From cc155ec460b672bb82a414d1e44e5d18e9ebbe25 Mon Sep 17 00:00:00 2001
From: "Unknown W. Brackets" <checkins@unknownbrackets.org>
Date: Sat, 15 Jan 2022 23:24:58 -0800
Subject: [PATCH] softgpu: Avoid texture/CLUT flush unless overlap.

Only need to flush here if there's some overlap in the target.
---
 GPU/Software/BinManager.cpp    |  2 +-
 GPU/Software/BinManager.h      |  2 +-
 GPU/Software/SoftGpu.cpp       | 19 +++++++++++--------
 GPU/Software/TransformUnit.cpp | 20 +++++++++++++++++++-
 GPU/Software/TransformUnit.h   |  3 ++-
 5 files changed, 34 insertions(+), 12 deletions(-)

diff --git a/GPU/Software/BinManager.cpp b/GPU/Software/BinManager.cpp
index e56f2553e2..cdb07f8bc9 100644
--- a/GPU/Software/BinManager.cpp
+++ b/GPU/Software/BinManager.cpp
@@ -174,7 +174,7 @@ void BinManager::UpdateState() {
 	}
 }
 
-void BinManager::UpdateClut(void *src) {
+void BinManager::UpdateClut(const void *src) {
 	if (cluts_.Full())
 		Flush();
 	clutIndex_ = (int)cluts_.Push(BinClut());
diff --git a/GPU/Software/BinManager.h b/GPU/Software/BinManager.h
index 138531bb81..eff0c4e6c9 100644
--- a/GPU/Software/BinManager.h
+++ b/GPU/Software/BinManager.h
@@ -153,7 +153,7 @@ public:
 	~BinManager();
 
 	void UpdateState();
-	void UpdateClut(void *src);
+	void UpdateClut(const void *src);
 
 	const Rasterizer::RasterizerState &State() {
 		return states_[stateIndex_];
diff --git a/GPU/Software/SoftGpu.cpp b/GPU/Software/SoftGpu.cpp
index 2e21e9dc07..44401d8ba7 100644
--- a/GPU/Software/SoftGpu.cpp
+++ b/GPU/Software/SoftGpu.cpp
@@ -548,6 +548,10 @@ void SoftGPU::ExecuteOp(u32 op, u32 diff) {
 
 	case GE_CMD_SCISSOR1:
 	case GE_CMD_SCISSOR2:
+		for (int i = 0; i < 8; ++i) {
+			// TODO: Narrow down the risk range, for now assuming 512x512 max texture.
+			drawEngine_->transformUnit.FlushIfOverlap(gstate.getTextureAddress(i), 4 * 512 * 512);
+		}
 		break;
 
 	case GE_CMD_MINZ:
@@ -576,8 +580,8 @@ void SoftGPU::ExecuteOp(u32 op, u32 diff) {
 	case GE_CMD_TEXADDR5:
 	case GE_CMD_TEXADDR6:
 	case GE_CMD_TEXADDR7:
-		// TODO: Try not flushing here, unless overlap with framebuf/depthbuf?
-		drawEngine_->transformUnit.Flush();
+		// TODO: Narrow down the risk range, for now assuming 512x512 max texture.
+		drawEngine_->transformUnit.FlushIfOverlap(gstate.getTextureAddress(cmd - GE_CMD_TEXADDR0), 4 * 512 * 512);
 		break;
 
 	case GE_CMD_TEXBUFWIDTH0:
@@ -588,8 +592,8 @@ void SoftGPU::ExecuteOp(u32 op, u32 diff) {
 	case GE_CMD_TEXBUFWIDTH5:
 	case GE_CMD_TEXBUFWIDTH6:
 	case GE_CMD_TEXBUFWIDTH7:
-		// TODO: Try not flushing here, unless overlap with framebuf/depthbuf?
-		drawEngine_->transformUnit.Flush();
+		// TODO: Narrow down the risk range, for now assuming 512x512 max texture.
+		drawEngine_->transformUnit.FlushIfOverlap(gstate.getTextureAddress(cmd - GE_CMD_TEXBUFWIDTH0), 4 * 512 * 512);
 		break;
 
 	case GE_CMD_CLUTADDR:
@@ -598,13 +602,12 @@ void SoftGPU::ExecuteOp(u32 op, u32 diff) {
 
 	case GE_CMD_LOADCLUT:
 		{
-			// Might be copying drawing into the CLUT, so flush.
-			// TODO: It seems worth copying the CLUT to state...
-			drawEngine_->transformUnit.Flush();
-
 			u32 clutAddr = gstate.getClutAddress();
 			u32 clutTotalBytes = gstate.getClutLoadBytes();
 
+			// Might be copying drawing into the CLUT, so flush.
+			drawEngine_->transformUnit.FlushIfOverlap(clutAddr, clutTotalBytes);
+
 			if (Memory::IsValidAddress(clutAddr)) {
 				u32 validSize = Memory::ValidSize(clutAddr, clutTotalBytes);
 				Memory::MemcpyUnchecked(clut, clutAddr, validSize);
diff --git a/GPU/Software/TransformUnit.cpp b/GPU/Software/TransformUnit.cpp
index 5d2a9a527a..c628a3c853 100644
--- a/GPU/Software/TransformUnit.cpp
+++ b/GPU/Software/TransformUnit.cpp
@@ -613,7 +613,25 @@ void TransformUnit::Flush() {
 	GPUDebug::NotifyDraw();
 }
 
-void TransformUnit::NotifyClutUpdate(void *src) {
+void TransformUnit::FlushIfOverlap(uint32_t addr, uint32_t sz) {
+	if (!Memory::IsVRAMAddress(addr))
+		return;
+	addr &= 0x0FFFFFFF;
+
+	uint32_t targetHeight = gstate.getScissorY2() + 1;
+	uint32_t target = gstate.getFrameBufAddress() & 0x0FFFFFFF;
+	uint32_t targetStride = gstate.FrameBufStride() * (gstate.FrameBufFormat() == GE_FORMAT_8888 ? 4 : 2);
+	uint32_t ztarget = gstate.getDepthBufAddress() & 0x0FFFFFFF;
+	uint32_t ztargetStride = gstate.DepthBufStride() * 2;
+
+	// TODO: Skip if the texture is between width and stride?
+	if (addr < target + targetHeight * targetStride && addr + sz >= target)
+		Flush();
+	else if (addr < ztarget + targetHeight * ztargetStride && addr + sz >= ztarget)
+		Flush();
+}
+
+void TransformUnit::NotifyClutUpdate(const void *src) {
 	binner_->UpdateClut(src);
 }
 
diff --git a/GPU/Software/TransformUnit.h b/GPU/Software/TransformUnit.h
index 8f8fcd72b2..fe8f757d9a 100644
--- a/GPU/Software/TransformUnit.h
+++ b/GPU/Software/TransformUnit.h
@@ -120,7 +120,8 @@ public:
 	bool GetCurrentSimpleVertices(int count, std::vector<GPUDebugVertex> &vertices, std::vector<u16> &indices);
 
 	void Flush();
-	void NotifyClutUpdate(void *src);
+	void FlushIfOverlap(uint32_t addr, uint32_t sz);
+	void NotifyClutUpdate(const void *src);
 
 private:
 	VertexData ReadVertex(VertexReader &vreader, bool &outside_range_flag);