rendering optimizations; work towards making the transition to VBOs eventually

Aquaria/DSQ.cpp

@@ -3512,10 +3512,9 @@ bool DSQ::loadTileset(std::string pack, const unsigned char *usedIdx, size_t use
 		const float cell = 64.0f/512.0f;
 		for (int i = 0; i < 27; i++)
 		{
-			ElementTemplate t;
+			ElementTemplate *et = new ElementTemplate;
-			t.idx = 1024+i;
+			et->idx = 1024+i;
-			t.tex = aqtex;
+			et->tex = aqtex;
-			t.loaded = true;
 			int x = i,y=0;
 			while (x >= 6)
 			{
@@ -3523,19 +3522,20 @@ bool DSQ::loadTileset(std::string pack, const unsigned char *usedIdx, size_t use
 				y++;
 			}
 
-			t.tu1 = x*cell;
+			et->tu1 = x*cell;
-			t.tv1 = y*cell;
+			et->tv1 = y*cell;
-			t.tu2 = t.tu1 + cell;
+			et->tu2 = et->tu1 + cell;
-			t.tv2 = t.tv1 + cell;
+			et->tv2 = et->tv1 + cell;
 
-			t.tv2 = 1 - t.tv2;
+			et->tv2 = 1 - et->tv2;
-			t.tv1 = 1 - t.tv1;
+			et->tv1 = 1 - et->tv1;
-			std::swap(t.tv1,t.tv2);
+			std::swap(et->tv1,et->tv2);
 
-			t.w = 512*cell;
+			et->w = 512*cell;
-			t.h = 512*cell;
+			et->h = 512*cell;
 
-			tilemgr.tileset.elementTemplates.push_back(t);
+			et->finalize();
+			tilemgr.tileset.elementTemplates.push_back(et);
 		}
 	}
 

Aquaria/SceneEditor.cpp

@@ -703,7 +703,7 @@ void SceneEditor::init()
 
 	if (curElement < dsq->tilemgr.tileset.elementTemplates.size())
 	{
-		placer->setTexture(dsq->tilemgr.tileset.elementTemplates[curElement].gfx);
+		placer->setTexture(dsq->tilemgr.tileset.elementTemplates[curElement]->gfx);
 		placer->scale = Vector(1,1);
 	}
 	else
@@ -888,7 +888,7 @@ void SceneEditor::editModeElements()
 	editType = ET_ELEMENTS;
 	if (curElement < dsq->tilemgr.tileset.elementTemplates.size())
 	{
-		placer->setTexture(dsq->tilemgr.tileset.elementTemplates[curElement].gfx);
+		placer->setTexture(dsq->tilemgr.tileset.elementTemplates[curElement]->gfx);
 		placer->scale = Vector(1,1);
 	}
 	placer->alpha = 0.5;
@@ -2002,7 +2002,6 @@ void SceneEditor::cycleSelectedTiles(int direction)
 	const int maxn = (int)dsq->tilemgr.tileset.elementTemplates.size();
 	if(!maxn)
 		return;
-	const ElementTemplate * const base = &dsq->tilemgr.tileset.elementTemplates[0];
 	for(size_t i = 0; i < n; ++i)
 	{
 		TileData& t = ts.tiles[selectedTiles[i]];
@@ -2093,9 +2092,9 @@ void SceneEditor::cyclePlacer(int direction)
 	if(nextidx >= maxn)
 		nextidx -= maxn;
 
-	if (maxn && dsq->tilemgr.tileset.elementTemplates[curElement].idx < 1024)
+	if (maxn && dsq->tilemgr.tileset.elementTemplates[curElement]->idx < 1024)
 	{
-		placer->setTexture(dsq->tilemgr.tileset.elementTemplates[curElement].gfx);
+		placer->setTexture(dsq->tilemgr.tileset.elementTemplates[curElement]->gfx);
 		curElement = nextidx;
 	}
 }
@@ -2108,7 +2107,7 @@ void SceneEditor::selectZero()
 	{
 		if (dsq->tilemgr.tileset.elementTemplates.empty()) return;
 		curElement = 0;
-		placer->setTexture(dsq->tilemgr.tileset.elementTemplates[curElement].gfx);
+		placer->setTexture(dsq->tilemgr.tileset.elementTemplates[curElement]->gfx);
 	}
 }
 
@@ -2122,14 +2121,14 @@ void SceneEditor::selectEnd()
 		size_t largest = 0;
 		for (size_t i = 0; i < dsq->tilemgr.tileset.elementTemplates.size(); i++)
 		{
-			ElementTemplate et = dsq->tilemgr.tileset.elementTemplates[i];
+			ElementTemplate *et = dsq->tilemgr.tileset.elementTemplates[i];
-			if (et.idx < 1024 && i > largest)
+			if (et->idx < 1024 && i > largest)
 			{
 				largest = i;
 			}
 		}
 		curElement = largest;
-		placer->setTexture(dsq->tilemgr.tileset.elementTemplates[curElement].gfx);
+		placer->setTexture(dsq->tilemgr.tileset.elementTemplates[curElement]->gfx);
 	}
 }
 
@@ -2139,7 +2138,7 @@ void SceneEditor::placeElement()
 	{
 		if (!core->getShiftState() && !core->getKeyState(KEY_LALT))
 		{
-			unsigned tilesetID = dsq->tilemgr.tileset.elementTemplates[curElement].idx;
+			unsigned tilesetID = dsq->tilemgr.tileset.elementTemplates[curElement]->idx;
 			dsq->tilemgr.createOneTile(tilesetID, bgLayer, placer->position.x, placer->position.y);
 			// FIXME: need to update grid or no?
 			updateText();

Aquaria/TileMgr.cpp

@@ -187,8 +187,8 @@ TileData* TileMgr::_createTile(unsigned tilesetID, unsigned layer,  float x, flo
 	t.texscaleY = 1;
 	t.scalex = 1;
 	t.scaley = 1;
-	t.beforeScaleOffsetX = 0;
+	//t.beforeScaleOffsetX = 0;
-	t.beforeScaleOffsetY = 0;
+	//t.beforeScaleOffsetY = 0;
 	t.flags = GetTileFlags(ef);
 	t.tag = 0;
 	t.et = tileset.getByIdx(tilesetID);

BBGE/OpenGLStubs.h

@@ -70,6 +70,7 @@ GL_FUNC(void,glCallList,(GLuint list),(list),)
 GL_FUNC(void,glClearDepth,(GLclampd x),(x),)
 GL_FUNC(void,glColor3f,(GLfloat r,GLfloat g,GLfloat b),(r,g,b),)
 GL_FUNC(void,glDeleteLists,(GLuint list, GLsizei range),(list,range),)
+GL_FUNC(void,glDrawArrays,(GLenum mode, GLint first, GLsizei count), (mode,first,count),)
 GL_FUNC(void,glDrawPixels,(GLsizei width, GLsizei height, GLenum format, GLenum type, const GLvoid *pixels),(width,height,format,type,pixels),)
 GL_FUNC(GLuint,glGenLists,(GLsizei range),(range),return)
 GL_FUNC(void,glNewList,(GLuint list, GLenum mode),(list,mode),)

BBGE/Tile.h

@@ -134,7 +134,7 @@ private:
 struct TileData
 {
 	float x, y, scalex, scaley, texscaleX, texscaleY;
-	float beforeScaleOffsetX, beforeScaleOffsetY; // almost always 0. // TODO: this is nasty, ideally get rid of this
+	//float beforeScaleOffsetX, beforeScaleOffsetY; // almost always 0. // TODO: this is nasty, ideally get rid of this
 	float rotation;
 	unsigned flags; // TileFlags
 	unsigned tag; // FIXME: make this int

BBGE/TileRender.cpp

@@ -41,9 +41,26 @@ static inline const Vector& getTagColor(int tag)
 
 }
 
+static const float s_quadVerts[] =
+{
+	-0.5f, +0.5f,
+	+0.5f, +0.5f,
+	+0.5f, -0.5f,
+	-0.5f, -0.5f,
+};
 
 void TileRender::onRender(const RenderState& rs) const
 {
+	if(storage.tiles.empty())
+		return;
+
+	glPushClientAttrib(GL_CLIENT_ALL_ATTRIB_BITS);
+	glEnableClientState(GL_TEXTURE_COORD_ARRAY);
+	glEnableClientState(GL_VERTEX_ARRAY);
+	glVertexPointer(2, GL_FLOAT, 0, s_quadVerts);
+
+	RenderState rx(rs);
+
 	// prepare. get parallax scroll factors
 	const RenderObjectLayer& rl = core->renderObjectLayers[this->layer];
 	const Vector M = rl.followCameraMult; // affected by parallaxLock
@@ -58,8 +75,10 @@ void TileRender::onRender(const RenderState& rs) const
 	unsigned lastTexRepeat = false;
 	unsigned lastTexId = 0;
 
-	const bool renderBorders = this->renderBorders;
+	const bool renderExtras = renderBorders || RenderObject::renderCollisionShape;
-	//bool mustSetColor = false;
+	const TileEffectData *prevEff = ((TileEffectData*)NULL)+1; // initial value is different from anything else
+	const RenderGrid *grid = NULL;
+	const float *lastTexcoordBuf = NULL;
 
 	for(size_t i = 0; i < storage.tiles.size(); ++i)
 	{
@@ -75,13 +94,12 @@ void TileRender::onRender(const RenderState& rs) const
 		}
 
 		const ElementTemplate * const et = tile.et;
+		const float sw = et->w * tile.scalex;
+		const float sh = et->h * tile.scaley;
 
 		// adapted from RenderObject::isOnScreen()
 		{
-			const float cw = et->w * tile.scalex;
+			const float cullRadiusSqr = ((sw*sw + sh*sh) * core->invGlobalScaleSqr) + core->cullRadiusSqr;
-			const float ch = et->h * tile.scaley;
-			const float cullRadiusSqr = ((cw*cw + ch*ch) * core->invGlobalScaleSqr) + core->cullRadiusSqr;
-
 			if ((pos - core->cullCenter).getSquaredLength2D() >= cullRadiusSqr)
 				continue;
 		}
@@ -107,104 +125,92 @@ void TileRender::onRender(const RenderState& rs) const
 		glTranslatef(pos.x, pos.y, pos.z);
 
 		glRotatef(tile.rotation, 0, 0, 1);
-		if(tile.flags & TILEFLAG_FH) // TODO: This is not necessary! Since we have no children, flipped texcoords are fine
+		if(tile.flags & TILEFLAG_FH)
 			glRotatef(180, 0, 1, 0);
 
 		// this is only relevant in editor mode and is always 0 otherwise
-		glTranslatef(tile.beforeScaleOffsetX, tile.beforeScaleOffsetY, 0);
+		//glTranslatef(tile.beforeScaleOffsetX, tile.beforeScaleOffsetY, 0);
 
-		glScalef(tile.scalex, tile.scaley, 1);
+		glScalef(sw, sh, 1);
-		//glScalef(tile.scalex * et->w, tile.scaley * et->h, 1); // TODO use this + fixed verts
 
-		BlendType blend = BLEND_DEFAULT;
 		float alpha = rs.alpha;
-		RenderGrid *grid = NULL;
 		const TileEffectData * const eff = tile.eff;
-		if(eff)
+		if(eff != prevEff) // effects between tiles are often shared so this works not only for NULL
 		{
-			grid = eff->grid;
+			prevEff = eff;
-			alpha *= eff->alpha.x;
+			BlendType blend = BLEND_DEFAULT;
-			blend = eff->blend;
+			alpha = rs.alpha;
+			grid = NULL;
+
+			if(eff)
+			{
+				grid = eff->grid;
+				alpha *= eff->alpha.x;
+				blend = eff->blend;
+			}
+
+			rs.gpu.setBlend(blend);
+			glColor4f(rs.color.x, rs.color.y, rs.color.z, alpha);
 		}
 
-		rs.gpu.setBlend(blend);
-
-		// TODO: only need to do this when prev. tile had different alpha
-		glColor4f(rs.color.x, rs.color.y, rs.color.z, alpha);
-
-		const Vector upperLeftTextureCoordinates(et->tu1, et->tv1);
-		const Vector lowerRightTextureCoordinates(et->tu2, et->tv2);
-
 		if(!grid)
 		{
-			const float _w2 = et->w * 0.5f;
+			const float *tcbuf = tile.et->texcoordQuadPtr;
-			const float _h2 = et->h * 0.5f;
+			assert(tcbuf);
-
+			if(lastTexcoordBuf != tcbuf)
-			glBegin(GL_QUADS);
 			{
-				glTexCoord2f(upperLeftTextureCoordinates.x, 1.0f-upperLeftTextureCoordinates.y);
+				lastTexcoordBuf = tcbuf;
-				glVertex2f(-_w2, +_h2);
+				glTexCoordPointer(2, GL_FLOAT, 0, tcbuf);
-
-				glTexCoord2f(lowerRightTextureCoordinates.x, 1.0f-upperLeftTextureCoordinates.y);
-				glVertex2f(+_w2, +_h2);
-
-				glTexCoord2f(lowerRightTextureCoordinates.x, 1.0f-lowerRightTextureCoordinates.y);
-				glVertex2f(+_w2, -_h2);
-
-				glTexCoord2f(upperLeftTextureCoordinates.x, 1.0f-lowerRightTextureCoordinates.y);
-				glVertex2f(-_w2, -_h2);
 			}
-			glEnd();
+			glDrawArrays(GL_QUADS, 0, 4);
 		}
 		else
 		{
-			glPushMatrix();
-			glScalef(et->w, et->h, 1);
-
-			RenderState rx(rs);
 			rx.alpha = alpha;
+			const Vector upperLeftTextureCoordinates(et->tu1, et->tv1);
+			const Vector lowerRightTextureCoordinates(et->tu2, et->tv2);
 			grid->render(rx, upperLeftTextureCoordinates, lowerRightTextureCoordinates);
-			if (RenderObject::renderCollisionShape)
+		}
+
+		if(renderExtras)
+		{
+			glBindTexture(GL_TEXTURE_2D, 0);
+			lastTexId = 0;
+			prevEff = ((TileEffectData*)NULL)+1;
+
+			if(grid && RenderObject::renderCollisionShape)
 			{
-				glBindTexture(GL_TEXTURE_2D, 0);
 				grid->renderDebugPoints(rs);
-				lastTexId = 0;
 			}
 
-			glPopMatrix();
+			if(renderBorders)
+			{
+				float c = (tile.flags & TILEFLAG_SELECTED) ? 1.0f : 0.5f;
+				Vector color(c,c,c);
+				color *= getTagColor(tile.tag);
+
+				glColor4f(color.x, color.y, color.z, 1.0f);
+				glPointSize(16);
+				glBegin(GL_POINTS);
+					glVertex2f(0,0);
+				glEnd();
+
+				glLineWidth(2);
+				glBegin(GL_LINE_STRIP);
+					glVertex2f(0.5f, 0.5f);
+					glVertex2f(0.5f, -0.5f);
+					glVertex2f(-0.5f, -0.5f);
+					glVertex2f(-0.5f, 0.5f);
+					glVertex2f(0.5f, 0.5f);
+				glEnd();
+			}
 		}
 
-		if(renderBorders)
-		{
-			lastTexId = 0;
-			glBindTexture(GL_TEXTURE_2D, 0);
-
-			float c = (tile.flags & TILEFLAG_SELECTED) ? 1.0f : 0.5f;
-			Vector color(c,c,c);
-			color *= getTagColor(tile.tag);
-			const float _w2 = et->w * 0.5f;
-			const float _h2 = et->h * 0.5f;
-
-			glColor4f(color.x, color.y, color.z, 1.0f);
-			glPointSize(16);
-			glBegin(GL_POINTS);
-				glVertex2f(0,0);
-			glEnd();
-
-			glLineWidth(2);
-			glBegin(GL_LINE_STRIP);
-				glVertex2f(_w2, _h2);
-				glVertex2f(_w2, -_h2);
-				glVertex2f(-_w2, -_h2);
-				glVertex2f(-_w2, _h2);
-				glVertex2f(_w2, _h2);
-			glEnd();
-		}
-
-
 		glPopMatrix();
 	}
 
+	glPopClientAttrib();
+
 	RenderObject::lastTextureApplied = lastTexId;
 	RenderObject::lastTextureRepeat = !!lastTexRepeat;
 }

BBGE/Tileset.cpp

@@ -16,7 +16,7 @@ Tileset::~Tileset()
 
 bool Tileset::loadFile(const char *fn, const unsigned char *usedIdx, size_t usedIdxLen)
 {
-	elementTemplates.clear();
+	clear();
 
 	InStream in(fn);
 	if(!in)
@@ -34,12 +34,12 @@ bool Tileset::loadFile(const char *fn, const unsigned char *usedIdx, size_t used
 		{
 			if(idx < 1024)
 			{
-				ElementTemplate t;
+				ElementTemplate *et = new ElementTemplate;
-				t.idx = idx;
+				et->idx = idx;
-				t.gfx = gfx;
+				et->gfx = gfx;
-				t.w = w;
+				et->w = w;
-				t.h = h;
+				et->h = h;
-				elementTemplates.push_back(t);
+				elementTemplates.push_back(et);
 			}
 			else
 				warn = true;
@@ -59,9 +59,9 @@ bool Tileset::loadFile(const char *fn, const unsigned char *usedIdx, size_t used
 
 	for (size_t i = 0; i < elementTemplates.size(); i++)
 	{
-		size_t idx = elementTemplates[i].idx;
+		size_t idx = elementTemplates[i]->idx;
 		if (!usedIdx || (idx < usedIdxLen && usedIdx[idx]))
-			usedTex.push_back(elementTemplates[i].gfx);
+			usedTex.push_back(elementTemplates[i]->gfx);
 	}
 	std::sort(usedTex.begin(), usedTex.end());
 	// drop duplicates
@@ -90,15 +90,16 @@ bool Tileset::loadFile(const char *fn, const unsigned char *usedIdx, size_t used
 	std::ostringstream failed;
 	for (size_t i = 0; i < elementTemplates.size(); i++)
 	{
-		ElementTemplate& et = elementTemplates[i];
+		ElementTemplate *et = elementTemplates[i];
 		// only check those that are actualy loaded; otherwise this would load in textures
 		// that we didn't bother to batch-load above
-		if(!usedIdx || (et.idx < usedIdxLen && usedIdx[et.idx]))
+		if(!usedIdx || (et->idx < usedIdxLen && usedIdx[et->idx]))
 		{
-			if(!et.getTexture()) // assigns width/height and caches texture pointer
+			et->finalize(); // assigns width/height and caches texture pointer
+			if(!et->tex)
 			{
 				++nfailed;
-				failed << et.gfx << " ";
+				failed << et->gfx << " ";
 			}
 		}
 	}
@@ -119,6 +120,9 @@ void Tileset::clear()
 	for(size_t i = 0; i < dummies.size(); ++i)
 		delete dummies[i];
 	dummies.clear();
+
+	for(size_t i = 0; i < elementTemplates.size(); ++i)
+		delete elementTemplates[i];
 	elementTemplates.clear();
 }
 
@@ -126,11 +130,11 @@ const ElementTemplate *Tileset::getByIdx(size_t idx)
 {
 	for (size_t i = 0; i < elementTemplates.size(); i++)
 	{
-		ElementTemplate& et = elementTemplates[i];
+		ElementTemplate *et = elementTemplates[i];
-		if (et.idx == idx)
+		if (et->idx == idx)
 		{
-			et.getTexture(); // HACK: make sure the texture is loaded before this gets used
+			et->finalize(); // HACK: make sure the texture is loaded before this gets used
-			return &et;
+			return et;
 		}
 	}
 
@@ -151,6 +155,7 @@ const ElementTemplate *Tileset::getByIdx(size_t idx)
 
 	ElementTemplate *dummy = new ElementTemplate;
 	dummy->idx = idx;
+	dummy->finalize();
 	dummies.push_back(dummy);
 
 	return dummy;
@@ -160,17 +165,22 @@ const ElementTemplate* Tileset::getAdjacent(size_t idx, int direction, bool wrap
 {
 	ElementTemplate *et = _getAdjacent(idx, direction, wraparound);
 	if(et)
-		et->getTexture(); // load just in case
+		et->finalize(); // load just in case
 	return et;
 }
 
-Texture* ElementTemplate::getTexture()
+static const float s_defaultTexcoordBuf[] =
 {
-	if(loaded)
+	0, 1,
-		return tex.content();
+	1, 1,
+	1, 0,
+	0, 0
+};
 
-	loaded = true;
+void ElementTemplate::finalize()
-	tex = core->getTexture(gfx); // may end up NULL
+{
+	if(!gfx.empty())
+		tex = core->getTexture(gfx); // may end up NULL
 	if(tex)
 	{
 		if(!w)
@@ -186,8 +196,20 @@ Texture* ElementTemplate::getTexture()
 			h = 64;
 	}
 
-	return tex.content();
+	if(tu1 == 0 && tv1 == 0 && tu2 == 1 && tv2 == 1)
-
+		texcoordQuadPtr = s_defaultTexcoordBuf;
+	else
+	{
+		texcoordQuadPtr = &texcoordQuadBuffer[0];
+		texcoordQuadBuffer[0] =      tu1;
+		texcoordQuadBuffer[1] = 1.0f-tv1;
+		texcoordQuadBuffer[2] =      tu2;
+		texcoordQuadBuffer[3] = 1.0f-tv1;
+		texcoordQuadBuffer[4] =      tu2;
+		texcoordQuadBuffer[5] = 1.0f-tv2;
+		texcoordQuadBuffer[6] =      tu1;
+		texcoordQuadBuffer[7] = 1.0f-tv2;
+	}
 }
 
 ElementTemplate * Tileset::_getAdjacent(size_t idx, int direction, bool wraparound)
@@ -198,21 +220,21 @@ ElementTemplate * Tileset::_getAdjacent(size_t idx, int direction, bool wraparou
 	int mindiff = 0;
 	for (size_t i = 0; i < maxn; i++)
 	{
-		if (elementTemplates[i].idx == idx)
+		if (elementTemplates[i]->idx == idx)
 		{
 			if(wraparound)
 			{
 				if(!i && direction < 0)
-					return &elementTemplates.back();
+					return elementTemplates.back();
 				if(i + direction >= maxn)
-					return &elementTemplates[0];
+					return elementTemplates[0];
 			}
 			else
 
 			i += direction; // may underflow
-			return i < maxn ? &elementTemplates[i] : NULL;
+			return i < maxn ? elementTemplates[i] : NULL;
 		}
-		int diff = labs((int)elementTemplates[i].idx - (int)idx);
+		int diff = labs((int)elementTemplates[i]->idx - (int)idx);
 		if(diff < mindiff || !mindiff)
 		{
 			mindiff = diff;
@@ -230,11 +252,11 @@ ElementTemplate * Tileset::_getAdjacent(size_t idx, int direction, bool wraparou
 	else if(wraparound)
 	{
 		if(!closest && direction < 0)
-			return &elementTemplates.back();
+			return elementTemplates.back();
 		if(closest + direction >= maxn)
-			return &elementTemplates[0];
+			return elementTemplates[0];
 	}
 
 	size_t i = closest + direction;
-	return i < maxn ? &elementTemplates[i] : NULL;
+	return i < maxn ? elementTemplates[i] : NULL;
 }

BBGE/Tileset.h

@@ -8,21 +8,24 @@
 class ElementTemplate
 {
 public:
-	ElementTemplate() { w=0; h=0; idx=-1; tu1=tv1=0; tu2=tv2=1; loaded=false; }
+	ElementTemplate() { w=0; h=0; idx=-1; tu1=tv1=0; tu2=tv2=1; texcoordQuadPtr=NULL; }
 	inline bool operator<(const ElementTemplate& o) const { return idx < o.idx; }
 
-	Texture *getTexture(); // loads if not already loaded
+	void finalize(); // call after settings params
 
 	// lazily assigned when tex is loaded
 	CountedPtr<Texture> tex; // NULL if failed to load or not yet loaded
 	float w,h; // custom size if used, otherwise texture size
+	const float *texcoordQuadPtr;
+	float texcoordQuadBuffer[8];
 
 	// fixed
 	float tu1, tu2, tv1, tv2; // texcoords
 	size_t idx;
 	std::string gfx;
 
-	bool loaded;
+private:
+	ElementTemplate(const ElementTemplate&); // no copy
 };
 
 class Tileset
@@ -43,7 +46,7 @@ public:
 	// never returns dummy ET. May return NULL.
 	const ElementTemplate *getAdjacent(size_t idx, int direction, bool wraparound);
 
-	std::vector<ElementTemplate> elementTemplates;
+	std::vector<ElementTemplate*> elementTemplates;
 
 private:
 	ElementTemplate *_getAdjacent(size_t idx, int direction, bool wraparound);