUBE_MAP);
 			mCurrTexture = cubeMap->mImages[0]->getTexName();
@@ -390,7 +387,7 @@ bool LLTexUnit::bind(LLRenderTarget* renderTarget, bool bindDepth)
 {
 	if (mIndex < 0) return false;
 
-	gGL.flush();
+	//gGL.flush();
 
 	if (bindDepth)
 	{
@@ -438,12 +435,16 @@ void LLTexUnit::unbind(eTextureType type)
 
 	//always flush and activate for consistency 
 	//   some code paths assume unbind always flushes and sets the active texture
-	gGL.flush();
-	activate();
+	if (gGL.mCurrTextureUnitIndex != mIndex || gGL.mDirty)
+	{
+		gGL.flush();
+		activate();
+	}
 
 	// Disabled caching of binding state.
-	if (mCurrTexType == type)
+	if (mCurrTexType == type && mCurrTexture != 0)
 	{
+		gGL.flush();
 		mCurrTexture = 0;
 		if (LLGLSLShader::sNoFixedFunction && type == LLTexUnit::TT_TEXTURE)
 		{
@@ -1044,6 +1045,7 @@ LLRender::LLRender()
     mMode(LLRender::TRIANGLES),
     mCurrTextureUnitIndex(0),
     mMaxAnisotropy(0.f),
+	mLineWidth(1.f),
 	mPrimitiveReset(false)
 {	
 	mTexUnits.reserve(LL_NUM_TEXTURE_LAYERS);
@@ -1663,8 +1665,6 @@ bool LLRender::unprojectf(const LLVector3& windowCoordinate, const LLMatrix4a& m
 
 void LLRender::pushMatrix()
 {
-	flush();
-	
 	{
 		if (mMatIdx[mMatrixMode] < LL_MATRIX_STACK_DEPTH-1)
 		{
@@ -1680,15 +1680,19 @@ void LLRender::pushMatrix()
 
 void LLRender::popMatrix()
 {
-	flush();
 	{
 		if (mMatIdx[mMatrixMode] > 0)
 		{
+			if ( memcmp(mMatrix[mMatrixMode][mMatIdx[mMatrixMode]].getF32ptr(), mMatrix[mMatrixMode][mMatIdx[mMatrixMode] - 1].getF32ptr(), sizeof(LLMatrix4a)) )
+			{
+				flush();
+			}
 			--mMatIdx[mMatrixMode];
 			mMatHash[mMatrixMode]++;
 		}
 		else
 		{
+			flush();
 			LL_WARNS() << "Matrix stack underflow." << LL_ENDL;
 		}
 	}
@@ -1697,6 +1701,7 @@ void LLRender::popMatrix()
 void LLRender::loadMatrix(const LLMatrix4a& mat)
 {
 	flush();
+
 	mMatrix[mMatrixMode][mMatIdx[mMatrixMode]] = mat;
 	mMatHash[mMatrixMode]++;
 }
@@ -1772,6 +1777,18 @@ void LLRender::scaleUI(F32 x, F32 y, F32 z)
 	mUIScale.back().mul(scale);
 }
 
+void LLRender::rotateUI(LLQuaternion& rot)
+{
+	if (mUIRotation.empty())
+	{
+		mUIRotation.push_back(rot);
+	}
+	else
+	{
+		mUIRotation.push_back(mUIRotation.back()*rot);
+	}
+}
+
 void LLRender::pushUIMatrix()
 {
 	if (mUIOffset.empty())
@@ -1791,6 +1808,10 @@ void LLRender::pushUIMatrix()
 	{
 		mUIScale.push_back(mUIScale.back());
 	}
+	if (!mUIRotation.empty())
+	{
+		mUIRotation.push_back(mUIRotation.back());
+	}
 }
 
 void LLRender::popUIMatrix()
@@ -1801,6 +1822,10 @@ void LLRender::popUIMatrix()
 	}
 	mUIOffset.pop_back();
 	mUIScale.pop_back();
+	if (!mUIRotation.empty())
+	{
+		mUIRotation.pop_back();
+	}
 }
 
 LLVector3 LLRender::getUITranslation()
@@ -1830,6 +1855,8 @@ void LLRender::loadUIIdentity()
 	}
 	mUIOffset.back().splat(0.f);
 	mUIScale.back().splat(1.f);
+	if (!mUIRotation.empty())
+		mUIRotation.push_back(LLQuaternion());
 }
 
 void LLRender::setColorMask(bool writeColor, bool writeAlpha)
@@ -1839,8 +1866,6 @@ void LLRender::setColorMask(bool writeColor, bool writeAlpha)
 
 void LLRender::setColorMask(bool writeColorR, bool writeColorG, bool writeColorB, bool writeAlpha)
 {
-	flush();
-
 	if (mCurrColorMask[0] != writeColorR ||
 		mCurrColorMask[1] != writeColorG ||
 		mCurrColorMask[2] != writeColorB ||
@@ -1851,6 +1876,7 @@ void LLRender::setColorMask(bool writeColorR, bool writeColorG, bool writeColorB
 		mCurrColorMask[2] = writeColorB;
 		mCurrColorMask[3] = writeAlpha;
 
+		flush();
 		glColorMask(writeColorR ? GL_TRUE : GL_FALSE, 
 					writeColorG ? GL_TRUE : GL_FALSE,
 					writeColorB ? GL_TRUE : GL_FALSE,
@@ -1891,8 +1917,6 @@ void LLRender::setSceneBlendType(eBlendType type)
 
 void LLRender::setAlphaRejectSettings(eCompareFunc func, F32 value)
 {
-	flush();
-
 	if (LLGLSLShader::sNoFixedFunction)
 	{ //glAlphaFunc is deprecated in OpenGL 3.3
 		return;
@@ -1901,6 +1925,7 @@ void LLRender::setAlphaRejectSettings(eCompareFunc func, F32 value)
 	if (mCurrAlphaFunc != func ||
 		mCurrAlphaFuncVal != value || mDirty)
 	{
+		flush();
 		mCurrAlphaFunc = func;
 		mCurrAlphaFuncVal = value;
 		if (func == CF_DEFAULT)
@@ -2030,6 +2055,23 @@ void LLRender::setAmbientLightColor(const LLColor4& color)
 	}
 }
 
+void LLRender::setLineWidth(F32 line_width)
+{
+	if (LLRender::sGLCoreProfile)
+	{
+		line_width = 1.f;
+	}
+	if (mLineWidth != line_width)
+	{
+		if (mMode == LLRender::LINES || LLRender::LINE_STRIP)
+		{
+			flush();
+		}
+		mLineWidth = line_width;
+		glLineWidth(line_width);
+	}
+}
+
 bool LLRender::verifyTexUnitActive(U32 unitToVerify)
 {
 	if (mCurrTextureUnitIndex == unitToVerify)
@@ -2092,8 +2134,13 @@ void LLRender::end()
 		mPrimitiveReset = true;
 	}
 }
+
+extern bool countFlushes;
+extern unsigned int flushCount;
+
 void LLRender::flush()
 {
+	if (countFlushes) ++flushCount;
 	if (mCount > 0)
 	{
 #if 0
@@ -2236,13 +2283,33 @@ void LLRender::vertex4a(const LLVector4a& vertex)
 
 	if (mUIOffset.empty())
 	{
-		mVerticesp[mCount]=vertex;
+		if (!mUIRotation.empty() && mUIRotation.back().isNotIdentity())
+		{
+			LLVector4 vert(vertex.getF32ptr());
+			mVerticesp[mCount].loadua((vert*mUIRotation.back()).mV);
+		}
+		else
+		{
+			mVerticesp[mCount] = vertex;
+		}
 	}
 	else
 	{
-		//LLVector3 vert = (LLVector3(x,y,z)+mUIOffset.back()).scaledVec(mUIScale.back());
-		mVerticesp[mCount].setAdd(vertex,mUIOffset.back());
-		mVerticesp[mCount].mul(mUIScale.back());
+		if (!mUIRotation.empty() && mUIRotation.back().isNotIdentity())
+		{
+			LLVector4 vert(vertex.getF32ptr());
+			vert = vert * mUIRotation.back();
+			LLVector4a postrot_vert;
+			postrot_vert.loadua(vert.mV);
+			mVerticesp[mCount].setAdd(postrot_vert, mUIOffset.back());
+			mVerticesp[mCount].mul(mUIScale.back());
+		}
+		else
+		{
+			//LLVector3 vert = (LLVector3(x,y,z)+mUIOffset.back()).scaledVec(mUIScale.back());
+			mVerticesp[mCount].setAdd(vertex, mUIOffset.back());
+			mVerticesp[mCount].mul(mUIScale.back());
+		}
 	}
 
 	mCount++;
diff --git a/indra/llrender/llrender.h b/indra/llrender/llrender.h
index 15be0d3cf..5f7bb74df 100644
--- a/indra/llrender/llrender.h
+++ b/indra/llrender/llrender.h
@@ -65,7 +65,6 @@ public:
 	{
 		TT_TEXTURE = 0,			// Standard 2D Texture
 		TT_CUBE_MAP,		// 6-sided cube map texture
-		//TT_MULTISAMPLE_TEXTURE, // see GL_ARB_texture_multisample  Do not use
 		TT_NONE 		// No texture type is currently enabled
 	} eTextureType;
 
@@ -380,6 +379,8 @@ public:
 
 	void translateUI(F32 x, F32 y, F32 z);
 	void scaleUI(F32 x, F32 y, F32 z);
+	// Rotates vertices, pre-translation/scale
+	void rotateUI(LLQuaternion& rot);
 	void pushUIMatrix();
 	void popUIMatrix();
 	void loadUIIdentity();
@@ -435,6 +436,8 @@ public:
 	LLLightState* getLight(U32 index);
 	void setAmbientLightColor(const LLColor4& color);
 
+	void setLineWidth(F32 line_width);
+
 	LLTexUnit* getTexUnit(U32 index);
 
 	U32	getCurrentTexUnitIndex(void) const { return mCurrTextureUnitIndex; }
@@ -475,6 +478,7 @@ private:
 	bool				mCurrColorMask[4];
 	eCompareFunc			mCurrAlphaFunc;
 	F32				mCurrAlphaFuncVal;
+	F32				mLineWidth;
 
 	LLPointer<LLVertexBuffer>	mBuffer;
 	LLStrider<LLVector4a>		mVerticesp;
@@ -493,6 +497,7 @@ private:
 
 	std::vector<LLVector4a, boost::alignment::aligned_allocator<LLVector4a, 64> > mUIOffset;
 	std::vector<LLVector4a, boost::alignment::aligned_allocator<LLVector4a, 64> > mUIScale;
+	std::vector<LLQuaternion> mUIRotation;
 
 	bool			mPrimitiveReset;
 } LL_ALIGN_POSTFIX(16);
diff --git a/indra/llrender/llrender2dutils.cpp b/indra/llrender/llrender2dutils.cpp
index 4960d5e8f..a82032f9c 100644
--- a/indra/llrender/llrender2dutils.cpp
+++ b/indra/llrender/llrender2dutils.cpp
@@ -490,7 +490,6 @@ void gl_draw_scaled_image_with_border(S32 x, S32 y, S32 width, S32 height, LLTex
 		gGL.getTexUnit(0)->bind(image, true);
 
 	gGL.color4fv(color.mV);
-	gGL.diffuseColor4fv(color.mV); //workaround: Intel HD 4000
 	
 	const S32 NUM_VERTICES = (2 + 2 * 3) * 3 + 4;
 	LLVector2 uv[NUM_VERTICES];
@@ -894,12 +893,11 @@ void gl_ring( F32 radius, F32 width, const LLColor4& center_color, const LLColor
 		if( render_center )
 		{
 			gGL.color4fv(center_color.mV);
-			gGL.diffuseColor4fv(center_color.mV);
 			gl_deep_circle( radius, width, steps );
 		}
 		else
 		{
-			gGL.diffuseColor4fv(side_color.mV);
+			gGL.color4fv(side_color.mV);
 			gl_washer_2d(radius - width, radius, steps, side_color, side_color);
 			gGL.translateUI(0.f, 0.f, width);
 			gl_washer_2d(radius, radius - width, steps, side_color, side_color);
@@ -1520,8 +1518,7 @@ void LLRender2D::setScaleFactor(const LLVector2 &scale_factor)
 //static
 void LLRender2D::setLineWidth(F32 width)
 {
-	gGL.flush();
-	glLineWidth(width * lerp(sGLScaleFactor.mV[VX], sGLScaleFactor.mV[VY], 0.5f));
+	gGL.setLineWidth(width * lerp(sGLScaleFactor.mV[VX], sGLScaleFactor.mV[VY], 0.5f));
 }
 
 //static
diff --git a/indra/llrender/llshadermgr.cpp b/indra/llrender/llshadermgr.cpp
index 6800bd2b4..8f697e72d 100644
--- a/indra/llrender/llshadermgr.cpp
+++ b/indra/llrender/llshadermgr.cpp
@@ -538,7 +538,7 @@ GLhandleARB LLShaderMgr::loadShaderFile(const std::string& filename, S32 & shade
 	range = mShaderObjects.equal_range(filename);
 	for (std::multimap<std::string, CachedObjectInfo>::iterator it = range.first; it != range.second;++it)
 	{
-		if((*it).second.mLevel == shader_level && (*it).second.mType == type && (*it).second.mDefinitions == (defines ? *defines : std::map<std::string, std::string>()))
+		if((*it).second.mLevel == shader_level && (*it).second.mType == type && (*it).second.mIndexChannels == texture_index_channels && (*it).second.mDefinitions == (defines ? *defines : std::map<std::string, std::string>()))
 		{
 			//LL_INFOS("ShaderLoading") << "Loading cached shader for " << filename << LL_ENDL;
 			return (*it).second.mHandle;
@@ -935,7 +935,7 @@ GLhandleARB LLShaderMgr::loadShaderFile(const std::string& filename, S32 & shade
 	if (ret)
 	{
 		// Add shader file to map
-		mShaderObjects.insert(make_pair(filename,CachedObjectInfo(ret,try_gpu_class,type,defines)));
+		mShaderObjects.insert(make_pair(filename,CachedObjectInfo(ret,try_gpu_class,type, texture_index_channels,defines)));
 		shader_level = try_gpu_class;
 	}
 	else
diff --git a/indra/llrender/llshadermgr.h b/indra/llrender/llshadermgr.h
index 02bcdf7b5..53e9abb2a 100644
--- a/indra/llrender/llshadermgr.h
+++ b/indra/llrender/llshadermgr.h
@@ -242,11 +242,12 @@ DISPLAY_GAMMA,
 public:
 	struct CachedObjectInfo
 	{
-		CachedObjectInfo(GLhandleARB handle, U32 level, GLenum type, std::map<std::string,std::string> *definitions) : 
-			mHandle(handle), mLevel(level), mType(type), mDefinitions(definitions ? *definitions : std::map<std::string,std::string>()){}
+		CachedObjectInfo(GLhandleARB handle, U32 level, GLenum type, U32 texture_index_channels, std::map<std::string,std::string> *definitions) : 
+			mHandle(handle), mLevel(level), mType(type), mIndexChannels(texture_index_channels), mDefinitions(definitions ? *definitions : std::map<std::string,std::string>()){}
 		GLhandleARB mHandle;	//Actual handle of the opengl shader object.
 		U32 mLevel;				//Level /might/ not be needed, but it's stored to ensure there's no change in behavior.
 		GLenum mType;			//GL_VERTEX_SHADER_ARB or GL_FRAGMENT_SHADER_ARB. Tracked because some utility shaders can be loaded as both types (carefully).
+		U32 mIndexChannels;     //LLShaderFeatures::mIndexedTextureChannels
 		std::map<std::string,std::string> mDefinitions;
 	};
 	// Map of shader names to compiled
diff --git a/indra/newview/app_settings/shaders/class1/deferred/dofCombineF.glsl b/indra/newview/app_settings/shaders/class1/deferred/dofCombineF.glsl
index 75be20aba..c0e4e791d 100644
--- a/indra/newview/app_settings/shaders/class1/deferred/dofCombineF.glsl
+++ b/indra/newview/app_settings/shaders/class1/deferred/dofCombineF.glsl
@@ -40,6 +40,7 @@ uniform float max_cof;
 uniform float res_scale;
 uniform float dof_width;
 uniform float dof_height;
+uniform float kern_scale;
 
 VARYING vec2 vary_fragcoord;
 
@@ -53,9 +54,9 @@ vec4 dofSample(sampler2D tex, vec2 tc)
 
 void main() 
 {
-	vec2 tc = vary_fragcoord.xy;
+	vec2 tc = vary_fragcoord.xy * res_scale;
 	
-	vec4 dof = dofSample(diffuseRect, vary_fragcoord.xy*res_scale);
+	vec4 dof = dofSample(diffuseRect, tc);
 	
 	vec4 diff = texture2D(lightMap, vary_fragcoord.xy);
 
@@ -66,10 +67,10 @@ void main()
 		float sc = a/res_scale;
 		
 		vec4 col;
-		col = texture2D(lightMap, vary_fragcoord.xy+vec2(sc,sc));
-		col += texture2D(lightMap, vary_fragcoord.xy+vec2(-sc,sc));
-		col += texture2D(lightMap, vary_fragcoord.xy+vec2(sc,-sc));
-		col += texture2D(lightMap, vary_fragcoord.xy+vec2(-sc,-sc));
+		col = texture2D(lightMap, vary_fragcoord.xy+vec2(sc,sc)*kern_scale);
+		col += texture2D(lightMap, vary_fragcoord.xy+vec2(-sc,sc)*kern_scale);
+		col += texture2D(lightMap, vary_fragcoord.xy+vec2(sc,-sc)*kern_scale);
+		col += texture2D(lightMap, vary_fragcoord.xy+vec2(-sc,-sc)*kern_scale);
 		
 		diff = mix(diff, col*0.25, a);
 	}
diff --git a/indra/newview/app_settings/shaders/class1/deferred/postDeferredF.glsl b/indra/newview/app_settings/shaders/class1/deferred/postDeferredF.glsl
index b81aa6b1b..57522e953 100644
--- a/indra/newview/app_settings/shaders/class1/deferred/postDeferredF.glsl
+++ b/indra/newview/app_settings/shaders/class1/deferred/postDeferredF.glsl
@@ -36,6 +36,7 @@ uniform sampler2D diffuseRect;
 uniform mat4 inv_proj;
 uniform float max_cof;
 uniform float res_scale;
+uniform vec2 kern_scale;
 
 VARYING vec2 vary_fragcoord;
 
@@ -76,7 +77,7 @@ void main()
 {
 	vec2 tc = vary_fragcoord.xy;
 	
-	vec4 diff = texture2D(diffuseRect, vary_fragcoord.xy);
+	vec4 diff = texture2D(diffuseRect, tc);
 	
 	{ 
 		float w = 1.0;
@@ -97,7 +98,7 @@ void main()
 					float samp_x = sc*sin(ang);
 					float samp_y = sc*cos(ang);
 					// you could test sample coords against an interesting non-circular aperture shape here, if desired.
-					dofSampleNear(diff, w, sc, vary_fragcoord.xy + vec2(samp_x,samp_y));
+					dofSampleNear(diff, w, sc, tc + vec2(samp_x,samp_y) * kern_scale);
 				}
 				sc -= 1.0;
 			}
@@ -114,7 +115,7 @@ void main()
 					float samp_x = sc*sin(ang);
 					float samp_y = sc*cos(ang);
 					// you could test sample coords against an interesting non-circular aperture shape here, if desired.
-					dofSample(diff, w, sc, vary_fragcoord.xy + vec2(samp_x,samp_y));
+					dofSample(diff, w, sc, tc + vec2(samp_x,samp_y) * kern_scale);
 				}
 				sc -= 1.0;
 			}
diff --git a/indra/newview/app_settings/shaders/class1/deferred/sunLightSSAOF.glsl b/indra/newview/app_settings/shaders/class1/deferred/sunLightSSAOF.glsl
index cc3a30768..55c780535 100644
--- a/indra/newview/app_settings/shaders/class1/deferred/sunLightSSAOF.glsl
+++ b/indra/newview/app_settings/shaders/class1/deferred/sunLightSSAOF.glsl
@@ -33,15 +33,16 @@ out vec4 frag_color;
 
 VARYING vec2 vary_fragcoord;
 
-uniform sampler2D depthMapDownsampled;
 uniform sampler2D depthMap;
 
 uniform sampler2D diffuseRect;
 
+uniform vec2 kern_scale;
+
 void main() 
 {	
 	frag_color[0] = 1.0;
-    frag_color[1] = texture2D(diffuseRect,vary_fragcoord.xy).r;
+    frag_color[1] = texture2D(diffuseRect,vary_fragcoord.xy * kern_scale).r;
 	frag_color[2] = 1.0;
 	frag_color[3] = 1.0;
 }
diff --git a/indra/newview/app_settings/shaders/class2/deferred/sunLightSSAOF.glsl b/indra/newview/app_settings/shaders/class2/deferred/sunLightSSAOF.glsl
index b38cd76bb..1f53c86ac 100644
--- a/indra/newview/app_settings/shaders/class2/deferred/sunLightSSAOF.glsl
+++ b/indra/newview/app_settings/shaders/class2/deferred/sunLightSSAOF.glsl
@@ -50,6 +50,8 @@ uniform vec4 shadow_clip;
 
 VARYING vec2 vary_fragcoord;
 
+uniform vec2 kern_scale;
+
 uniform mat4 inv_proj;
 uniform vec2 proj_shadow_res;
 uniform vec3 sun_dir;
@@ -62,6 +64,7 @@ uniform float shadow_offset;
 uniform float spot_shadow_bias;
 uniform float spot_shadow_offset;
 
+
 vec2 encode_normal(vec3 n)
 {
 	float f = sqrt(8 * n.z + 8);
@@ -241,7 +244,7 @@ void main()
 	}
 	
 	frag_color[0] = shadow;
-	frag_color[1] = texture2D(diffuseRect,vary_fragcoord).r;
+	frag_color[1] = texture2D(diffuseRect,vary_fragcoord.xy * kern_scale).r;
 	
 	spos = vec4(shadow_pos+norm*spot_shadow_offset, 1.0);
 	
diff --git a/indra/newview/lldrawpoolalpha.cpp b/indra/newview/lldrawpoolalpha.cpp
index b3897005b..587c0418e 100644
--- a/indra/newview/lldrawpoolalpha.cpp
+++ b/indra/newview/lldrawpoolalpha.cpp
@@ -125,6 +125,7 @@ void LLDrawPoolAlpha::beginPostDeferredPass(S32 pass)
 
 		//prime simple shader (loads shadow relevant uniforms)
 		gPipeline.bindDeferredShader(*simple_shader);
+		gPipeline.unbindDeferredShader(*simple_shader);
 
 		//simple_shader->uniform1f(LLShaderMgr::DISPLAY_GAMMA, (gamma > 0.1f) ? 1.0f / gamma : (1.0f/2.2f));
 	}
@@ -154,7 +155,11 @@ void LLDrawPoolAlpha::beginPostDeferredPass(S32 pass)
 }
 
 void LLDrawPoolAlpha::endPostDeferredPass(S32 pass) 
-{ 
+{
+	if (current_shader)
+	{
+		gPipeline.unbindDeferredShader(*current_shader);
+	}
 	if (pass == 1 && !LLPipeline::sImpostorRender)
 	{
 		gPipeline.mDeferredDepth.flush();
@@ -437,6 +442,8 @@ void LLDrawPoolAlpha::renderAlpha(U32 mask, S32 pass)
 						// (this way we won't rebind shaders unnecessarily).
 						if (current_shader != target_shader)
 						{
+							if(current_shader)
+								gPipeline.unbindDeferredShader(*current_shader);
 							gPipeline.bindDeferredShader(*target_shader);
 						}
 					}
@@ -448,6 +455,8 @@ void LLDrawPoolAlpha::renderAlpha(U32 mask, S32 pass)
 						// (this way we won't rebind shaders unnecessarily).
 						if(current_shader != target_shader)
 						{
+							if(current_shader)
+								gPipeline.unbindDeferredShader(*current_shader);
 							target_shader->bind();
 						}
 					}
diff --git a/indra/newview/llnetmap.cpp b/indra/newview/llnetmap.cpp
index 431db3206..d36c521e4 100644
--- a/indra/newview/llnetmap.cpp
+++ b/indra/newview/llnetmap.cpp
@@ -268,21 +268,12 @@ void LLNetMap::draw()
 	// Prepare a scissor region
 	F32 rotation = 0;
 
-	gGL.pushMatrix();
 	gGL.pushUIMatrix();
-	
-	LLVector3 offset = gGL.getUITranslation();
-	LLVector3 scale = gGL.getUIScale();
 
-	gGL.loadIdentity();
-	gGL.loadUIIdentity();
-	gGL.scalef(scale.mV[0], scale.mV[1], scale.mV[2]);
-	gGL.translatef(offset.mV[0], offset.mV[1], offset.mV[2]);
 	{
 		LLLocalClipRect clip(getLocalRect());
 		{
 			gGL.getTexUnit(0)->unbind(LLTexUnit::TT_TEXTURE);
-			gGL.matrixMode(LLRender::MM_MODELVIEW);
 
 			// Draw background rectangle.
 			if(isBackgroundVisible())
@@ -297,15 +288,16 @@ void LLNetMap::draw()
 		S32 center_sw_left = getRect().getWidth() / 2 + llfloor(mCurPan.mV[VX]);
 		S32 center_sw_bottom = getRect().getHeight() / 2 + llfloor(mCurPan.mV[VY]);
 
-		gGL.pushMatrix();
-		gGL.translatef( (F32) center_sw_left, (F32) center_sw_bottom, 0.f);
+		gGL.pushUIMatrix();
+		gGL.translateUI( (F32) center_sw_left, (F32) center_sw_bottom, 0.f);
 
 		static LLCachedControl<bool> rotate_map("MiniMapRotate", true);
 		if (rotate_map)
 		{
 			// Rotate subsequent draws to agent rotation.
 			rotation = atan2( LLViewerCamera::getInstance()->getAtAxis().mV[VX], LLViewerCamera::getInstance()->getAtAxis().mV[VY] );
-			gGL.rotatef( rotation * RAD_TO_DEG, 0.f, 0.f, 1.f);
+			LLQuaternion rot(rotation, LLVector3(0.f, 0.f, 1.f));
+			gGL.rotateUI(rot);
 		}
 
 		// Figure out where agent is.
@@ -537,7 +529,7 @@ void LLNetMap::draw()
 		}
 // [/SL:KB]
 
-		gGL.popMatrix();
+		gGL.popUIMatrix();
 
 		// Mouse pointer in local coordinates
 		S32 local_mouse_x;
@@ -709,14 +701,15 @@ void LLNetMap::draw()
 
 		LLColor4 c = rotate_map ? map_frustum_color() : map_frustum_rotating_color();
 
-		gGL.pushMatrix();
+		gGL.pushUIMatrix();
 
-		gGL.translatef(ctr_x, ctr_y, 0);
+		gGL.translateUI(ctr_x, ctr_y, 0);
 
 		// If we don't rotate the map, we have to rotate the frustum.
 		if (!rotate_map)
 		{
-			gGL.rotatef(atan2(LLViewerCamera::getInstance()->getAtAxis().mV[VX], LLViewerCamera::getInstance()->getAtAxis().mV[VY]) * RAD_TO_DEG, 0.f, 0.f, -1.f);
+			LLQuaternion rot(atan2(LLViewerCamera::getInstance()->getAtAxis().mV[VX], LLViewerCamera::getInstance()->getAtAxis().mV[VY]), LLVector3(0.f, 0.f, -1.f));
+			gGL.rotateUI(rot);
 		}
 
 		gGL.begin( LLRender::TRIANGLES  );
@@ -728,7 +721,7 @@ void LLNetMap::draw()
 			gGL.vertex2f( -half_width_pixels, far_clip_pixels );
 		gGL.end();
 
-		gGL.popMatrix();
+		gGL.popUIMatrix();
 
 		// <exodus> Draw mouse radius
 		static const LLCachedControl<LLColor4> map_avatar_rollover_color("ExodusMapRolloverCircleColor");
@@ -737,8 +730,7 @@ void LLNetMap::draw()
 		gl_circle_2d(local_mouse_x, local_mouse_y, min_pick_dist, 32, true);
 		// </exodus>
 	}
-	
-	gGL.popMatrix();
+	;
 	gGL.popUIMatrix();
 
 	// Rotation of 0 means that North is up
diff --git a/indra/newview/pipeline.cpp b/indra/newview/pipeline.cpp
index 52a064dba..2362670de 100644
--- a/indra/newview/pipeline.cpp
+++ b/indra/newview/pipeline.cpp
@@ -7194,7 +7194,7 @@ void LLPipeline::renderBloom(BOOL for_snapshot, F32 zoom_factor, int subfield, b
 
 				drawFullScreenRect();
 
-				unbindDeferredShader(*shader);
+				unbindDeferredShader(*shader, &mFinalScreen);
 				mDeferredLight.flush();
 			}
 
@@ -7214,7 +7214,7 @@ void LLPipeline::renderBloom(BOOL for_snapshot, F32 zoom_factor, int subfield, b
 
 				drawFullScreenRect();
 
-				unbindDeferredShader(*shader);
+				unbindDeferredShader(*shader, &mDeferredLight);
 				mFinalScreen.flush();
 				gGL.setColorMask(true, true);
 			}
@@ -7252,8 +7252,8 @@ void LLPipeline::renderBloom(BOOL for_snapshot, F32 zoom_factor, int subfield, b
 
 				shader->uniform1f(LLShaderMgr::DOF_MAX_COF, CameraMaxCoF);
 				shader->uniform1f(LLShaderMgr::DOF_RES_SCALE, CameraDoFResScale);
-				shader->uniform1f(LLShaderMgr::DOF_WIDTH, dof_width-1);
-				shader->uniform1f(LLShaderMgr::DOF_HEIGHT, dof_height-1);
+				shader->uniform1f(LLShaderMgr::DOF_WIDTH, CameraDoFResScale - CameraDoFResScale / dof_width);
+				shader->uniform1f(LLShaderMgr::DOF_HEIGHT, CameraDoFResScale - CameraDoFResScale / dof_height);
 
 				drawFullScreenRect();
 
@@ -7262,7 +7262,7 @@ void LLPipeline::renderBloom(BOOL for_snapshot, F32 zoom_factor, int subfield, b
 					gGL.getTexUnit(channel)->setTextureFilteringOption(LLTexUnit::TFO_POINT);
 				}
 
-				unbindDeferredShader(*shader);
+				unbindDeferredShader(*shader, &mFinalScreen);
 
 				if (multisample)
 				{
@@ -7300,7 +7300,7 @@ void LLPipeline::renderBloom(BOOL for_snapshot, F32 zoom_factor, int subfield, b
 				gGL.getTexUnit(channel)->setTextureFilteringOption(LLTexUnit::TFO_POINT);
 			}
 
-			unbindDeferredShader(*shader);
+			unbindDeferredShader(*shader, &mFinalScreen);
 
 			if (multisample)
 			{
@@ -7335,7 +7335,7 @@ void LLPipeline::renderBloom(BOOL for_snapshot, F32 zoom_factor, int subfield, b
 				gGL.getTexUnit(channel)->setTextureFilteringOption(LLTexUnit::TFO_POINT);
 			}
 
-			unbindDeferredShader(*shader);
+			unbindDeferredShader(*shader, &render_target);
 			
 			mFXAABuffer.flush();
 
@@ -7570,7 +7570,7 @@ void LLPipeline::bindDeferredShader(LLGLSLShader& shader, LLRenderTarget* diffus
 
 	for (U32 i = 0; i < 4; i++)
 	{
-		channel = shader.enableTexture(LLShaderMgr::DEFERRED_SHADOW0+i, LLTexUnit::TT_TEXTURE);
+		channel = shader.enableTexture(LLShaderMgr::DEFERRED_SHADOW0+i);
 		stop_glerror();
 		if (channel > -1)
 		{
@@ -7671,8 +7671,6 @@ void LLPipeline::bindDeferredShader(LLGLSLShader& shader, LLRenderTarget* diffus
 		norm_mat.transpose();
 		shader.uniformMatrix4fv(LLShaderMgr::DEFERRED_NORM_MATRIX, 1, FALSE, norm_mat.getF32ptr());
 	}
-
-	//shader.uniform1f(LLShaderMgr::DEFERRED_DOWNSAMPLED_DEPTH_SCALE, llclamp(RenderSSAOResolutionScale.get(),.01f,1.f));
 }
 
 static LLFastTimer::DeclareTimer FTM_GI_TRACE("Trace");
@@ -7796,6 +7794,9 @@ void LLPipeline::renderDeferredLighting()
 				mDeferredLight.clear(GL_COLOR_BUFFER_BIT);
 				glClearColor(0,0,0,0);
 
+				F32 ssao_scale = llclamp(RenderSSAOResolutionScale.get(), .01f, 1.f);
+				gDeferredSunProgram.uniform2f(LLShaderMgr::DEFERRED_KERN_SCALE, ssao_scale, ssao_scale);
+
 				//Enable bilinear filtering, as the screen tex resolution may not match current framebuffer resolution. Eg, half-res SSAO
 				// diffuse map should only be found if the sun shader is the SSAO variant.
 				S32 channel = gDeferredSunProgram.enableTexture(LLShaderMgr::DEFERRED_DIFFUSE, mScreen.getUsage());
@@ -7838,7 +7839,7 @@ void LLPipeline::renderDeferredLighting()
 			F32 dist_factor = RenderShadowBlurDistFactor;
 
 			// sample symmetrically with the middle sample falling exactly on 0.0
-			F32 x = 0.f;
+			//F32 x = 0.f;
 
 			gDeferredBlurLightProgram.uniform2f(sDelta, (blur_size * (kern_length / 2.f - 0.5f)) / mScreen.getWidth(), 0.f);
 			gDeferredBlurLightProgram.uniform1f(sDistFactor, dist_factor);
@@ -7863,7 +7864,7 @@ void LLPipeline::renderDeferredLighting()
 				drawFullScreenRect();
 			}
 			mDeferredLight.flush();
-			unbindDeferredShader(gDeferredBlurLightProgram);
+			unbindDeferredShader(gDeferredBlurLightProgram, &mScreen, &mScreen);
 		}
 
 		stop_glerror();
@@ -8374,6 +8375,10 @@ void LLPipeline::renderDeferredLightingToRT(LLRenderTarget* target)
 			{ //paint shadow/SSAO light map (direct lighting lightmap)
 				LLFastTimer ftm(FTM_SUN_SHADOW);
 				bindDeferredShader(gDeferredSunProgram);
+
+				F32 ssao_scale = llclamp(RenderSSAOResolutionScale.get(), .01f, 1.f);
+				gDeferredSunProgram.uniform2f(LLShaderMgr::DEFERRED_KERN_SCALE, ssao_scale, ssao_scale);
+
 				glClearColor(1,1,1,1);
 				mDeferredLight.clear(GL_COLOR_BUFFER_BIT);
 				glClearColor(0,0,0,0);
@@ -8935,14 +8940,17 @@ void LLPipeline::setupSpotLight(LLGLSLShader& shader, LLDrawable* drawablep)
 		
 }
 
-void LLPipeline::unbindDeferredShader(LLGLSLShader &shader)
+void LLPipeline::unbindDeferredShader(LLGLSLShader &shader, LLRenderTarget* diffuse_source, LLRenderTarget* light_source)
 {
+	diffuse_source = diffuse_source ? diffuse_source : &mDeferredScreen;
+	light_source = light_source ? light_source : &mDeferredLight;
+
 	stop_glerror();
 	shader.disableTexture(LLShaderMgr::DEFERRED_NORMAL, mDeferredScreen.getUsage());
-	shader.disableTexture(LLShaderMgr::DEFERRED_DIFFUSE, mDeferredScreen.getUsage());
+	shader.disableTexture(LLShaderMgr::DEFERRED_DIFFUSE, diffuse_source->getUsage());
 	shader.disableTexture(LLShaderMgr::DEFERRED_SPECULAR, mDeferredScreen.getUsage());
-	shader.disableTexture(LLShaderMgr::DEFERRED_DEPTH, mDeferredScreen.getUsage());
-	shader.disableTexture(LLShaderMgr::DEFERRED_LIGHT, mDeferredLight.getUsage());
+	shader.disableTexture(LLShaderMgr::DEFERRED_DEPTH, mDeferredDepth.getUsage());
+	shader.disableTexture(LLShaderMgr::DEFERRED_LIGHT, light_source->getUsage());
 	shader.disableTexture(LLShaderMgr::DIFFUSE_MAP);
 	shader.disableTexture(LLShaderMgr::DEFERRED_BLOOM);
 
diff --git a/indra/newview/pipeline.h b/indra/newview/pipeline.h
index 4629a9839..a94a49745 100644
--- a/indra/newview/pipeline.h
+++ b/indra/newview/pipeline.h
@@ -289,7 +289,7 @@ public:
 	void bindDeferredShader(LLGLSLShader& shader, LLRenderTarget* diffuse_source = NULL, LLRenderTarget* light_source = NULL);
 	void setupSpotLight(LLGLSLShader& shader, LLDrawable* drawablep);
 
-	void unbindDeferredShader(LLGLSLShader& shader);
+	void unbindDeferredShader(LLGLSLShader& shader, LLRenderTarget* diffuse_source = NULL, LLRenderTarget* light_source = NULL);
 	void renderDeferredLighting();
 	void renderDeferredLightingToRT(LLRenderTarget* target);