Merge pull request #5808 from unknownbrackets/perf

More small performance tweaks based on ARM profiling

GPU/GLES/TransformPipeline.cpp

@@ -255,9 +255,13 @@ VertexDecoder *TransformDrawEngine::GetVertexDecoder(u32 vtype) {
 }
 
 void TransformDrawEngine::SetupVertexDecoder(u32 vertType) {
+	SetupVertexDecoderInternal(vertType);
+}
+
+inline void TransformDrawEngine::SetupVertexDecoderInternal(u32 vertType) {
 	// As the decoder depends on the UVGenMode when we use UV prescale, we simply mash it
 	// into the top of the verttype where there are unused bits.
-	u32 vertTypeID = (vertType & 0xFFFFFF) | (gstate.getUVGenMode() << 24);
+	const u32 vertTypeID = (vertType & 0xFFFFFF) | (gstate.getUVGenMode() << 24);
 
 	// If vtype has changed, setup the vertex decoder.
 	// TODO: Simply cache the setup decoders instead.
@@ -280,7 +284,7 @@ void TransformDrawEngine::SubmitPrim(void *verts, void *inds, GEPrimitiveType pr
 	}
 	prevPrim_ = prim;
 
-	SetupVertexDecoder(vertType);
+	SetupVertexDecoderInternal(vertType);
 
 	dec_->IncrementStat(STAT_VERTSSUBMITTED, vertexCount);
 
@@ -373,33 +377,45 @@ void TransformDrawEngine::DecodeVertsStep() {
 
 		// 1. Look ahead to find the max index, only looking as "matching" drawcalls.
 		//    Expand the lower and upper bounds as we go.
-		int j = i + 1;
 		int lastMatch = i;
-		while (j < numDrawCalls) {
-			if (drawCalls[j].verts != dc.verts)
-				break;
-			if (uvScale && memcmp(&uvScale[j], &uvScale[i], sizeof(uvScale[0])) != 0)
-				break;
+		const int total = numDrawCalls;
+		if (uvScale) {
+			for (int j = i + 1; j < total; ++j) {
+				if (drawCalls[j].verts != dc.verts)
+					break;
+				if (memcmp(&uvScale[j], &uvScale[i], sizeof(uvScale[0])) != 0)
+					break;
 
-			indexLowerBound = std::min(indexLowerBound, (int)drawCalls[j].indexLowerBound);
-			indexUpperBound = std::max(indexUpperBound, (int)drawCalls[j].indexUpperBound);
-			lastMatch = j;
-			j++;
-		}
+				indexLowerBound = std::min(indexLowerBound, (int)drawCalls[j].indexLowerBound);
+				indexUpperBound = std::max(indexUpperBound, (int)drawCalls[j].indexUpperBound);
+				lastMatch = j;
+			}
+		} else {
+			for (int j = i + 1; j < total; ++j) {
+				if (drawCalls[j].verts != dc.verts)
+					break;
 
-		// 2. Loop through the drawcalls, translating indices as we go.
-		for (j = i; j <= lastMatch; j++) {
-			switch (indexType) {
-			case GE_VTYPE_IDX_8BIT >> GE_VTYPE_IDX_SHIFT:
-				indexGen.TranslatePrim(drawCalls[j].prim, drawCalls[j].vertexCount, (const u8 *)drawCalls[j].inds, indexLowerBound);
-				break;
-			case GE_VTYPE_IDX_16BIT >> GE_VTYPE_IDX_SHIFT:
-				indexGen.TranslatePrim(drawCalls[j].prim, drawCalls[j].vertexCount, (const u16 *)drawCalls[j].inds, indexLowerBound);
-				break;
+				indexLowerBound = std::min(indexLowerBound, (int)drawCalls[j].indexLowerBound);
+				indexUpperBound = std::max(indexUpperBound, (int)drawCalls[j].indexUpperBound);
+				lastMatch = j;
 			}
 		}
 
-		int vertexCount = indexUpperBound - indexLowerBound + 1;
+		// 2. Loop through the drawcalls, translating indices as we go.
+		switch (indexType) {
+		case GE_VTYPE_IDX_8BIT >> GE_VTYPE_IDX_SHIFT:
+			for (int j = i; j <= lastMatch; j++) {
+				indexGen.TranslatePrim(drawCalls[j].prim, drawCalls[j].vertexCount, (const u8 *)drawCalls[j].inds, indexLowerBound);
+			}
+			break;
+		case GE_VTYPE_IDX_16BIT >> GE_VTYPE_IDX_SHIFT:
+			for (int j = i; j <= lastMatch; j++) {
+				indexGen.TranslatePrim(drawCalls[j].prim, drawCalls[j].vertexCount, (const u16 *)drawCalls[j].inds, indexLowerBound);
+			}
+			break;
+		}
+
+		const int vertexCount = indexUpperBound - indexLowerBound + 1;
 		// 3. Decode that range of vertex data.
 		dec_->DecodeVerts(decoded + decodedVerts_ * (int)dec_->GetDecVtxFmt().stride,
 			dc.verts, indexLowerBound, indexUpperBound);

GPU/GLES/TransformPipeline.h

@@ -124,6 +124,7 @@ public:
 	void ClearTrackedVertexArrays();
 
 	void SetupVertexDecoder(u32 vertType);
+	inline void SetupVertexDecoderInternal(u32 vertType);
 
 	// This requires a SetupVertexDecoder call first.
 	int EstimatePerVertexCost() {

GPU/GPUCommon.cpp

@@ -559,17 +559,19 @@ void GPUCommon::SlowRunLoop(DisplayList &list)
 // The newPC parameter is used for jumps, we don't count cycles between.
 void GPUCommon::UpdatePC(u32 currentPC, u32 newPC) {
 	// Rough estimate, 2 CPU ticks (it's double the clock rate) per GPU instruction.
-	int executed = (currentPC - cycleLastPC) / 4;
+	u32 executed = (currentPC - cycleLastPC) / 4;
 	cyclesExecuted += 2 * executed;
-	gpuStats.otherGPUCycles += 2 * executed;
-	cycleLastPC = newPC == 0 ? currentPC : newPC;
+	cycleLastPC = currentPC;
 
-	gpuStats.gpuCommandsAtCallLevel[std::min(currentList->stackptr, 3)] += executed;
+	if (g_Config.bShowDebugStats) {
+		gpuStats.otherGPUCycles += 2 * executed;
+		gpuStats.gpuCommandsAtCallLevel[std::min(currentList->stackptr, 3)] += executed;
+	}
 
 	// Exit the runloop and recalculate things.  This happens a lot in some games.
 	easy_guard innerGuard(listLock);
 	if (currentList)
-		downcount = currentList->stall == 0 ? 0x0FFFFFFF : (currentList->stall - cycleLastPC) / 4;
+		downcount = currentList->stall == 0 ? 0x0FFFFFFF : (currentList->stall - currentPC) / 4;
 	else
 		downcount = 0;
 }

GPU/GPUCommon.h

@@ -81,7 +81,10 @@ protected:
 	// To avoid virtual calls to PreExecuteOp().
 	virtual void FastRunLoop(DisplayList &list) = 0;
 	void SlowRunLoop(DisplayList &list);
-	void UpdatePC(u32 currentPC, u32 newPC = 0);
+	void UpdatePC(u32 currentPC, u32 newPC);
+	void UpdatePC(u32 currentPC) {
+		UpdatePC(currentPC, currentPC);
+	}
 	void UpdateState(GPUState state);
 	void PopDLQueue();
 	void CheckDrawSync();