Doer fonts fast(er)! (better batching for font glyph rendering)

indra/llrender/llfontgl.cpp

@@ -268,9 +268,15 @@ S32 LLFontGL::render(const LLWString &wstr, S32 begin_offset, F32 x, F32 y, cons
 
 	const LLFontGlyphInfo* next_glyph = NULL;
 
-	// Remember last-used texture to avoid unnecesssary bind calls.
+	const S32 GLYPH_BATCH_SIZE = 30;
-	LLImageGL *last_bound_texture = NULL;
+	static LL_ALIGN_16(LLVector4a vertices[GLYPH_BATCH_SIZE * 4]);
+	static LLVector2 uvs[GLYPH_BATCH_SIZE * 4];
+	static 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];
@@ -299,11 +305,7 @@ S32 LLFontGL::render(const LLWString &wstr, S32 begin_offset, F32 x, F32 y, cons
 				break;
 			}
 
-			if (last_bound_texture != ext_image)
+			gGL.getTexUnit(0)->bind(ext_image);
-			{
-				gGL.getTexUnit(0)->bind(ext_image);
-				last_bound_texture = ext_image;
-			}
 
 			// snap origin to whole screen pixel
 			const F32 ext_x = (F32)llround(cur_render_x + (EXT_X_BEARING * sScaleX));
@@ -311,7 +313,23 @@ S32 LLFontGL::render(const LLWString &wstr, S32 begin_offset, F32 x, F32 y, cons
 
 			LLRectf uv_rect(0.f, 1.f, 1.f, 0.f);
 			LLRectf screen_rect(ext_x, ext_y + ext_height, ext_x + ext_width, ext_y);
-			drawGlyph(screen_rect, uv_rect, LLColor4::white, style, shadow, drop_shadow_strength);
+
+			if (glyph_count > 0)
+			{
+				gGL.begin(LLRender::QUADS);
+				{
+					gGL.vertexBatchPreTransformed(vertices, uvs, colors, glyph_count * 4);
+				}
+				gGL.end();
+				glyph_count = 0;
+			}
+			renderQuad(vertices, uvs, colors, screen_rect, uv_rect, LLColor4U::white, 0);
+			//No batching here. It will never happen.
+			gGL.begin(LLRender::QUADS);
+			{
+				gGL.vertexBatchPreTransformed(vertices, uvs, colors, 4);
+			}
+			gGL.end();
 
 			if (!label.empty())
 			{
@@ -320,13 +338,11 @@ S32 LLFontGL::render(const LLWString &wstr, S32 begin_offset, F32 x, F32 y, cons
 									 /*llfloor*/(ext_x / sScaleX) + ext_image->getWidth() + EXT_X_BEARING - sCurOrigin.mX, 
 									 /*llfloor*/(cur_render_y / sScaleY) - sCurOrigin.mY,
 									 color,
-									 halign, BASELINE, NORMAL, NO_SHADOW, S32_MAX, S32_MAX, NULL,
+									 halign, BASELINE, UNDERLINE, NO_SHADOW, S32_MAX, S32_MAX, NULL,
 									 TRUE );
 				gGL.popMatrix();
 			}
 
-			gGL.color4fv(color.mV);
-
 			chars_drawn++;
 			cur_x += ext_advance;
 			if (((i + 1) < length) && wstr[i+1])
@@ -349,11 +365,24 @@ S32 LLFontGL::render(const LLWString &wstr, S32 begin_offset, F32 x, F32 y, cons
 				break;
 			}
 			// Per-glyph bitmap texture.
-			LLImageGL *image_gl = font_bitmap_cache->getImageGL(fgi->mBitmapNum);
+			S32 next_bitmap_num = fgi->mBitmapNum;
-			if (last_bound_texture != image_gl)
+			if (next_bitmap_num != bitmap_num)
 			{
-				gGL.getTexUnit(0)->bind(image_gl);
+				// Actually draw the queued glyphs before switching their texture;
-				last_bound_texture = image_gl;
+				// otherwise the queued glyphs will be taken from wrong textures.
+				if (glyph_count > 0)
+				{
+					gGL.begin(LLRender::QUADS);
+					{
+						gGL.vertexBatchPreTransformed(vertices, uvs, colors, glyph_count * 4);
+					}
+					gGL.end();
+					glyph_count = 0;
+				}
+
+				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))
@@ -368,13 +397,24 @@ S32 LLFontGL::render(const LLWString &wstr, S32 begin_offset, F32 x, F32 y, cons
 					(fgi->mYBitmapOffset + fgi->mHeight + PAD_UVY) * inv_height,
 					(fgi->mXBitmapOffset + fgi->mWidth) * inv_width,
 				(fgi->mYBitmapOffset - PAD_UVY) * inv_height);
-		// snap glyph origin to whole screen pixel
+			// snap glyph origin to whole screen pixel
-		LLRectf screen_rect((F32)llround(cur_render_x + (F32)fgi->mXBearing),
+			LLRectf screen_rect((F32)llround(cur_render_x + (F32)fgi->mXBearing),
 				    (F32)llround(cur_render_y + (F32)fgi->mYBearing),
 				    (F32)llround(cur_render_x + (F32)fgi->mXBearing) + (F32)fgi->mWidth,
 				    (F32)llround(cur_render_y + (F32)fgi->mYBearing) - (F32)fgi->mHeight);
 			
-			drawGlyph(screen_rect, uv_rect, color, style, 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, shadow, drop_shadow_strength);
 
 			chars_drawn++;
 			cur_x += fgi->mXAdvance;
@@ -400,6 +440,13 @@ S32 LLFontGL::render(const LLWString &wstr, S32 begin_offset, F32 x, F32 y, cons
 		}
 	}
 
+	gGL.begin(LLRender::QUADS);
+	{
+		gGL.vertexBatchPreTransformed(vertices, uvs, colors, glyph_count * 4);
+	}
+	gGL.end();
+
+
 	if (right_x)
 	{
 		*right_x = (cur_x - origin.mV[VX]) / sScaleX;
@@ -1243,95 +1290,95 @@ 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(LLVector4a* 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);
+	S32 index = 0;
-	gGL.vertex2f(llfont_round_x(screen_rect.mRight), 
-				llfont_round_y(screen_rect.mTop));
 
-	gGL.texCoord2f(uv_rect.mLeft, uv_rect.mTop);
+	vertex_out[index].set(screen_rect.mRight, screen_rect.mTop, 0.f);
-	gGL.vertex2f(llfont_round_x(screen_rect.mLeft), 
+	uv_out[index] = LLVector2(uv_rect.mRight, uv_rect.mTop);
-				llfont_round_y(screen_rect.mTop));
+	colors_out[index] = color;
+	index++;
 
-	gGL.texCoord2f(uv_rect.mLeft, uv_rect.mBottom);
+	vertex_out[index].set(screen_rect.mLeft, screen_rect.mTop, 0.f);
-	gGL.vertex2f(llfont_round_x(screen_rect.mLeft + slant_amt), 
+	uv_out[index] = LLVector2(uv_rect.mLeft, uv_rect.mTop);
-				llfont_round_y(screen_rect.mBottom));
+	colors_out[index] = color;
+	index++;
 
-	gGL.texCoord2f(uv_rect.mRight, uv_rect.mBottom);
+	vertex_out[index].set(screen_rect.mLeft, screen_rect.mBottom, 0.f);
-	gGL.vertex2f(llfont_round_x(screen_rect.mRight + slant_amt), 
+	uv_out[index] = LLVector2(uv_rect.mLeft, uv_rect.mBottom);
-				llfont_round_y(screen_rect.mBottom));
+	colors_out[index] = color;
+	index++;
+
+	vertex_out[index].set(screen_rect.mRight, 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, LLVector4a* 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
+		for (S32 pass = 0; pass < 2; pass++)
-		//Allow both when we need them.
-		if (style & BOLD)
 		{
-			gGL.color4fv(color.mV);
+			LLRectf screen_rect_offset = screen_rect;
-			for (S32 pass = 0; pass < 2; pass++)
-			{
-				LLRectf screen_rect_offset = screen_rect;
 
-				screen_rect_offset.translate((F32)(pass * BOLD_OFFSET), 0.f);
+			screen_rect_offset.translate((F32)(pass * BOLD_OFFSET), 0.f);
-				renderQuad(screen_rect_offset, 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, color, slant_offset);
-			}
+			glyph_count++;
 		}
-		else if (shadow == DROP_SHADOW_SOFT)
-		{
-			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;
-
-				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);
-			}
-			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);
-		}
-
 	}
-	gGL.end();
+	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++)
+		{
+			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(&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++;
+	}
 }

indra/llrender/llfontgl.h

@@ -232,8 +232,8 @@ protected:
 	LLFontDescriptor mFontDescriptor;
 	LLPointer<LLFontFreetype> mFontFreetype;
 
-	void renderQuad(const LLRectf& screen_rect, const LLRectf& uv_rect, F32 slant_amt) const;
+	void renderQuad(LLVector4a* 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(const LLRectf& screen_rect, const LLRectf& uv_rect, const LLColor4& color, U8 style, ShadowType shadow, F32 drop_shadow_fade) const;
+	void drawGlyph(S32& glyph_count, LLVector4a* 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;