From 028a341cc8f5715f764291c8c24c118c0a7cf260 Mon Sep 17 00:00:00 2001
From: "Unknown W. Brackets" <checkins@unknownbrackets.org>
Date: Sat, 17 Sep 2022 16:28:15 -0700
Subject: [PATCH] softgpu: Explicitly flush on sync and output.

We could in theory skip flush on FinishDeferred, and allow some CPU/GPU
overlap.  If we did, we'd still want to flush at these times.
---
 GPU/Software/SoftGpu.cpp       | 13 +++++++++++++
 GPU/Software/SoftGpu.h         |  2 ++
 GPU/Software/TransformUnit.cpp |  8 ++++++++
 GPU/Software/TransformUnit.h   |  1 +
 4 files changed, 24 insertions(+)

diff --git a/GPU/Software/SoftGpu.cpp b/GPU/Software/SoftGpu.cpp
index 9de5fe5cbd..48c5258cd5 100644
--- a/GPU/Software/SoftGpu.cpp
+++ b/GPU/Software/SoftGpu.cpp
@@ -639,6 +639,7 @@ void SoftGPU::CopyToCurrentFboFromDisplayRam(int srcwidth, int srcheight) {
 }
 
 void SoftGPU::CopyDisplayToOutput(bool reallyDirty) {
+	drawEngine_->transformUnit.Flush("output");
 	// The display always shows 480x272.
 	CopyToCurrentFboFromDisplayRam(FB_WIDTH, FB_HEIGHT);
 	MarkDirty(displayFramebuf_, displayStride_, 272, displayFormat_, SoftGPUVRAMDirty::CLEAR);
@@ -1138,6 +1139,18 @@ void SoftGPU::FinishDeferred() {
 	drawEngine_->transformUnit.Flush("finish");
 }
 
+int SoftGPU::ListSync(int listid, int mode) {
+	// Take this as a cue that we need to finish drawing.
+	drawEngine_->transformUnit.Flush("listsync");
+	return GPUCommon::ListSync(listid, mode);
+}
+
+u32 SoftGPU::DrawSync(int mode) {
+	// Take this as a cue that we need to finish drawing.
+	drawEngine_->transformUnit.Flush("drawsync");
+	return GPUCommon::DrawSync(mode);
+}
+
 void SoftGPU::GetStats(char *buffer, size_t bufsize) {
 	drawEngine_->transformUnit.GetStats(buffer, bufsize);
 }
diff --git a/GPU/Software/SoftGpu.h b/GPU/Software/SoftGpu.h
index 11d5dd16f3..54880bd477 100644
--- a/GPU/Software/SoftGpu.h
+++ b/GPU/Software/SoftGpu.h
@@ -131,6 +131,8 @@ public:
 	void InitClear() override {}
 	void ExecuteOp(u32 op, u32 diff) override;
 	void FinishDeferred() override;
+	int ListSync(int listid, int mode) override;
+	u32 DrawSync(int mode) override;
 
 	void SetDisplayFramebuffer(u32 framebuf, u32 stride, GEBufferFormat format) override;
 	void CopyDisplayToOutput(bool reallyDirty) override;
diff --git a/GPU/Software/TransformUnit.cpp b/GPU/Software/TransformUnit.cpp
index 0cb91dd280..dfc8c6a58e 100644
--- a/GPU/Software/TransformUnit.cpp
+++ b/GPU/Software/TransformUnit.cpp
@@ -484,6 +484,7 @@ void TransformUnit::SubmitPrimitive(const void* vertices, const void* indices, G
 	// then resolve the indices. This lets us avoid transforming shared vertices twice.
 
 	binner_->UpdateState(vreader.isThrough());
+	hasDraws_ = true;
 
 	static TransformState transformState;
 	if (binner_->HasDirty(SoftDirty::LIGHT_ALL | SoftDirty::TRANSFORM_ALL)) {
@@ -826,8 +827,12 @@ void TransformUnit::SendTriangle(CullType cullType, const VertexData *verts, int
 }
 
 void TransformUnit::Flush(const char *reason) {
+	if (!hasDraws_)
+		return;
+
 	binner_->Flush(reason);
 	GPUDebug::NotifyDraw();
+	hasDraws_ = false;
 }
 
 void TransformUnit::GetStats(char *buffer, size_t bufsize) {
@@ -836,6 +841,9 @@ void TransformUnit::GetStats(char *buffer, size_t bufsize) {
 }
 
 void TransformUnit::FlushIfOverlap(const char *reason, bool modifying, uint32_t addr, uint32_t stride, uint32_t w, uint32_t h) {
+	if (!hasDraws_)
+		return;
+
 	if (binner_->HasPendingWrite(addr, stride, w, h))
 		Flush(reason);
 	if (modifying && binner_->HasPendingRead(addr, stride, w, h))
diff --git a/GPU/Software/TransformUnit.h b/GPU/Software/TransformUnit.h
index fd6e047350..2b35501b44 100644
--- a/GPU/Software/TransformUnit.h
+++ b/GPU/Software/TransformUnit.h
@@ -149,6 +149,7 @@ private:
 	// This is the index of the next vert in data (or higher, may need modulus.)
 	int data_index_ = 0;
 	GEPrimitiveType prev_prim_ = GE_PRIM_POINTS;
+	bool hasDraws_ = false;
 };
 
 class SoftwareDrawEngine : public DrawEngineCommon {