merge heads. whew.

12 files changed, 445 insertions, 166 deletions

diff --git a/indra/llrender/llfontfreetype.cpp b/indra/llrender/llfontfreetype.cpp
index 163b79813e..b84e696e2d 100644
--- a/indra/llrender/llfontfreetype.cpp
+++ b/indra/llrender/llfontfreetype.cpp
@@ -253,10 +253,10 @@ F32 LLFontFreetype::getXAdvance(llwchar wch) const
 	}
 	else
 	{
-		gi = get_if_there(mCharGlyphInfoMap, (llwchar)0, (LLFontGlyphInfo*)NULL);
-		if (gi)
+		char_glyph_info_map_t::iterator found_it = mCharGlyphInfoMap.find((llwchar)0);
+		if (found_it != mCharGlyphInfoMap.end())
 		{
-			return gi->mXAdvance;
+			return found_it->second->mXAdvance;
 		}
 	}
 
diff --git a/indra/llrender/llfontfreetype.h b/indra/llrender/llfontfreetype.h
index f3f62533da..f1b23f22d5 100644
--- a/indra/llrender/llfontfreetype.h
+++ b/indra/llrender/llfontfreetype.h
@@ -27,7 +27,7 @@
 #ifndef LL_LLFONTFREETYPE_H
 #define LL_LLFONTFREETYPE_H
 
-#include <map>
+#include <boost/unordered_map.hpp>
 #include "llpointer.h"
 #include "llstl.h"
 
@@ -164,7 +164,7 @@ private:
 
 	BOOL mValid;
 
-	typedef std::map<llwchar, LLFontGlyphInfo*> char_glyph_info_map_t;
+	typedef boost::unordered_map<llwchar, LLFontGlyphInfo*> char_glyph_info_map_t;
 	mutable char_glyph_info_map_t mCharGlyphInfoMap; // Information about glyph location in bitmap
 
 	mutable LLPointer<LLFontBitmapCache> mFontBitmapCachep;
diff --git a/indra/llrender/llfontgl.cpp b/indra/llrender/llfontgl.cpp
index a548e618b3..386bb987f9 100644
--- a/indra/llrender/llfontgl.cpp
+++ b/indra/llrender/llfontgl.cpp
@@ -141,6 +141,8 @@ S32 LLFontGL::render(const LLWString &wstr, S32 begin_offset, const LLRect& rect
 S32 LLFontGL::render(const LLWString &wstr, S32 begin_offset, F32 x, F32 y, const LLColor4 &color, HAlign halign, VAlign valign, U8 style, 
 					 ShadowType shadow, S32 max_chars, S32 max_pixels, F32* right_x, BOOL use_ellipses) const
 {
+	LLFastTimer _(FTM_RENDER_FONTS);
+
 	if(!sDisplayFont) //do not display texts
 	{
 		return wstr.length() ;
@@ -175,16 +177,18 @@ S32 LLFontGL::render(const LLWString &wstr, S32 begin_offset, F32 x, F32 y, cons
 
 	gGL.loadUIIdentity();
 	
-	gGL.translateUI(floorf(sCurOrigin.mX*sScaleX), floorf(sCurOrigin.mY*sScaleY), sCurOrigin.mZ);
+	//gGL.translateUI(floorf(sCurOrigin.mX*sScaleX), floorf(sCurOrigin.mY*sScaleY), sCurOrigin.mZ);
 
 	// this code snaps the text origin to a pixel grid to start with
-	F32 pixel_offset_x = llround((F32)sCurOrigin.mX) - (sCurOrigin.mX);
-	F32 pixel_offset_y = llround((F32)sCurOrigin.mY) - (sCurOrigin.mY);
-	gGL.translateUI(-pixel_offset_x, -pixel_offset_y, 0.f);
+	//F32 pixel_offset_x = llround((F32)sCurOrigin.mX) - (sCurOrigin.mX);
+	//F32 pixel_offset_y = llround((F32)sCurOrigin.mY) - (sCurOrigin.mY);
+	//gGL.translateUI(-pixel_offset_x, -pixel_offset_y, 0.f);
 
-	LLFastTimer t(FTM_RENDER_FONTS);
+	LLVector2 origin(floorf(sCurOrigin.mX*sScaleX), floorf(sCurOrigin.mY*sScaleY));
+	// snap the text origin to a pixel grid to start with
+	origin.mV[VX] -= llround((F32)sCurOrigin.mX) - (sCurOrigin.mX);
+	origin.mV[VY] -= llround((F32)sCurOrigin.mY) - (sCurOrigin.mY);
 
-	gGL.color4fv( color.mV );
 
 	S32 chars_drawn = 0;
 	S32 i;
@@ -204,8 +208,8 @@ S32 LLFontGL::render(const LLWString &wstr, S32 begin_offset, F32 x, F32 y, cons
  	// Not guaranteed to be set correctly
 	gGL.setSceneBlendType(LLRender::BT_ALPHA);
 	
-	cur_x = ((F32)x * sScaleX);
-	cur_y = ((F32)y * sScaleY);
+	cur_x = ((F32)x * sScaleX) + origin.mV[VX];
+	cur_y = ((F32)y * sScaleY) + origin.mV[VY];
 
 	// Offset y by vertical alignment.
 	switch (valign)
@@ -270,6 +274,15 @@ S32 LLFontGL::render(const LLWString &wstr, S32 begin_offset, F32 x, F32 y, cons
 
 	const LLFontGlyphInfo* next_glyph = NULL;
 
+	const S32 GLYPH_BATCH_SIZE = 30;
+	LLVector3 vertices[GLYPH_BATCH_SIZE * 4];
+	LLVector2 uvs[GLYPH_BATCH_SIZE * 4];
+	LLColor4U colors[GLYPH_BATCH_SIZE * 4];
+
+	LLColor4U text_color(color);
+
+	S32 bitmap_num = -1;
+	S32 glyph_count = 0;
 	for (i = begin_offset; i < begin_offset + length; i++)
 	{
 		llwchar wch = wstr[i];
@@ -286,8 +299,13 @@ S32 LLFontGL::render(const LLWString &wstr, S32 begin_offset, F32 x, F32 y, cons
 			break;
 		}
 		// Per-glyph bitmap texture.
-		LLImageGL *image_gl = mFontFreetype->getFontBitmapCache()->getImageGL(fgi->mBitmapNum);
-		gGL.getTexUnit(0)->bind(image_gl);
+		S32 next_bitmap_num = fgi->mBitmapNum;
+		if (next_bitmap_num != bitmap_num)
+		{
+			bitmap_num = next_bitmap_num;
+			LLImageGL *font_image = font_bitmap_cache->getImageGL(bitmap_num);
+			gGL.getTexUnit(0)->bind(font_image);
+		}
 	
 		if ((start_x + scaled_max_pixels) < (cur_x + fgi->mXBearing + fgi->mWidth))
 		{
@@ -307,7 +325,18 @@ S32 LLFontGL::render(const LLWString &wstr, S32 begin_offset, F32 x, F32 y, cons
 				    llround(cur_render_x + (F32)fgi->mXBearing) + (F32)fgi->mWidth,
 				    llround(cur_render_y + (F32)fgi->mYBearing) - (F32)fgi->mHeight);
 		
-		drawGlyph(screen_rect, uv_rect, color, style_to_add, shadow, drop_shadow_strength);
+		if (glyph_count >= GLYPH_BATCH_SIZE)
+		{
+			gGL.begin(LLRender::QUADS);
+			{
+				gGL.vertexBatchPreTransformed(vertices, uvs, colors, glyph_count * 4);
+			}
+			gGL.end();
+
+			glyph_count = 0;
+		}
+
+		drawGlyph(glyph_count, vertices, uvs, colors, screen_rect, uv_rect, text_color, style_to_add, shadow, drop_shadow_strength);
 
 		chars_drawn++;
 		cur_x += fgi->mXAdvance;
@@ -332,11 +361,19 @@ S32 LLFontGL::render(const LLWString &wstr, S32 begin_offset, F32 x, F32 y, cons
 		cur_render_y = cur_y;
 	}
 
+	gGL.begin(LLRender::QUADS);
+	{
+		gGL.vertexBatchPreTransformed(vertices, uvs, colors, glyph_count * 4);
+	}
+	gGL.end();
+
+
 	if (right_x)
 	{
-		*right_x = cur_x / sScaleX;
+		*right_x = (cur_x - origin.mV[VX]) / sScaleX;
 	}
 
+	//FIXME: add underline as glyph?
 	if (style_to_add & UNDERLINE)
 	{
 		F32 descender = mFontFreetype->getDescenderHeight();
@@ -356,7 +393,7 @@ S32 LLFontGL::render(const LLWString &wstr, S32 begin_offset, F32 x, F32 y, cons
 		gGL.pushUIMatrix();
 		renderUTF8(std::string("..."), 
 				0,
-				cur_x / sScaleX, (F32)y,
+				(cur_x - origin.mV[VX]) / sScaleX, (F32)y,
 				color,
 				LEFT, valign,
 				style_to_add,
@@ -1085,95 +1122,95 @@ LLFontGL &LLFontGL::operator=(const LLFontGL &source)
 	return *this;
 }
 
-void LLFontGL::renderQuad(const LLRectf& screen_rect, const LLRectf& uv_rect, F32 slant_amt) const
+void LLFontGL::renderQuad(LLVector3* vertex_out, LLVector2* uv_out, LLColor4U* colors_out, const LLRectf& screen_rect, const LLRectf& uv_rect, const LLColor4U& color, F32 slant_amt) const
 {
-	gGL.texCoord2f(uv_rect.mRight, uv_rect.mTop);
-	gGL.vertex2f(llfont_round_x(screen_rect.mRight), 
-				llfont_round_y(screen_rect.mTop));
-
-	gGL.texCoord2f(uv_rect.mLeft, uv_rect.mTop);
-	gGL.vertex2f(llfont_round_x(screen_rect.mLeft), 
-				llfont_round_y(screen_rect.mTop));
-
-	gGL.texCoord2f(uv_rect.mLeft, uv_rect.mBottom);
-	gGL.vertex2f(llfont_round_x(screen_rect.mLeft + slant_amt), 
-				llfont_round_y(screen_rect.mBottom));
-
-	gGL.texCoord2f(uv_rect.mRight, uv_rect.mBottom);
-	gGL.vertex2f(llfont_round_x(screen_rect.mRight + slant_amt), 
-				llfont_round_y(screen_rect.mBottom));
+	S32 index = 0;
+
+	vertex_out[index] = LLVector3(llfont_round_x(screen_rect.mRight), llfont_round_y(screen_rect.mTop), 0.f);
+	uv_out[index] = LLVector2(uv_rect.mRight, uv_rect.mTop);
+	colors_out[index] = color;
+	index++;
+
+	vertex_out[index] = LLVector3(llfont_round_x(screen_rect.mLeft), llfont_round_y(screen_rect.mTop), 0.f);
+	uv_out[index] = LLVector2(uv_rect.mLeft, uv_rect.mTop);
+	colors_out[index] = color;
+	index++;
+
+	vertex_out[index] = LLVector3(llfont_round_x(screen_rect.mLeft), llfont_round_y(screen_rect.mBottom), 0.f);
+	uv_out[index] = LLVector2(uv_rect.mLeft, uv_rect.mBottom);
+	colors_out[index] = color;
+	index++;
+
+	vertex_out[index] = LLVector3(llfont_round_x(screen_rect.mRight), llfont_round_y(screen_rect.mBottom), 0.f);
+	uv_out[index] = LLVector2(uv_rect.mRight, uv_rect.mBottom);
+	colors_out[index] = color;
 }
 
-void LLFontGL::drawGlyph(const LLRectf& screen_rect, const LLRectf& uv_rect, const LLColor4& color, U8 style, ShadowType shadow, F32 drop_shadow_strength) const
+void LLFontGL::drawGlyph(S32& glyph_count, LLVector3* vertex_out, LLVector2* uv_out, LLColor4U* colors_out, const LLRectf& screen_rect, const LLRectf& uv_rect, const LLColor4U& color, U8 style, ShadowType shadow, F32 drop_shadow_strength) const
 {
 	F32 slant_offset;
 	slant_offset = ((style & ITALIC) ? ( -mFontFreetype->getAscenderHeight() * 0.2f) : 0.f);
 
-	gGL.begin(LLRender::QUADS);
+	//FIXME: bold and drop shadow are mutually exclusive only for convenience
+	//Allow both when we need them.
+	if (style & BOLD)
 	{
-		//FIXME: bold and drop shadow are mutually exclusive only for convenience
-		//Allow both when we need them.
-		if (style & BOLD)
+		for (S32 pass = 0; pass < 2; pass++)
 		{
-			gGL.color4fv(color.mV);
-			for (S32 pass = 0; pass < 2; pass++)
-			{
-				LLRectf screen_rect_offset = screen_rect;
+			LLRectf screen_rect_offset = screen_rect;
 
-				screen_rect_offset.translate((F32)(pass * BOLD_OFFSET), 0.f);
-				renderQuad(screen_rect_offset, uv_rect, slant_offset);
-			}
+			screen_rect_offset.translate((F32)(pass * BOLD_OFFSET), 0.f);
+			renderQuad(&vertex_out[glyph_count * 4], &uv_out[glyph_count * 4], &colors_out[glyph_count * 4], screen_rect_offset, uv_rect, color, slant_offset);
+			glyph_count++;
 		}
-		else if (shadow == DROP_SHADOW_SOFT)
+	}
+	else if (shadow == DROP_SHADOW_SOFT)
+	{
+		LLColor4U shadow_color = LLFontGL::sShadowColor;
+		shadow_color.mV[VALPHA] = U8(color.mV[VALPHA] * drop_shadow_strength * DROP_SHADOW_SOFT_STRENGTH);
+		for (S32 pass = 0; pass < 5; pass++)
 		{
-			LLColor4 shadow_color = LLFontGL::sShadowColor;
-			shadow_color.mV[VALPHA] = color.mV[VALPHA] * drop_shadow_strength * DROP_SHADOW_SOFT_STRENGTH;
-			gGL.color4fv(shadow_color.mV);
-			for (S32 pass = 0; pass < 5; pass++)
-			{
-				LLRectf screen_rect_offset = screen_rect;
+			LLRectf screen_rect_offset = screen_rect;
 
-				switch(pass)
-				{
-				case 0:
-					screen_rect_offset.translate(-1.f, -1.f);
-					break;
-				case 1:
-					screen_rect_offset.translate(1.f, -1.f);
-					break;
-				case 2:
-					screen_rect_offset.translate(1.f, 1.f);
-					break;
-				case 3:
-					screen_rect_offset.translate(-1.f, 1.f);
-					break;
-				case 4:
-					screen_rect_offset.translate(0, -2.f);
-					break;
-				}
-			
-				renderQuad(screen_rect_offset, uv_rect, slant_offset);
+			switch(pass)
+			{
+			case 0:
+				screen_rect_offset.translate(-1.f, -1.f);
+				break;
+			case 1:
+				screen_rect_offset.translate(1.f, -1.f);
+				break;
+			case 2:
+				screen_rect_offset.translate(1.f, 1.f);
+				break;
+			case 3:
+				screen_rect_offset.translate(-1.f, 1.f);
+				break;
+			case 4:
+				screen_rect_offset.translate(0, -2.f);
+				break;
 			}
-			gGL.color4fv(color.mV);
-			renderQuad(screen_rect, uv_rect, slant_offset);
-		}
-		else if (shadow == DROP_SHADOW)
-		{
-			LLColor4 shadow_color = LLFontGL::sShadowColor;
-			shadow_color.mV[VALPHA] = color.mV[VALPHA] * drop_shadow_strength;
-			gGL.color4fv(shadow_color.mV);
-			LLRectf screen_rect_shadow = screen_rect;
-			screen_rect_shadow.translate(1.f, -1.f);
-			renderQuad(screen_rect_shadow, uv_rect, slant_offset);
-			gGL.color4fv(color.mV);
-			renderQuad(screen_rect, uv_rect, slant_offset);
-		}
-		else // normal rendering
-		{
-			gGL.color4fv(color.mV);
-			renderQuad(screen_rect, uv_rect, slant_offset);
+		
+			renderQuad(&vertex_out[glyph_count * 4], &uv_out[glyph_count * 4], &colors_out[glyph_count * 4], screen_rect_offset, uv_rect, shadow_color, slant_offset);
+			glyph_count++;
 		}
-
+		renderQuad(&vertex_out[glyph_count * 4], &uv_out[glyph_count * 4], &colors_out[glyph_count * 4], screen_rect, uv_rect, color, slant_offset);
+		glyph_count++;
+	}
+	else if (shadow == DROP_SHADOW)
+	{
+		LLColor4U shadow_color = LLFontGL::sShadowColor;
+		shadow_color.mV[VALPHA] = U8(color.mV[VALPHA] * drop_shadow_strength);
+		LLRectf screen_rect_shadow = screen_rect;
+		screen_rect_shadow.translate(1.f, -1.f);
+		renderQuad(&vertex_out[glyph_count * 4], &uv_out[glyph_count * 4], &colors_out[glyph_count * 4], screen_rect_shadow, uv_rect, shadow_color, slant_offset);
+		glyph_count++;
+		renderQuad(&vertex_out[glyph_count * 4], &uv_out[glyph_count * 4], &colors_out[glyph_count * 4], screen_rect, uv_rect, color, slant_offset);
+		glyph_count++;
+	}
+	else // normal rendering
+	{
+		renderQuad(&vertex_out[glyph_count * 4], &uv_out[glyph_count * 4], &colors_out[glyph_count * 4], screen_rect, uv_rect, color, slant_offset);
+		glyph_count++;
 	}
-	gGL.end();
 }
diff --git a/indra/llrender/llfontgl.h b/indra/llrender/llfontgl.h
index 224ba51609..dc8d848ed2 100644
--- a/indra/llrender/llfontgl.h
+++ b/indra/llrender/llfontgl.h
@@ -211,8 +211,8 @@ private:
 	LLFontDescriptor mFontDescriptor;
 	LLPointer<LLFontFreetype> mFontFreetype;
 
-	void renderQuad(const LLRectf& screen_rect, const LLRectf& uv_rect, F32 slant_amt) const;
-	void drawGlyph(const LLRectf& screen_rect, const LLRectf& uv_rect, const LLColor4& color, U8 style, ShadowType shadow, F32 drop_shadow_fade) const;
+	void renderQuad(LLVector3* vertex_out, LLVector2* uv_out, LLColor4U* colors_out, const LLRectf& screen_rect, const LLRectf& uv_rect, const LLColor4U& color, F32 slant_amt) const;
+	void drawGlyph(S32& glyph_count, LLVector3* vertex_out, LLVector2* uv_out, LLColor4U* colors_out, const LLRectf& screen_rect, const LLRectf& uv_rect, const LLColor4U& color, U8 style, ShadowType shadow, F32 drop_shadow_fade) const;
 
 	// Registry holds all instantiated fonts.
 	static LLFontRegistry* sFontRegistry;
diff --git a/indra/llrender/llgl.cpp b/indra/llrender/llgl.cpp
index ec1bd353cb..c0edd92bc1 100644
--- a/indra/llrender/llgl.cpp
+++ b/indra/llrender/llgl.cpp
@@ -179,6 +179,9 @@ PFNGLRENDERBUFFERSTORAGEMULTISAMPLEEXTPROC glRenderbufferStorageMultisampleEXT =
 // GL_EXT_framebuffer_blit
 PFNGLBLITFRAMEBUFFEREXTPROC glBlitFramebufferEXT = NULL;
 
+// GL_EXT_blend_func_separate
+PFNGLBLENDFUNCSEPARATEEXTPROC glBlendFuncSeparateEXT = NULL;
+
 // GL_ARB_draw_buffers
 PFNGLDRAWBUFFERSARBPROC glDrawBuffersARB = NULL;
 
@@ -318,6 +321,7 @@ LLGLManager::LLGLManager() :
 	mHasCompressedTextures(FALSE),
 	mHasFramebufferObject(FALSE),
 	mHasFramebufferMultisample(FALSE),
+	mHasBlendFuncSeparate(FALSE),
 
 	mHasVertexBufferObject(FALSE),
 	mHasPBuffer(FALSE),
@@ -346,6 +350,8 @@ LLGLManager::LLGLManager() :
 
 	mHasSeparateSpecularColor(FALSE),
 
+	mDebugGPU(FALSE),
+
 	mDriverVersionMajor(1),
 	mDriverVersionMinor(0),
 	mDriverVersionRelease(0),
@@ -513,11 +519,23 @@ bool LLGLManager::initGL()
 		return false;
 	}
 	
+	setToDebugGPU();
 
 	initGLStates();
 	return true;
 }
 
+void LLGLManager::setToDebugGPU()
+{
+	//"MOBILE INTEL(R) 965 EXPRESS CHIP", 
+	if (mGLRenderer.find("INTEL") != std::string::npos && mGLRenderer.find("965") != std::string::npos)
+	{
+		mDebugGPU = TRUE ;
+	}
+
+	return ;
+}
+
 void LLGLManager::getGLInfo(LLSD& info)
 {
 	info["GLInfo"]["GLVendor"] = std::string((const char *)glGetString(GL_VENDOR));
@@ -627,6 +645,11 @@ void LLGLManager::initExtensions()
 #else
 	mHasDrawBuffers = FALSE;
 # endif
+# if GL_EXT_blend_func_separate
+	mHasBlendFuncSeparate = TRUE;
+#else
+	mHasBlendFuncSeparate = FALSE;
+# endif
 	mHasMipMapGeneration = FALSE;
 	mHasSeparateSpecularColor = FALSE;
 	mHasAnisotropic = FALSE;
@@ -653,6 +676,7 @@ void LLGLManager::initExtensions()
 		&& ExtensionExists("GL_EXT_packed_depth_stencil", gGLHExts.mSysExts);
 	mHasFramebufferMultisample = mHasFramebufferObject && ExtensionExists("GL_EXT_framebuffer_multisample", gGLHExts.mSysExts);
 	mHasDrawBuffers = ExtensionExists("GL_ARB_draw_buffers", gGLHExts.mSysExts);
+	mHasBlendFuncSeparate = ExtensionExists("GL_EXT_blend_func_separate", gGLHExts.mSysExts);
 	mHasTextureRectangle = ExtensionExists("GL_ARB_texture_rectangle", gGLHExts.mSysExts);
 #if !LL_DARWIN
 	mHasPointParameters = !mIsATI && ExtensionExists("GL_ARB_point_parameters", gGLHExts.mSysExts);
@@ -676,6 +700,7 @@ void LLGLManager::initExtensions()
 		mHasFramebufferObject = FALSE;
 		mHasFramebufferMultisample = FALSE;
 		mHasDrawBuffers = FALSE;
+		mHasBlendFuncSeparate = FALSE;
 		mHasMipMapGeneration = FALSE;
 		mHasSeparateSpecularColor = FALSE;
 		mHasAnisotropic = FALSE;
@@ -700,6 +725,7 @@ void LLGLManager::initExtensions()
 		mHasShaderObjects = FALSE;
 		mHasVertexShader = FALSE;
 		mHasFragmentShader = FALSE;
+		mHasBlendFuncSeparate = FALSE;
 		LL_WARNS("RenderInit") << "GL extension support forced to SIMPLE level via LL_GL_BASICEXT" << LL_ENDL;
 	}
 	if (getenv("LL_GL_BLACKLIST"))	/* Flawfinder: ignore */
@@ -728,7 +754,8 @@ void LLGLManager::initExtensions()
 		if (strchr(blacklist,'r')) mHasDrawBuffers = FALSE;//S
 		if (strchr(blacklist,'s')) mHasFramebufferMultisample = FALSE;
 		if (strchr(blacklist,'t')) mHasTextureRectangle = FALSE;
-
+		if (strchr(blacklist,'u')) mHasBlendFuncSeparate = FALSE;//S
+		
 	}
 #endif // LL_LINUX || LL_SOLARIS
 	
@@ -776,6 +803,14 @@ void LLGLManager::initExtensions()
 	{
 		LL_INFOS("RenderInit") << "Couldn't initialize GL_ARB_fragment_shader" << LL_ENDL;
 	}
+	if (!mHasBlendFuncSeparate)
+	{
+		LL_INFOS("RenderInit") << "Couldn't initialize GL_EXT_blend_func_separate" << LL_ENDL;
+	}
+	if (!mHasDrawBuffers)
+	{
+		LL_INFOS("RenderInit") << "Couldn't initialize GL_ARB_draw_buffers" << LL_ENDL;
+	}
 
 	// Disable certain things due to known bugs
 	if (mIsIntel && mHasMipMapGeneration)
@@ -846,6 +881,10 @@ void LLGLManager::initExtensions()
 	{
 		glDrawBuffersARB = (PFNGLDRAWBUFFERSARBPROC) GLH_EXT_GET_PROC_ADDRESS("glDrawBuffersARB");
 	}
+	if (mHasBlendFuncSeparate)
+	{
+		glBlendFuncSeparateEXT = (PFNGLBLENDFUNCSEPARATEEXTPROC) GLH_EXT_GET_PROC_ADDRESS("glBlendFuncSeparateEXT");
+	}
 #if (!LL_LINUX && !LL_SOLARIS) || LL_LINUX_NV_GL_HEADERS
 	// This is expected to be a static symbol on Linux GL implementations, except if we use the nvidia headers - bah
 	glDrawRangeElements = (PFNGLDRAWRANGEELEMENTSPROC)GLH_EXT_GET_PROC_ADDRESS("glDrawRangeElements");
@@ -1008,24 +1047,9 @@ void flush_glerror()
 	glGetError();
 }
 
-void assert_glerror()
+void do_assert_glerror()
 {
-	if (!gGLActive)
-	{
-		//llwarns << "GL used while not active!" << llendl;
-
-		if (gDebugSession)
-		{
-			//ll_fail("GL used while not active");
-		}
-	}
-
-	if (gNoRender || !gDebugGL) 
-	{
-		return;
-	}
-	
-	if (!gGLManager.mInited)
+	if (LL_UNLIKELY(!gGLManager.mInited))
 	{
 		LL_ERRS("RenderInit") << "GL not initialized" << LL_ENDL;
 	}
@@ -1033,10 +1057,9 @@ void assert_glerror()
 	GLenum error;
 	error = glGetError();
 	BOOL quit = FALSE;
-	while (error)
+	while (LL_UNLIKELY(error))
 	{
 		quit = TRUE;
-#ifndef LL_LINUX // *FIX: !  This should be an error for linux as well.
 		GLubyte const * gl_error_msg = gluErrorString(error);
 		if (NULL != gl_error_msg)
 		{
@@ -1060,7 +1083,6 @@ void assert_glerror()
 			}
 		}
 		error = glGetError();
-#endif
 	}
 
 	if (quit)
@@ -1076,6 +1098,25 @@ void assert_glerror()
 	}
 }
 
+void assert_glerror()
+{
+	if (!gGLActive)
+	{
+		//llwarns << "GL used while not active!" << llendl;
+
+		if (gDebugSession)
+		{
+			//ll_fail("GL used while not active");
+		}
+	}
+
+	if (!gNoRender && gDebugGL) 
+	{
+		do_assert_glerror();
+	}
+}
+	
+
 void clear_glerror()
 {
 	//  Create or update texture to be used with this data 
@@ -1089,7 +1130,7 @@ void clear_glerror()
 //
 
 // Static members
-std::map<LLGLenum, LLGLboolean> LLGLState::sStateMap;
+boost::unordered_map<LLGLenum, LLGLboolean> LLGLState::sStateMap;
 
 GLboolean LLGLDepthTest::sDepthEnabled = GL_FALSE; // OpenGL default
 GLenum LLGLDepthTest::sDepthFunc = GL_LESS; // OpenGL default
@@ -1137,7 +1178,7 @@ void LLGLState::resetTextureStates()
 void LLGLState::dumpStates() 
 {
 	LL_INFOS("RenderState") << "GL States:" << LL_ENDL;
-	for (std::map<LLGLenum, LLGLboolean>::iterator iter = sStateMap.begin();
+	for (boost::unordered_map<LLGLenum, LLGLboolean>::iterator iter = sStateMap.begin();
 		 iter != sStateMap.end(); ++iter)
 	{
 		LL_INFOS("RenderState") << llformat(" 0x%04x : %s",(S32)iter->first,iter->second?"TRUE":"FALSE") << LL_ENDL;
@@ -1173,7 +1214,7 @@ void LLGLState::checkStates(const std::string& msg)
 		}
 	}
 	
-	for (std::map<LLGLenum, LLGLboolean>::iterator iter = sStateMap.begin();
+	for (boost::unordered_map<LLGLenum, LLGLboolean>::iterator iter = sStateMap.begin();
 		 iter != sStateMap.end(); ++iter)
 	{
 		LLGLenum state = iter->first;
diff --git a/indra/llrender/llgl.h b/indra/llrender/llgl.h
index 46c57c7585..5e8965c06a 100644
--- a/indra/llrender/llgl.h
+++ b/indra/llrender/llgl.h
@@ -30,7 +30,7 @@
 // This file contains various stuff for handling gl extensions and other gl related stuff.
 
 #include <string>
-#include <map>
+#include <boost/unordered_map.hpp>
 #include <list>
 
 #include "llerror.h"
@@ -81,6 +81,7 @@ public:
 	BOOL mHasCompressedTextures;
 	BOOL mHasFramebufferObject;
 	BOOL mHasFramebufferMultisample;
+	BOOL mHasBlendFuncSeparate;
 	
 	// ARB Extensions
 	BOOL mHasVertexBufferObject;
@@ -113,6 +114,9 @@ public:
 
 	// Misc extensions
 	BOOL mHasSeparateSpecularColor;
+
+	//whether this GPU is in the debug list.
+	BOOL mDebugGPU;
 	
 	S32 mDriverVersionMajor;
 	S32 mDriverVersionMinor;
@@ -141,6 +145,7 @@ private:
 	void initExtensions();
 	void initGLStates();
 	void initGLImages();
+	void setToDebugGPU();
 };
 
 extern LLGLManager gGLManager;
@@ -234,7 +239,7 @@ public:
 	static void checkClientArrays(const std::string& msg = "", U32 data_mask = 0x0001);
 	
 protected:
-	static std::map<LLGLenum, LLGLboolean> sStateMap;
+	static boost::unordered_map<LLGLenum, LLGLboolean> sStateMap;
 	
 public:
 	enum { CURRENT_STATE = -2 };
diff --git a/indra/llrender/llglheaders.h b/indra/llrender/llglheaders.h
index 436c861ab4..5a34b46d0c 100644
--- a/indra/llrender/llglheaders.h
+++ b/indra/llrender/llglheaders.h
@@ -210,6 +210,9 @@ extern PFNGLGETCOMPRESSEDTEXIMAGEARBPROC glGetCompressedTexImageARB;
 
 extern PFNGLCOLORTABLEEXTPROC glColorTableEXT;
 
+//GL_EXT_blend_func_separate
+extern PFNGLBLENDFUNCSEPARATEEXTPROC glBlendFuncSeparateEXT;
+
 //GL_EXT_framebuffer_object
 extern PFNGLISRENDERBUFFEREXTPROC glIsRenderbufferEXT;
 extern PFNGLBINDRENDERBUFFEREXTPROC glBindRenderbufferEXT;
@@ -243,7 +246,10 @@ extern PFNGLGENERATEMIPMAPEXTPROC glGenerateMipmapEXT;
 # include "GL/glh_extensions.h"
 # undef __APPLE__
 
-#elif LL_LINUX 
+#elif LL_LINUX
+//----------------------------------------------------------------------------
+// LL_LINUX
+
 //----------------------------------------------------------------------------
 // Linux, MESA headers, but not necessarily assuming MESA runtime.
 // quotes so we get libraries/.../GL/ version
@@ -279,6 +285,7 @@ extern PFNGLGENERATEMIPMAPEXTPROC glGenerateMipmapEXT;
 # define LL_LINUX_NV_GL_HEADERS 0
 #endif // LL_LINUX && defined(WINGDIAPI)
 
+
 #if LL_LINUX_NV_GL_HEADERS
 // Missing functions when using nvidia headers:
 extern PFNGLACTIVETEXTUREARBPROC	glActiveTextureARB;
@@ -439,6 +446,9 @@ extern PFNGLGETATTRIBLOCATIONARBPROC glGetAttribLocationARB;
 extern PFNGLCOMPRESSEDTEXIMAGE2DARBPROC glCompressedTexImage2DARB;
 extern PFNGLGETCOMPRESSEDTEXIMAGEARBPROC glGetCompressedTexImageARB;
 
+//GL_EXT_blend_func_separate
+extern PFNGLBLENDFUNCSEPARATEEXTPROC glBlendFuncSeparateEXT;
+
 //GL_EXT_framebuffer_object
 extern PFNGLISRENDERBUFFEREXTPROC glIsRenderbufferEXT;
 extern PFNGLBINDRENDERBUFFEREXTPROC glBindRenderbufferEXT;
@@ -467,7 +477,10 @@ extern PFNGLBLITFRAMEBUFFEREXTPROC glBlitFramebufferEXT;
 //GL_ARB_draw_buffers
 extern PFNGLDRAWBUFFERSARBPROC glDrawBuffersARB;
 
+
 #elif LL_WINDOWS
+//----------------------------------------------------------------------------
+// LL_WINDOWS
 
 // windows gl headers depend on things like APIENTRY, so include windows.
 #define WIN32_LEAN_AND_MEAN
@@ -635,6 +648,9 @@ extern PFNGLBINDATTRIBLOCATIONARBPROC glBindAttribLocationARB;
 extern PFNGLGETACTIVEATTRIBARBPROC glGetActiveAttribARB;
 extern PFNGLGETATTRIBLOCATIONARBPROC glGetAttribLocationARB;
 
+//GL_EXT_blend_func_separate
+extern PFNGLBLENDFUNCSEPARATEEXTPROC glBlendFuncSeparateEXT;
+
 //GL_EXT_framebuffer_object
 extern PFNGLISRENDERBUFFEREXTPROC glIsRenderbufferEXT;
 extern PFNGLBINDRENDERBUFFEREXTPROC glBindRenderbufferEXT;
@@ -663,6 +679,7 @@ extern PFNGLBLITFRAMEBUFFEREXTPROC glBlitFramebufferEXT;
 //GL_ARB_draw_buffers
 extern PFNGLDRAWBUFFERSARBPROC glDrawBuffersARB;
 
+
 #elif LL_DARWIN
 //----------------------------------------------------------------------------
 // LL_DARWIN
@@ -679,6 +696,9 @@ extern PFNGLDRAWBUFFERSARBPROC glDrawBuffersARB;
 // Note that they also must not be called on 10.3.9.  This should be taken care of by a runtime check for the existence of the GL extension.
 #include <AvailabilityMacros.h>
 
+//GL_EXT_blend_func_separate
+extern void glBlendFuncSeparateEXT(GLenum srcRGB, GLenum dstRGB, GLenum srcAlpha, GLenum dstAlpha) AVAILABLE_MAC_OS_X_VERSION_10_4_AND_LATER;
+
 // GL_EXT_framebuffer_object
 extern GLboolean glIsRenderbufferEXT(GLuint renderbuffer) AVAILABLE_MAC_OS_X_VERSION_10_4_AND_LATER;
 extern void glBindRenderbufferEXT(GLenum target, GLuint renderbuffer) AVAILABLE_MAC_OS_X_VERSION_10_4_AND_LATER;
diff --git a/indra/llrender/llimagegl.cpp b/indra/llrender/llimagegl.cpp
index 7860b32f34..90fcf6ab3f 100644
--- a/indra/llrender/llimagegl.cpp
+++ b/indra/llrender/llimagegl.cpp
@@ -144,12 +144,12 @@ void LLImageGL::checkTexSize(bool forced) const
 			if (gDebugSession)
 			{
 				gFailLog << "wrong texture size and discard level!" << 
-					mWidth << " Height: " << mHeight << " Current Level: " << mCurrentDiscardLevel << std::endl;
+					mWidth << " Height: " << mHeight << " Current Level: " << (S32)mCurrentDiscardLevel << std::endl;
 			}
 			else
 			{
 				llerrs << "wrong texture size and discard level: width: " << 
-					mWidth << " Height: " << mHeight << " Current Level: " << mCurrentDiscardLevel << llendl ;
+					mWidth << " Height: " << mHeight << " Current Level: " << (S32)mCurrentDiscardLevel << llendl ;
 			}
 		}
 
@@ -1051,8 +1051,12 @@ BOOL LLImageGL::setSubImageFromFrameBuffer(S32 fb_x, S32 fb_y, S32 x_pos, S32 y_
 {
 	if (gGL.getTexUnit(0)->bind(this, false, true))
 	{
-		checkTexSize(true) ;
-		llcallstacks << fb_x << " : " << fb_y << " : " << x_pos << " : " << y_pos << " : " << width << " : " << height << llcallstacksendl ;
+		if(gGLManager.mDebugGPU)
+		{
+			llinfos << "Calling glCopyTexSubImage2D(...)" << llendl ;
+			checkTexSize(true) ;
+			llcallstacks << fb_x << " : " << fb_y << " : " << x_pos << " : " << y_pos << " : " << width << " : " << height << llcallstacksendl ;
+		}
 
 		glCopyTexSubImage2D(GL_TEXTURE_2D, 0, fb_x, fb_y, x_pos, y_pos, width, height);
 		mGLTextureCreated = true;
@@ -1648,7 +1652,7 @@ void LLImageGL::calcAlphaChannelOffsetAndStride()
 	}
 }
 
-void LLImageGL::analyzeAlpha(const void* data_in, S32 w, S32 h)
+void LLImageGL::analyzeAlpha(const void* data_in, U32 w, U32 h)
 {
 	if(!mNeedsAlphaAndPickMask)
 	{
@@ -1656,26 +1660,91 @@ void LLImageGL::analyzeAlpha(const void* data_in, S32 w, S32 h)
 	}
 
 	U32 length = w * h;
-	const GLubyte* current = ((const GLubyte*) data_in) + mAlphaOffset ;
+	U32 alphatotal = 0;
 	
-	S32 sample[16];
-	memset(sample, 0, sizeof(S32)*16);
-
-	for (U32 i = 0; i < length; i++)
+	U32 sample[16];
+	memset(sample, 0, sizeof(U32)*16);
+
+	// generate histogram of quantized alpha.
+	// also add-in the histogram of a 2x2 box-sampled version.  The idea is
+	// this will mid-skew the data (and thus increase the chances of not
+	// being used as a mask) from high-frequency alpha maps which
+	// suffer the worst from aliasing when used as alpha masks.
+	if (w >= 2 && h >= 2)
+	{
+		llassert(w%2 == 0);
+		llassert(h%2 == 0);
+		const GLubyte* rowstart = ((const GLubyte*) data_in) + mAlphaOffset;
+		for (U32 y = 0; y < h; y+=2)
+		{
+			const GLubyte* current = rowstart;
+			for (U32 x = 0; x < w; x+=2)
+			{
+				const U32 s1 = current[0];
+				alphatotal += s1;
+				const U32 s2 = current[w * mAlphaStride];
+				alphatotal += s2;
+				current += mAlphaStride;
+				const U32 s3 = current[0];
+				alphatotal += s3;
+				const U32 s4 = current[w * mAlphaStride];
+				alphatotal += s4;
+				current += mAlphaStride;
+
+				++sample[s1/16];
+				++sample[s2/16];
+				++sample[s3/16];
+				++sample[s4/16];
+
+				const U32 asum = (s1+s2+s3+s4);
+				alphatotal += asum;
+				sample[asum/(16*4)] += 4;
+			}
+			
+			rowstart += 2 * w * mAlphaStride;
+		}
+		length *= 2; // we sampled everything twice, essentially
+	}
+	else
 	{
-		++sample[*current/16];
-		current += mAlphaStride ;
+		const GLubyte* current = ((const GLubyte*) data_in) + mAlphaOffset;
+		for (U32 i = 0; i < length; i++)
+		{
+			const U32 s1 = *current;
+			alphatotal += s1;
+			++sample[s1/16];
+			current += mAlphaStride;
+		}
 	}
+	
+	// if more than 1/16th of alpha samples are mid-range, this
+	// shouldn't be treated as a 1-bit mask
 
-	U32 total = 0;
+	// also, if all of the alpha samples are clumped on one half
+	// of the range (but not at an absolute extreme), then consider
+	// this to be an intentional effect and don't treat as a mask.
+
+	U32 midrangetotal = 0;
 	for (U32 i = 4; i < 11; i++)
 	{
-		total += sample[i];
+		midrangetotal += sample[i];
+	}
+	U32 lowerhalftotal = 0;
+	for (U32 i = 0; i < 8; i++)
+	{
+		lowerhalftotal += sample[i];
+	}
+	U32 upperhalftotal = 0;
+	for (U32 i = 8; i < 16; i++)
+	{
+		upperhalftotal += sample[i];
 	}
 
-	if (total > length/16)
+	if (midrangetotal > length/16 || // lots of midrange, or
+	    (lowerhalftotal == length && alphatotal != 0) || // all close to transparent but not all totally transparent, or
+	    (upperhalftotal == length && alphatotal != 255*length)) // all close to opaque but not all totally opaque
 	{
-		mIsMask = FALSE;
+		mIsMask = FALSE; // not suitable for masking
 	}
 	else
 	{
diff --git a/indra/llrender/llimagegl.h b/indra/llrender/llimagegl.h
index 95379298ef..87a835cdcc 100644
--- a/indra/llrender/llimagegl.h
+++ b/indra/llrender/llimagegl.h
@@ -85,7 +85,7 @@ public:
 protected:
 	virtual ~LLImageGL();
 
-	void analyzeAlpha(const void* data_in, S32 w, S32 h);
+	void analyzeAlpha(const void* data_in, U32 w, U32 h);
 	void calcAlphaChannelOffsetAndStride();
 
 public:
diff --git a/indra/llrender/llrender.cpp b/indra/llrender/llrender.cpp
index a829616df2..e1f6964f83 100644
--- a/indra/llrender/llrender.cpp
+++ b/indra/llrender/llrender.cpp
@@ -772,8 +772,10 @@ LLRender::LLRender()
 
 	mCurrAlphaFunc = CF_DEFAULT;
 	mCurrAlphaFuncVal = 0.01f;
-	mCurrBlendSFactor = BF_UNDEF;
-	mCurrBlendDFactor = BF_UNDEF;
+	mCurrBlendColorSFactor = BF_UNDEF;
+	mCurrBlendAlphaSFactor = BF_UNDEF;
+	mCurrBlendColorDFactor = BF_UNDEF;
+	mCurrBlendAlphaDFactor = BF_UNDEF;
 }
 
 LLRender::~LLRender()
@@ -843,9 +845,9 @@ void LLRender::translateUI(F32 x, F32 y, F32 z)
 		llerrs << "Need to push a UI translation frame before offsetting" << llendl;
 	}
 
-	mUIOffset.front().mV[0] += x;
-	mUIOffset.front().mV[1] += y;
-	mUIOffset.front().mV[2] += z;
+	mUIOffset.back().mV[0] += x;
+	mUIOffset.back().mV[1] += y;
+	mUIOffset.back().mV[2] += z;
 }
 
 void LLRender::scaleUI(F32 x, F32 y, F32 z)
@@ -855,27 +857,27 @@ void LLRender::scaleUI(F32 x, F32 y, F32 z)
 		llerrs << "Need to push a UI transformation frame before scaling." << llendl;
 	}
 
-	mUIScale.front().scaleVec(LLVector3(x,y,z));
+	mUIScale.back().scaleVec(LLVector3(x,y,z));
 }
 
 void LLRender::pushUIMatrix()
 {
 	if (mUIOffset.empty())
 	{
-		mUIOffset.push_front(LLVector3(0,0,0));
+		mUIOffset.push_back(LLVector3(0,0,0));
 	}
 	else
 	{
-		mUIOffset.push_front(mUIOffset.front());
+		mUIOffset.push_back(mUIOffset.back());
 	}
 	
 	if (mUIScale.empty())
 	{
-		mUIScale.push_front(LLVector3(1,1,1));
+		mUIScale.push_back(LLVector3(1,1,1));
 	}
 	else
 	{
-		mUIScale.push_front(mUIScale.front());
+		mUIScale.push_back(mUIScale.back());
 	}
 }
 
@@ -885,8 +887,8 @@ void LLRender::popUIMatrix()
 	{
 		llerrs << "UI offset stack blown." << llendl;
 	}
-	mUIOffset.pop_front();
-	mUIScale.pop_front();
+	mUIOffset.pop_back();
+	mUIScale.pop_back();
 }
 
 LLVector3 LLRender::getUITranslation()
@@ -895,7 +897,7 @@ LLVector3 LLRender::getUITranslation()
 	{
 		llerrs << "UI offset stack empty." << llendl;
 	}
-	return mUIOffset.front();
+	return mUIOffset.back();
 }
 
 LLVector3 LLRender::getUIScale()
@@ -904,7 +906,7 @@ LLVector3 LLRender::getUIScale()
 	{
 		llerrs << "UI scale stack empty." << llendl;
 	}
-	return mUIScale.front();
+	return mUIScale.back();
 }
 
 
@@ -914,8 +916,8 @@ void LLRender::loadUIIdentity()
 	{
 		llerrs << "Need to push UI translation frame before clearing offset." << llendl;
 	}
-	mUIOffset.front().setVec(0,0,0);
-	mUIScale.front().setVec(1,1,1);
+	mUIOffset.back().setVec(0,0,0);
+	mUIScale.back().setVec(1,1,1);
 }
 
 void LLRender::setColorMask(bool writeColor, bool writeAlpha)
@@ -989,15 +991,44 @@ void LLRender::blendFunc(eBlendFactor sfactor, eBlendFactor dfactor)
 {
 	llassert(sfactor < BF_UNDEF);
 	llassert(dfactor < BF_UNDEF);
-	if (mCurrBlendSFactor != sfactor || mCurrBlendDFactor != dfactor)
+	if (mCurrBlendColorSFactor != sfactor || mCurrBlendColorDFactor != dfactor ||
+	    mCurrBlendAlphaSFactor != sfactor || mCurrBlendAlphaDFactor != dfactor)
 	{
-		mCurrBlendSFactor = sfactor;
-		mCurrBlendDFactor = dfactor;
+		mCurrBlendColorSFactor = sfactor;
+		mCurrBlendAlphaSFactor = sfactor;
+		mCurrBlendColorDFactor = dfactor;
+		mCurrBlendAlphaDFactor = dfactor;
 		flush();
 		glBlendFunc(sGLBlendFactor[sfactor], sGLBlendFactor[dfactor]);
 	}
 }
 
+void LLRender::blendFunc(eBlendFactor color_sfactor, eBlendFactor color_dfactor,
+			 eBlendFactor alpha_sfactor, eBlendFactor alpha_dfactor)
+{
+	llassert(color_sfactor < BF_UNDEF);
+	llassert(color_dfactor < BF_UNDEF);
+	llassert(alpha_sfactor < BF_UNDEF);
+	llassert(alpha_dfactor < BF_UNDEF);
+	if (!gGLManager.mHasBlendFuncSeparate)
+	{
+		LL_WARNS_ONCE("render") << "no glBlendFuncSeparateEXT(), using color-only blend func" << llendl;
+		blendFunc(color_sfactor, color_dfactor);
+		return;
+	}
+	if (mCurrBlendColorSFactor != color_sfactor || mCurrBlendColorDFactor != color_dfactor ||
+	    mCurrBlendAlphaSFactor != alpha_sfactor || mCurrBlendAlphaDFactor != alpha_dfactor)
+	{
+		mCurrBlendColorSFactor = color_sfactor;
+		mCurrBlendAlphaSFactor = alpha_sfactor;
+		mCurrBlendColorDFactor = color_dfactor;
+		mCurrBlendAlphaDFactor = alpha_dfactor;
+		flush();
+		glBlendFuncSeparateEXT(sGLBlendFactor[color_sfactor], sGLBlendFactor[color_dfactor],
+				       sGLBlendFactor[alpha_sfactor], sGLBlendFactor[alpha_dfactor]);
+	}
+}
+
 LLTexUnit* LLRender::getTexUnit(U32 index)
 {
 	if (index < mTexUnits.size())
@@ -1173,18 +1204,79 @@ void LLRender::vertex3f(const GLfloat& x, const GLfloat& y, const GLfloat& z)
 	}
 	else
 	{
-		LLVector3 vert = (LLVector3(x,y,z)+mUIOffset.front()).scaledVec(mUIScale.front());
+		LLVector3 vert = (LLVector3(x,y,z)+mUIOffset.back()).scaledVec(mUIScale.back());
 		mVerticesp[mCount] = vert;
 	}
 
 	mCount++;
-	if (mCount < 4096)
+	mVerticesp[mCount] = mVerticesp[mCount-1];
+	mColorsp[mCount] = mColorsp[mCount-1];
+	mTexcoordsp[mCount] = mTexcoordsp[mCount-1];
+}
+
+void LLRender::vertexBatchPreTransformed(LLVector3* verts, S32 vert_count)
+{
+	if (mCount + vert_count > 4094)
 	{
-		mVerticesp[mCount] = mVerticesp[mCount-1];
-		mColorsp[mCount] = mColorsp[mCount-1];
+		//	llwarns << "GL immediate mode overflow.  Some geometry not drawn." << llendl;
+		return;
+	}
+
+	for (S32 i = 0; i < vert_count; i++)
+	{
+		mVerticesp[mCount] = verts[i];
+
+		mCount++;
 		mTexcoordsp[mCount] = mTexcoordsp[mCount-1];
+		mColorsp[mCount] = mColorsp[mCount-1];
 	}
+
+	mVerticesp[mCount] = mVerticesp[mCount-1];
 }
+
+void LLRender::vertexBatchPreTransformed(LLVector3* verts, LLVector2* uvs, S32 vert_count)
+{
+	if (mCount + vert_count > 4094)
+	{
+		//	llwarns << "GL immediate mode overflow.  Some geometry not drawn." << llendl;
+		return;
+	}
+
+	for (S32 i = 0; i < vert_count; i++)
+	{
+		mVerticesp[mCount] = verts[i];
+		mTexcoordsp[mCount] = uvs[i];
+
+		mCount++;
+		mColorsp[mCount] = mColorsp[mCount-1];
+	}
+
+	mVerticesp[mCount] = mVerticesp[mCount-1];
+	mTexcoordsp[mCount] = mTexcoordsp[mCount-1];
+}
+
+void LLRender::vertexBatchPreTransformed(LLVector3* verts, LLVector2* uvs, LLColor4U* colors, S32 vert_count)
+{
+	if (mCount + vert_count > 4094)
+	{
+		//	llwarns << "GL immediate mode overflow.  Some geometry not drawn." << llendl;
+		return;
+	}
+
+	for (S32 i = 0; i < vert_count; i++)
+	{
+		mVerticesp[mCount] = verts[i];
+		mTexcoordsp[mCount] = uvs[i];
+		mColorsp[mCount] = colors[i];
+
+		mCount++;
+	}
+
+	mVerticesp[mCount] = mVerticesp[mCount-1];
+	mTexcoordsp[mCount] = mTexcoordsp[mCount-1];
+	mColorsp[mCount] = mColorsp[mCount-1];
+}
+
 void LLRender::vertex2i(const GLint& x, const GLint& y)
 {
 	vertex3f((GLfloat) x, (GLfloat) y, 0);	
diff --git a/indra/llrender/llrender.h b/indra/llrender/llrender.h
index fb3a4d214d..2767aa64a8 100644
--- a/indra/llrender/llrender.h
+++ b/indra/llrender/llrender.h
@@ -311,13 +311,21 @@ public:
 	void color3fv(const GLfloat* c);
 	void color4ubv(const GLubyte* c);
 
+	void vertexBatchPreTransformed(LLVector3* verts, S32 vert_count);
+	void vertexBatchPreTransformed(LLVector3* verts, LLVector2* uvs, S32 vert_count);
+	void vertexBatchPreTransformed(LLVector3* verts, LLVector2* uvs, LLColor4U*, S32 vert_count);
+
 	void setColorMask(bool writeColor, bool writeAlpha);
 	void setColorMask(bool writeColorR, bool writeColorG, bool writeColorB, bool writeAlpha);
 	void setSceneBlendType(eBlendType type);
 
 	void setAlphaRejectSettings(eCompareFunc func, F32 value = 0.01f);
 
+	// applies blend func to both color and alpha
 	void blendFunc(eBlendFactor sfactor, eBlendFactor dfactor);
+	// applies separate blend functions to color and alpha
+	void blendFunc(eBlendFactor color_sfactor, eBlendFactor color_dfactor,
+		       eBlendFactor alpha_sfactor, eBlendFactor alpha_dfactor);
 
 	LLTexUnit* getTexUnit(U32 index);
 
@@ -356,13 +364,15 @@ private:
 	std::vector<LLTexUnit*>		mTexUnits;
 	LLTexUnit*			mDummyTexUnit;
 
-	eBlendFactor mCurrBlendSFactor;
-	eBlendFactor mCurrBlendDFactor;
+	eBlendFactor mCurrBlendColorSFactor;
+	eBlendFactor mCurrBlendColorDFactor;
+	eBlendFactor mCurrBlendAlphaSFactor;
+	eBlendFactor mCurrBlendAlphaDFactor;
 
 	F32				mMaxAnisotropy;
 
-	std::list<LLVector3> mUIOffset;
-	std::list<LLVector3> mUIScale;
+	std::vector<LLVector3> mUIOffset;
+	std::vector<LLVector3> mUIScale;
 
 };
 
diff --git a/indra/llrender/llvertexbuffer.cpp b/indra/llrender/llvertexbuffer.cpp
index 0f41e278b6..de4501dd0f 100644
--- a/indra/llrender/llvertexbuffer.cpp
+++ b/indra/llrender/llvertexbuffer.cpp
@@ -376,6 +376,11 @@ LLVertexBuffer::LLVertexBuffer(U32 typemask, S32 usage) :
 	{
 		mUsage = 0 ; 
 	}
+
+	if (mUsage == GL_STREAM_DRAW_ARB && !sUseStreamDraw)
+	{
+		mUsage = 0;
+	}
 	
 	if (mUsage == GL_STREAM_DRAW_ARB && !sUseStreamDraw)
 	{