Misc GPU stats etc changes

Core/HLE/sceDisplay.cpp

@@ -211,8 +211,9 @@ void hleEnterVblank(u64 userdata, int cyclesLate) {
 			"Draw calls: %i, flushes %i\n"
 			"Cached Draw calls: %i\n"
 			"Num Tracked Vertex Arrays: %i\n"
-			"Vertices Transformed: %i\n"
+			"Vertices Submitted: %i\n"
 			"Cached Vertices Drawn: %i\n"
+			"Uncached Vertices Drawn: %i\n"
 			"FBOs active: %i\n"
 			"Textures active: %i, decoded: %i\n"
 			"Texture invalidations: %i\n"
@@ -230,8 +231,9 @@ void hleEnterVblank(u64 userdata, int cyclesLate) {
 			gpuStats.numFlushes,
 			gpuStats.numCachedDrawCalls,
 			gpuStats.numTrackedVertexArrays,
-			gpuStats.numVertsTransformed,
+			gpuStats.numVertsSubmitted,
 			gpuStats.numCachedVertsDrawn,
+			gpuStats.numUncachedVertsDrawn,
 			gpuStats.numFBOs,
 			gpuStats.numTextures,
 			gpuStats.numTexturesDecoded,
@@ -241,10 +243,11 @@ void hleEnterVblank(u64 userdata, int cyclesLate) {
 			gpuStats.numShaders
 			);
 
-		float zoom = 0.5f; /// g_Config.iWindowZoom;
+		float zoom = 0.3f; /// g_Config.iWindowZoom;
+		float soff = 0.3f;
 		PPGeBegin();
-		PPGeDrawText(stats, 1, 1, 0, zoom, 0xFF000000);
-		PPGeDrawText(stats, -1, -1, 0, zoom, 0xFF000000);
+		PPGeDrawText(stats, soff, soff, 0, zoom, 0xCC000000);
+		PPGeDrawText(stats, -soff, -soff, 0, zoom, 0xCC000000);
 		PPGeDrawText(stats, 0, 0, 0, zoom, 0xFFFFFFFF);
 		PPGeEnd();
 

GPU/GLES/DisplayListInterpreter.cpp

@@ -166,8 +166,8 @@ const int flushBeforeCommandList[] = {
 GLES_GPU::GLES_GPU(int renderWidth, int renderHeight)
 :		interruptsEnabled_(true),
 		displayFramebufPtr_(0),
-		prevDisplayFramebufPtr_(0),
-		prevPrevDisplayFramebufPtr_(0),
+		prevDisplayFramebuf_(0),
+		prevPrevDisplayFramebuf_(0),
 		renderWidth_(renderWidth),
 		renderHeight_(renderHeight)
 {
@@ -257,8 +257,6 @@ void GLES_GPU::BeginFrame() {
 void GLES_GPU::SetDisplayFramebuffer(u32 framebuf, u32 stride, int format) {
 	if (framebuf & 0x04000000) {
 		//DEBUG_LOG(G3D, "Switch display framebuffer %08x", framebuf);
-		prevPrevDisplayFramebufPtr_ = prevDisplayFramebufPtr_;
-		prevDisplayFramebufPtr_ = displayFramebufPtr_;
 		displayFramebufPtr_ = framebuf;
 		displayStride_ = stride;
 		displayFormat_ = format;
@@ -275,6 +273,9 @@ void GLES_GPU::CopyDisplayToOutput() {
 	EndDebugDraw();
 
 	VirtualFramebuffer *vfb = GetDisplayFBO();
+	prevPrevDisplayFramebuf_ = prevDisplayFramebuf_;
+	prevDisplayFramebuf_ = displayFramebuf_;
+	displayFramebuf_ = vfb;
 	fbo_unbind();
 
 	glstate.viewport.set(0, 0, PSP_CoreParameter().pixelWidth, PSP_CoreParameter().pixelHeight);
@@ -307,17 +308,12 @@ void GLES_GPU::CopyDisplayToOutput() {
 	BeginDebugDraw();
 }
 
-static bool MaskedEqual(u32 addr1, u32 addr2) {
-	return (addr1 & 0x3FFFFFF) == (addr2 & 0x3FFFFFF);
-}
-
-
 void GLES_GPU::DecimateFBOs() {
 	for (auto iter = vfbs_.begin(); iter != vfbs_.end();) {
 		VirtualFramebuffer *v = *iter;
-		if (MaskedEqual(v->fb_address, displayFramebufPtr_) ||
-				MaskedEqual(v->fb_address, prevDisplayFramebufPtr_) ||
-				MaskedEqual(v->fb_address, prevPrevDisplayFramebufPtr_)) {
+		if (v == displayFramebuf_ ||
+				v == prevDisplayFramebuf_ ||
+				v == prevPrevDisplayFramebuf_) {
 			++iter;
 			continue;
 		}
@@ -330,6 +326,10 @@ void GLES_GPU::DecimateFBOs() {
 	}
 }
 
+static bool MaskedEqual(u32 addr1, u32 addr2) {
+	return (addr1 & 0x3FFFFFF) == (addr2 & 0x3FFFFFF);
+}
+
 GLES_GPU::VirtualFramebuffer *GLES_GPU::GetDisplayFBO() {
 	for (auto iter = vfbs_.begin(); iter != vfbs_.end(); ++iter) {
 		if (MaskedEqual((*iter)->fb_address, displayFramebufPtr_) && (*iter)->format == displayFormat_) {
@@ -338,7 +338,7 @@ GLES_GPU::VirtualFramebuffer *GLES_GPU::GetDisplayFBO() {
 		}
 	}
 	DEBUG_LOG(HLE, "Finding no FBO matching address %08x", displayFramebufPtr_);
-#ifdef _DEBUG
+#if 0  // defined(_DEBUG)
 	std::string debug = "FBOs: ";
 	for (auto iter = vfbs_.begin(); iter != vfbs_.end(); ++iter) {
 		char temp[256];
@@ -376,7 +376,7 @@ void GLES_GPU::SetRenderFrameBuffer() {
 	VirtualFramebuffer *vfb = 0;
 	for (auto iter = vfbs_.begin(); iter != vfbs_.end(); ++iter) {
 		VirtualFramebuffer *v = *iter;
-		if (v->fb_address == fb_address && v->width == drawing_width && v->height == drawing_height && v->format == fmt) {
+		if (v->fb_address == fb_address && v->width == drawing_width && v->height == drawing_height /* && v->format == fmt */) {
 			// Let's not be so picky for now. Let's say this is the one.
 			vfb = v;
 			// Update fb stride in case it changed
@@ -406,7 +406,7 @@ void GLES_GPU::SetRenderFrameBuffer() {
 		case GE_FORMAT_8888: vfb->colorDepth = FBO_8888;
 		}
 //#ifdef ANDROID
-//		vfb->colorDepth = FBO_5551;
+//	vfb->colorDepth = FBO_8888;
 //#endif
 
 		vfb->fbo = fbo_create(vfb->width * renderWidthFactor_, vfb->height * renderHeightFactor_, 1, true, vfb->colorDepth);

GPU/GLES/DisplayListInterpreter.h

@@ -73,8 +73,6 @@ private:
 	bool interruptsEnabled_;
 
 	u32 displayFramebufPtr_;
-	u32 prevDisplayFramebufPtr_;
-	u32 prevPrevDisplayFramebufPtr_;
 	u32 displayStride_;
 	int displayFormat_;
 
@@ -107,6 +105,10 @@ private:
 		FBO *fbo;
 	};
 
+	VirtualFramebuffer *displayFramebuf_;
+	VirtualFramebuffer *prevDisplayFramebuf_;
+	VirtualFramebuffer *prevPrevDisplayFramebuf_;
+
 	void SetRenderFrameBuffer();  // Uses parameters computed from gstate
 	// TODO: Break out into some form of FBO manager
 	VirtualFramebuffer *GetDisplayFBO();

GPU/GLES/TransformPipeline.cpp

@@ -691,7 +691,7 @@ void TransformDrawEngine::SubmitPrim(void *verts, void *inds, int prim, int vert
 		gpuStats.numJoins++;
 	}
 	gpuStats.numDrawCalls++;
-	gpuStats.numVertsTransformed += vertexCount;
+	gpuStats.numVertsSubmitted += vertexCount;
 
 	DeferredDrawCall &dc = drawCalls[numDrawCalls++];
 	dc.verts = verts;
@@ -921,6 +921,7 @@ void TransformDrawEngine::Flush() {
 						if (vai->ebo)
 							glBindBuffer(GL_ELEMENT_ARRAY_BUFFER, vai->ebo);
 						useElements = vai->ebo ? true : false;
+						gpuStats.numCachedVertsDrawn += vai->numVerts;
 					}
 					vbo = vai->vbo;
 					ebo = vai->ebo;
@@ -934,6 +935,7 @@ void TransformDrawEngine::Flush() {
 				{
 					vai->numDraws++;
 					gpuStats.numCachedDrawCalls++;
+					gpuStats.numCachedVertsDrawn += vai->numVerts;
 					// DecodeVerts(); // TODO : Remove
 					vbo = vai->vbo;
 					ebo = vai->ebo;
@@ -955,6 +957,7 @@ void TransformDrawEngine::Flush() {
 		} else {
 			DecodeVerts();
 rotateVBO:
+			gpuStats.numUncachedVertsDrawn += indexGen.VertexCount();
 			useElements = !indexGen.SeenOnlyPurePrims();
 			if (g_Config.bUseVBO) {
 				// Just rotate VBO.
@@ -991,6 +994,7 @@ rotateVBO:
 			glBindBuffer(GL_ARRAY_BUFFER, 0);
 	} else {
 		DecodeVerts();
+		gpuStats.numUncachedVertsDrawn += indexGen.VertexCount();
 		prim = indexGen.Prim();
 		DEBUG_LOG(G3D, "Flush prim %i SW! %i verts in one go", prim, indexGen.VertexCount());
 

GPU/GPUState.h

@@ -259,8 +259,9 @@ struct GPUStatistics
 		numJoins = 0;
 		numDrawCalls = 0;
 		numCachedDrawCalls = 0;
-		numVertsTransformed = 0;
+		numVertsSubmitted = 0;
 		numCachedVertsDrawn = 0;
+		numUncachedVertsDrawn = 0;
 		numTrackedVertexArrays = 0;
 		numTextureInvalidations = 0;
 		numTextureSwitches = 0;
@@ -275,8 +276,9 @@ struct GPUStatistics
 	int numDrawCalls;
 	int numCachedDrawCalls;
 	int numFlushes;
-	int numVertsTransformed;
+	int numVertsSubmitted;
 	int numCachedVertsDrawn;
+	int numUncachedVertsDrawn;
 	int numTrackedVertexArrays;
 	int numTextureInvalidations;
 	int numTextureSwitches;