Background raster: Optimize a little, add more stats.

hrydgard · Jan 2, 2025 · 6d5eced · 6d5eced
1 parent e538b13
commit 6d5eced
Show file tree

Hide file tree

Showing 4 changed files with 20 additions and 9 deletions.
diff --git a/GPU/Common/DrawEngineCommon.cpp b/GPU/Common/DrawEngineCommon.cpp
@@ -1060,13 +1060,9 @@ void DrawEngineCommon::EnqueueDepthDraw(const DepthDraw &draw) {
 }
 
 // Returns true if we actually waited.
-bool DrawEngineCommon::WaitForDepthPassFinish() {
+void DrawEngineCommon::WaitForDepthPassFinish() {
 	{
 		std::lock_guard<std::mutex> lock(depthEnqueueMutex_);
-		if (!inDepthDrawPass_) {
-			_dbg_assert_(depthIndexCount_ == 0 && depthVertexCount_ == 0 && depthDraws_.empty());
-			return false;
-		}
 		// In case the depth raster thread is idle, we need to nudge it.
 		_dbg_assert_(!finishedSubmitting_);
 		finishedSubmitting_ = true;
@@ -1078,7 +1074,6 @@ bool DrawEngineCommon::WaitForDepthPassFinish() {
 	while (!finishedDrawing_) {
 		depthFinishCond_.wait(lock);
 	}
-	return true;
 }
 
 void DrawEngineCommon::FlushQueuedDepth() {
@@ -1087,7 +1082,8 @@ void DrawEngineCommon::FlushQueuedDepth() {
 		depthRasterPassStart_ = 0.0;
 	}
 
-	if (WaitForDepthPassFinish()) {
+	if (inDepthDrawPass_) {
+		WaitForDepthPassFinish();
 		// At this point, we know that the depth thread is paused.
 
 		// Reset queue
@@ -1114,10 +1110,14 @@ void DrawEngineCommon::DepthThreadFunc() {
 		{
 			std::unique_lock<std::mutex> lock(depthEnqueueMutex_);
 			if (depthDraws_.size() == curDraw_) {
+				gpuStats.numDepthThreadCaughtUp++;
 				// We've drawn all we can. Let's check if we're finished.
 				// If we reach here, we've drawn everything we can. And if that's the last
 				// that will come in this batch, we notify.
 				if (finishedSubmitting_) {
+					gpuStats.numDepthThreadFinished++;
+					gpuStats.numDepthDraws += (int)depthDraws_.size();
+
 					// lock.unlock();  // possible optimization?
 					std::lock_guard<std::mutex> flock(depthFinishMutex_);
 					finishedDrawing_ = true;

diff --git a/GPU/Common/DrawEngineCommon.h b/GPU/Common/DrawEngineCommon.h
@@ -368,7 +368,7 @@ class DrawEngineCommon {
 private:
 	// Internal implementation details.
 	void DepthThreadFunc();
-	bool WaitForDepthPassFinish();
+	void WaitForDepthPassFinish();
 
 	inline void EnqueueDepthDraw(const DepthDraw &draw);
 	inline void ProcessDepthDraw(const DepthDraw &draw);

diff --git a/GPU/GPU.h b/GPU/GPU.h
@@ -118,6 +118,9 @@ struct GPUStatistics {
 		numDepthRasterTooSmall = 0;
 		numDepthRasterZCulled = 0;
 		numDepthEarlyBoxCulled = 0;
+		numDepthThreadCaughtUp = 0;
+		numDepthThreadFinished = 0;
+		numDepthDraws = 0;
 		vertexGPUCycles = 0;
 		otherGPUCycles = 0;
 	}
@@ -169,6 +172,9 @@ struct GPUStatistics {
 	int numDepthRasterTooSmall;
 	int numDepthRasterZCulled;
 	int numDepthEarlyBoxCulled;
+	int numDepthThreadCaughtUp;
+	int numDepthThreadFinished;
+	int numDepthDraws;
 	// Flip count. Doesn't really belong here.
 	int numFlips;
 };

diff --git a/GPU/GPUCommonHW.cpp b/GPU/GPUCommonHW.cpp
@@ -1802,7 +1802,9 @@ size_t GPUCommonHW::FormatGPUStatsCommon(char *buffer, size_t size) {
 		"Cpy: depth %d, color %d, reint %d, blend %d, self %d\n"
 		"GPU cycles: %d (%0.1f per vertex)\n"
 		"Z-rast: %0.2f+%0.2f+%0.2f (total %0.2f/%0.2f) ms\n"
-		"Z-rast: %d prim, %d nopix, %d small, %d earlysize, %d zcull, %d box\n%s",
+		"Z-rast: %d prim, %d nopix, %d small, %d earlysize, %d zcull, %d box\n"
+		"Z-rast: %d draws, %d thread caught up, %d thread finishes\n"
+		"%s",
 		gpuStats.msProcessingDisplayLists * 1000.0f,
 		gpuStats.numDrawSyncs,
 		gpuStats.numListSyncs,
@@ -1850,6 +1852,9 @@ size_t GPUCommonHW::FormatGPUStatsCommon(char *buffer, size_t size) {
 		gpuStats.numDepthRasterEarlySize,
 		gpuStats.numDepthRasterZCulled,
 		gpuStats.numDepthEarlyBoxCulled,
+		gpuStats.numDepthDraws,
+		gpuStats.numDepthThreadCaughtUp,
+		gpuStats.numDepthThreadFinished,
 		debugRecording_ ? "(debug-recording)" : ""
 	);
 }